Partial discharge separation method based on FPGA multi-dimensional feature parallel calculation

Abstract

The application relates to the technical field of power equipment state monitoring and fault diagnosis, and particularly discloses a partial discharge separation method based on FPGA multi-dimensional feature parallel calculation. The method performs real-time parallel calculation of three characteristic values in the FPGA: equivalent frequency bandwidth EBW representing the dispersion degree of signal spectrum energy, signal equivalent steepness ES representing the transient change degree of the signal, and time-frequency energy concentration TFEC representing the aggregation degree of signal energy on a time-frequency plane. The three characteristic calculation units simultaneously operate in parallel architecture, and the pipeline cross scheduling is adopted to make the calculation of multiple pulses overlap and execute, so that the calculation efficiency is improved by more than 2 times compared with the serial mode. The FPGA transmits the three characteristic values to an upper computer to perform three-dimensional space signal clustering and realize the separation of different discharge sources.

Description

Technical Field

[0001] This invention relates to the technical field of power equipment condition monitoring and fault diagnosis, and more specifically, to a partial discharge separation method based on FPGA multidimensional feature parallel computing. Background Technology

[0002] Partial discharge (PD) is a significant indicator of insulation degradation in power equipment, and its effective monitoring is crucial for ensuring the safe operation of power systems. When insulation defects exist within power equipment, partial discharge occurs under the influence of a strong electric field, releasing transient pulse signals ranging from nanoseconds to microseconds. By detecting and analyzing these pulse signals, the insulation condition of the equipment can be assessed, preventing insulation breakdown accidents.

[0003] However, in actual operating environments, power equipment such as substations and switchgear often have multiple partial discharge sources, accompanied by various electromagnetic interference signals (such as corona discharge, switching operation pulses, periodic narrowband interference, white noise, etc.). This results in the sensor receiving a signal that is a mixture of signals from multiple discharge sources and interference sources. How to accurately separate partial discharge signals from different sources from the mixed signal has been a long-standing technical challenge in this field.

[0004] In existing technologies, partial discharge signal separation mainly employs the following methods:

[0005] (1) Phase Distribution Map (PRPD) based method: Separation is achieved by statistically analyzing the phase distribution characteristics of the discharge signal within the power frequency cycle. This type of method utilizes the differences in phase distribution of different discharge sources within the power frequency cycle (such as the symmetry of the positive and negative half-cycles, the concentration of phase intervals, etc.) to distinguish them. However, this method requires the accumulation of a large number of discharge events to form an effective statistical distribution map, has poor real-time performance, and the response delay is usually on the order of tens of seconds to several minutes. Furthermore, it is difficult to effectively distinguish between different types of discharge sources with similar phase characteristics.

[0006] (2) Time-frequency analysis-based methods: Wavelet transform, short-time Fourier transform, and other techniques are used to extract the joint features of the signal in the time and frequency domains. This type of method can reveal the local time-frequency characteristics of the signal and has a certain effect on distinguishing signals with different waveforms. However, time-frequency analysis algorithms have high computational complexity, involving a large number of multiplication-addition operations and matrix operations. Traditionally, when implemented serially in FPGA or host computer software, the processing speed is slow and it is difficult to meet the real-time processing requirements of high-speed pulse streams.

[0007] (3) Methods based on pulse waveform features: Extract traditional time-domain parameters such as pulse rise time, pulse width, peak amplitude, and fall time for cluster analysis. This type of method is simple to calculate and easy to implement, but these traditional features only describe the signal from a single dimension, have a single physical meaning, and have limited distinguishability. They are also ineffective at separating discharge sources with similar waveforms but different origins.

[0008] The aforementioned existing technologies share the following common problems:

[0009] First, feature extraction efficiency is low. In existing methods, feature extraction is usually implemented serially in host computer software, calculating pulse by pulse. When the pulse repetition frequency is high (e.g., thousands to tens of thousands of times per second), the system's processing capacity is insufficient, resulting in the loss of a large number of pulses.

[0010] Second, the distinguishability of features is insufficient. Traditional features such as rise time and pulse width have only a single physical meaning and can only reflect one aspect of the signal. Different types of discharge sources overlap significantly in a single feature space, resulting in low classification accuracy.

