A partial discharge pulse extraction method based on noise reduction clustering

Abstract

The present invention relates to a local discharge pulse extraction method based on noise reduction clustering, which is based on a wavelet algorithm and includes: collecting original power generation signals; determining wavelet decomposition coefficients at each level; extracting wavelet decomposition coefficients at each level; clustering the extracted wavelet decomposition coefficient pulse segments at each level to obtain a subsequence of an aliased pulse sequence, which corresponds to independent signals of different discharge sources or different propagation paths. Based on the diversity and complexity of environmental interference in the current partial discharge detection of power equipment, the present invention improves the signal-to-noise ratio and multi-discharge clustering calculation of interference signals in real time during the detection process, thereby effectively improving the anti-interference performance of the partial discharge detection device.

Description

Technical Field

[0001] The invention relates to the technical field of partial discharge high-frequency current detection, in particular to a pulse extraction method based on noise reduction clustering. Background Art

[0002] Partial discharge is the main factor leading to insulation degradation and an important indicator of insulation degradation. It can effectively reflect the latent defects and faults inside the equipment and is an important indicator for evaluating the insulation status of the equipment. Partial discharge has a better detection effect on the early detection of sudden faults of insulators and has obvious advantages in detecting insulation defects or faults of power equipment. Therefore, the effective detection of partial discharge signals is of great significance to ensure the safe and stable operation of power equipment.

[0003] The Chinese patent application with publication number CN106771905B discloses a pulse extraction method suitable for high-frequency current partial discharge detection, which processes the signal based on wavelet decomposition, including:

[0004] Step 1: First, perform wavelet decomposition on the digital signal sequence collected by AD;

[0005] Step 2: Calculate the characteristic quantity of the signal after wavelet decomposition to obtain the peak value of the wavelet decomposition coefficient and the root mean square value corresponding to the peak value;

[0006] Step 3: According to the ratio of the peak value to the effective value of each level of wavelet decomposition coefficient, determine the signal type and shrink the wavelet decomposition coefficient;

[0007] Step 4: Reconstruct the wavelet coefficients containing the partial discharge pulse signal;

[0008] Step 5: Based on the reconstructed signal, the pulse segments are extracted by automatic threshold calculation.

[0009] Perform wavelet decomposition on the digital signal sequence collected by AD, including:

[0010] a) Downsample the original sequence S(n) to obtain the odd sequence S(n / 2-1) and the even sequence S(n / 2), where S(n / 2) and S(n / 2-1) are half of S(n) respectively;

[0011] b) S(n / 2-1) is multiplied and added with the high-pass filter h(n / 2-1) to obtain the sequence S1(n / 2-1) and the high-frequency coefficient cD1;

[0012] c) S(n / 2) is multiplied and added with the low-pass filter g(n / 2) to obtain the sequence S1(n / 2) and the low-frequency coefficient cA1;

[0013] d) S1(n / 2) is extracted downward to obtain the odd number sequence S1(n / 4-1) and the even number series S1(n / 4), where S1(n / 4-1) and S1(n / 4) are respectively half of the S1(n / 2) series;

[0014] e), S(n / 4-1) and the high-pass filter h(n / 4-1) are multiplied and added to obtain the sequence S2(n / 4-1) and the high-frequency coefficient cD2;

[0015] f), S(n / 4) and the low-pass filter g(n / 4) are multiplied and added to obtain the sequence S2(n / 4) and the low-frequency coefficient cA2;

[0016] g) Decompose the low-frequency sequence in turn, and finally obtain the wavelet coefficients: low-frequency coefficient cA6, high-frequency coefficients cD6~cD1.

[0017] This technical solution adopts the wavelet decomposition method, which can separate the signal and the noise to achieve the purpose of improving the signal-to-noise ratio and realize real-time high-performance noise suppression at a high sampling rate.

[0018] However, the above technical solution has weak anti-interference capability. Summary of the invention

[0019] The purpose of the present application is to provide a local discharge pulse extraction method based on noise reduction clustering to solve the problem that the existing solution has weak anti-interference ability.

