Next Article in Journal
Matching Aerial Images to 3D Building Models Using Context-Based Geometric Hashing
Next Article in Special Issue
Microsoft Kinect Visual and Depth Sensors for Breathing and Heart Rate Analysis
Previous Article in Journal
Systematic Calibration for Ultra-High Accuracy Inertial Measurement Units
Previous Article in Special Issue
Estimating 3D Leaf and Stem Shape of Nursery Paprika Plants by a Novel Multi-Camera Photography System
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

An Ultrahigh Frequency Partial Discharge Signal De-Noising Method Based on a Generalized S-Transform and Module Time-Frequency Matrix

1
School of Electrical Engineering, Wuhan University, No. 299, Bayi Road, Wuhan 430072, China
2
Guangzhou Power Supply Bureau Co. Ltd., No. 38, Huangshidong Road, Guangzhou 510620, China
*
Author to whom correspondence should be addressed.
Sensors 2016, 16(6), 941; https://doi.org/10.3390/s16060941
Submission received: 18 May 2016 / Revised: 17 June 2016 / Accepted: 20 June 2016 / Published: 22 June 2016
(This article belongs to the Special Issue Non-Contact Sensing)

Abstract

:
Due to electromagnetic interference in power substations, the partial discharge (PD) signals detected by ultrahigh frequency (UHF) antenna sensors often contain various background noises, which may hamper high voltage apparatus fault diagnosis and localization. This paper proposes a novel de-noising method based on the generalized S-transform and module time-frequency matrix to suppress noise in UHF PD signals. The sub-matrix maximum module value method is employed to calculate the frequencies and amplitudes of periodic narrowband noise, and suppress noise through the reverse phase cancellation technique. In addition, a singular value decomposition de-noising method is employed to suppress Gaussian white noise in UHF PD signals. Effective singular values are selected by employing the fuzzy c-means clustering method to recover the PD signals. De-noising results of simulated and field detected UHF PD signals prove the feasibility of the proposed method. Compared with four conventional de-noising methods, the results show that the proposed method can suppress background noise in the UHF PD signal effectively, with higher signal-to-noise ratio and less waveform distortion.

Graphical Abstract

1. Introduction

Partial discharge (PD) is a major cause and manifestation of insulation degradation. PD detection has been utilized for insulation condition assessment of high voltage (HV) apparatus [1]. When PD occurs in a HV apparatus, ultra-high frequency (UHF) electromagnetic waves are radiated from the PD source and propagate to the space of the substation through objects with no shielding effect [2]. External UHF antenna sensors can be utilized to detect UHF PD signals [3], however, in field PD tests, the UHF PD signals detected by antenna sensors are vulnerable to interference due to the numerous electromagnetic waves present in space [4]. The background noise interferes with the PD detection and causes PD pulse shapes to be distorted, which may negatively affect HV apparatus fault diagnosis and localization accuracy [5,6].Therefore, noise suppression is a vital issue of field UHF PD signal testing.
The background noise mainly includes the periodic narrowband noise and white Gaussian noise [7]. The periodic narrowband noise generally originates from wireless communication systems and may overwhelm the original PD signals. Several techniques have been proposed to suppress the narrowband noise [8,9,10,11,12,13]. Methods based on adaptive filters have been proposed to effectively suppress the narrowband noise [9,10], however, when the frequencies of the interferences and the PD signal are mixed, the de-noised PD signal is distorted due to the calculation error of the Fourier transform spectral analysis and the characteristics of the impulse response filter [11]. The methods based on wavelet transform (WT) were reported to extract PD signals from excessive narrowband noise [12], but due to the diversity of detected UHF PD signals, it is difficult to select the suitable mother wavelet and the number of decomposition-reconstruction levels [13].
The Gaussian white noise is generated by the heating effect of HV apparatus and will distort the UHF PD signals. Many methods have been proposed to suppress the white noise, such as the WT method [14], empirical mode decomposition method [15], and mathematical morphology method [16]. Recently, a novel method based on singular value decomposition (SVD) has been utilized for Gaussian white noise suppression, and more accurate recovery of the original PD signal was obtained [17]. A hard thresholding method based on standard deviation was employed to select effective singular values in this method.
In this paper, a novel de-noising method based on the module time-frequency matrix (MTFM) is proposed to suppress periodic narrowband noise and Gaussian white noise. The MTFM of detected UHF PD signals is constructed by employing the generalized S-transform time-frequency analysis technique. Frequencies and amplitudes of periodic narrowband noise can be calculated by utilizing the sub-matrix maximum module value method. The periodic narrowband noise is separated from the noisy UHF PD signals through the reverse phase superposition of the MTFM. Different from the conventional SVD de-noising method in [17,18], the singular values is calculated through decomposing the 2-dimensional MTFM. The fuzzy c-means (FCM) clustering method is employed to select effective singular values for recovering the UHF PD signals without Gaussian white noise. The de-noising results of simulated and field test UHF PD signals demonstrate the validity of the proposed method.

