A hierarchical frequency reduction and sparsification real-time processing method and system for partial discharge signals

Abstract

The application discloses a kind of partial discharge signal grading frequency reduction and sparse real-time processing method and system, belong to electric power equipment insulation on-line monitoring technical field.In method, first, the original sampling signal is carried out anti-aliasing filter and is reduced sampling with first downsampling factor, obtains the signal in downsampling domain;Baseline estimation is carried out in downsampling domain, extracts pulse component and calculates adaptive threshold;Candidate pulse detection is carried out based on pulse component and adaptive threshold;Only the local window of each candidate pulse is carried out time-frequency domain multi-scale analysis to verify effective pulse;Effective pulse position is mapped back full sampling rate domain accurate positioning, and waveform feature verification is carried out in full sampling rate domain, and finally output effective pulse and its signal-to-noise ratio.The application gradually decreases data quantity along processing link, different processing steps are carried out in the sampling rate domain satisfying respective accuracy demand, without special hardware, real-time processing can be realized on general processor, and high detection accuracy and low computing expense are considered.

Description

Technical Field

[0001] This invention belongs to the field of power system online monitoring and signal processing technology, and relates to a method and system for real-time processing of graded frequency reduction and sparsification of partial discharge signals. Background Technology

[0002] Ultra-high frequency partial discharge (UHF-PD) online monitoring is an important means of assessing the insulation condition of high-voltage power equipment. Its core lies in the real-time analysis and pulse extraction of broadband transient signals acquired by a high-speed ADC. In actual operation, partial discharge pulses exhibit significant suddenness and randomness. Effective pulse detection not only relies on amplitude thresholds but also requires consideration of time-frequency domain characteristics to distinguish between genuine discharges and complex on-site environments such as periodic narrowband interference and random noise. To ensure the timeliness and effectiveness of diagnosis, the system needs to complete frame-by-frame processing of massive amounts of continuously sampled data within tens of minutes or even less, which places extremely high demands on the computational efficiency of the algorithm and the hardware processing capabilities.

[0003] Existing technologies face significant bottlenecks when processing such high-throughput data. When using a full-sampling-rate direct processing approach, the computational complexity of core algorithms such as continuous wavelet transform and morphological filtering is extremely high, with processing times for a single frame reaching several seconds under typical configurations, completely failing to meet the demands of real-time online monitoring. While GPU parallel acceleration and FPGA hardware pipelines can improve throughput through parallelization or hardware implementation, the former is limited by the power consumption, cost, and data transmission overhead of embedded monitoring environments, while the latter faces practical constraints such as long development cycles, poor algorithm iteration flexibility, and high floating-point resource consumption. While using discrete wavelet transform instead of continuous wavelet transform or introducing a fixed-rate downsampling simplification strategy can reduce computational load, it comes at the cost of sacrificing frequency resolution, scale selection flexibility, or pulse detection accuracy, making it difficult to balance efficiency and reliability.

[0004] In-depth analysis reveals that the common limitation of the aforementioned solutions lies in their adherence to a single sampling rate processing paradigm throughout the entire process. Regardless of the acceleration or simplification methods employed, each algorithm module must complete its computation on a unified raw sampling rate data stream, and the total computational load of the system is determined by the cumulative complexity of the serially connected modules. This fixed pattern results in a constant total amount of data throughout the process. There is a theoretical lower limit to shortening the time per step through hardware acceleration, while sacrificing speed through algorithm simplification inevitably compromises detection performance, creating an inherent contradiction between efficiency and accuracy.

[0005] To address this issue, this invention constructs a multi-level sampling rate adaptive processing architecture for UHF partial discharge monitoring, based on the differentiated requirements of different signal processing stages for data accuracy and information integrity. Summary of the Invention

[0006] The purpose of this invention is to solve the problems in the prior art and provide a method and system for real-time processing of graded frequency reduction and sparsification of partial discharge signals.

[0007] To achieve the above objectives, the present invention employs the following technical solution: A method for graded frequency reduction and sparsification of partial discharge signals in real time includes the following steps: After applying anti-aliasing filtering to the original sampled signal, downsampling is performed using the first downsampling factor to obtain the downsampled domain signal. Baseline estimation is performed on the downsampled domain signal, the pulse component signal is extracted, and an adaptive threshold is calculated on the downsampled domain signal; Based on the pulse component signal and adaptive threshold, candidate pulse detection is performed in the downsampling domain to obtain a set of candidate pulse positions; For each candidate pulse in the candidate pulse location set, perform time-frequency domain multi-scale analysis in the downsampling domain to verify and filter effective pulses; The positions of the selected valid pulses are mapped from the downsampling domain to the full sampling rate domain for precise location. The waveform characteristics of the precisely located pulse are verified across the entire sampling rate domain, and the final valid pulse is output. The signal-to-noise ratio of the validated pulses is calculated as an auxiliary quality indicator. The amount of data along the processing link signal decreases step by step, and different processing steps are executed in the sampling rate domain that meets their respective accuracy requirements.

[0008] The anti-aliasing filter uses a low-pass filter to filter the original signal. Perform filtering to obtain the filtered signal The cutoff frequency of the low-pass filter satisfy ,in The original sampling rate, The first downsampling factor; downsampling processing is performed on the filtered signal. conduct Decimation is performed to obtain the downsampled domain signal. ,in The original signal length, Group delay compensation for the anti-aliasing filter; the low-pass filter is a 4th-order Butterworth IIR low-pass filter with a cutoff frequency of The first downsampling factor The value range is 5 to 20.