[0020] To achieve the above object, the present invention proposes a local discharge pulse extraction method based on noise reduction clustering, comprising the following steps:

[0021] Collecting original power generation signals;

[0022] Determine the wavelet decomposition coefficients at each level;

[0023] Extract the partial discharge pulse segments from the wavelet decomposition coefficients at each level;

[0024] Clustering the pulse segments of wavelet decomposition coefficients extracted at each level to obtain subsequences of the aliased pulse sequence, which correspond to independent signals of different discharge sources or different propagation paths;

[0025] In the software implementation of clustering, it is necessary to extract the equivalent duration and equivalent bandwidth parameters according to the configured pulse extraction width. The calculation formulas are:

[0026] V (n) is the vertical scale of the pulse signal, T is the equivalent time length;

[0027] S (n)is the frequency domain distribution sequence after FFT processing of the pulse signal, and F is the equivalent bandwidth.

[0028] Furthermore, the method also includes a step of diagnosing the clustered discharge statistical phase spectrum.

[0029] Furthermore, the method further includes calculating the signal-to-noise ratio of the wavelet decomposition coefficients at each level. If the signal-to-noise ratio is lower than a predetermined limit, the wavelet decomposition coefficients at this level are set to zero.

[0030] Furthermore, the pulse segments exceeding the corresponding set threshold are extracted from the wavelet decomposition coefficients of each level; the set threshold is determined according to a calibration experiment.

[0031] Furthermore, the set threshold is a multiple of the root mean square value of the wavelet decomposition coefficients of the corresponding level.

[0032] Furthermore, the noise suppression ratio is used as an evaluation index to evaluate the interference suppression effect.

[0033] Furthermore, the noise suppression ratio is expressed as: NRR =10(lgσ1 2 -lgσ2 2 ), where σ1 and σ2 are the standard deviations of the original signal and the denoised signal, respectively.

[0034] The beneficial effects of the present invention are as follows: compared with the prior art, the present invention adds a step of clustering the extracted pulse segments to obtain a subsequence of the aliased pulse sequence, which corresponds to independent signals of different discharge sources or different propagation paths; it can improve the on-site anti-interference ability of the method. The present invention aims at the key information such as the source, time-frequency characteristics, and statistical characteristics of the high-frequency electromagnetic interference of the substation, in order to solve the complex problem of the inability to set thresholds and the coexistence of multiple pulse signals in a strong electromagnetic interference environment, in the detection process, the interference signal is improved in real time to achieve signal-to-noise ratio and multi-discharge clustering, and an efficient and real-time integrated algorithm based on wavelet decomposition multi-level intelligent noise reduction clustering is proposed, which effectively improves the anti-interference performance of the partial discharge detection device, and is the key to promoting the popularization and application of high-frequency current partial discharge detection technology and exerting good effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method of the present invention;

[0036] Figure 2 It is a logical schematic diagram of the present invention;

[0037] Figure 3 The clustering separation effect of 3-10 PD sources;

[0038] Figure 4 This is a verification diagram of the wavelet denoising effect of the actual measured signal of the C-phase core grounding of the No. 1 main transformer in a 500kV substation according to the present invention;

[0039] Figure 5 This is a verification diagram of the wavelet denoising effect of the present invention on the grounding signal collected from the A-phase clamp of the No. 1 main transformer in a certain 500kV substation;

[0040] Figure 6 This is the structure diagram of the Data_prc module;

[0041] Figure 7 This is the structure diagram of the Frame_sample_tf module;

[0042] Figure 8 This is the structure diagram of the Fft_cal module;

[0043] Fig. 9 This is the structure diagram of N_cal module;

[0044] Fig.10 This is the structure diagram of the T_cal module;

[0045] Fig.11 It is a pulse waveform diagram;

[0046] Fig.12 for Fig.11 Frequency domain distribution diagram of the signal;

[0047] Explanation of reference numerals: Sn-original signal; DFH-wavelet decomposition high-pass filter; DFL-wavelet decomposition low-pass filter; CD-high frequency signal; CA-low frequency signal; N-number of sampling points. DETAILED DESCRIPTION

[0048] The method flow of the present invention is as follows Figure 1 and Figure 2 As shown, the specific steps are as follows:

[0049] (1) The original signal S(i), i=0, 1, ..., N, is downsampled to obtain a half-length odd sequence SA(i) and a half-length even sequence SD(i), where SA(i)=S(2i); SD(i)=S(2i+1), i=0, 1, ..., N / 2;