2. Algorithm of the Generalized S-Transform

As a non-stationary signal, the localized feature information of an UHF PD signal cannot be expressed only in time domain or frequency domain [19,20]. To extract information related to the time-frequency variation of a PD signal, Stockwell proposed the S-transform [21] to map time domain signal into the time-frequency domain, the S-transform of signal x(t) is defined as:
S ( τ , f ) = x ( t ) w ( t τ , f ) e j 2 π f t d t
w ( t τ , f ) = | f | 2 π e f 2 ( t τ ) 2 2
where t and τ are time, f is frequency, w(t − τ, f) is the Gaussian window function. The inverse S-transform is expressed as:
x ( t ) = [ S ( τ , f ) ] e j 2 π f t d f
The S-transform combines the separate strengths of the short-time Fourier transform and WT, which has provided an alternative approach to process non-stationary signals [22]. Because the height and the width of Gaussian window are varied by changing the frequency, the S-transform overcomes the defect of constant time-frequency resolution of the short time Fourier transform [23].
Because the form of the S-transform window function is invariant, the application is limited in some cases. Pinnegar proposed a generalized S-transform [24] by adding an adjustable factor λ to the Gaussian window function as the Equation (4) shown:
w ( t τ , f , λ ) = | λ f | 2 π e λ 2 f 2 ( t τ ) 2 2
where λ is the adjustable factor and λ > 0. Then the generalized S-transform is obtained as:
G ( τ , f , λ ) = x ( t ) w ( t τ , f , λ ) e j 2 π f t d t
In the generalized S-transform, λ < 1 corresponds to high frequency resolution, λ > 1 corresponds to high time resolution, λ = 1 corresponds to the standard S-transform window. In the application of generalized S-transform, an appropriate value of λ to the actual situation can be chosen to modify the time-frequency resolution.
According to Equation (5), as fn/NT, τiT, the discrete form of the generalized S-transform is expressed as:
{ G ( i T , n N T , λ ) = m = 0 N 1 X [ m + n N T ] e 2 π 2 m 2 / λ 2 n 2 e j 2 π m i / N , n 0 G ( i T , 0 ) = 1 N m = 0 N 1 x ( m N T ) , n = 0
where T is the sampling interval, N is the total number of sample points.
The generalized S-transform of x(t) is a 2-dimensional complex matrix in time-frequency domain, where the columns correspond to time sampling points and the rows correspond to time frequency sampling points [25]. For simplifying the calculations, the MTFM can be obtained by modeling every element in the complex matrix. The time-frequency distribution based on the MTFM reflects the time-frequency features of UHF PD signal. These features contribute to the analysis of the periodic narrowband noise [26].

3. De-Noising Method of Periodic Narrowband Noise

3.1. Simulated Signals

Due to the noise interference in field-detected UHF PD signals, the original PD signal cannot be obtained for illustrating the procedure and verifying the feasibility of the proposed method. An UHF PD signal is a type of transient signal with a short time duration [11]. Therefore, the single exponential decay oscillating impulses (pulses 1 and 2) and double exponential decay oscillating impulses (pulses 3 and 4) are utilized to simulate four types of UHF PD signals [12,13]. Equations (7) and (8) give the corresponding mathematical models, respectively:
Z 1 ( t ) = A 1 e t τ sin 2 π f c t
Z 2 ( t ) = A 2 ( e 1.3 t τ e 2.2 t τ ) sin 2 π f c t
where A is the amplitude, τ is the decay coefficient and fc is the oscillation frequency. Table 1 gives the parameters of each PD pulse. The simulated sampling frequency of each pulse is 10 Gs/s. Figure 1a shows the four types of UHF PD signals.
The waveform of practical periodic narrowband noise is sinusoidal with constant amplitude [27]. Equation (9) gives the mathematical model of periodic narrowband noise, where A is the amplitude assumed to be 2 mV. Based on the frequencies of wireless communication signals in China, fi is the frequency assumed to be 470 MHz, 900 MHz, 1800 MHz, respectively. The simulated UHF PD signals with periodic narrowband noise and Gaussian white noise are given in Figure 1b. The SNR of these noisy PD signals is −5.983 dB:
Z 3 ( t ) = A i = 1 n sin ( 2 π f i t )

3.2. Calculation Method of Frequency and Amplitude

3.2.1. Sub-Matrix Maximum Module Value Method

When the frequencies of the periodic narrowband noise and PD signals are partially overlapped, the PD pulse(s) are mixed with strong narrowband noise in a time-frequency distribution calculated by the S-transform, meaning that it is difficult to separate the narrowband noise from the PD pulse(s) [26]. The MTFM of the noisy PD signals calculated by the S-transform and the generalized S-transform (λ = 0.3), respectively, are drawn as a time-frequency distribution contour map in Figure 2. The comparison results show that higher frequency resolution can be obtained by utilizing the generalized S-transform (λ < 1), which may contribute to extract the narrowband noise features [26].
The time-frequency distribution of periodic narrowband noise is concentrated, with long duration, and the frequency distribution of a PD signal is discrete, with short duration. Therefore, both types of signals can be recognized in the time-frequency distribution. Due to the influence of Gaussian white noise, the accurate frequencies and amplitudes of periodic narrowband noise cannot be obtained directly from the time-frequency distribution. In this paper, a novel sub-matrix maximum module value (SMMV) method is proposed for calculating the frequencies and amplitudes of narrowband noise. The SMMV method is described as follows:
Step 1:
As shown in Figure 2b, parts of the periodic narrowband noise are divided into Regions A, B, C, and the corresponding localized time-frequency sub-matrices are extracted;
Step 2:
Search the coordinate positions of the maximum module value points in each column of each sub-matrix;
Step 3:
Based on the MTFM calculated by the generalized S-transform (λ < 1), record the corresponding frequency values of the maximum module value points (f1, f2, f3, …, fn) and the calculated time of each frequency value (t1, t2, t3, …, tn);
Step 4:
Calculate the frequencies fp of each periodic narrowband noise based on Equation (10):
f p = k = 1 n f k t k n
Step 5:
Record the module value of each sampling point (m1, m2, …, mn) in time on the calculated frequencies in Step (4), and calculate the module value mp of each corresponding narrowband noise in the MTFM based on Equation (11):
m p = k = 1 n m k n
Step 6:
Calculate the amplitudes of periodic narrowband noise by utilizing the inverse generalized S-transform, and reconstruct time domain waveform based on Equation (3).