[0009] The baseline estimation employs a morphological Top-Hat operation: in the downsampled domain signal... Perform one-dimensional morphological opening operation on ,in The width of the structural element. and These are the minimum and maximum values ​​within the sliding window, respectively; the result of the opening operation... Linear interpolation upsampling to the original length To obtain the baseline estimate Calculate pulse component signals And estimate the noise standard deviation from the opening operation results. ,in It is the absolute median difference function.

[0010] The calculation of the adaptive threshold includes: [the process involves] processing the pulse component signal... Divided into Each data block Calculate the mean value for each of the 100 sampling points. and standard deviation CUSUM change point detection is performed on the block mean sequence. When a change point is detected, the reference value is automatically updated, the block-level threshold is calculated, and the data is linearly interpolated and upsampled to a point-by-point threshold array. Block statistics are calculated in the downsampling domain, and the corresponding block size is... One downsampling point.

[0011] The candidate pulse detection includes: downsampling the pulse component signal to obtain the downsampled pulse component. Calculate the pre-screening threshold A peak detection algorithm is used to detect peak values ​​exceeding the pre-screening threshold in the downsampling domain, thereby obtaining a set of candidate pulse positions. ; Utilizing the corresponding value of the adaptive threshold in the downsampling domain Filter and retain only Candidates; set minimum peak spacing To ensure that adjacent candidate pulses do not overlap, if the distance between candidate pulses is less than... Only those with larger amplitudes are retained.

[0012] The time-frequency domain multi-scale analysis includes: for each candidate pulse Extracting a local window from the downsampled signal The half-width of the local window , For the largest analytical scale, These are the expansion coefficients; within a local window, FFT convolution is used for calculation. Morlet wavelet coefficients at each scale ,in For the first Wavelet mother function of scale, For scale parameters, Calculate the energy at each scale and perform scale entropy verification, then select the pulses that pass the verification.

[0013] The step of mapping the filtered valid pulse positions from the downsampling domain to the full sampling rate domain for precise positioning includes: mapping the pulse positions confirmed in the downsampling domain... Multiply by the first downsampling factor To obtain a rough position across the entire sampling rate domain. Within the full sampling rate domain of the original signal, the pulse component signal is searched for within the search range, centered on a coarse location. maximum value ,in The search radius is used; after precise positioning, the pulses are deduplicated. If the distance between two pulses is less than twice the minimum pulse width, only the one with the larger amplitude is retained; finally, an adaptive threshold is used for verification, retaining only the pulses with the larger amplitude. The pulse.

[0014] The waveform feature verification of the precisely located pulse across the full sampling rate domain includes: pulse component signals across the full sampling rate domain. Above, for each candidate pulse after precise positioning Cut length is Calculate the time-domain kurtosis using a local window. and frequency domain spectral flatness ,in , In the formula The sampled values ​​of the pulse components within the local window. and These are the mean and standard deviation of the sampled values ​​within the window, respectively. The Fourier spectrum amplitude of the local window signal is used; a joint criterion is employed. and Determine if a pulse is valid.

[0015] The formula for calculating the signal-to-noise ratio is as follows: ,in For the first time after baseline removal A local window waveform, Estimation of global noise standard deviation; first downsampling factor Adaptively adjusts based on the signal-to-noise ratio or pulse density: increases when the signal-to-noise ratio is higher than the first threshold or the pulse density is lower than the second threshold. To improve processing speed, the value is reduced when the signal-to-noise ratio is below the third threshold or the pulse density is above the fourth threshold. Values ​​are set to ensure detection accuracy; adaptive adjustment is performed before processing each frame of data. The value can be adjusted from 5 to 20.

[0016] A real-time processing system for graded frequency reduction and sparsification of partial discharge signals includes: The anti-aliasing filter module is used to perform anti-aliasing low-pass filtering on the original sampled signal; The graded downsampling module is used to downsample the filtered signal to the downsampling domain using a first downsampling factor; The baseline estimation module is used to perform morphological opening operations on the downsampled domain signal to estimate the baseline and subtract the baseline from the original signal to extract the pulse component; An adaptive threshold calculation module is used to calculate adaptive thresholds in the downsampling domain based on block statistics and CUSUM change point detection. The candidate pulse detection module is used to detect candidate pulses in the downsampling domain based on pulse components and adaptive thresholds, and to set a minimum peak spacing to handle overlapping pulses; The time-frequency domain analysis module is used to perform multi-scale Morlet wavelet transform by applying FFT convolution to a local window of each candidate pulse in the downsampling domain, and to filter valid pulses by verifying the scale entropy. The cross-domain mapping and positioning module is used to map the verified pulse position from the downsampling domain to the full sampling rate domain, and achieve precise pulse positioning through local search; The waveform verification module is used to perform joint verification of kurtosis and spectral flatness on the precisely located pulse across the entire sampling rate domain, and output the final valid pulse. The signal-to-noise ratio calculation module is used to calculate auxiliary quality indicators for verified pulses.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention reduces the total computational load from the extremely high levels of traditional solutions to a level that can be handled in real time by a general-purpose processor through a collaborative architecture of hierarchical downsampling and progressive sparsification. Compared with existing solutions that require GPU parallel acceleration or FPGA hardware pipelines, no additional hardware is required, significantly reducing equipment cost, power consumption, and deployment complexity, making it suitable for various embedded online monitoring environments.

[0018] This invention employs a core strategy of performing as many calculations as possible at the coarsest sampling rate and performing the most refined calculations with the least amount of data. In the downsampling domain, coarse-grained processing such as morphological baseline estimation, CUSUM threshold calculation, and candidate pulse pre-screening are performed. The computationally intensive CWT multi-scale analysis is performed only within a local window of candidate pulses, while operations sensitive to waveform details, such as kurtosis and spectral flatness verification, are still performed in the full sampling rate domain. The detection accuracy is not significantly different from that of direct processing at the full sampling rate, truly achieving a balance between efficiency and accuracy.

