Partial discharge signal sharpening method based on VMD decomposition-wavelet denoising-deep refinement

Abstract

The invention relates to the technical field of signal processing, in particular to a partial discharge signal sharpening method based on VMD decomposition-wavelet denoising-depth refinement, which provides a cascade processing framework of VMD decomposition-wavelet adaptive threshold-depth residual refinement and combines a parameter adaptive selection and uncertainty evaluation mechanism, so that the partial discharge signal sharpening method based on VMD decomposition-wavelet denoising-depth refinement is realized. And the robustness and auditing performance of the method under different sensing channels and variable working conditions are improved, so that a reliable signal basis is provided for diagnosis of the insulation state of the power equipment.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a method for clarifying partial discharge signals based on VMD decomposition-wavelet denoising-depth refinement. Background Technology

[0002] Partial discharge (PD) signal detection is a core method for early diagnosis of insulation defects in power equipment. However, its field measurement is often affected by complex interference: PD transient pulses are easily superimposed with power frequency harmonics, narrowband radio interference, mechanical noise, random pulse noise, and dispersion attenuation caused by long-distance transmission, resulting in low signal-to-noise ratio, overlapping spectral lines, and obscuring of timing characteristics, which seriously restricts the accuracy of equipment insulation condition diagnosis.

[0003] Existing partial discharge (PD) signal processing techniques have significant limitations: traditional bandpass / notch filtering struggles to handle broadband non-stationary interference; wavelet thresholding denoising relies on empirical thresholds, which can easily lead to over-threshold clipping and PD spike distortion; empirical mode decomposition and its improvements (EMD / EEMD / CEEMDAN) suffer from mode aliasing and endpoint effects, resulting in poor decomposition stability; key parameters (number of modes, bandwidth penalty factor) of variational mode decomposition (VMD) methods need to be manually set, and improper parameters can easily cause mode overlap or misallocation of PD energy; end-to-end deep denoising networks are sensitive to data scale and operating condition coverage, lack generalization ability, and are uninterpretable.

[0004] Furthermore, existing solutions generally lack joint constraints on "morphological fidelity, spectral consistency, and timing accuracy," and have not established an uncertainty assessment mechanism adapted to engineering. Under variable conditions such as the coexistence of multiple types of noise, differences in sensing channels, and environmental noise drift, it is difficult to balance interference suppression and preservation of PD transient pulse characteristics. There is an urgent need for a method to clarify partial discharge signals with high fidelity and strong robustness. Summary of the Invention

[0005] The purpose of this invention is to propose a method for clarifying partial discharge signals based on VMD decomposition, wavelet denoising, and depth refinement, in order to solve the technical problems of weak anti-interference, high parameter dependence, and insufficient fidelity in existing partial discharge signal processing methods.

[0006] To achieve this objective, the present invention adopts the following technical solution: A partial discharge signal decryption method based on VMD decomposition, wavelet denoising, and deep refinement includes the following steps: S1. Preprocess and standardize the original partial discharge observation signal to obtain the bandpass signal, narrowband interference frequency set and pulse suspected time set; S2. Using the set of narrowband interference frequencies as the initial center frequency, perform variational mode decomposition on the bandpass signal, and adaptively select decomposition parameters by combining the joint criterion function to identify narrowband interference modes, low-frequency trend modes and candidate partial discharge modes; S3. Based on the pulse suspected time set, the threshold is corrected, and wavelet threshold denoising is performed on the candidate partial discharge modes to obtain the denoised mode components. S4. Remove narrowband interference modes and low-frequency trend modes, superimpose and reconstruct the denoised mode components to obtain the intermediate sharpened signal, and calculate the difference between the bandpass signal and the intermediate sharpened signal to construct the residual signal. S5. Input the intermediate sharpening signal and the residual signal into the deep residual network, and output the final sharpening signal after network processing. S6. Perform pulse detection and quality assessment on the final clear signal; S7. Adjust the detection threshold based on uncertainty estimation and complete the alarm classification.

[0007] Preferably, in S1, the preprocessing and standardization specifically includes the following steps: S11. Subtract the mean from the original partial discharge observation signal to remove the DC component, and then select a bandpass filter of the corresponding frequency band according to the sensor type to filter it to obtain the bandpass signal. The ultra-high frequency sensing channel corresponds to the 300MHz-1.5GHz frequency band, the ultrasonic sensing channel corresponds to the 20kHz-200kHz frequency band, and the current pulse channel corresponds to the 100kHz-20MHz frequency band. S12. The power spectrum of the bandpass signal is estimated using the Welch method. Narrowband peak regions with continuous frequency points in the power spectrum whose amplitude is more than 3dB higher than the average value of the whole spectrum and whose bandwidth is less than 1% of the sampling frequency are detected. The frequencies corresponding to these regions are divided into narrowband interference frequency sets. S13. Extract the time index of the envelope energy or instantaneous amplitude in the bandpass signal that exceeds twice the global mean, and form a set of suspected pulse moments.