3.2.2. Selection of the Adjustable Factor λ

In general, the generalized S-transform is a modified Fourier transform method, and the time-frequency resolution is still restricted by the Heisenberg uncertainty principle. This means that the time resolution decreases when the frequency resolution increases in a time-frequency distribution. Therefore, an appropriate value of λ contributes to the calculation of more accurate frequencies and amplitudes of periodic narrowband noise based on the SMMV method. Different values of λ were used to calculate the amplitudes and frequencies of each narrowband noise, with the results given in Table 2.
Compared with the original assumed value in the simulation, the relative errors of different calculated narrowband noise are given in Figure 3. Due to the improvement of frequency resolution, the relative errors of the calculated frequencies are reduced with the decrease in the value of λ. However, due to the reduction of time resolution, the relative errors of the calculated amplitudes are increased when the value of λ is lower than 0.3. The relative error is defined as:
Relative error = | O C | O × 100 %
where O and C are the amplitude (frequency) of the original and calculated periodical noise, respectively. The results show that the value of λ can be set between 0.3 and 0.4 to accurately calculate the frequencies and amplitudes of periodic narrowband noise. The value of λ is set to 0.3 in this paper.

3.3. Method of Periodic Narrowband Noise Suppression

To suppress the periodic narrowband noise, the MTFM based on the generalized S-transform is applied as follows:
Step 1:
Use the generalized S-transform to obtain the MTFM SM×N and time-frequency distribution of noisy UHF PD signals, where SM×N is the matrix with M rows and N columns ;
Step 2:
Extract the localized matrices of periodic narrowband noise without PD signals based on the time-frequency distribution;
Step 3:
Obtain the amplitudes and frequencies of periodic narrowband noise utilizing the above-described SMMV method;
Step 4:
Use the generalized S-transform to calculate the MTFM P1, P2, …, Pn of periodic narrowband noise based on Equation (6) and the results of Step (3), where Pk is the matrix with M rows and N columns;
Step 5:
Suppress the periodic narrowband noise by utilizing the reverse phase cancellation method is given in Equation (13), where WM×N is the MTFM without periodic narrowband noise:
W M × N = S M × N - k = 1 n P k
Step 6:
Use the inverse generalized S-transform to obtain the de-noised UHF PD signals in time domain of the WM×N.
Figure 4 gives the de-noising results of simulated noisy UHF PD signals utilizing the proposed method. The de-noising results indicate that periodic narrowband noise has been suppressed effectively and only Gaussian white noise exists in the UHF PD signals. The SNR of this signal is 7.085 dB.

4. De-Noising Method of Gaussian White Noise

4.1. Singular Value Decomposition De-Noising Method

Suppression of Gaussian white noise is vital to extract accurate UHF PD pulse waveforms. The singular value decomposition (SVD) de-noising method is a nonlinear filtering method, which was shown to be capable of effectively suppressing white noise [17,28]. The conventional SVD de-noising method procedure is given in Figure 5 and is illustrated as follows.
Step 1:
For a signal sequence x(i), i = 1, 2, …, N, a trajectory matrix of x(i) is defined as
A = [ x ( 1 ) x ( 2 ) x ( n ) x ( 2 ) x ( 3 ) x ( n + 1 ) x ( N n + 1 ) x ( N n + 2 ) x ( N ) ]
where ARm×n, 1 < n < N, and m = Nn + 1.
Step 2:
The SVD of this real matrix ARm×n is defined as:
A = U Λ V T
where U and V are the orthogonal matrix, URm×n, VRm×n. Λ = [diag(a1, a2, …, aq), O] or its transposition matrix, which is determined by m < n or mn. O is the zero matrix, q = min(m, n), a1a2 ≥…≥ aq. ai, i = 1, 2, …, q, are the singular values of matrix.
Step 3:
In matrix Λ, the effective singular values corresponding to a PD signal are reserved while the corresponding white noise are set to zero, and thus a new matrix Λre can be obtained.
Step 4:
A new trajectory matrix Are is reconstructed by utilizing Λre, the original U and V through Equation (16), and the de-noised PD signals are recovered from the trajectory matrix Are:
A r e = U Λ r e V T