[0019] This invention progressively reduces the data size along the processing chain: the original data is downsampled to one-tenth of its original size; block statistics compress the data to scalar level; candidate pre-screening reduces the input for subsequent computationally intensive operations to a very small number of candidate positions; CWT local windows further compress the amount of data actually involved in time-frequency analysis to an extremely low level; finally, full sampling rate domain verification processes only a very small number of valid pulses within a local window. The data size of each module decreases progressively, and the most computationally intensive operation always deals with the least amount of data, fundamentally compressing the overall computational scale. Furthermore, each module can be independently upgraded and replaced, exhibiting good scalability. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the real-time processing method for graded frequency reduction and sparsification of partial discharge signals according to the present invention. Figure 2 This is a structural diagram of the real-time processing system for graded frequency reduction and sparsification of partial discharge signals according to the present invention. Detailed Implementation

[0022] 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. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1This invention provides a method for graded frequency reduction and sparsification of partial discharge signals in real time, comprising the following steps: After applying anti-aliasing filtering to the original sampled signal, it is downsampled using a first downsampling factor to obtain the downsampled domain signal; the anti-aliasing filtering uses a low-pass filter on the original signal. Perform filtering to obtain the filtered signal The cutoff frequency of the low-pass filter satisfy ,in The original sampling rate, The first downsampling factor; downsampling processing is performed on the filtered signal. conduct Decimation is performed to obtain the downsampled domain signal. ,in The original signal length, Group delay compensation for the anti-aliasing filter; the low-pass filter is a 4th-order Butterworth IIR low-pass filter with a cutoff frequency of The first downsampling factor The value range is 5 to 20.

[0026] Baseline estimation is performed on the downsampled domain signal to extract the pulse component signal, and an adaptive threshold is calculated on the downsampled domain signal; the baseline estimation employs morphological Top-Hat operation: on the downsampled domain signal... Perform one-dimensional morphological opening operation on ,in The width of the structural element. and These are the minimum and maximum values ​​within the sliding window, respectively; the result of the opening operation... Linear interpolation upsampling to the original length To obtain the baseline estimate Calculate pulse component signals And estimate the noise standard deviation from the opening operation results. ,in It is the absolute median difference function.

[0027] The calculation of the adaptive threshold includes: [the process involves] processing the pulse component signal... Divided into Each data block Calculate the mean value for each of the 100 sampling points. and standard deviation CUSUM change point detection is performed on the block mean sequence. When a change point is detected, the reference value is automatically updated, the block-level threshold is calculated, and the data is linearly interpolated and upsampled to a point-by-point threshold array. Block statistics are calculated in the downsampling domain, and the corresponding block size is... One downsampling point.

[0028] Based on the pulse component signal and an adaptive threshold, candidate pulse detection is performed in the downsampled domain to obtain a set of candidate pulse positions; the candidate pulse detection includes: downsampling the pulse component signal to obtain the pulse components in the downsampled domain. Calculate the pre-screening threshold A peak detection algorithm is used to detect peak values ​​exceeding the pre-screening threshold in the downsampling domain, thereby obtaining a set of candidate pulse positions. ; Utilizing the corresponding value of the adaptive threshold in the downsampling domain Filter and retain only Candidates; set minimum peak spacing To ensure that adjacent candidate pulses do not overlap, if the distance between candidate pulses is less than... Only those with larger amplitudes are retained.

[0029] For each candidate pulse in the candidate pulse position set, a time-frequency domain multi-scale analysis is performed in the downsampling domain to verify and filter valid pulses; the time-frequency domain multi-scale analysis includes: for each candidate pulse Extracting a local window from the downsampled signal The half-width of the local window , For the largest analytical scale, These are the expansion coefficients; within a local window, FFT convolution is used for calculation. Morlet wavelet coefficients at each scale ,in For the first Wavelet mother function of scale, For scale parameters, Calculate the energy at each scale and perform scale entropy verification, then select the pulses that pass the verification.

[0030] The filtered valid pulse positions are mapped from the downsampling domain to the full sampling rate domain for precise location; this includes: mapping the pulse positions confirmed in the downsampling domain. Multiply by the first downsampling factor To obtain a rough position across the entire sampling rate domain. Within the full sampling rate domain of the original signal, the pulse component signal is searched for within the search range, centered on a coarse location. maximum value ,in The search radius is used; after precise positioning, the pulses are deduplicated. If the distance between two pulses is less than twice the minimum pulse width, only the one with the larger amplitude is retained; finally, an adaptive threshold is used for verification, retaining only the pulses with the larger amplitude. The pulse.

[0031] The waveform characteristics of the precisely located pulse are verified across the entire sampling rate domain, and the final valid pulse is output, including the pulse component signal across the entire sampling rate domain. Above, for each candidate pulse after precise positioning Cut length is Calculate the time-domain kurtosis using a local window. and frequency domain spectral flatness ,in , In the formula The sampled values ​​of the pulse components within the local window. and These are the mean and standard deviation of the sampled values ​​within the window, respectively. The Fourier spectrum amplitude of the local window signal is used; a joint criterion is employed. and Determine if a pulse is valid.

[0032] The signal-to-noise ratio (SNR) of the validated pulses is calculated as an auxiliary quality indicator; the formula for calculating the SNR is as follows: ,in For the first time after baseline removal A local window waveform, Estimation of global noise standard deviation; first downsampling factor Adaptively adjusts based on the signal-to-noise ratio or pulse density: increases when the signal-to-noise ratio is higher than the first threshold or the pulse density is lower than the second threshold. To improve processing speed, the value is reduced when the signal-to-noise ratio is below the third threshold or the pulse density is above the fourth threshold. Values ​​are set to ensure detection accuracy; adaptive adjustment is performed before processing each frame of data. The value can be adjusted from 5 to 20.