[0050] (2) Calculate the convolution of SA(i) and SD(i) with the db2 wavelet decomposition low-pass filter DFL and high-pass filter DFH to obtain wavelet coefficients CA1(i) and CD1(i), i = 0, 1, ..., N / 2;

[0051] (3) The wavelet coefficient CA1(i) is downsampled to obtain its half-length odd sequence CA1A(i) and half-length even sequence CA1D(i), where:

[0052] CA1A(i)=CA1(2i), i=0, 1,..., N / 4

[0053] CA1D(i)=CA1(2i+1), i=0, 1,..., N / 4

[0054] (4) Sequences CA1A(i) and CA1D(i) are convolved with the wavelet decomposition low-pass filter DFL and the wavelet decomposition high-pass filter DFH, respectively, to obtain wavelet decomposition coefficients CA2(i) and CD2(i), i = 0, 1, ..., N / 4;

[0055] (5) The wavelet coefficient CA2(i) is downsampled to obtain its half-length odd sequence CA2A(i) and half-length even sequence CA2D(i), where:

[0056] CA2A(i)=CA2(2i), i=0, 1,..., N / 8

[0057] CA2D(i)=CA2(2i+1), i=0, 1,..., N / 8

[0058] (6) Sequences CA2A(i) and CA2D(i) are convolved with the wavelet decomposition low-pass filter DFL and the wavelet decomposition high-pass filter DFH, respectively, to obtain wavelet decomposition coefficients CA3(i) and CD3(i), i = 0, 1, ..., N / 8;

[0059] (7) The wavelet coefficient CA3(i) is downsampled to obtain its half-length odd sequence CA3A(i) and half-length even sequence CA3D(i), where:

[0060] CA3A(i)=CA3(2i), i=0, 1,..., N / 16

[0061] CA3D(i)=CA3(2i+1), i=0, 1,..., N / 16;

[0062] (8) Sequences CA3A(i) and CA3D(i) are convolved with the wavelet decomposition low-pass filter DFL and the wavelet decomposition high-pass coefficient DFH, respectively, to obtain wavelet decomposition coefficients CA4(i) and CD4(i), i = 0, 1, ..., N / 16;

[0063] (9) The wavelet coefficient CA4(i) is downsampled to obtain its half-length odd sequence CA4A(i) and half-length even sequence CA4D(i), where:

[0064] CA4A(i)=CA4(2i), i=0, 1,..., N / 32

[0065] CA4D(i)=CA3(2i+1), i=0, 1,..., N / 32

[0066] (10) Sequences CA4A(i) and CA4D(i) are convolved with the wavelet decomposition low-pass filter DFL and the wavelet decomposition high-pass coefficient DFH, respectively, to obtain wavelet decomposition coefficients CA5(i) and CD5(i), i = 0, 1, ..., N / 32.

[0067] Finally, the wavelet decomposition coefficients CA5, CD5, CD4, CD3, CD2 and CD1 are obtained.

[0068] (11) Signal-to-noise ratio calculation:

[0069] Calculate the effective values ​​(root mean square values) of the wavelet coefficients at each level of CA5, CD5, CD4, CD3, CD2, and CD1, and record them as Vrms(CD5), Vrms(CD4), Vrms(CD3), Vrms(CD2), and Vrms(CD1), respectively;

[0070] Extract the maximum values ​​of wavelet coefficients at each level, which are Vmax(CD5), Vmax(CD4), Vmax(CD3), Vmax(CD2) and Vmax(CD1);

[0071] Calculate the signal-to-noise ratio of wavelet coefficients at each level, using the The ratio is used as an indicator to judge the signal-to-noise ratio. If the signal-to-noise ratio of a certain level of wavelet coefficients is less than 3, the wavelet coefficients of this level are set to 0 and do not participate in the reconstruction.

[0072] like Figure 1 and Figure 2 The wavelet decomposition multi-level noise reduction clustering algorithm for partial discharge of power equipment shown in the figure performs hierarchical optimization based on the optimal signal-to-noise ratio principle of the wavelet decomposition coefficients at each level. The higher the signal-to-noise ratio, the lower the background interference of the signal component at this level, and the more significant the amplitude of the discharge pulse waveform signal. Similarly, the hardware overhead and algorithm efficiency are comprehensively considered.