4.2. Selection Method of Effective Singular Values

Because the trajectory matrix is created by truncating the signal in the time domain, SVD results only contain time-amplitude information. It means that the recovered de-noised PD signal may be partially distorted [18]. The MTFM of the noisy PD signals is set as the trajectory matrix A in this paper. Eigenvectors of frequency and time are represented by the orthogonal matrices U and V, respectively. SVD results of the MTFM contain time-frequency-amplitude information, meaning that some characteristic quantities of another dimension can be obtained in the U, V and singular values. Therefore, the recovered PD signal after de-noising will be more accurate compared with the conventional algorithm ones.
For the SVD de-noising method, the most important step is how to select the number of effective singular values, which largely determines the de-noising effect. The recovered PD signal is distorted when too small a number of effective singular values is selected, and residual white noise would remain when selecting over many effective singular values [17]. The methods proposed in [17,18] selected effective singular values by the hard thresholding technique, which may deteriorate the original UHF PD signal. In the MTFM, as Figure 4b showed, the module values corresponding to PD signal are greater with a concentrated distribution. On the contrary, the module values corresponding to Gaussian white noise are lower with a scattered distribution. Therefore, the magnitudes of singular values corresponding to PD signals are greater with a large difference. In this paper, a novel method for selecting effective singular values based on the fuzzy c-means (FCM) clustering algorithm is proposed. Features fe1 and fe2 are extracted as the features for clustering when selecting effective singular values. fe1 and fe2 are shown in Equation (17):
{ f e 1 = a 1 , a 2 , ... , a n f e 2 = b 1 , b 2 , ... , b n , b i = a i a i 1 , i = 1 , 2 , ... , q 1
The basic principle of FCM can be described briefly as follows. Given a dataset of PD features F = {fe1, fe2, ···, feN}, where fej = (fe1, fe2, ···, fem)T. FCM aims to minimize the c-means function, which is defined as:
min f ( U U , V V ) = i = 1 c j = 1 N ( μ i j ) m f e j - v i 2
s . t . { i = 1 c μ i j = 1 , ( 1 j N ) 0 μ i j 1 , ( 1 i c , 1 j N ) 0 j = 1 N μ i k N , ( 1 i c )
where c is the pre-defined cluster number, VV indicates the cluster center matrix, vi is the cluster center of class i, 1 ≤ i ≤ c, UU = [μij]c×N represents the fuzzy membership degree matrix, μij is the fuzzy membership degree of fj belong to class i, m is the weighting coefficient, and m is set as 2 in this paper, dij = //fj-vi//2 represents the Euclidean distance between sample fj and vi. In order to obtain the minimum of f (UU, VV) and find the optimal UU = [μij]c×N and vi, a Lagrange multiplier is formed to solve the optimization problem:
F = i = 1 c ( μ i j ) m ( d i j ) 2 - λ ( i = 1 c μ i j - 1 )
The optimal conditions of Equation (20) are:
{ F λ = i = 1 c μ i j 1 = 0 F μ i j = [ m ( μ i j ) m 1 ( d i j ) 2 λ ] = 0
According to Equation (21), μij is calculated uing Equation (22). Then by solving f ( U U , V V ) v i = 0 , vi is calculated as in Equation (23). More details of the FCM method can be found in [29]:
μ i j = k = 1 c ( d i j d k j ) 2 1 m
v i = j = 1 N ( μ i j ) m · f e j j = 1 N ( μ i j ) m

4.3. De-Noising Method of Gaussian White Noise

The suppression procedure of Gaussian white noise based on the MTFM and FCM clustering involves the following steps:
Step 1:
Calculate singular values of the MTFM of the noisy PD signals by SVD;
Step 2:
Calculate features of singular values sequence based on Equation (17);
Step 3:
Classify singular values into two groups utilizing FCM clustering algorithm, which represent UHF PD signals and Gaussian white noise, respectively;
Step 4:
Select the group whose maximum value is greater as the effective singular values, the number of effective singular values is defined as k in this paper;
Step 5:
Recover the de-noised PD signal by the conventional SVD de-noising method.
The calculated singular values of the MTFM are given in Figure 6, and the classification result utilizing the FCM clustering algorithm is given in Figure 7. The value of k was set as 4 in this paper. The de-noising results are given in Figure 8. The results show that Gaussian white noise is suppressed successfully by using the proposed method. When the number of effective singular values was set to k − 1 and k + 1, the de-noised PD signals are given in the Figure 9. The first pulse of recovered PD signals is distorted severely when k = 3, and the SNR of the de-noised signals will be lowered when k = 5. The comparison of results indicates that the proposed method based on FCM clustering algorithm is capable of selecting effective singular values.

5. Simulation De-Noising Results and Discussions

The procedure of the proposed de-noising method is summarized in Figure 10. For comparison, four conventional de-noising methods [9,10,11,12,13,17] were employed to suppress the noisy UHF PD signals shown in Figure 1b. The types of noise and de-noising techniques are shown in Table 3. The db2 and db8 wavelets are selected as the mother wavelet in Methods C and D where the decomposition level was set as 10. The de-noised PD signals are given in Figure 11. To compare the performance of various de-noising methods, the following evaluation parameters are introduced [14,30,31].
(I) The SNR illustrates the effectiveness of noise suppression, which is defined as:
SNR = 10 log 10 ( i = 1 N | s ( i ) | 2 i = 1 N | r ( i ) s ( i ) | 2 )
where s(i) is the original signal and r(i) is the de-noised signal. Higher SNR value means more effective background noise suppression.
(II) The root-mean-square error (RMSE) is defined as:
RMSE = 1 N i = 1 N | r ( i ) s ( i ) | 2
RMSE is used to evaluate waveform distortion of de-noised signal compared with the original signal. Lower RMSE value means less waveform distortion of UHF PD signal waveform.
(III) The normalized correlation coefficient (NCC) is defined as:
NCC = i = 1 N s ( i ) · r ( i ) ( i = 1 N s 2 ( i ) ) · ( i = 1 N r 2 ( i ) )
NCC is used to evaluate waveforms similarity between the original and de-noised signal. The NCC value is between the negative one (−1) and positive one (+1), −1 means fore-and-aft waveform reverse, zero means orthogonal, and +1 means almost the same.
(IV) The variation trend parameter (VTP) is composed of rise variation trend parameter (RVTP) and fall variation trend parameter (FVTP).
RVTP = i = 2 N [ r ( i ) r ( i 1 ) ] i = 2 N [ s ( i ) s ( i 1 ) ]
where s(i) > s(i − 1), r(i) > r(i − 1).
FVTP = i = 2 N [ r ( i 1 ) r ( i ) ] i = 2 N [ s ( i 1 ) s ( i ) ]
where s(i) < s(i − 1), r(i) < r(i − 1).
VTP is the mean value of RVTP and FVTP.
VTP = RVTP + FVTP 2
VTP is used to describe the similarity of the wave variation tendency, which measures the oscillating situation of the waveform. When the VTP value is close to 1, the variation tendency of two waveforms is the most similar.
Calculated results of various evaluation parameters are given in Table 4. The comparison results of de-noised PD signals and evaluation parameters show that:
(1)
The adaptive filtering method is capable of suppressing periodic narrowband noise. However, because the frequencies of narrowband noise and UHF PD signals are mixed, the de-noised PD signal waveforms are distorted.
(2)
Due to the large variation in time-frequency characteristics of UHF PD signals, de-noising results are affected by the mother wavelet selection [11]. In this paper, the de-noised signals with lower SNR and larger waveform distortion are obtained by utilizing Method C (select db2 wavelet); the de-noised signals with higher SNR and less waveform distortion are obtained by utilizing Method D (select db8 wavelet). Therefore, the de-noising effectiveness when using wavelet decomposition may become worse when an unsuitable mother wavelet is selected.
(3)
Because the effective singular values calculated from the MTFM represent more time-frequency-amplitude information, the proposed method in this paper is capable of obtaining the de-noised UHF PD signals with higher SNR and less waveform distortion compared with conventional SVD de-noising method.
(4)
Compared with four conventional de-noising methods, the proposed method based on the MTFM is capable of suppressing the periodic narrowband noise and Gaussian white noise successfully. Hence, by the proposed method, the de-noised UHF PD signals are similar to the original signals.
The pulse shapes of de-noised UHF PD signal utilizing Method A are given in Figure 12. The original and the de-noised PD pulses were drawn in the black dashed line and red solid line, respectively. The results show that the de-noising method proposed in this paper can recover the original PD pulses successfully and thus can assist in feature parameters extraction of UHF PD signals for insulation defect type recognition.