[0033] The amount of data along the processing link signal decreases step by step, and different processing steps are executed in the sampling rate domain that meets their respective accuracy requirements.

[0034] like Figure 2 The diagram shown is a module structure diagram of a real-time processing system for graded frequency reduction and sparsification of partial discharge signals according to the present invention, specifically including: The anti-aliasing filter module is used to perform anti-aliasing low-pass filtering on the original sampled signal; The graded downsampling module is used to downsample the filtered signal to the downsampling domain using a first downsampling factor; The baseline estimation module is used to perform morphological opening operations on the downsampled domain signal to estimate the baseline and subtract the baseline from the original signal to extract the pulse component; An adaptive threshold calculation module is used to calculate adaptive thresholds in the downsampling domain based on block statistics and CUSUM change point detection. The candidate pulse detection module is used to detect candidate pulses in the downsampling domain based on pulse components and adaptive thresholds, and to set a minimum peak spacing to handle overlapping pulses; The time-frequency domain analysis module is used to perform multi-scale Morlet wavelet transform by applying FFT convolution to a local window of each candidate pulse in the downsampling domain, and to filter valid pulses by verifying the scale entropy. The cross-domain mapping and positioning module is used to map the verified pulse position from the downsampling domain to the full sampling rate domain, and achieve precise pulse positioning through local search; The waveform verification module is used to perform joint verification of kurtosis and spectral flatness on the precisely located pulse across the entire sampling rate domain, and output the final valid pulse. The signal-to-noise ratio calculation module is used to calculate auxiliary quality indicators for verified pulses.

[0035] This invention proposes a real-time processing method for ultra-high frequency partial discharge signals based on hierarchical downsampling and progressive sparsification. This method comprises two synergistic architectural strategies: hierarchical downsampling, where different processing modules execute at different sampling rate domains according to their accuracy requirements; and progressive sparsification, which uses a funnel-shaped filtering pipeline to gradually reduce the computational load at each step. Specific measures include: Step A: Anti-aliasing low-pass filtering. Before downsampling, the original signal x[n] is subjected to anti-aliasing low-pass filtering to obtain the filtered signal. : =LPF(x[n]) Where LPF is a low-pass filter, and its cutoff frequency satisfies: A 4th-order Butterworth IIR low-pass filter is used, with a cutoff frequency of =0.8× Taking D=10 as an example, =65MHz, Nyquist frequency is / (2D)=3.25MHz, set =2.6MHz.

[0036] Moving average pooling is essentially a rectangular window FIR filter with a sinc function frequency response and only about -13dB sidelobe attenuation, offering very limited aliasing suppression. UHF-PD signals can have a spectrum of tens of MHz; after downsampling, the Nyquist frequency drops to 3.25MHz, causing frequency components above 3.25MHz in the original signal to alias into the baseband. A 4th-order Butterworth filter provides approximately 80dB / decathlon attenuation in the stopband, keeping aliasing to an acceptable level. Cutoff frequency. =0.8× A balance is struck between passband flatness and stopband attenuation.

[0037] Aliasing Effect Analysis Table

[0038] Step B1: Mean pooling downsampling is applied to the filtered signal. Perform a D-fold decimation to obtain the downsampled signal: Where D is the downsampling factor, with a value ranging from 5 to 20, preferably D=10. L is the group delay compensation of the anti-aliasing filter.

[0039] Mean pooling preserves the DC component and low-frequency characteristics of the signal, so that subsequent morphological opening operations and CUSUM statistics calculations are not affected by the aliasing effect introduced by downsampling.

[0040] The sampling rate after downsampling is / D=6.5MHz, and the data volume is N / D=2,600,000 points. In a preferred embodiment, the downsampling factor D can be adaptively adjusted according to the signal-to-noise ratio or pulse density of the signal.

[0041] Step B2: Morphological baseline correction in the downsampled domain, in the downsampled signal Perform morphological Top-Hat baseline estimation: (1) One-dimensional morphological opening operation: in Width of structural element (preferred) =5), min and max are the minimum and maximum values ​​within the sliding window, respectively.

[0042] (2) The opening operation result is linearly interpolated and upsampled to the original length N to obtain the baseline estimate b[n].

[0043] (3) Calculate the pulse component signal: (4) Estimate the noise standard deviation from the opening operation signal (excluding pulses): The time complexity of morphological opening is O(N). When performed in the downsampling domain, the amount of data is reduced by a factor of D, and the computation time is reduced by a factor of D. The struct element width Ws of the morphological operation corresponds to a wider time window in the downsampling domain, yet it can still effectively separate the impulse from the baseline.

[0044] The purpose of morphological opening is to separate the pulse from the baseline; it is not necessary to preserve the fine waveform of the pulse. The pulse rise edge width changes from ~10 full samples to ~1 downsample in the downsampling domain, but the structuring element width... =5 is still greater than the pulse width, so the opening operation result can still accurately estimate the baseline. The final pulse component p[n] is obtained by calculating the difference across the entire sampling rate domain, preserving the fine waveform of the pulse.

[0045] Step B3: Downsampling domain block statistics and CUSUM thresholding. Perform adaptive CUSUM threshold calculation on the downsampling signal: (1) Divide the pulse component signal p[n] into K=[N / B] data blocks, each block has B sampling points (preferably B=4096), and calculate the mean of each block. and standard deviation .

[0046] Block statistics calculation can be performed in the downsampling domain (block size corresponds to B / D=410 downsampling points), further reducing the amount of computation.

[0047] (2) Perform CUSUM change point detection on the block mean sequence and automatically update the reference value when a change point is detected.