[0073] (12) Adaptive threshold setting: After a large number of tests, a certain multiple of the root mean square value of the wavelet decomposition level coefficient is used as the threshold for pulse signal extraction. This multiple is pre-set based on calibration experiments.

[0074] Based on a preset adaptive threshold, the pulse segments exceeding the threshold are extracted from the wavelet decomposition coefficients of the above preferred level, so as to further reduce the hardware storage resource expenditure and the consumption of computing resources.

[0075] Steps (1)-(12) are the same as the technical solutions adopted in the comparative documents cited in the background technology.

[0076] (13) Based on the level of wavelet decomposition, the extracted pulse segments are clustered to obtain subsequences of the aliased pulse sequence, which correspond to independent signals from different discharge sources or different propagation paths. The clustering effect is shown in Figure 3 shown.

[0077] (14) Based on conventional partial discharge analysis methods, the discharge statistical phase spectrum PRPD after cluster separation is drawn, and conventional pattern recognition and other diagnostic algorithms are further executed.

[0078] like Figure 4 and Figure 5 The verification method of the wavelet denoising effect of the field measured signal of the wavelet decomposition multi-level denoising clustering algorithm of partial discharge of power equipment shown in the figure is to denoise the field collected signal. It can be seen from the figure that the denoising method adopted can effectively remove the periodic narrowband interference and white noise in the signal. After the denoising process, the pulse signal submerged by the noise is revealed, and the high-frequency components of the signal are also revealed.

[0079] Since the field test signals contain different degrees of noise interference, the parameters such as SNR, CC, MSE and ED cannot be obtained. In order to evaluate the denoising effect of the adopted method on the field measured signals, the noise suppression ratio parameter is introduced as an evaluation index. Its specific expression is shown in the following formula:

[0080] λ NRR =10(lgσ1 2 -lgσ2 2 )

[0081] Where σ1 and σ2 are the standard deviations of the original signal and the denoised signal respectively; λ NRR The size of reflects the prominence of the useful signal after interference suppression.

[0082] After calculation, λ NRR The values ​​are 18.37dB and 20.72dB respectively. Figure 3 and Figure 4 It can be seen from the reconstructed signal waveform and reconstructed signal spectrum in that this method can effectively remove the interference in the test signal and obtain a more realistic partial discharge signal waveform and signal spectrum diagram.

[0083] In the above embodiment, the limit value of the signal-to-noise ratio in step (11) is 3. As other implementations, other values ​​may also be used. Step (12) sets an adaptive threshold for each level to extract pulse segments exceeding the threshold in order to reduce resource consumption. As other implementations, other methods may be used to extract (e.g., fixed threshold, or full extraction) to retain the required information.

[0084] The improvement focus of this embodiment is on the implementation of clustering, which specifically includes hardware design and software design.

[0085] The hardware is implemented using FPGA, including the Data_prc module, such as Figure 6 As shown. The Data_prc module adds a tf calculation module (such as Figure 6 The leftmost box), you can use the register setting to turn on the tf calculation enable and send one channel of data to the tf calculation module for calculation, and change the input of the original processing module of this channel to 0. The output logic of Data_prc is to see which frame_sample module has data and output the data of that channel.

[0086] Frame_sample_tf module such as Figure 7 As shown, the tf calculation module mentioned above, according to the tf calculation formula, the only difference between t and f calculation is that the f value needs to be calculated first, so the fft calculation is a separate module. Although the calculation formula after fft in tf calculation is the same, the bit width of the value to be calculated is different. In order to reduce resource utilization, the effective n value is calculated in two separate modules (n_cal, s_n_cal), and the process of calculating the tf value after obtaining the effective n value is another two modules (t_cal, s_t_cal). The calculation adopts a non-backpressure pipeline design, and the start signal at the front ensures that the interval is greater than 2048, so that all calculation steps have sufficient time. Ensure that the delay of the three required results is consistent, based on the maximum delay path (f calculation).