[0008] Preferably, in S2, the optimization objective of the variational mode decomposition is: ; The constraints are: ; in, For modal number, Represents the time variable The first derivative is calculated. The impulse function is used to construct the Hilbert transform kernel. The imaginary unit, For the time variable corresponding to the discrete sampled signal, For narrowband eigenmodes, The center frequency corresponding to the narrowband eigenmode. It is a bandpass signal; The joint criterion function for: ; in, As a bandwidth penalty factor, For pulse kurtosis, The spectral entropy of the modal center frequency distribution. Modal overlap rate, weighting coefficient ∈[0,1] and satisfy Through multiple sets of comparative experiments, the signal-to-noise ratio improvement and pulse reconstruction error under different weight combinations were calculated, and the weight combination with the best overall performance was selected. The narrowband interference mode is one whose center frequency is located in the neighborhood of the narrowband interference frequency set and whose bandwidth satisfies the following conditions. The narrowband eigenmodes, among which, For modal bandwidth, The sampling frequency; The low-frequency trend mode is a narrowband intrinsic mode in which the main frequency energy is concentrated in the lowest 5% of the bandpass signal power spectrum or the instantaneous energy change rate is less than 10% of the global mean change rate. The candidate partial discharge modes are narrowband intrinsic modes other than narrowband interference modes and low-frequency trend modes.

[0009] Preferably, in S3, the wavelet threshold denoising specifically includes the following steps: S31. Perform discrete wavelet transform on the candidate partial discharge modes to obtain approximation coefficients. and detail coefficient ; S32. Calculate the noise standard deviation of the detail coefficients using the median absolute deviation. : ; in, For the first The candidate localization mode The detail factor of the layer, The candidate localization mode number. The number of decomposition levels; S33. Set hierarchical thresholds based on noise standard deviation. : ; in, For the first Number of layer coefficients; S34, Combining the pulse suspected time set The hierarchical threshold is corrected by adjusting the formula. : ; S35. Apply the corrected hierarchy threshold to the detail coefficients. Soft thresholding is performed to obtain the detail coefficients after denoising. Combined with approximation coefficients Denoising mode components are obtained from reconstruction. .

[0010] Preferably, in S4, the narrowband interference mode and low-frequency trend mode are removed, and the reconstructed denoised mode components are superimposed. Obtain intermediate sharpening signal : ; And calculate the bandpass signal With intermediate clarity signal The difference is used to construct the residual signal. : .

[0011] Preferably, in S5, the deep residual network adopts a one-dimensional convolutional neural network or 1DU-Net structure, and the deep residual network includes an input layer, an encoder, a decoder and an output layer. The input layer receives the intermediate decryption signal. With residual signal A two-dimensional vector; The encoder contains four residual blocks. Each residual block consists of two one-dimensional convolutional layers with 3×1 convolutional kernels and stride of 1, one batch normalization layer, and one ReLU activation function. Feature downsampling is performed at each encoding stage. The decoder contains four upsampling layers and has a structure symmetrical to the encoder; The output layer uses the Sigmoid function to output the final clear signal. :

[0012] in, For the trained deep residual network; The training process of the deep residual network is as follows: Using simulated signals of corona discharge, surface discharge, and internal discharge as samples, with a total sample size ≥ 10,000, the data is augmented by time axis flipping and noise injection with a signal-to-noise ratio of -10dB to 5dB. Training is performed using the Adam optimizer, with an initial learning rate set to 0.001 and a loss function of:

[0013] in, For the final clear signal, For an ideal pulse signal, For envelope function or energy operator, For pulse kurtosis, The pulse arrival time, , , and These are the weighting coefficients.

[0014] Preferably, in S6, the pulse detection employs envelope analysis or Teager energy operator analysis to extract the pulse arrival time of the final clear signal; The quality assessment uses the signal-to-noise ratio improvement, pulse arrival time error, morphological peak signal-to-noise ratio, and SSIM as indicators.

[0015] Preferably, in S7, the uncertainty estimation estimates the variance of the final clear signal by Monte Carlo Dropout or deep model ensemble. :

[0016] in, For the first The prediction results of the random sampling For the number of samples, Output the mean of all samples; The process of adjusting the detection threshold is as follows: setting a baseline detection threshold. Calculate the confidence correction factor based on the uncertainty. :

[0017] in, With an adjustment coefficient of 0.2-0.5, the final detection threshold is... ; The alarm classification criteria are as follows: Level I is defined as variance < 0.5 times the global mean, Level II is defined as variance ≤ 0.5 times the global mean and Level III is defined as variance > 1 times the global mean.

[0018] One of the above technical solutions has the following beneficial effects: In summary, the beneficial effects of this technical solution include: 1. Strong interference suppression capability and outstanding robustness: Under multi-source interference such as power frequency harmonics, narrowband interference, random pulses and environmental noise, it achieves signal recovery with morphological fidelity, accurate timing and effective noise suppression of partial discharge transient pulses; by proposing a cascaded processing framework of "VMD decomposition - wavelet adaptive threshold - deep residual refinement", combined with parameter adaptive selection and uncertainty assessment mechanism, the robustness and auditability of the method under different sensing channels and variable operating conditions are improved, thus providing a reliable signal basis for the insulation condition diagnosis of power equipment.

[0019] 2. High fidelity of partial discharge characteristics, avoiding signal distortion: Through the dual guarantee of "pulse timing correction wavelet threshold" and "deep residual network skip connection", the problem of partial discharge spike distortion caused by over-thresholding and peak clipping in traditional wavelet denoising and insufficient generalization of deep networks is effectively solved. The processed signal not only retains the time-domain spike shape of the partial discharge pulse, but its spectral characteristics are also more than 95% consistent with the original effective signal, with a timing error of less than 10μs, satisfying the joint constraints of "morphological fidelity - spectral consistency - timing accuracy", providing a reliable feature basis for insulation defect diagnosis.