[0048] (3) Calculate the block-level threshold and linearly interpolate to upsample to the point-by-point threshold array T[n].

[0049] The amount of data processed by CUSUM was reduced from N=26M to K=635 blocks, a reduction of approximately 40,000 times. The time complexity of CUSUM's recursive computation is O(K), which is almost negligible.

[0050] Step B4: Pre-screening of candidate pulses in the downsampled domain, performing fast candidate pulse detection on the downsampled signal: (1) Downsample the pulse component signal to obtain the pulse component in the downsampled domain. [m].

[0051] (2) Calculate the pre-screening threshold in the downsampling domain: (3) Use a peak detection algorithm (such as scipy.signal.find_peaks) to detect peaks exceeding the pre-screening threshold in the downsampling domain to obtain the candidate pulse position set {c1,c2,...,c m}

[0052] (4) Use the CUSUM threshold array for filtering, retaining only pds[c i ]>Tds[c i [Candidates].

[0053] Peak detection has a time complexity of O(N / D), which is D times faster than the full sampling rate domain. After pre-screening and CUSUM filtering, the number of candidates M is usually <200, providing extremely sparse input for subsequent computationally intensive operations (CWT).

[0054] When multiple pulses may exist within a local window (dense pulse train scenario), the following strategy is adopted: (1) Set a minimum peak spacing during the peak detection stage. This ensures that adjacent candidate pulses do not overlap.

[0055] (2) If the spacing of the candidate pulses in the downsampling domain is less than Only candidates with larger amplitudes are retained.

[0056] (3) In the time-frequency domain multi-scale analysis stage, if multiple energy peaks are detected in the local window, they are verified separately.

[0057] Step B5: Local CWT multiscale analysis in the downsampling domain, performing CWT multiscale analysis only on M candidate pulses (not the entire signal): (1) For each candidate pulse c i Extracting a local window from the downsampled signal: Where the window half-width W= 4×amax×Kext (Preferred) =12).

[0058] (2) Within a local window, use FFT convolution to calculate the Morlet wavelet coefficients at S scales: (3) Calculate the energy at each scale and verify the scale entropy.

[0059] Acceleration principle (product effect of three-stage acceleration):

[0060] Step B6: Cross-domain coordinate mapping and precise positioning, mapping the pulse position confirmed in the downsampling domain back to the full sampling rate domain: (1) Rough location: (2) In the full sampling rate domain of the original signal, with the coarse position as the center, search for the maximum value of the pulse component signal p[n] within the search range [ni^{coarse}-R, ni^{coarse}+R]: The search radius R = D × 2 = 20 (all sampling points).

[0061] (3) Deduplication: If the distance between two precisely positioned pulses is less than twice the minimum pulse width, only the one with the larger amplitude is retained.

[0062] (4) Final CUSUM threshold verification: Only retain pulses where p[ni^{fine}]>T[ni^{fine}].

[0063] Because downsampling uses mean pooling, the peak position of the pulse may have a deviation of ±1 downsampling point in the downsampling domain, corresponding to a deviation of ±D = ±10 points in the full sampling rate domain. A search radius of R = 2D = 20 points is sufficient to cover this deviation range, ensuring that the accurate positioning error is less than 1 full sampling point.

[0064] Step B7: Full sampling rate domain pulse verification. On the pulse component signal p[n] in the full sampling rate domain, perform joint verification of kurtosis and spectral flatness on the precisely located candidate pulse: (1) For each candidate pulse n i Extract a length of p[n] from p[n]. Local window (preferred) =64).

[0065] (2) Calculate the time-domain kurtosis κ i Spectral flatness (SFM) in the frequency domain i .

[0066] (3) Joint criterion: κ i ≥κth and SFMi≥SFMth.

[0067] Why must it be performed across the full sampling rate domain? Kurtosis calculation depends on the fine shape of the waveform's rising edge (fourth-order statistical moments), and spectral flatness calculation depends on the high-frequency details of the spectrum. Both of these characteristics are sensitive to the waveform's sampling density; calculations performed in the downsampling domain will result in under-kurtosis and distorted spectral flatness.

[0068] The table below shows the decreasing data volume at each stage of the complete processing pipeline:

[0069] The computationally intensive CWT operation is performed on only 0.46% of the data, while the full sampling rate domain validation operation only processes 82 local windows (64 points each, totaling 5,248 points, accounting for 0.02%).

[0070] Comparison of computational complexity

[0071] Example 1: Standard UHF-PD Signal Processing Case This embodiment uses a typical UHF partial discharge online monitoring scenario as an example to describe the method of the present invention in detail. The UHF antenna receives the electromagnetic wave signal radiated by partial discharge. After amplitude detection by an AD8318 detector, the analog signal is converted into a digital signal by an ADC with a sampling rate of 65MHz. The signal parameters are as follows: sampling rate fs = 65MHz, each frame of data contains 20 power frequency cycles, the data length is N = 26,000,000 sampling points, and 60 analog PD pulses are injected into the original signal.

[0072] The first downsampling factor D = 10, the downsampling rate is 6.5MHz, and the downsampling domain data size is 2,600,000 points; the morphological structuring element width Ws = 5; the CUSUM block size B = 4096, the number of blocks K = 635; the CWT scale number S = 12; and the kurtosis threshold κ for impulse validation. th = 5.0, spectral flatness threshold (SFM) th = 0.20.

[0073] First, the anti-aliasing filter module is configured. A 4th-order Butterworth IIR low-pass filter is used, with a cutoff frequency set to fc = 0.8 × 65 MHz / (2 × 10) = 2.6 MHz. After the original signal x[n] is filtered by this low-pass filter, the filtered signal x is obtained. filt [n]. Then, mean pooling downsampling with a factor of D=10 is performed to obtain the downsampled domain signal x. ds [m], the data volume decreased to 2,600,000 points.