[0011] Third, the utilization rate of hardware resources is low. Existing methods perform feature calculations one by one in a serial manner, which fails to fully utilize the parallel computing capabilities of hardware platforms such as FPGAs, resulting in a waste of hardware resources.

[0012] Fourth, data transmission is a heavy burden. A large amount of raw waveform data needs to be transmitted to the host computer, which consumes communication bandwidth and increases system latency, especially in multi-channel monitoring scenarios.

[0013] Therefore, there is an urgent need for a partial discharge signal processing method that can perform real-time parallel computation of multiple highly discriminative feature values ​​on an FPGA hardware platform to overcome the shortcomings of the existing technologies. Summary of the Invention

[0014] To address the aforementioned technical problems in related technologies, this invention provides a partial discharge separation method based on parallel computation of multidimensional features of FPGA, which can solve the above problems.

[0015] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows:

[0016] A partial discharge separation method based on FPGA multidimensional feature parallel computing includes the following steps:

[0017] Step S1: Signal Acquisition and Preprocessing

[0018] The partial discharge signal is sampled using a high-speed ADC (analog-to-digital converter) with a sampling frequency of no less than 200 MSPS and a sampling bit width of no less than 12 bits. The acquired digital signal is then subjected to digital bandpass filtering, with the bandpass range set according to the on-site interference conditions, preferably 500 kHz to 50 MHz, to filter out out-of-band noise and power frequency components.

[0019] The filtered signal is then analyzed using a sliding window pulse detection method to identify valid pulse events: a trigger threshold V is set. th When the signal amplitude continuously exceeds V th When the number of sampling points reaches the preset minimum pulse width, it is determined as the start of a valid pulse; when the signal amplitude is continuously lower than V... th When the number of sampling points reaches the preset termination condition, the pulse is determined to have ended. After a valid pulse is detected, the pulse waveform data (including several pre-trigger points before the actual trigger) is captured and sent to the feature calculation module.

[0020] Step S2: Calculate the first eigenvalue—Equivalent Bandwidth EBW—in real time within the FPGA.

[0021] For each detected partial discharge pulse signal, its equivalent bandwidth (EBW) is calculated in real time within the FPGA. The EBW characterizes the dispersion of the partial discharge signal's spectral energy and is a measure of the signal's energy distribution width in the frequency domain. A larger EBW value indicates a more dispersed spectral energy distribution and richer frequency components; a smaller EBW value indicates a more concentrated signal spectrum and more homogeneous frequency components.

[0022] The formula for calculating the equivalent bandwidth (EBW) is:

[0023] ;

[0024] in, The amplitude of the pulse signal's spectrum. The center frequency of the spectrum. Let f(i) be the frequency of the i-th frequency point (assuming the pulse waveform of the x(i) sequence has 1024 points and the sampling rate Fs is 100MHz, then performing a 1024-point FFT on this x(i) sequence to obtain X(fi) would give a frequency resolution of f = 100MHz / 1024). i =f*i, where i = 0~1023, f i It is actually X(f) i (The frequency value corresponding to the i-th point in the sequence). Number of spectrum points;

[0025] The center frequency of the spectrum The calculation formula is:

[0026] .

[0027] The physical significance of EBW lies in the fact that different types of partial discharge sources (such as internal air gap discharge, surface discharge, corona discharge, etc.) will generate pulse signals with different frequency components. Air gap discharge typically produces pulses with steep rise edges and wide frequencies, resulting in larger EBW values; while surface discharge pulses have relatively flat rise edges and narrower frequencies, leading to smaller EBW values. EBW characteristics can effectively distinguish discharge sources with different physical mechanisms.

[0028] Step S3: Calculate the second eigenvalue—the equivalent steepness of the signal ES—in real time within the FPGA.

[0029] For each partial discharge pulse signal, its equivalent steepness ES is calculated in real time inside the FPGA. The equivalent steepness ES is used to characterize the severity of transient changes in the partial discharge signal and is an improvement and enhancement of the traditional rise time characteristic.