6. De-Noising Results of Field Detected Signals

To verify the effectiveness of the proposed de-noising method for field detected PD signals, two UHF PD signals were detected from two 500 kV substations in Guangzhou, China. The picture of the field detection setup is given in Figure 13. An external horn antenna sensor (AInfo-LB530NF, A-Info Technology Co., Ltd, Chengdu, China) was employed to detect the PD signals. The operation bandwidth of antenna sensor was 0.5–3 GHz, and the maximum antenna gain was 11 dBi. The detected UHF PD signals were recorded using a digital oscilloscope (Lecory-WavePro740, Teledyne LeCroy, New York, NY, USA), which has 4 GHz bandwidth, 20 GSamples/s sampling rate, central processing unit (CPU) with 2.6 GHz dominant frequency and random access memory (RAM) with 2 G memory. The phase resolved partial discharge (PRPD) patterns corresponding to two detected UHF signals are given in Figure 14, and according to [20,31], the detected signals were originated from the PD sources. The time-domain waveforms and time-frequency distributions of two detected PD signals are given in Figure 15. It is obviously that narrowband noise and white noise interfered with the detected PD signals severely.
The calculated frequencies and amplitude of periodic narrowband noise employing the SMMV method are given in Table 5.
Periodic narrowband noise originates in local mobile phone wireless communication signals. The de-noising results of field detected PD signals by the methods in Table 3 are given in Figure 16. Because the original PD signal without noise is not obtained as a reference in field detection, the evaluation parameters introduced in Equations (24)–(29) are not available. Therefore, noise reduction ratio (NRR) and the amplitude reduction ratio (ARR) are proposed to evaluate the de-noising results [32], which are defined as:
NRR = 10 ( log 10 σ 1 2 log 10 σ 2 2 )
ARR = Y Z Y × 100 %
where σ1 and σ2 are the standard deviation of detected signal and de-noised signal, Y and Z are the maximum amplitude of detected noisy signal and de-noised signal. High NRR and low ARR means effective de-noising results. The evaluation parameters using each de-noising method are given in Table 6. The result shows that the proposed method is capable of suppressing noise effectively with less amplitude reduction.
The computation of proposed de-noising method was carried out by MATLAB codes in computer platform of oscilloscope. The calculating time of de-noising two UHF PD signals is 4.751 s and 4.897 s, respectively. It means that the proposed de-noising method is feasible in field test, and the de-noised results can be obtained in real time.

7. Conclusions

To suppress the periodic narrowband noise and Gaussian white noise in UHF PD signal, a novel de-noising method is proposed based on the generalized S-transform and MTFM in this paper. The results are concluded as follows:
(1)
To suppress periodic narrowband noise, the SMMV method based on generalized S-transform and MTFM is employed. By calculating the frequencies and amplitudes of the narrowband noise, the corresponding MTFMs are obtained and removed.
(2)
To suppress Gaussian white noise, singular values are calculated through decomposing the 2-dimensional MTFM based on the generalized S-transform. Effective singular values can be selected successfully by employing the FCM clustering method.
(3)
De-noising results of simulated and field detected UHF PD signals validate the feasibility of this method. Compared with some conventional de-noising methods, the proposed method obtains de-noised signals with high SNR and less waveform distortion.

Acknowledgments

This research was supported by Guangzhou Power Supply Co. Ltd. (Guangzhou, China), the field tests were performed in the Guangnan and Ruibao 500 kV substations in Guangzhou, China.

Author Contributions

Yushun Liu and Wenjun Zhou conceived the algorithm and wrote the manuscript. Yushun Liu, Pengfei Li and Shuai Yang designed and performed the experiment. Yan Tian provided the field detection equipment. All authors have read and approved the final manuscript.

Conflicts of Interest

The authors declare no conflict of interest.