[0074] Second, baseline estimation is performed in the downsampling domain. For x ds A morphological opening operation is performed on [m], with the structuring element width Ws = 5. The opening result xopen[m] is linearly interpolated and upsampled to the original length N to obtain the baseline estimate b[n]. Then, the pulse component signal p[n] = max(x[n] - b[n], 0) is calculated. At the same time, the global noise standard deviation is estimated from xopen[m].

[0075] Adaptive threshold calculation is then performed. The pulse component signal is divided into 635 data blocks, each with 4096 sampling points, and the mean and standard deviation of each block are calculated. CUSUM change point detection is performed on the block mean sequence. When a change point is detected, the reference value is automatically updated, the block-level threshold is calculated, and the data is interpolated and upsampled to the point-by-point threshold array T[n].

[0076] During the candidate pulse detection phase, the pulse component signal is downsampled to obtain p. ds[m] After calculating the pre-screening threshold, a peak detection algorithm is used to detect peaks exceeding the threshold in the downsampling domain. After CUSUM threshold filtering and minimum peak spacing deduplication, approximately 200 candidate pulse positions are obtained.

[0077] Subsequently, time-frequency domain multi-scale analysis was performed. A local window was truncated for each candidate pulse, with the half-width determined based on the maximum analysis scale and the expansion coefficient Kext=12. Within each local window, FFT convolution was used to calculate Morlet wavelet coefficients at 12 scales, the energy at each scale was calculated, and scale entropy verification was performed. Approximately 82 pulses passed the verification.

[0078] The verified pulse position is mapped from the downsampling domain to the full sampling rate domain. Using c i After obtaining a rough location using ×D, the maximum value of the pulse component signal is searched within a search radius R=20 to achieve precise positioning. The located pulses are then deduplicated, and an adaptive threshold is used for final verification.

[0079] Finally, waveform feature verification is performed on the precisely located candidate pulses across the full sampling rate domain. A local window of length Nw=64 is truncated for each candidate pulse, and the temporal kurtosis κ is calculated. i Spectral flatness (SFM) in the frequency domain i The joint criterion of κi≥5.0 and SFMi≥0.20 is used to determine the final valid pulse, and the signal-to-noise ratio of each pulse is calculated as an auxiliary quality index.

[0080] Morphological Top-Hat baseline estimation took 45ms, CUSUM threshold calculation took 18ms, CWT ridge detection took 102ms, pulse verification took 8ms, and auxiliary processing (phase extraction and width estimation) took 12ms, for a total processing time of approximately 185ms per frame. The results showed that 55 valid pulses were detected, representing a detection rate of 91.7% compared to the injected 60 pulses. The number of false alarms was 0, and the processing time per frame was 185ms, meeting the real-time requirements of online monitoring.

[0081] In comparison, when processing the same signal using the traditional full-sampling-rate direct processing scheme (without downsampling and sparsification), the processing time per frame is approximately 12,000 ms, the detection rate is approximately 90%, and the number of false alarms is 3. The comparative results show that the present invention outperforms the traditional scheme in both detection rate and false alarm rate, with a processing speed improvement of approximately 65 times.

[0082] Example 2: Experiment on the effect of different downsampling factors. This example aims to examine the effect of different downsampling factors D on processing performance and detection accuracy, so as to verify the rationality of the selection of downsampling factors in this invention and the necessity of adaptive adjustment.

[0083] Keeping the signal conditions and the number of injected pulses exactly the same as in Example 1, three sets of experiments were set up with downsampling factors D = 5, D = 10, and D = 20 respectively. Except for the downsampling factor, all other algorithm parameters in each set of experiments were kept consistent with those in Example 1. The differences in processing time, detection rate, false alarms, and positioning error among the sets were compared.

[0084] Experiment 1: D = 5. The sampling rate after downsampling is 13MHz, and the downsampling domain data volume is 5,200,000 points. Processing results: Single frame processing time is approximately 320ms, detection rate is 93.3%, false alarm count is 1, and the root mean square error of positioning is 0.8 full sampling points. This configuration offers the highest detection accuracy and most accurate positioning, but due to the relatively small downsampling factor, the processing time is relatively long. It is suitable for offline analysis or fault diagnosis scenarios with high accuracy requirements.

[0085] Experiment 2: D = 10. The sampling rate after downsampling is 6.5MHz, and the downsampling domain data volume is 2,600,000 points. As mentioned earlier, the processing time is approximately 185ms, the detection rate is 91.7%, the false alarm rate is 0, and the root mean square error of the positioning error is 1.5 full sampling points. This configuration achieves the best balance between detection accuracy and processing efficiency, is suitable for conventional online monitoring scenarios, and is the preferred embodiment of this invention.

[0086] Experiment 3: D = 20. The sampling rate after downsampling is 3.25MHz, and the downsampling domain data volume is 1,300,000 points. Processing results: Single frame processing time is approximately 95ms, detection rate is 85.0%, false alarms are 0, and the root mean square error of the positioning error is 3.2 full sampling points. This configuration offers the fastest processing speed, but the detection rate decreases significantly, and the positioning error increases. It is suitable for special scenarios with extremely high real-time requirements and relatively low accuracy requirements, or for rapid inspection under conditions of extremely high signal-to-noise ratio.

[0087] The results of the three sets of experiments above show that: the smaller the downsampling factor D, the more complete the original information is preserved in the downsampling domain signal, resulting in higher detection and positioning accuracy, but the processing time increases accordingly; the larger D is, the faster the processing speed, but the accuracy decreases. Considering both the real-time requirements of online monitoring (single frame <200ms) and the detection accuracy requirements, D=10 is the optimal choice to balance the two.