[0030] The formula for calculating the equivalent steepness ES of a signal is:

[0031] ;

[0032] in, Let be the voltage value at the i-th sampling point. The voltage value at the (i-1)th sampling point. These are adaptive weighting coefficients based on signal amplitude. The time it takes for the signal to rise from 10% peak value to 90% peak value. for The maximum peak amplitude of the signal at each sampling point This represents the number of sampling points within the pulse duration.

[0033] The adaptive weighting coefficient The calculation formula is:

[0034] .

[0035] Introducing adaptive weights Subsequently, the ES (Eigenvalue Analysis) is no longer a simple first-order difference accumulation, but rather assigns different weights based on the amplitude of the sampling points: larger amplitudes near the rising edge of the main pulse have higher weights and contribute more to the ES; while low-amplitude noise regions and secondary oscillation regions have lower weights and contribute less to the ES. This design allows the ES to focus more on the main rising structure of the pulse, reducing the interference of low-amplitude noise and waveform tail oscillations on the eigenvalues, and significantly improving the robustness and noise resistance of the features.

[0036] The physical significance of ES lies in the fact that different types of discharge sources have different pulse rise edge steepness. Corona discharge typically has an extremely steep rise edge (nanosecond level) and a large ES value; while the rise edge of internal air gap discharge is relatively gentle and the ES value is moderate; the rise edge of noise signals is usually even gentler and the ES value is small. The ES characteristic can effectively distinguish discharge sources with different steepness.

[0037] Step S4: Calculate the third eigenvalue—Time-Frequency Energy Concentration (TFEC)—in real time within the FPGA.

[0038] For each partial discharge pulse signal, the Time-Frequency Energy Concentration (TFEC) is calculated in real time inside the FPGA. The TFEC is used to characterize the degree of concentration of signal energy in the time-frequency plane and is a characteristic quantity that describes the signal energy distribution pattern from the time-frequency joint domain.

[0039] The formula for calculating the Time-Frequency Concentration Energy (TFEC) is as follows:

[0040] ;

[0041] in, This refers to the energy in the core region on the time-frequency plane that exceeds a threshold. This represents the total energy across the entire time-frequency analysis region.

[0042] Specifically, a Short-Time Fourier Transform (STFT) is performed on the pulse signal. The STFT is as follows: the original pulse discrete signal is denoted as x[n], with a length of N points. The frame length is L, the frame shift is R, and the total number of frames is... The conventional FFT only performs an FFT of length N on x[n]. Short-Time Fourier Transform (SFT), however, involves taking a fixed segment of data of length L from each of the N points, shifting it by R points to obtain a frame of data. Each frame requires FFT processing, so M frames need to be processed (see Example 4 for details) to obtain the time-frequency distribution. Time-frequency energy .

[0043] The total energy of the entire time-frequency analysis region mentioned above The sum of the energies of all time-frequency points in the entire time-frequency analysis region is calculated using the following formula:

[0044] ;

[0045] The threshold θ is set as a quantile-based adaptive threshold: all time-frequency energy values ​​E(m,k) are sorted in ascending order, and the p-quantile (preferably p=0.90) is taken as the threshold. The core region energy Ecore is the sum of the energies of all time-frequency points whose energy values ​​exceed the threshold, and its calculation formula is as follows:

[0046] , ;

[0047] in The energy at each time-frequency point in the entire time-frequency analysis region. This refers to the m-th time frame (range 0~M-1). For the k-th frequency point (k is the k-th frequency point, and f as explained earlier) i Similarly, replacing k with i conveys the same meaning, ranging from 0 to K-1. Short-Time Fourier Transform (SFT) involves performing an FFT on the L sequence after each R-point shift. The FFT computation length is configurable; in general engineering, it's assumed to be equal to L, but for mathematical rigor, we use the variable K to represent the FFT computation length (total frames). , Indicates the number of pulse points. The number of points per frame. The number of frame shift points. The number of points for performing an FFT on each windowed signal frame. The energy threshold is set to an adaptive threshold based on quantiles.