[0020] 3. Parameter Adaptation, Reduced Manual Dependence: Addressing the shortcomings of traditional VMD and wavelet denoising that rely on empirical parameters, this solution achieves automatic optimization of key parameters through adaptive design of "VMD decomposition guided by narrowband interference frequency set", "parameter optimization by joint criterion function", and "threshold correction for pulse suspected time". It can adapt to the partial discharge signal processing needs of different equipment (transformers, switchgear, cables) and different operating conditions without manual intervention, thus lowering the threshold for engineering applications. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a partial discharge signal de-sharpening method based on VMD decomposition, wavelet denoising, and depth refinement. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] A partial discharge signal decryption method based on VMD decomposition, wavelet denoising, and deep refinement includes the following steps: S1. Preprocess and standardize the original partial discharge observation signal to obtain the bandpass signal, narrowband interference frequency set and pulse suspected time set; S2. Using the set of narrowband interference frequencies as the initial center frequency, perform variational mode decomposition on the bandpass signal, and adaptively select decomposition parameters by combining the joint criterion function to identify narrowband interference modes, low-frequency trend modes and candidate partial discharge modes; S3. Based on the pulse suspected time set, the threshold is corrected, and wavelet threshold denoising is performed on the candidate partial discharge modes to obtain the denoised mode components. S4. Remove narrowband interference modes and low-frequency trend modes, superimpose and reconstruct the denoised mode components to obtain the intermediate sharpened signal, and calculate the difference between the bandpass signal and the intermediate sharpened signal to construct the residual signal. S5. Input the intermediate sharpening signal and the residual signal into the deep residual network, and output the final sharpening signal after network processing. S6. Perform pulse detection and quality assessment on the final clear signal; S7. Adjust the detection threshold based on uncertainty estimation and complete the alarm classification.

[0024] To achieve clarity in partial discharge signals under complex operating conditions, this invention proposes a three-stage cascaded method combining variational mode decomposition (VMD), wavelet adaptive threshold denoising, and a deep residual network. This method can maintain the authenticity of pulse morphology and temporal characteristics in environments with strong noise and multiple interferences, thus effectively solving the problems of insufficient fidelity and parameter dependence in existing technologies. The specific working principle is as follows: S1 mainly involves preprocessing and standardization: Clearly defining the processing criteria: First, preprocessing the original partial discharge signal such as baseline correction and amplitude normalization is performed. Then, effective frequency bands are filtered through bandpass filtering to obtain the bandpass signal. Simultaneously, spectral analysis is used to identify the characteristic frequencies of narrowband interference, forming a set of narrowband interference frequencies. Time-domain algorithms (such as thresholding) are used to locate suspected partial discharge pulse times, generating a set of suspected pulse times. These three results provide crucial basis for subsequent steps, avoiding blind processing.

[0025] SS primarily focuses on adaptive VMD decomposition: addressing the issues of traditional VMD requiring manual parameter setting and being prone to aliasing, it uses a narrowband interference frequency set as the initial center frequency to drive VMD decomposition. Simultaneously, a joint criterion function is used to automatically optimize parameters, ultimately separating three types of modes: narrowband interference modes containing narrowband interference and power frequency harmonics; low-frequency trend modes containing baseline drift and low-frequency noise; and candidate partial discharge modes containing core partial discharge characteristics, thus separating interference from effective signals at the source.

[0026] S3 primarily addresses pulse-time correction wavelet denoising: Traditional wavelet denoising relies on empirical thresholds, which are prone to peak clipping and distortion. This technical solution uses a set of suspected pulse times to dynamically correct the threshold: increasing the threshold sensitivity during suspected partial discharge pulse times (to prevent noise misjudgment) and decreasing the threshold during non-pulse times (to prevent signal clipping), thus denoising candidate partial discharge modes, suppressing random noise while preserving transient peaks of partial discharge to the greatest extent.

[0027] S4 mainly involves mode reconstruction and residual construction: After removing narrowband interference modes and low-frequency trend modes, the denoised candidate partial discharge modes are superimposed and reconstructed to obtain an intermediate sharpened signal. Considering that decomposition and denoising may miss weak partial discharge signals or residual interference, the difference between the bandpass signal and the intermediate sharpened signal is calculated to generate a residual signal, which supplements the missing information for subsequent fine-tuning.

[0028] S5 primarily refines deep residual networks: addressing the issues of poor generalization and difficulty in interpretation in end-to-end deep networks, it inputs both the intermediate clarified signal (containing core features) and the residual signal (containing supplementary information) into the deep residual network. The intermediate signal provides prior information on partial discharge (PD) features, while the residual signal guides the network in processing residual problems. Combined with the skip connections in the residual blocks, it enhances interference suppression while avoiding PD feature distortion, thus integrating the interpretability of traditional methods with the fitting capabilities of deep learning.

[0029] S6 and S7 are mainly for quality assessment and closed-loop adjustment: pulse detection (to verify integrity) and quality assessment (to check signal-to-noise ratio, fidelity, and spectral consistency) are performed on the final clear signal. Then, based on the assessment results, uncertainty estimation is performed, and errors in the quantification, decomposition, and denoising processes are decomposed. The detection threshold is dynamically adjusted to achieve alarm classification, which ensures signal reliability and adapts to the response requirements of different insulation defects on site.

[0030] To further explain, in S1, the preprocessing and standardization specifically includes the following steps: S11. Subtract the mean from the original partial discharge observation signal to remove the DC component, and then select a bandpass filter of the corresponding frequency band according to the sensor type to filter it to obtain the bandpass signal. The ultra-high frequency sensing channel corresponds to the 300MHz-1.5GHz frequency band, the ultrasonic sensing channel corresponds to the 20kHz-200kHz frequency band, and the current pulse channel corresponds to the 100kHz-20MHz frequency band. S12. The power spectrum of the bandpass signal is estimated using the Welch method. Narrowband peak regions with continuous frequency points in the power spectrum whose amplitude is more than 3dB higher than the average value of the whole spectrum and whose bandwidth is less than 1% of the sampling frequency are detected. The frequencies corresponding to these regions are divided into narrowband interference frequency sets. S13. Extract the time index of the envelope energy or instantaneous amplitude in the bandpass signal that exceeds twice the global mean, and form a set of suspected pulse moments.