Abbreviations

The following abbreviations are used in this manuscript:
PDPartial discharge
UHFUltrahigh frequency
HVHigh voltage
WTWavelet transform
SNRSignal noise ratio
MTFMModule time-frequency matrix
FCMFuzzy c-means
SMMVSub-matrix maximum module value method
SVDSingular value decomposition
PRPDPhase resolved partial discharge
CPUCentral processing unit
RAMRandom access memory
RMSERoot mean square error
NCCNormalized correlation coefficient
VTPVariation trend parameter
RVTPRise variation trend parameter
FVTPFall variation trend parameter
NRRNoise reduction ratio
ARRAmplitude reduction ratio

References

  1. Álvarez, F.; Garnacho, F.; Ortego, J.; Sánchez-Urán, M.A. Application of HFCT and UHF sensors in on-line partial discharge measurements for insulation diagnosis of high voltage equipment. Sensors 2015, 15, 7360–7387. [Google Scholar] [CrossRef] [PubMed]
  2. Hou, H.; Sheng, G.; Jiang, X. Robust time delay estimation method for locating UHF signals of partial discharge in substation. IEEE Trans. Power Deliv. 2013, 28, 1960–1968. [Google Scholar]
  3. Hou, H.; Sheng, G.; Jiang, X. Localization algorithm for the PD source in substation based on L-shaped antenna array signal processing. IEEE Trans. Power Deliv. 2015, 30, 472–479. [Google Scholar] [CrossRef]
  4. Robles, G.; Sánchez-Fernández, M.; Albarracin-Sanchez, R.; Rojas-Moreno, M.V.; Rajo-Iglesias, E.; Martínez-Tarifa, J.M. Antenna parametrization for the detection of partial discharges. IEEE Trans. Instrum. Meas. 2013, 62, 932–941. [Google Scholar] [CrossRef]
  5. Robles, G.; Fresno, J.M.; Martínez-Tarifa, J.M. Separation of radio-frequency sources and localization of partial discharges in noisy environments. Sensors 2015, 15, 9882–9898. [Google Scholar] [CrossRef] [PubMed]
  6. Hou, H.; Sheng, G.; Li, S.; Jiang, X. A novel algorithm for separating multiple PD sources in a substation based on spectrum reconstruction of UHF signals. IEEE Trans. Power Deliv. 2015, 30, 809–817. [Google Scholar] [CrossRef]
  7. Sriram, S.; Nitin, S.; Prabhu, K.M.M.; Bastiaans, M.J. Signal denoising techniques for partial discharge measurements. IEEE Trans. Dielectr. Electr. Insul. 2005, 12, 1182–1191. [Google Scholar] [CrossRef]
  8. Luo, G.; Zhang, D.; Tseng, K.J.; He, J. Impulsive noise reduction for transient Earth voltage-based partial discharge using Wavelet-entropy. IET Sci. Meas. Technol. 2016, 10, 69–76. [Google Scholar] [CrossRef]
  9. Khan, S.Z.; Zhu, H.; Jin, X.; Tan, K. A new adaptive technique for on-line partial discharge monitoring. IEEE Trans. Dielectr. Electr. Insul. 1995, 2, 700–707. [Google Scholar] [CrossRef]
  10. Köpf, U.; Feser, K. Rejection of narrow-band noise and repetitive pulses in on-site PD measurements. IEEE Trans. Dielectr. Electr. Insul. 1995, 2, 433–446. [Google Scholar] [CrossRef]
  11. Satish, L.; Nazneen, B. Wavelet-based denoising of partial discharge signals buried in excessive noise and interference. IEEE Trans. Dielectr. Electr. Insul. 2003, 10, 354–367. [Google Scholar] [CrossRef]
  12. Ma, X.; Zhou, C.; Kemp, I.J. Interpretation of wavelet analysis and its application in partial discharge detection. IEEE Trans. Dielectr. Electr. Insul. 2002, 9, 446–457. [Google Scholar] [CrossRef] [Green Version]
  13. Ma, X.; Zhou, C.; Kemp, I.J. Automated wavelet selection and thresholding for PD detection. IEEE Elect. Insul. Mag. 2002, 18, 37–45. [Google Scholar] [CrossRef]
  14. Xu, Z.; Tang, J.; Caixin, S. Application of complex wavelet transform to suppress white noise in GIS UHF PD signals. IEEE Trans. Power Deliv. 2007, 22, 1498–1504. [Google Scholar]
  15. Tang, Y.; Tai, C.; Su, C.; Chen, C.; Chen, J. A correlated empirical mode decomposition method for partial discharge signal denoising. Meas. Sci. Technol. 2010, 21, 085106. [Google Scholar] [CrossRef]
  16. Ashtiani, M.; Shahrtash, S. Feature-oriented de-noising of partial discharge signals employing mathematical morphology filters. IEEE Trans. Dielectr. Electr. Insul. 2012, 19, 2128–2136. [Google Scholar] [CrossRef]
  17. Ashtiani, M.; Shahrtash, S. Partial discharge de-noising employing adaptive singular value decomposition. IEEE Trans. Dielectr. Electr. Insul. 2014, 21, 775–782. [Google Scholar] [CrossRef]
  18. Zhao, X.; Ye, B. Selection of effective singular values using difference spectrum and its application to fault diagnosis of headstock. Mech. Syst. Signal Process. 2011, 25, 1617–1631. [Google Scholar] [CrossRef]
  19. Wang, K.; Liao, R.; Yang, L.; Li, J.; Grzybowski, S.; Hao, J. Optimal features selected by NSGA-II for partial discharge pulses separation based on time-frequency representation and matrix decomposition. IEEE Trans. Dielectr. Electr. Insul. 2013, 20, 825–838. [Google Scholar] [CrossRef]