[0048] The physical significance of TFEC lies in the fact that the energy concentration patterns of pulse signals generated by different partial discharge sources differ in the time-frequency plane. Narrowband interference signals have their energy highly concentrated at a few frequency points, resulting in a TFEC value close to 1; while the energy of partial discharge pulses exhibits a wide bandwidth distribution in the time-frequency plane, leading to a moderate TFEC value; and white noise has its energy evenly distributed across the entire time-frequency plane, resulting in a relatively small TFEC value. TFEC can effectively distinguish partial discharge signals from various types of interference signals.

[0049] Step S5: Parallel computation and data transmission of three eigenvalues

[0050] The calculation of the three eigenvalues ​​EBW, ES, and TFEC is implemented within the FPGA using a parallel computing architecture. Specifically:

[0051] (1) The three feature calculation units (EBW calculation unit, ES calculation unit, TFEC calculation unit) share the input data buffer and each has independent computing resources (DSP unit, BRAM, logic unit), and can extract features from the same pulse signal at the same time without queuing.

[0052] (2) Pipeline cross-scheduling method: Each computing unit is divided into multiple pipeline stages (such as data loading stage, operation stage, and output stage). When the first pulse enters the second pipeline stage, the second pulse can start entering the first pipeline stage, realizing the overlapping execution of multiple pulses in the time dimension. This method significantly shortens the average processing time of continuous pulses and improves the pulse throughput of the system.

[0053] After the calculation is completed, the three characteristic values, along with auxiliary information such as the pulse timestamp and pulse peak amplitude, are packaged into a data frame and transmitted to the host computer via a communication interface (such as Ethernet, PCIe, UART, etc.). Because only the characteristic values ​​are transmitted instead of the original waveform data, the amount of communication data is reduced by more than 90%, significantly reducing communication bandwidth requirements and transmission latency.

[0054] The host computer performs signal clustering processing on the received multi-pulse feature points in a three-dimensional feature space (with EBW, ES, and TFEC as the three coordinate axes). It adopts density-based clustering algorithms (such as DBSCAN) or partition-based clustering algorithms (such as K-means). Each cluster corresponds to a discharge source or interference source, thereby achieving the separation of signals from different sources.

[0055] The beneficial effects of this invention are:

[0056] (1) The three innovative feature values ​​proposed in this invention characterize the partial discharge signal from different dimensions: the equivalent bandwidth EBW characterizes the degree of spectral energy dispersion from the frequency domain perspective; the equivalent steepness ES characterizes the degree of transient change from the time domain perspective; and the time-frequency energy concentration TFEC characterizes the energy accumulation mode from the time-frequency joint domain perspective. The three features are independent of each other and have excellent separability for different types of discharge sources in the three-dimensional feature space.

[0057] (2) This invention uses FPGA hardware to implement the parallel calculation of three feature values. The three feature calculation units run in parallel at the hardware level, and a pipeline cross-scheduling strategy is adopted to make the calculation of multiple pulses overlap in time. Compared with the serial calculation method, the overall calculation efficiency is improved by more than 2 times.

[0058] (3) The calculation of the three feature values ​​is completed inside the FPGA. Only the extracted feature data is transmitted to the host computer, which greatly reduces the amount of communication data, reduces system latency, and meets the requirements of real-time online monitoring. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is an overall block diagram of the FPGA feature calculation core described in the embodiments of the present invention;

[0061] Figure 2 This is a block diagram of the equivalent bandwidth (EBW) calculation unit according to an embodiment of the present invention;

[0062] Figure 3 This is a block diagram of the signal equivalent steepness (ES) calculation unit according to an embodiment of the present invention;

[0063] Figure 4 This is a block diagram of the time-frequency energy concentration TFEC calculation unit as described in an embodiment of the present invention. Detailed Implementation

[0064] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0065] Example 1:

[0066] like Figure 1 As shown, the overall architecture of the FPGA feature calculation core in this application includes the following functional modules:

[0067] ADC Interface Module: Receives the digital signal output from a high-speed ADC (preferably LVDS or JESD204B interface), performs data alignment, serial-to-parallel conversion, and buffering, and outputs a parallel digital signal to subsequent modules. This module supports ADC chips with a sampling rate of at least 200 MSPS and a data bit width of at least 12 bits.