[0031] Specifically, S11 mainly involves DC component removal and frequency band focusing: Since the original partial discharge observation signal is often mixed with a DC component, which has no partial discharge information but may affect the accuracy of subsequent signal analysis, the DC component is first removed by subtracting the signal mean. Different types of sensors (ultra-high frequency, ultrasonic, current pulse) have different sensitive frequency bands and partial discharge signal propagation characteristics, so dedicated bandpass filters need to be matched: the ultra-high frequency sensing channel selects the 300MHz-1.5GHz frequency band (adapting to the ultra-high frequency partial discharge pulse propagation characteristics), the ultrasonic sensing channel selects the 20kHz-200kHz frequency band (matching the ultrasonic partial discharge signal frequency band), and the current pulse channel selects the 100kHz-20MHz frequency band (fitting the current pulse partial discharge signal range). By filtering, the effective partial discharge frequency band signal is retained, and irrelevant interference outside the frequency band is filtered out to obtain the bandpass signal that focuses the core information.

[0032] S12 primarily focuses on the accurate identification of narrowband interference frequencies: Since narrowband radio interference and other narrow-frequency interference may still exist within bandpass signals, this interference manifests as local peaks in the power spectrum. The Welch method (using segmented windowing and overlapping sampling to improve power spectrum estimation accuracy) is employed to analyze the power spectrum of the bandpass signal. A dual criterion is set: "continuous frequency point amplitude is more than 3dB above the full spectrum average" (ensuring significant interference) and "bandwidth is less than 1% of the sampling frequency" (defining narrowband characteristics). This allows for the detection and locking of narrowband peak regions in the power spectrum. The frequencies corresponding to these regions are then integrated into a set of narrowband interference frequencies, providing frequency priors for subsequent VMD decomposition and accurate separation of interference modes.

[0033] S13 primarily focuses on locating suspected pulse moments: Partial discharge (PD) signals are transient pulse signals, whose envelope energy or instantaneous amplitude is significantly higher than the background noise. By extracting time points in the bandpass signal where "envelope energy exceeds twice the global mean" or "instantaneous amplitude exceeds twice the global mean," the indices corresponding to these time points are integrated into a set of suspected pulse moments. This clarifies the possible time range of PD signals, providing a locational basis for protecting PD spikes and avoiding clipping in the subsequent wavelet denoising stage.

[0034] To further explain, in S2, the optimization objective of the variational mode decomposition is: ; The constraints are: ; in, For modal number, Represents the time variable The first derivative is calculated. The impulse function is used to construct the Hilbert transform kernel. The imaginary unit, For the time variable corresponding to the discrete sampled signal, For the first A narrowband eigenmode, The center frequency corresponding to the narrowband eigenmode. It is a bandpass signal; The joint criterion function for: ; in, As a bandwidth penalty factor, For pulse kurtosis, The spectral entropy of the modal center frequency distribution. Modal overlap rate, weighting coefficient ∈[0,1] and satisfy Through multiple sets of comparative experiments, the signal-to-noise ratio improvement and pulse reconstruction error under different weight combinations were calculated, and the weight combination with the best overall performance was selected. The narrowband interference mode is one whose center frequency is located in the neighborhood of the narrowband interference frequency set and whose bandwidth satisfies the following conditions. The narrowband eigenmodes, among which, For modal bandwidth, The sampling frequency; The low-frequency trend mode is a narrowband intrinsic mode in which the main frequency energy is concentrated in the lowest 5% of the bandpass signal power spectrum or the instantaneous energy change rate is less than 10% of the global mean change rate. The candidate partial discharge modes are narrowband intrinsic modes other than narrowband interference modes and low-frequency trend modes.

[0035] First, by constructing the objective function and constraints, the bandpass signal is... Decomposed into several with definite center frequencies Narrowband eigenmodes The optimization objective is to minimize the sum of the bandwidths of all modes after the Hilbert transform, ensuring that the frequencies of each mode are concentrated; the constraint condition forces that the superposition of all modes equals the original bandpass signal. This ensures the integrity of the decomposition and avoids the loss of partial discharge energy.

[0036] Secondly, through the joint criterion function Dynamically determine the optimal number of modes With bandwidth penalty factor Among them, pulse kurtosis Spectral entropy measures the sharpness of partial discharge pulses in a mode (higher kurtosis indicates a greater probability of containing partial discharge information). The modal overlap rate reflects the energy concentration of the modal center frequency (the lower the spectral entropy, the more concentrated the frequency distribution, and the better the decomposition effect). Assess the independence of different modes (lower overlap indicates more thorough mode separation); weighting coefficients. Through multiple sets of comparative experiments (using signal-to-noise ratio improvement and pulse reconstruction error as evaluation indicators), we ensure that the criterion function can comprehensively balance the three major requirements of "partial discharge feature preservation, frequency concentration, and mode independence", thus avoiding the blindness of traditional VMD that relies on manually set parameters.