[0087] Among them, the Fft_cal module is as follows Figure 8 As shown, the main function is to calculate FFT, and it also has the function of calculating the maximum value. Xilinx's IPcore is used to calculate FFT. When the input data is greater than the set threshold, 56 data are taken forward to start the FFT calculation of 2048 data. The calculation results in two data, the real part and the imaginary part. It is necessary to square the real part and the imaginary part and calculate the square of the amplitude before subsequent calculations.

[0088] Among them, N_cal module is as follows Fig. 9 As shown, its function is to calculate the effective n value, and it consists of an accumulator, a multiplier, and a divider.

[0089] Among them, the T_cal module is as follows Fig.10 As shown, the function is the last step of calculating the t value or the f value, including an accumulator, a divider and a square root module.

[0090] Software algorithms include:

[0091] The host computer configures the width of pulse extraction, and the FPGA extracts TF parameters according to the configured width. The following example shows the case where the pulse width is 2048 points.

[0092] The calculation input of T and F is a complete pulse waveform. Fig.11 Taking the pulse shown as an example, the pulse waveform has 2048 points in total, the ordinate is V, the abscissa is n, and the waveform sequence array is defined as V(n), where n ranges from 0 to 2048.

[0093] The calculation steps of parameter T are:

[0094] (1) Signal time domain standardization processing:

[0095]

[0096] (2) Find the time center of gravity of the signal:

[0097]

[0098] (3) Find the equivalent duration

[0099] Combining formula (1-1) to formula (1-2), the calculation of the equivalent time length T is as follows:

[0100]

[0101] The calculation steps of parameter F are:

[0102] (1) First, perform FFT processing on the time domain signal to obtain the frequency domain distribution as follows Fig.12 As shown: The obtained frequency domain distribution sequence is recorded as S (n) ;

[0103] (2) Standardization of frequency domain signals:

[0104]

[0105] (3) Find the frequency domain centroid of the frequency domain:

[0106]

[0107] (4) Calculate the equivalent bandwidth

[0108] Combining equations (1-4) to (1-5), the calculation of the equivalent bandwidth F is as follows:

[0109]

Claims

1. A local discharge pulse extraction method based on noise reduction clustering, characterized in that: Here are the steps: Collecting original power generation signals; Determine the wavelet decomposition coefficients at each level; Extract the partial discharge pulse segments from the wavelet decomposition coefficients at each level; Clustering the pulse segments of the extracted wavelet decomposition coefficients at each level to obtain a subsequence of the aliased pulse sequence, which corresponds to independent signals of different discharge sources or different propagation paths; The implementation of clustering includes hardware design and software design. The hardware is implemented using FPGA. In the software design, FPGA extracts the equivalent duration and equivalent bandwidth parameters according to the configuration width. The calculation formulas are: V (n) is the vertical scale of the pulse signal, T is the equivalent time length; S (n) is the frequency domain distribution sequence after FFT processing of the pulse signal, and F is the equivalent bandwidth.

2. The method for extracting partial discharge pulses based on noise reduction clustering according to claim 1, characterized in that: The method also includes a step of diagnosing the clustered discharge statistical phase spectrum.

3. The method for extracting partial discharge pulses based on noise reduction clustering according to claim 1 or 2, characterized in that: The method also includes calculating the signal-to-noise ratio of the wavelet decomposition coefficients at each level. If the signal-to-noise ratio is lower than a predetermined limit, the wavelet decomposition coefficients at this level are set to zero.

4. The method for extracting partial discharge pulses based on noise reduction clustering according to claim 3, characterized in that: For each level of wavelet decomposition coefficients, pulse segments exceeding a corresponding set threshold are extracted; the set threshold is determined according to a calibration experiment.

5. The method for extracting partial discharge pulses based on noise reduction clustering according to claim 4, characterized in that: The set threshold is a multiple of the root mean square value of the wavelet decomposition coefficients of the corresponding level.

6. The method for extracting partial discharge pulses based on noise reduction clustering according to claim 3, characterized in that: The noise suppression ratio is used as an evaluation index to evaluate the interference suppression effect.

7. The method for extracting partial discharge pulses based on noise reduction clustering according to claim 6, characterized in that: The noise suppression ratio is expressed as: NRR =10(lgσ1 2 -lgσ2 2 ), where σ1 and σ2 are the standard deviations of the original signal and the denoised signal, respectively.

Images