[0068] Digital filtering module: Performs bandpass filtering on the input signal, using an FIR filter structure with an order of 128, a passband range of 500kHz to 50MHz, and a stopband attenuation of no less than 60dB, filtering out out-of-band noise and power frequency components. The filter coefficients are pre-stored in the FPGA's ROM.

[0069] Pulse detection module: Detects valid pulse events based on threshold comparison. Sets the positive trigger threshold V. th_high and negative trigger threshold V th_low (usually V)th_high =+3σ,V th_low =-3σ, where σ is the root mean square noise value). When the signal amplitude continuously exceeds V... th_high or below V th_low When the number of sampling points reaches the preset minimum pulse width (e.g., 10 sampling points), it is determined as the start of a valid pulse; when the number of sampling points whose absolute value of the signal amplitude is continuously lower than the noise threshold reaches the preset end condition (e.g., 50 sampling points), it is determined as the end of the pulse. After a valid pulse is detected, the complete waveform (preferably 2048 points in total length) from 50 points before the pre-trigger to 100 points after the pulse ends is extracted as pulse data.

[0070] Data distribution module: Simultaneously broadcasts the detected pulse data to the three feature calculation units to ensure that the three units receive exactly the same input data, thus guaranteeing the synchronization and consistency of feature calculation.

[0071] EBW Calculation Unit: Real-time calculation of equivalent bandwidth characteristic values; its internal structure is referenced. Figure 2 It includes an FFT module (preferred radix-2 FFT algorithm), an amplitude calculation module (CORDIC algorithm or multiplication and addition method), a center frequency calculation module (multiply-accumulate), a weighted variance calculation module (multiply-accumulate), and a square root module (CORDIC algorithm).

[0072] ES Computation Unit: Real-time calculation of the equivalent steepness characteristic value of the signal; its internal structure is referenced. Figure 3 It includes a peak detection module (comparator + register), a threshold zero-crossing detection module (comparator + counter), a difference calculation module (subtractor + absolute value), a weight calculation module (divider), a weighted accumulation module (multiply-accumulate accumulator), and a normalized division module (divider).

[0073] TFEC Calculation Unit: Calculates time-frequency energy concentration characteristics in real time; its internal structure is referenced... Figure 4 It includes a sliding window framing module (shift register), a window function weighting module (multiplier), a short-time FFT module (radix-2 FFT), an energy calculation module (multiply-accumulate), a total energy accumulation module (accumulator), an adaptive threshold calculation module (sorting comparator), a threshold comparison module (comparator), a core energy accumulation module (accumulator + comparator), and a ratio calculation module (divider).

[0074] Data Packaging Module: Packs three characteristic values, pulse timestamp, pulse peak amplitude, and pulse sequence number into a fixed-format data frame. The frame header contains a synchronization word and frame length information.

[0075] Communication interface module: Transmits feature data frames to the host computer via Gigabit Ethernet (UDP protocol) or PCIe interface at a transmission rate of not less than 100Mbps.

[0076] The overall data processing flow is as follows: ADC data, after digital filtering, enters the pulse detection module. Upon detecting a valid pulse, data interception is triggered. The data distribution module simultaneously sends the pulse data to three feature calculation units for parallel processing. After each unit independently completes its calculation, it outputs the feature values ​​to the data packaging module, which then transmits them to the host computer via the communication interface module. The three feature calculation units work in parallel, sharing input data and independently outputting results, achieving efficient real-time feature extraction.

[0077] Example 2:

[0078] like Figure 2 As shown, the equivalent bandwidth (EBW) calculation unit includes an FFT module, an amplitude calculation module, a center frequency calculation module, a weighted variance calculation module, and a square root module. The specific calculation steps are as follows:

[0079] Step 2.1: FFT operation

[0080] Perform an N-point FFT operation on the input pulse signal to obtain the real part Re[i] and the imaginary part Im[i] of the spectrum, where i=0,1,2,…,N / 2-1 (taking the single-sided spectrum), and N is the number of spectrum points.