[0037] Finally, combining the narrowband interference frequency set obtained from S1 with the mode characteristics themselves, the narrowband eigenmodes obtained from the decomposition are... Classification: Narrowband interference modes must simultaneously satisfy the following conditions: "the center frequency is located in the neighborhood of the narrowband interference frequency set" (matching known interference frequencies) and "bandwidth". (Conforming to narrowband characteristics); low-frequency trend modes must meet the following conditions: "the main frequency energy is concentrated in the lowest 5% of the bandpass signal power spectrum" (energy is distributed in the low frequency range, with no partial discharge information) or "the instantaneous energy change rate is less than 10% of the global average change rate" (the signal is stable and has no transient pulse characteristics); candidate partial discharge modes are the remaining narrowband intrinsic modes after excluding the above two types of interference modes, that is, the effective modes containing partial discharge information are initially screened out.

[0038] To further explain, in S3, the wavelet threshold denoising specifically includes the following steps: S31. Perform discrete wavelet transform on the candidate partial discharge modes to obtain approximation coefficients. and detail coefficient ; S32. Calculate the noise standard deviation of the detail coefficients using the median absolute deviation. : ; in, For the first The candidate localization mode The detail factor of the layer, The candidate localization mode number. The number of decomposition levels; S33. Set hierarchical thresholds based on noise standard deviation. : ; in, For the first Number of layer coefficients; S34, Combining the pulse suspected time set The hierarchical threshold is corrected by adjusting the formula. : ; S35. Apply the corrected hierarchy threshold to the detail coefficients. Soft thresholding is performed to obtain the detail coefficients after denoising. Combined with approximation coefficients Denoising mode components are obtained from reconstruction. .

[0039] Specifically, S31 mainly involves wavelet coefficient decomposition: performing a discrete wavelet transform (DWT) on each candidate partial discharge mode to decompose the modal signal into approximate coefficients. and detail coefficient : where the approximation coefficient Low-frequency global characteristics reflecting the mode (including partial discharge pulse information), detail coefficients It reflects the high-frequency local characteristics of the mode (mainly including noise and partial discharge pulse edge details), and achieves a preliminary distinction between noise and effective information through coefficient decomposition. S32 is primarily for accurate estimation of noise standard deviation: based on detail coefficients. Based on the statistical characteristics, the median absolute deviation (MAD) was used to calculate the noise standard deviation of each decomposition layer. Because noise exhibits a Gaussian distribution in the wavelet domain, it accounts for a high proportion of the detail coefficients. This can be addressed using the formula... , where 0.6745 is the coefficient corresponding to the 0.75 quantile of the standard normal distribution, which can robustly estimate the noise intensity of each layer without the need for prior knowledge of noise, and provide a quantitative basis for subsequent threshold setting.

[0040] S33 primarily sets the threshold for the basic level: based on the estimated noise standard deviation. Number of coefficients in each layer Through formula Set the threshold for the basic level. This formula is derived from the "asymptotic optimal criterion for threshold selection under Gaussian noise" in statistics. The threshold can be dynamically adjusted according to the noise intensity and number of coefficients of each level to ensure that the noise suppression strength of different decomposition levels is adapted to their noise levels.

[0041] S34 is primarily for pulse-protective threshold correction: combining the pulse potential time set obtained from S1. By correcting the formula Threshold adjustment: During non-impulse-like moments, maintain the baseline threshold to effectively suppress noise; during impulse-like moments, adjust the threshold using the coefficient ( Lowering the threshold avoids the loss of high-frequency details (included in the detail coefficient) at the edge of the partial discharge pulse due to excessive thresholding, thus achieving a balance between "noise suppression" and "pulse fidelity".

[0042] S35 is primarily for noise reduction and reconstruction: it uses a corrected hierarchical threshold to adjust the detail coefficients. Perform soft thresholding (i.e., set coefficients with absolute values ​​less than the threshold to zero, and subtract the threshold from coefficients with absolute values ​​greater than the threshold) to obtain the denoised detail coefficients. Then adjust the detail coefficient after noise reduction. Compared with unprocessed approximation coefficients Performing an inverse wavelet transform, the reconstructed denoised mode components with significantly reduced noise and complete preservation of partial discharge characteristics are obtained. .

[0043] To further explain, in S4, the narrowband interference mode and low-frequency trend mode are removed, and the reconstructed denoised mode components are superimposed. Obtain intermediate sharpening signal : ; And calculate the bandpass signal With intermediate clarity signal The difference is used to construct the residual signal. : .

[0044] Specifically, firstly, two types of interference modes without effective partial discharge information (narrowband interference mode and low-frequency trend mode) are completely excluded from all decomposed modes, retaining only the denoised mode components processed by S3. To avoid interference from being introduced into subsequent signal reconstruction processes, all signals are then aligned according to the timeline principle. Direct superposition, using formulas Reconstruction yields intermediate clarified signal This process focuses on effective partial discharge information, integrates partial discharge characteristics from various modes, and achieves preliminary signal clarification after interference suppression. Finally, the bandpass signal obtained in S1 is used... Based on the formula The amplitude difference between the two at corresponding time points is calculated to construct the residual signal r(t). The essence of the residual signal is a comprehensive reflection of the weak interference not completely removed by the previous "VMD decomposition-wavelet denoising-effective mode reconstruction" process, the residual error of mode decomposition, and the subtle loss of partial discharge signal. It accurately captures the "missed" interference and signal deviation that were not resolved by the previous processing, and provides a clear target for the targeted correction of the S5 deep residual network.