  20. Wang, K.; Li, J.; Zhang, S.; Liao, R.; Wu, F.; Yang, L.; Li, J.; Grzybowski, S.; Yan, J. A hybrid algorithm based on s transform and affinity propagation clustering for separation of two simultaneously artificial partial discharge sources. IEEE Trans. Dielectr. Electr. Insul. 2015, 22, 1042–1060. [Google Scholar] [CrossRef]
  21. Stockwell, R.G.; Mansinha, L.; Lowe, R.P. Localization of the complex spectrum: The S transform. IEEE Trans. Signal Process. 1996, 44, 998–1001. [Google Scholar] [CrossRef]
  22. Dash, P.K.; Panigrahi, B.K.; Panda, G. Power quality analysis using S-transform. IEEE Trans. Power Deliv. 2003, 18, 406–411. [Google Scholar] [CrossRef]
  23. Huang, N.; Zhang, S.; Cai, G.; Xu, D. Power quality disturbances recognition based on a multiresolution generalized S-transform and a PSO-improved decision tree. Energies 2015, 8, 549–572. [Google Scholar] [CrossRef]
  24. Pinnegar, C.R.; Mansinha, L. The S-transform with windows of arbitrary and varying shape. Geophysics 2003, 68, 381–385. [Google Scholar] [CrossRef]
  25. Xiao, X.; Xu, F.; Yang, H. Short duration disturbance classifying based on S-transform maximum similarity. Int. J. Electr. Power 2009, 31, 374–378. [Google Scholar] [CrossRef]
  26. Li, B.; Zhang, P.; Liu, D.; Mi, S.; Ren, G.; Tian, H. Feature extraction for rolling element bearing fault diagnosis utilizing generalized S transform and two-dimensional non-negative matrix factorization. J. Sound Vib. 2011, 330, 2388–2399. [Google Scholar] [CrossRef]
  27. Sturm, C.; Wiesbeck, W. Waveform design and signal processing aspects for fusion of wireless communications and radar sensing. IEEE Proc. 2011, 99, 1236–1259. [Google Scholar] [CrossRef]
  28. Zhao, X.; Ye, B. Similarity of signal processing effect between Hankel matrix-based SVD and wavelet transform and its mechanism analysis. Mech. Syst. Signal Process. 2009, 23, 1062–1075. [Google Scholar] [CrossRef]
  29. Bezdek, J.C.; Ehrlich, R.; Full, W. FCM: The fuzzy c-means clustering algorithm. Comput. Geosci. UK 1984, 10, 191–203. [Google Scholar] [CrossRef]
  30. Zhang, X.; Zhou, J.; Li, N.; Wang, Y. Suppression of UHF partial discharge signals buried in white-noise interference based on block thresholding spatial correlation combinative de-noising method. IET Gener. Transm. Distrib. 2012, 6, 353–362. [Google Scholar] [CrossRef]
  31. Ardila-Rey, J.A.; Martinez-Tarifa, J.A.; Robles, G.; Rojas-Moreno, M.V. Partial discharge and noise separation by means of spectral-power clustering techniques. IEEE Trans. Dielectr. Electr. Insul. 2013, 20, 1436–1443. [Google Scholar] [CrossRef]
  32. Hussein, R.; Shaban, K.B.; El-Hag, A.H. Wavelet transform with histogram-based threshold estimation for online partial discharge signal denoising. IEEE Trans. Instrum. Meas. 2015, 64, 3601–3614. [Google Scholar] [CrossRef]
Figure 1. Simulated UHF PD signals: (a) Simulated PD signals; (b) Simulated noisy PD signals.
Figure 1. Simulated UHF PD signals: (a) Simulated PD signals; (b) Simulated noisy PD signals.
Sensors 16 00941 g001
Figure 2. The time-frequency distribution of the noisy PD signals calculated by (a) the S-transform; and (b) the generalized S-transform.
Figure 2. The time-frequency distribution of the noisy PD signals calculated by (a) the S-transform; and (b) the generalized S-transform.
Sensors 16 00941 g002
Figure 3. Relative errors by using different value of λ: (a) frequency; (b) amplitude.
Figure 3. Relative errors by using different value of λ: (a) frequency; (b) amplitude.
Sensors 16 00941 g003
Figure 4. UHF PD signals with suppressed periodic narrowband noise: (a) in time domain; (b) in time-frequency domain.
Figure 4. UHF PD signals with suppressed periodic narrowband noise: (a) in time domain; (b) in time-frequency domain.
Sensors 16 00941 g004
Figure 5. Procedure of the conventional SVD de-noising method.
Figure 5. Procedure of the conventional SVD de-noising method.
Sensors 16 00941 g005
Figure 6. Calculated singular values by decomposing the MTFM.
Figure 6. Calculated singular values by decomposing the MTFM.
Sensors 16 00941 g006
Figure 7. Selection results of effective singular values based on the FCM clustering algorithm.
Figure 7. Selection results of effective singular values based on the FCM clustering algorithm.
Sensors 16 00941 g007
Figure 8. De-noised PD signals: (a) time-domain signals; (b) time-frequency distribution.
Figure 8. De-noised PD signals: (a) time-domain signals; (b) time-frequency distribution.
Sensors 16 00941 g008
Figure 9. De-noised PD signals employing SVD when (a) k = 3; and (b) k = 5.
Figure 9. De-noised PD signals employing SVD when (a) k = 3; and (b) k = 5.
Sensors 16 00941 g009
Figure 10. Procedure of the proposed de-noising method.