[0081] Step 2.2: Calculation of Spectral Amplitude

[0082] Calculate the spectral amplitude at each frequency point:

[0083] .

[0084] The amplitude calculation uses the CORDIC algorithm to perform the square root operation, which saves hardware resources.

[0085] Step 2.3: Calculation of the center frequency of the spectrum

[0086] Calculate the center frequency of the spectrum :

[0087] ;

[0088] Where the frequency of the i-th frequency point , The sampling frequency is used; the center frequency calculation module can use two accumulators to calculate the numerator and denominator in parallel, and finally obtain the result through a divider. To prevent division by zero, a lower limit protection must be set for the denominator.

[0089] Step 2.4: Weighted Variance Calculation

[0090] Calculate the weighted variance of the spectral energy:

[0091] ;

[0092] Step 2.5: Square root output

[0093] Calculating equivalent bandwidth:

[0094] .

[0095] Embodiment 3:

[0096] As shown in Figure 3 , the signal equivalent steepness ES calculation unit comprises a peak detection module, a threshold zero-crossing detection module, a difference calculation module, a weight calculation module, a weighted accumulation module and a normalized division module. The specific calculation steps are as follows:

[0097] Step 3.1: Peak detection

[0098] Traverse all sampling points of the pulse signal (T points), detect and record the signal peak V peak = max(|x(t i )|), is the voltage value of the i-th sampling point. The peak detection module adopts a pipelined comparator, when a new sampling point is input in each clock cycle, it is compared with the current peak and updated.

[0099] Step 3.2: Threshold calculation

[0100] Calculate the 10% threshold and 90% threshold: wherein V 10 = 0.1×V peak ; V 90 = 0.9×V peak .

[0101] Step 3.3: Rise time detection

[0102] Detect the time t1 when the rising edge of the signal crosses V 10 for the first time and the time t2 when it crosses V 90 for the first time, calculate the rise time: T 90 = t2-t1.

[0103] To ensure detection reliability, both t1 and t2 are required to be within the range of the main rising edge of the pulse (that is, the monotonic rising interval of the signal). If the detection fails (e.g., t2<t1), use the default value T 90 = a preset minimum value (e.g., 2ns).

[0104] Step 3.4: Difference and weight calculation

[0105] For each sampling point within the pulse duration (i=1,2,…,T), calculate the absolute difference of adjacent points and the adaptive weight:

[0106] ;

[0107] ;

[0108] in, Let be the voltage value at the i-th sampling point. This represents the voltage value at the (i-1)th sampling point.

[0109] The difference calculation module uses a subtractor and an absolute value circuit; the weight calculation module uses a divider, and the output is a floating-point number between 0 and 1. The starting point is defined as diff[0] = 0.

[0110] Step 3.5: Weighted Accumulation

[0111] Calculate the weighted difference summation:

[0112] ;

[0113] Introduction Subsequently, ES is no longer a simple first-order difference accumulation, but focuses more on the changes in the main pulse structure: sampling points near the rising edge of the main pulse have large amplitudes and high weights, contributing significantly to s; while low-amplitude noise regions and tail oscillation regions have small amplitudes and low weights, contributing little to s. This design effectively reduces the interference of low-amplitude noise and secondary oscillations on the calculation results, improving the robustness of the features.

[0114] Step 3.6: Normalize the output

[0115] Calculate the equivalent steepness of the signal:

[0116] .

[0117] Example 4:

[0118] like Figure 4 As shown, the Time-Frequency Energy Concentration (TFEC) calculation unit includes a sliding window framing module, a window function weighting module, a short-time FFT module, an energy calculation module, a total energy accumulation module, an adaptive threshold calculation module, a threshold comparison module, a core energy accumulation module, and a ratio calculation module. The input signal is a fixed-length real-valued pulse waveform, with the pulse length N preferably being 2048 points. TFEC extraction is completed within a single fixed-length pulse.