[0045] To further explain, in S5, the deep residual network adopts a one-dimensional convolutional neural network or 1DU-Net structure, and the deep residual network includes an input layer, an encoder, a decoder and an output layer. The input layer receives the intermediate decryption signal. With residual signal A two-dimensional vector; The encoder contains four residual blocks. Each residual block consists of two one-dimensional convolutional layers with 3×1 convolutional kernels and stride of 1, one batch normalization layer, and one ReLU activation function. Feature downsampling is performed at each encoding stage. The decoder contains four upsampling layers and has a structure symmetrical to the encoder; The output layer uses the Sigmoid function to output the final clear signal. :

[0046] in, For the trained deep residual network; The training process of the deep residual network is as follows: Using simulated signals of corona discharge, surface discharge, and internal discharge as samples, with a total sample size ≥ 10,000, the data is augmented by time axis flipping and noise injection with a signal-to-noise ratio of -10dB to 5dB. Training is performed using the Adam optimizer, with an initial learning rate set to 0.001 and a loss function of:

[0047] in, For the final clear signal, For an ideal pulse signal, For envelope function or energy operator, For pulse kurtosis, The pulse arrival time, , , and These are the weighting coefficients.

[0048] First, the input layer of the deep residual network receives the "intermediate sharpening signal". With residual signal A two-dimensional vector composed of [variables], where the intermediate sharpened signal [is present]. Provide the partially denoised main characteristics of the partial discharge as the basis for correction; residual signal It provides information on subtle interferences and signal deviations that were not completely removed in the preceding sequence, providing a clear basis for the network's targeted learning error and avoiding blind fitting by the network.

[0049] Next, the deep residual network adopts a one-dimensional convolutional neural network or 1DU-Net structure to adapt to the temporal characteristics of the partial discharge signal: The encoder consists of four residual blocks, each constructed using two 3×1 convolutional kernels (with a stride of 1), a batch normalization layer, and a ReLU activation function. The 3×1 convolutional kernels capture local temporal features while reducing parameter redundancy, the batch normalization layer accelerates training convergence, and the ReLU activation function enhances nonlinear fitting capabilities. Each encoding stage compresses the feature map size through downsampling (e.g., pooling) to extract deep features of global interference and signal bias.

[0050] Decoder: Contains 4 upsampling layers symmetrical to the encoder (such as deconvolution and linear interpolation), which gradually restore the feature map size and map the deep features extracted by the encoder back to the original signal dimension, ensuring that the output signal is time-aligned with the input signal.

[0051] Output layer: The Sigmoid function is used to normalize the output amplitude to the [0,1] interval to ensure a clear final signal. Amplitude stability.

[0052] Finally, the deep residual network is trained using a multi-constraint training strategy to obtain the trained deep residual network: Sample and data augmentation: Using three typical partial discharge simulation signals (corona discharge, surface discharge, and internal discharge) as samples (total ≥10,000 groups), the sample diversity is expanded by time axis flipping (enhancing timing robustness) and -10dB to 5dB noise injection (simulating complex noise environment on site) to avoid network overfitting.

[0053] Optimizer and learning rate: The Adam optimizer (adaptive learning rate adjustment) is used, with an initial learning rate of 0.001 to balance training speed and convergence stability.

[0054] Multi-dimensional loss function: Constraining the output signal through the loss function L — (L1 norm) guarantees amplitude fidelity. (The L1 norm of the spectrum) ensures spectrum consistency. (Krunic difference) ensures the sharpness of the pulse shape. (Temporal L1 norm) ensures accurate pulse arrival time; weighting coefficient , , and Adjust the importance of each constraint to ensure that the network output takes into account the four core characteristics of "amplitude-spectrum-morphology-timing".

[0055] To further explain, in S6, the pulse detection uses envelope analysis or Teager energy operator analysis to extract the pulse arrival time of the final clear signal; The quality assessment uses the signal-to-noise ratio improvement, pulse arrival time error, morphological peak signal-to-noise ratio, and SSIM as indicators.

[0056] Regarding pulse detection: Envelope analysis detection: for the final clear signal Envelope extraction (such as Hilbert transform or sliding window to find extrema) is performed to obtain the amplitude envelope curve of the signal; an adaptive threshold is set (such as 1.5-2 times the mean of the envelope). When the envelope curve exceeds the threshold, it is the arrival time of the partial discharge pulse. This method is suitable for partial discharge signals with significant amplitude differences.

[0057] Teager energy operator analysis and detection: using the Teager energy operator ,in, For the final clear signal, These are the first and second derivatives, respectively, used to calculate the instantaneous energy of the signal. This operator is highly sensitive to changes in the energy of transient pulses. When the instantaneous energy exceeds a set threshold, the corresponding moment is the pulse arrival time. It is suitable for weak pulses or partial discharge signals with superimposed subtle noise.

[0058] Regarding multi-dimensional quality assessment: Signal-to-noise ratio improvement (ΔSNR): Calculates the final sharp signal The difference between the signal-to-noise ratio (SNR) of the extracted signal and the SNR of the original observed signal quantifies the overall noise suppression effect; a larger difference indicates more thorough interference removal. Pulse arrival time error: The arrival time of the extracted final clear signal pulse is compared with the arrival time of the ideal pulse signal. By comparing the actual arrival times of the partial discharge pulses, the mean absolute value of the time deviation is calculated to evaluate the fidelity of the timing characteristics. The smaller the error, the more accurately the timing information of the partial discharge pulses is preserved.

[0059] Peak Signal-to-Noise Ratio (PSNR): The maximum amplitude of the final sharp signal is given by the formula, and the MSE is... and The mean square error (PSNR) is used to assess the similarity of signal amplitude shapes; a higher PSNR value indicates less amplitude distortion.

[0060] Structural Similarity Index (SSIM): Calculated from three dimensions: brightness, contrast, and structure. and The similarity (value range [0,1]) indicates that the closer SSIM is to 1, the more consistent the overall structure of the final clear signal (such as pulse rising edge, falling edge, and peak shape) is with the ideal signal, and the higher the morphological fidelity.