[0088] The above experiments also verified the rationality of the adaptive downsampling factor adjustment mechanism of this invention. In practical engineering applications, the D value can be dynamically adjusted according to the signal-to-noise ratio and pulse density of the field signal: when the field electromagnetic environment is good and the signal-to-noise ratio is high, the D value can be appropriately increased to improve processing speed and free up computing resources for other tasks; when the field interference is severe, the signal-to-noise ratio is low, or the discharge is dense, the D value can be decreased to ensure the reliability of detection and the accuracy of positioning. The adjustment range of the D value is 5 to 20, which can cover various application needs from offline precision analysis to online rapid monitoring.

[0089] Example 3: Verification of the anti-aliasing filtering effect. This example verifies the necessity and effectiveness of the anti-aliasing filtering stage in this invention through comparative experiments. Two comparative schemes were set up: Scheme A, the scheme of this invention, uses a 4th-order Butterworth low-pass filter (fc = 2.6MHz) for anti-aliasing filtering before downsampling; Scheme B, the comparative scheme, omits the anti-aliasing filtering stage and directly uses mean pooling downsampling (equivalent to rectangular window FIR filtering, with sidelobe attenuation of approximately -13dB). The downsampling factor for both schemes is D=10, and the other parameters are the same as in Example 1.

[0090] Experimental results show that in Scheme B, frequency components above 3.25MHz in the original signal are directly aliased into the baseband, resulting in significant waveform distortion of the pulse component signal in the downsampled domain. This reduces the accuracy of CWT ridge detection, ultimately achieving a detection rate of approximately 85% and two false alarms. In Scheme A, the 4th-order Butterworth filter provides approximately 80dB / decibels of attenuation in the stopband, effectively suppressing aliasing components. The downsampled domain signal remains well-preserved, with a stable detection rate of 91.7% and zero false alarms.

[0091] This embodiment verifies the crucial role of anti-aliasing filtering in ensuring the signal quality of the downsampled domain. While mean pooling downsampling without anti-aliasing filtering is computationally simple, waveform distortion caused by aliasing propagates and accumulates along the processing chain, ultimately compromising the reliability of pulse detection and the accuracy of pulse localization. This invention introduces an anti-aliasing filtering step before downsampling, achieving signal quality assurance in the downsampled domain with relatively low computational overhead, which is a crucial prerequisite for the establishment of a hierarchical downsampling architecture.

[0092] Example 4: Verification of the replaceability and scalability of each module. This example aims to verify the independent replaceability and upgradeability of each processing module of the present invention. Baseline estimation module replacement experiment: The morphological Top-Hat baseline estimation is replaced with the wavelet baseline estimation method. The downsampled signal is decomposed into 6 levels using the db4 wavelet, and the baseline estimate is obtained after reconstructing by setting the approximation coefficients to zero. After the replacement, the processing time increases by approximately 8ms (the computational cost of wavelet decomposition is slightly greater than that of morphological operations), but the baseline estimation accuracy is slightly improved under low signal-to-noise ratio conditions. Since the candidate pulse detection and subsequent verification modules remain unchanged, the overall detection accuracy is basically consistent with the original scheme. This experiment shows that the baseline estimation module can be replaced independently without affecting the normal operation of other modules.

[0093] Experimental replacement of the threshold calculation module: The CUSUM adaptive threshold was replaced with the EMA (Exponential Moving Average) threshold method. The EMA threshold calculates the dynamic threshold after smoothing the block mean sequence, and its computational cost is lower than CUSUM, but its response speed to signal abrupt changes is slightly slower. The processing time is reduced by approximately 5ms after the replacement. In sparse pulse scenarios, the detection performance is comparable to the original scheme, while the detection rate decreases slightly in dense pulse scenarios. Users can flexibly choose the threshold method according to the pulse density on site.

[0094] Experiments on adding / removing verification modules: A pulse width consistency verification module was added after the waveform feature verification stage. The half-width at half-maximum (WHM) of each candidate pulse was calculated, and candidates with widths exceeding the typical PD pulse range (20ns–100ns) were eliminated. Adding this verification module increased processing time by approximately 2ms, further reducing the false alarm rate to zero, but no missed detections occurred (the widths of all real pulses were within a reasonable range). This experiment demonstrates that users can flexibly add or remove verification modules according to actual needs to optimize the balance between false alarm rate and processing speed.

[0095] The three substitution and addition / reduction experiments described above demonstrate that the processing modules in this invention maintain loose coupling, and the modules interact through standardized data interfaces (downsampled domain signals, pulse position sets, threshold arrays, etc.). This scalable design enables the invention to be compatible with different algorithm implementations, facilitating engineers to customize configurations and continuously optimize according to specific application scenarios and signal characteristics without redesigning the entire processing pipeline.

[0096] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for graded frequency reduction and sparsification of partial discharge signals in real time, characterized in that, Includes the following steps: After applying anti-aliasing filtering to the original sampled signal, downsampling is performed using the first downsampling factor to obtain the downsampled domain signal. Baseline estimation is performed on the downsampled domain signal, the pulse component signal is extracted, and an adaptive threshold is calculated on the downsampled domain signal; Based on the pulse component signal and adaptive threshold, candidate pulse detection is performed in the downsampling domain to obtain a set of candidate pulse positions; For each candidate pulse in the candidate pulse location set, perform time-frequency domain multi-scale analysis in the downsampling domain to verify and filter effective pulses; The positions of the selected valid pulses are mapped from the downsampling domain to the full sampling rate domain for precise location. The waveform characteristics of the precisely located pulse are verified across the entire sampling rate domain, and the final valid pulse is output. The signal-to-noise ratio of the validated pulses is calculated as an auxiliary quality indicator. The amount of data along the processing link signal decreases step by step, and different processing steps are executed in the sampling rate domain that meets their respective accuracy requirements.

2. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The anti-aliasing filter uses a low-pass filter to filter the original signal. Perform filtering to obtain the filtered signal The cutoff frequency of the low-pass filter satisfy ,in The original sampling rate, The first downsampling factor; downsampling processing is performed on the filtered signal. conduct Decimation is performed to obtain the downsampled domain signal. ,in The original signal length, Group delay compensation for the anti-aliasing filter; the low-pass filter is a 4th-order Butterworth IIR low-pass filter with a cutoff frequency of The first downsampling factor The value range is 5 to 20.

3. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The baseline estimation employs a morphological Top-Hat operation: in the downsampled domain signal... Perform one-dimensional morphological opening operation on ,in The width of the structural element. and These are the minimum and maximum values ​​within the sliding window, respectively; the result of the opening operation... Linear interpolation upsampling to the original length To obtain the baseline estimate Calculate pulse component signals And estimate the noise standard deviation from the opening operation results. ,in It is the absolute median difference function.

4. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The calculation of the adaptive threshold includes: [the process involves] processing the pulse component signal... Divided into Each data block Calculate the mean value for each of the 100 sampling points. and standard deviation CUSUM change point detection is performed on the block mean sequence. When a change point is detected, the reference value is automatically updated, the block-level threshold is calculated, and the data is linearly interpolated and upsampled to a point-by-point threshold array. Block statistics are calculated in the downsampling domain, and the corresponding block size is... One downsampling point.

5. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The candidate pulse detection includes: downsampling the pulse component signal to obtain the downsampled pulse component. Calculate the pre-screening threshold A peak detection algorithm is used to detect peak values ​​exceeding the pre-screening threshold in the downsampling domain, thereby obtaining a set of candidate pulse positions. ; Utilizing the corresponding value of the adaptive threshold in the downsampling domain Filter and retain only Candidates; set minimum peak spacing To ensure that adjacent candidate pulses do not overlap, if the distance between candidate pulses is less than... Only those with larger amplitudes are retained.

6. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The time-frequency domain multi-scale analysis includes: for each candidate pulse Extracting a local window from the downsampled signal The half-width of the local window , For the largest analytical scale, These are the expansion coefficients; within a local window, FFT convolution is used for calculation. Morlet wavelet coefficients at each scale ,in For the first Wavelet mother function of scale, For scale parameters, Calculate the energy at each scale and perform scale entropy verification, then select the pulses that pass the verification.

7. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The step of mapping the filtered valid pulse positions from the downsampling domain to the full sampling rate domain for precise positioning includes: mapping the pulse positions confirmed in the downsampling domain... Multiply by the first downsampling factor To obtain a rough position across the entire sampling rate domain. Within the full sampling rate domain of the original signal, the pulse component signal is searched for within the search range, centered on a coarse location. maximum value ,in The search radius is used; after precise positioning, the pulses are deduplicated. If the distance between two pulses is less than twice the minimum pulse width, only the one with the larger amplitude is retained; finally, an adaptive threshold is used for verification, retaining only the pulses with the larger amplitude. The pulse.

8. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The waveform feature verification of the precisely located pulse across the full sampling rate domain includes: pulse component signals across the full sampling rate domain. Above, for each candidate pulse after precise positioning Cut length is Calculate the time-domain kurtosis using a local window. and frequency domain spectral flatness ,in , In the formula These are the sampled values ​​of the pulse components within a local window. and These are the mean and standard deviation of the sampled values ​​within the window, respectively. The Fourier spectrum amplitude of the local window signal is used; a joint criterion is employed. and Determine if a pulse is valid.

9. The method for graded frequency reduction and sparsification of partial discharge signals in real time as described in claim 1, characterized in that, The formula for calculating the signal-to-noise ratio is as follows: ,in For the first time after baseline removal A local window waveform, Estimation of global noise standard deviation; first downsampling factor Adaptively adjusts based on the signal-to-noise ratio or pulse density: increases when the signal-to-noise ratio is higher than the first threshold or the pulse density is lower than the second threshold. To improve processing speed, the value is reduced when the signal-to-noise ratio is below the third threshold or the pulse density is above the fourth threshold. Values ​​are set to ensure detection accuracy; Adaptive adjustments are performed before each frame of data is processed. The value can be adjusted from 5 to 20.

10. A real-time processing system for graded frequency reduction and sparsification of partial discharge signals, characterized in that, include: The anti-aliasing filter module is used to perform anti-aliasing low-pass filtering on the original sampled signal; The graded downsampling module is used to downsample the filtered signal to the downsampling domain using a first downsampling factor; The baseline estimation module is used to perform morphological opening operations on the downsampled domain signal to estimate the baseline and subtract the baseline from the original signal to extract the pulse component; An adaptive threshold calculation module is used to calculate adaptive thresholds in the downsampling domain based on block statistics and CUSUM change point detection. The candidate pulse detection module is used to detect candidate pulses in the downsampling domain based on pulse components and adaptive thresholds, and to set a minimum peak spacing to handle overlapping pulses; The time-frequency domain analysis module is used to perform multi-scale Morlet wavelet transform by applying FFT convolution to a local window of each candidate pulse in the downsampling domain, and to filter valid pulses by verifying the scale entropy. The cross-domain mapping and positioning module is used to map the verified pulse position from the downsampling domain to the full sampling rate domain, and achieve precise pulse positioning through local search; The waveform verification module is used to perform joint verification of kurtosis and spectral flatness on the precisely located pulse across the entire sampling rate domain, and output the final valid pulse. The signal-to-noise ratio calculation module is used to calculate auxiliary quality indicators for verified pulses.

Images