[0119] The specific calculation steps are as follows:

[0120] Step 4.1: Short-time frame processing

[0121] Let the length of each frame be L points and the frame shift be R points. Then, the nth sampling point within the m-th time frame (essentially, after cutting a long discrete-time sequence x[n] into many segments (short-time frames), uses a two-dimensional index (m,n) to mark "the nth point in the nth segment"). After short-time framing, x[n] becomes many small segments (frames) of length L. m is used to represent the frame number (frame 0, frame 1, ...); n is used to represent the position number within the frame (the 0th point to the L-1th point in this frame). This is represented in two-dimensional form as follows: ):

[0122] n=0,1,…,L-1;

[0123] The total number of frames is:

[0124] ;

[0125] The frame length L is determined based on the actual pulse length and time-frequency resolution requirements, and is preferably 512 points. Adjacent frames are overlapped, preferably with 50% overlap, i.e., R = L / 2 = 256 points. 50% overlap ensures the continuity of temporal coverage and avoids information loss.

[0126] Step 4.2: Apply a Hamming window to each frame of signal.

[0127] The window function is a Hamming window:

[0128] n=0,1,…,L-1;

[0129] Receive windowing signal:

[0130] .

[0131] Step 4.3: Short-Time Fourier Transform

[0132] Perform a K-point FFT (K is preferably equal to L, i.e., 512 points) on each windowed signal frame to obtain the short-time spectrum:

[0133] ;

[0134] in The imaginary unit (used to represent the "imaginary part" of complex numbers). In electrical / signal processing, it is often represented by j. The imaginary unit is defined as: j 2 =−1), since the input signal is a real signal, only the single-sided spectrum k=0,1,…,K / 2 is taken for subsequent processing.

[0135] Step 4.4: Time-Frequency Energy Calculation

[0136] Calculating the energy of each time-frequency point:

[0137] ;

[0138] The energy calculation module uses two multipliers to separately calculate the squares of the real part and the imaginary part, and then adds them to obtain the energy value.

[0139] Step 4.5: Calculate the quantile energy threshold

[0140] Sort the energy values E(m,k) of all time-frequency points from smallest to largest, and denote the sorted sequence as:

[0141] ;

[0142] wherein Q is the total number of time-frequency points;

[0143] Given a preset quantile proportion p (0<p<1), the index position corresponding to the p-th quantile is:

[0144] , wherein represents rounding up;

[0145] The quantile threshold is defined as:

[0146] ;

[0147] wherein represents the p-th quantile of the time-frequency energy set, preferably p=0.90, that is, the 90th percentile is taken as the threshold.

[0148] When p=0.90, the top 10% time-frequency points in energy ranking are retained;

[0149] When p=0.95, the top 5% time-frequency points in energy ranking are retained;

[0150] When p=0.98, the top 2% time-frequency points in energy ranking are retained.

[0151] Based on the threshold, time-frequency points satisfying E(m,k)≥θ form a candidate core time-frequency region.

[0152] The selection of p-value needs to be adjusted according to actual signal characteristics: for scenarios dominated by narrowband interference, it is recommended to use a higher p-value (0.95~0.98); for scenarios dominated by broadband discharge signals, it is recommended to use a lower p-value (0.85~0.90).

[0153] In FPGA implementation, due to the large hardware overhead of sorting operation, an approximate method is adopted: quantile estimation is realized by histogram statistics or iterative threshold approximation, which can reduce resource consumption.

[0154] Step 4.6: Calculate time-frequency energy concentration

[0155] The total time-frequency energy is the sum of the energies at all time-frequency points:

[0156] ;

[0157] The time-frequency points satisfying E(m,k)>θ are defined as the core time-frequency region (i.e., the high-energy region), and their core energy is:

[0158] , ;

[0159] Time-frequency energy concentration is defined as:

[0160] ;

[0161] The ratio calculation module uses a divider to calculate E. core / E total The output is a single-precision floating-point number between 0 and 1.

[0162] The closer the TFEC value is to 1, the more concentrated the signal energy is in the time-frequency plane (such as narrowband sinusoidal signals); the smaller the TFEC value, the more dispersed the energy distribution (such as white noise and broadband pulses). The TFEC value of partial discharge pulses is usually between 0.3 and 0.7, depending on the time-frequency characteristics of the pulse.