[0061] To further explain, in S7, the uncertainty estimate is performed by estimating the variance of the final clear signal through Monte Carlo Dropout or deep model ensemble. :

[0062] in, For the first The prediction results of the random sampling For the number of samples, Output the mean of all samples; The process of adjusting the detection threshold is as follows: setting a baseline detection threshold. Calculate the confidence correction factor based on the uncertainty. :

[0063] in, With an adjustment coefficient of 0.2-0.5, the final detection threshold is... ; The alarm classification criteria are as follows: Level I is defined as variance < 0.5 times the global mean, Level II is defined as variance ≤ 0.5 times the global mean and Level III is defined as variance > 1 times the global mean.

[0064] Regarding the quantitative estimation of uncertainty: Monte Carlo Dropout or deep model ensemble methods are used to calculate the final clear signal. variance To characterize uncertainty: through Multiple random samplings (such as Monte Carlo Dropout randomly shutting down some neurons during network inference, or deep model ensemble calling multiple trained sub-models) yield M prediction results. ; through formula It quantifies the fluctuation of signal output under different sampling conditions. The larger the variance, the lower the confidence of the signal, and vice versa.

[0065] Regarding dynamic adjustment of the detection threshold: based on the baseline threshold for pulse detection. Based on this, the confidence correction coefficient is calculated in conjunction with the uncertainty. Normalize the variance to the [0,1] interval, and adjust the coefficient. (0.2-0.5) Control the threshold adjustment range to achieve "the higher the uncertainty, the higher the correction coefficient". The logic of "the larger the value", the higher the final detection threshold. As uncertainty increases dynamically, it avoids false alarms caused by low-confidence signals.

[0066] Regarding the alarm classification logic: Based on the global variance mean, alarms are divided into three levels: Signals with a variance < 0.5 times the global mean are classified as Level I (high confidence), indicating stable signal output, minimal interference, and high reliability of the detection results; Signals with a variance ≤ 0.5 times the global mean and ≤ 1.5 times the global mean are classified as Level II (medium confidence), indicating some fluctuation in the signal, requiring further analysis based on other operating conditions; Signals with a variance > 1 times the global mean are classified as Level III (low confidence), indicating significant signal fluctuation, high risk of the detection results, and requiring retesting or manual review.

[0067] In summary, this invention can effectively recover the morphological and temporal characteristics of partial discharge transient pulses under complex electromagnetic and environmental interference conditions. Numerical simulations and typical example analyses show that this method can effectively recover the morphological and temporal characteristics of partial discharge transient pulses under different initial signal-to-noise ratios (SNRs). Within the range of 15dB to 5dB, compared to traditional wavelet thresholding denoising and single VMD methods, the signal-to-noise ratio improvement can be stably maintained within the range of 6–12dB, with a pulse arrival time error of less than 0.5μs. In typical power frequency harmonic and random pulse interference scenarios, the false alarm rate and false negative rate are reduced by approximately 20%–35% compared to traditional methods, respectively, significantly suppressing noise while maintaining the fidelity of pulse timing characteristics.

[0068] To verify the above performance, numerical simulations were performed under different signal-to-noise ratio conditions, and the results are shown in the table below (the data are used to illustrate the performance trend):

[0069] This invention can significantly improve the signal-to-noise ratio, reduce false alarm and false negative rates, and maintain the pulse timing characteristics without distortion. As can be seen from the table, the method of this invention can still maintain high signal reconstruction quality and detection accuracy under low signal-to-noise ratio conditions, which verifies the stability and robustness of the algorithm in multi-noise and multi-interference scenarios.

[0070] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0071] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A partial discharge signal clarification method based on VMD decomposition-wavelet denoising-deep refinement, characterized in that, The method comprises the following steps: S1, preprocessing and standardizing the original partial discharge observation signal to obtain a passband signal, a narrow-band interference frequency set and a pulse suspected time set; S2, performing variational mode decomposition on the passband signal with the narrow-band interference frequency set as the initial center frequency, and adaptively selecting the decomposition parameters and identifying the narrow-band interference mode, low-frequency trend mode and candidate partial discharge mode by combining the joint criterion function; S3, correcting the threshold based on the pulse suspected time set, and performing wavelet threshold denoising on the candidate partial discharge mode to obtain a denoised mode component; S4, removing the narrow-band interference mode and the low-frequency trend mode, superimposing the reconstructed denoised mode component to obtain an intermediate clear signal, and calculating the difference between the passband signal and the intermediate clear signal to construct a residual signal; S5, inputting the intermediate clear signal and the residual signal into a deep residual network, and outputting a final clear signal after network processing; S6, performing pulse detection and quality evaluation on the final clear signal; S7, adjusting the detection threshold and completing the alarm grading in combination with the uncertainty estimation.

2. The partial discharge signal clarification method based on VMD decomposition-wavelet denoising-deep refinement according to claim 1, characterized in that, In S1, the preprocessing and standardization specifically comprises the following steps: S11, subtracting the mean value of the original partial discharge observation signal to remove the direct current component, and then selecting a band-pass filter of a corresponding frequency band according to the sensor type to filter, to obtain a passband signal; wherein the ultra-high frequency sensor channel corresponds to a 300MHz-1.5GHz frequency band, the ultrasonic sensor channel corresponds to a 20kHz-200kHz frequency band, and the current pulse channel corresponds to a 100kHz-20MHz frequency band; S12, estimating the power spectrum of the passband signal by using the Welch method, detecting the narrow-band peak region in the power spectrum where the amplitude of the continuous frequency points is higher than 3dB of the average value of the whole spectrum and the bandwidth is less than 1% of the sampling frequency, and dividing the frequencies corresponding to the region into a narrow-band interference frequency set; S13, extracting the time index where the envelope energy or instantaneous amplitude of the passband signal exceeds twice the global average value to form a pulse suspected time set.