[0163] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A partial discharge separation method based on FPGA multidimensional feature parallel computing, characterized in that, Includes the following steps: The equivalent bandwidth EBW of each pulse is calculated in real time within the FPGA. The equivalent bandwidth EBW is used to characterize the dispersion of the signal spectrum energy, and the calculation formula is as follows: ； in, The amplitude of the pulse signal's spectrum. The center frequency of the spectrum. Let i be the frequency of the i-th frequency point. Number of spectrum points; The equivalent steepness ES of the signal for each pulse is calculated in real time within the FPGA. The equivalent steepness ES is used to characterize the drastic nature of the transient change of the signal, and the calculation formula is as follows: ； in, Let be the voltage value at the i-th sampling point. The voltage value at the (i-1)th sampling point. These are adaptive weighting coefficients based on signal amplitude. The time it takes for the signal to rise from 10% peak value to 90% peak value. for The maximum peak amplitude of the signal at each sampling point This represents the number of sampling points within the pulse duration. The time-frequency energy concentration (TFEC) of each pulse is calculated in real time within the FPGA. The TFEC characterizes the degree of concentration of signal energy in the time-frequency plane, and the calculation formula is as follows: ； in, This refers to the energy in the core region on the time-frequency plane that exceeds a threshold. This represents the total energy across the entire time-frequency analysis region. The calculation of the three eigenvalues ​​EBW, ES, and TFEC adopts a parallel computing architecture. The three eigenvalue calculation units share the input data buffer and run simultaneously. A pipelined cross-scheduling method is used to make the calculation of multiple pulses overlap in time. The calculated three eigenvalues ​​are transmitted to the host computer for three-dimensional spatial signal clustering to achieve the separation of different power sources.

2. The partial discharge separation method based on FPGA multi-dimensional feature parallel computing according to claim 1, characterized in that, The center frequency of the spectrum The calculation formula is: 。 3. The partial discharge separation method based on FPGA multi-dimensional feature parallel computing according to claim 1, characterized in that, The adaptive weighting coefficient The calculation formula is: 。 4. The partial discharge separation method based on FPGA multi-dimensional feature parallel computing according to claim 1, characterized in that, The total energy of the entire time-frequency analysis region The sum of the energies of all time-frequency points in the entire time-frequency analysis region is calculated using the following formula: ； The time-frequency points satisfying E(m,k)>θ are defined as the core time-frequency region, and the energy in the core region on the time-frequency plane exceeding the threshold is defined as the energy of the core region. The calculation formula is: ， ； in The energy at each time-frequency point in the entire time-frequency analysis region. For the m-th time frame, For the k-th frequency point, the total number of frames , Indicates the number of pulse points. The number of points per frame. The number of frame shift points. The number of points for performing an FFT on each windowed signal frame. The energy threshold is set to an adaptive threshold based on quantiles.

5. The partial discharge separation method based on FPGA multidimensional feature parallel computing according to claim 1, characterized in that, It also includes signal acquisition and preprocessing steps: using a high-speed ADC to sample the partial discharge signal, performing digital bandpass filtering and sliding window pulse detection on the acquired signal, identifying valid pulse events, and simultaneously distributing the detected pulse data to three feature calculation units.

6. A partial discharge signal separation system based on FPGA real-time parallel computing of multi-dimensional features, characterized in that, For performing the method as described in any one of claims 1 to 5, comprising: The front-end acquisition module is used to sample and digitize partial discharge signals; The FPGA feature calculation module is used to receive the digitized signal and calculate three feature values ​​for each pulse in parallel within the FPGA: the equivalent bandwidth EBW, the equivalent steepness ES, and the time-frequency energy concentration TFEC. The three feature values ​​are calculated simultaneously using a parallel computing architecture. The data transmission module is used to package and transmit the three calculated feature values ​​to the host computer. The host computer analysis module is used to perform signal clustering processing on the received feature values ​​in the three-dimensional feature space to separate signals from different power sources.

7. The partial discharge signal separation system based on FPGA real-time parallel computing of multi-dimensional features according to claim 1, characterized in that, The FPGA feature calculation module uses a pipelined cross-scheduling method to overlap the feature calculations of multiple pulses in time.

Images