3. The partial discharge signal clarification method based on VMD decomposition-wavelet denoising-deep refinement of claim 1, characterized in that, In S2, the optimization target of the variational mode decomposition is: ； The constraint condition is: ； wherein, is the modal number, denotes the first derivative with respect to the time variable , is the impulse function used to construct the Hilbert transform kernel, is the imaginary unit, is the time variable corresponding to the discrete sampled signal, is the narrowband eigenmode, is the center frequency corresponding to the narrowband eigenmode, is the bandpass signal; The joint criterion function is: ； wherein, is a bandwidth penalty factor, is a pulse kurtosis, is a spectral entropy of the modal center frequency distribution, is a modal overlap ratio, a weight coefficient ∈ [0, 1] and satisfies The signal-to-noise ratio improvement and pulse reconstruction error under different weight combinations are calculated through multiple sets of comparative experiments, and the weight combination with the best comprehensive performance is selected. The narrowband interference modal is a narrowband eigenmodal with a center frequency located in the neighborhood of the set of narrowband interference frequencies and a bandwidth satisfying , is the modal bandwidth, is the sampling frequency; The low-frequency trend mode is a narrow-band intrinsic mode whose main frequency energy is concentrated in the lowest 5% frequency band of the passband signal power spectrum or whose instantaneous energy change rate is lower than 10% of the global average change rate; The candidate partial discharge mode is a narrow-band intrinsic mode other than the narrow-band interference mode and the low-frequency trend mode.

4. The partial discharge signal clarification method based on VMD decomposition-wavelet denoising-deep refinement of claim 3, characterized in that, In S3, the wavelet threshold denoising specifically comprises the following steps: S31, performing discrete wavelet transform on the candidate bruit modality to obtain approximation coefficients and detail coefficients ; S32, calculate the noise standard deviation of the detail coefficient by the median absolute deviation : ； wherein, is the number of candidate local modes, is the detail coefficient of the first layer of the candidate local mode, is the candidate local mode number, is the number of decomposition layers; S33, setting a level threshold based on the noise standard deviation : ； wherein is the first number of layer coefficients; S34, combine the pulse suspected time set , correct the tier threshold by the correction formula : ； S35, using the corrected hierarchical threshold to the detail coefficient performing soft threshold shrinkage to obtain the denoised detail coefficient , and combining the approximate coefficient reconstructing to obtain the denoised modal component .

5. The partial discharge signal clarification method based on VMD decomposition-wavelet denoising-deep refinement of claim 4, characterized in that, In S4, the narrowband interference modal and the low frequency trend modal are rejected, and the reconstructed denoised modal component is superimposed An intermediate sharpened signal is obtained : ； and calculating a difference between the bandpass signal and the intermediate clarifying signal to construct a residual signal : 。 6. The partial discharge signal clarification method based on VMD decomposition-wavelet denoising-depth refinement of claim 5, characterized in that, In S5, the deep residual network adopts a one-dimensional convolutional neural network or a 1DU-Net structure, and the deep residual network comprises an input layer, an encoder, a decoder and an output layer; The input layer receives an intermediate sharpening signal with the residual signal comprising a two-dimensional vector; The encoder comprises 4 residual blocks, each residual block is composed of 2 one-dimensional convolutional layers with a 3x1 convolution kernel and a step of 1, 1 batch normalization layer and 1 ReLU activation function, and feature down-sampling is performed at each encoding stage; The decoder comprises 4 up-sampling layers, and the structure is symmetrical to the encoder; The output layer adopts a Sigmoid function to output the final clear signal : wherein, is the trained deep residual network; The training process of the deep residual network is as follows: taking a corona discharge, a surface discharge, and an internal discharge simulation signal as a sample, a total number of samples is greater than or equal to 10,000 groups, the sample is enhanced by time axis flipping and signal-to-noise ratio (SNR) -10 dB to 5 dB noise injection, an Adam optimizer is used for training, an initial learning rate is set to 0.001, and a loss function is as follows: wherein is the final clean signal, is the ideal pulse signal, is the envelope function or energy operator, is the pulse kurtosis, is the pulse arrival time, , , and are weight coefficients.

7. The partial discharge signal clarification method based on VMD decomposition-wavelet denoising-deep refinement of claim 1, characterized in that, In S6, the pulse detection adopts envelope analysis or Teager energy operator analysis to extract a pulse arrival time of the final clear signal; The quality evaluation takes a signal-to-noise ratio (SNR) improvement amount, a pulse arrival time error, a shape peak SNR, and an SSIM as indexes.

8. The partial discharge signal clarification method based on VMD decomposition-wavelet denoising-deep refinement of claim 6, characterized in that, In S7, the uncertainty estimate estimates the variance of the final clean signal by Monte Carlo Dropout or deep model ensemble : wherein, is the prediction result of the is the number of samples, is the mean output of all samples; The process of adjusting the detection threshold is: setting a reference detection threshold , calculating a confidence correction coefficient according to the uncertainty : wherein, a scaling factor of 0.2-0.5, the final detection threshold being ; The alarm grading standard is that variance less than 0.5 times a global mean value is classified as grade I, 0.5 times to 1.5 times the global mean value is classified as grade II, and variance greater than 1 times the global mean value is classified as grade III.

Images