Figure 10. Procedure of the proposed de-noising method.
Sensors 16 00941 g010
Figure 11. De-noising results by employing each method: (a) Method B; (b) Method C; (c) Method D; and (d) Method E.
Figure 11. De-noising results by employing each method: (a) Method B; (b) Method C; (c) Method D; and (d) Method E.
Sensors 16 00941 g011
Figure 12. Comparisons of the original and the de-noised PD pulse shapes. (a) Pulse 1; (b) Pulse 2; (c) Pulse 3; (d) Pulse 4.
Figure 12. Comparisons of the original and the de-noised PD pulse shapes. (a) Pulse 1; (b) Pulse 2; (c) Pulse 3; (d) Pulse 4.
Sensors 16 00941 g012
Figure 13. Field detection setup in substation.
Figure 13. Field detection setup in substation.
Sensors 16 00941 g013
Figure 14. PRPD patterns corresponding to the detected UHF signals: (a) signal #1; (b) signal #2.
Figure 14. PRPD patterns corresponding to the detected UHF signals: (a) signal #1; (b) signal #2.
Sensors 16 00941 g014
Figure 15. Waveforms and time-frequency distributions detected UHF PD signals: (a) signal #1; (b) signal #2.
Figure 15. Waveforms and time-frequency distributions detected UHF PD signals: (a) signal #1; (b) signal #2.
Sensors 16 00941 g015
Figure 16. De-noised PD signals in field detection using each method: (a) signal #1; (b) signal #2.
Figure 16. De-noised PD signals in field detection using each method: (a) signal #1; (b) signal #2.
Sensors 16 00941 g016
Table 1. Parameters of the simulated UHF PD signals.
Table 1. Parameters of the simulated UHF PD signals.
PD Pulse Sequence1234
τ/ns0.510.51
fc/MHz60060010001000
A/mV10401040
Table 2. Calculated amplitudes and frequencies with changing values of λ.
Table 2. Calculated amplitudes and frequencies with changing values of λ.
λNarrowband Noise #ANarrowband Noise #BNarrowband Noise #C
Frequency (MHz)Amplitude (mV)Frequency (MHz)Amplitude (mV)Frequency (MHz)Amplitude (mV)
0.1470.0851.949900.0321.9191800.1941.842
0.2470.0851.954900.0111.9251800.1941.883
0.3470.0851.986900.0761.9311800.8941.912
0.4470.3121.984899.3911.9331802.1401.924
0.5470.4761.979901.3281.9261802.9541.927
0.6470.8511.974898.3981.9151803.8231.925
0.7470.9941.976901.8521.9291804.9741.916
0.8471.2651.968902.3471.9241805.2931.917
0.9471.3251.968903.4811.9191806.9051.913
1.0471.7851.962903.9111.9181807.4881.904
1.1472.0461.953904.0561.9111809.2261.901
Table 3. Conventional de-noising methods.
Table 3. Conventional de-noising methods.
MethodNoise TypeTechnique
ANarrowband noise + White noiseProposed method in this paper
BNarrowband noise + White noiseAdaptive IIR filter
CNarrowband noise + White noiseAdaptive threshold WT de-noising (db2)
DNarrowband noise + White noiseAdaptive threshold WT de-noising (db8)
EWhite noiseConventional SVD de-noising method
Table 4. De-noising evaluation parameters using each method.
Table 4. De-noising evaluation parameters using each method.
Evaluation ParameterDe-Noising MethodPulse 1Pulse 2Pulse 3Pulse 4All
SNRMethod A14.44519.48216.79420.98716.473
Method B2.8751.8323.8922.9682.712
Method C−0.871−0.362−1.796−1.704−1.647
Method D0.5850.3151.6071.8610.941
Method E6.25487.6298.6629.0298.342
RMSEMethod A0.00138.3 × 10−47.6 × 10−43.7 × 10−43.4 × 10−4
Method B0.02480.04180.01960.03270.0198
Method C0.05870.06970.02340.02120.0432
Method D0.05510.04970.01890.02120.0347
Method E0.01240.01330.01660.03900.0134
NCCMethod A0.98980.99420.99340.99700.9907
Method B0.74160.64460.81180.72560.5773
Method C0.35030.45060.51220.54180.4301
Method D0.41520.57380.65770.74790.6209
Method E0.88320.91120.80350.53890.8251
VTPMethod A1.07451.03821.02601.05151.0476
Method B0.78140.67670.70420.72370.7432
Method C1.76451.28871.85541.75921.6346
Method D1.14841.06061.43271.30841.2636
Method E1.08521.07900.71570.78800.8956
Table 5. Calculated frequencies and amplitude of narrowband noise in field detection.
Table 5. Calculated frequencies and amplitude of narrowband noise in field detection.
SignalFrequency (MHz)Amplitude (mV)
UHF signal 1#875.91.03
941.12.16
UHF signal 2#876.43.53
961.31.24
Table 6. De-noising evaluation parameters of field detected signals.
Table 6. De-noising evaluation parameters of field detected signals.
SignalDe-Noising MethodEvaluation Parameter
NRRARR
PD Signal #1Method A15.9621.4%
Method B10.5934.6%
Method C13.2127.6%
Method D13.3641.3%
PD Signal #2Method A19.2332.5%
Method B17.2548.3%
Method C16.6843.4%
Method D17.6636.6%

Share and Cite

MDPI and ACS Style

Liu, Y.; Zhou, W.; Li, P.; Yang, S.; Tian, Y. An Ultrahigh Frequency Partial Discharge Signal De-Noising Method Based on a Generalized S-Transform and Module Time-Frequency Matrix. Sensors 2016, 16, 941. https://doi.org/10.3390/s16060941

AMA Style

Liu Y, Zhou W, Li P, Yang S, Tian Y. An Ultrahigh Frequency Partial Discharge Signal De-Noising Method Based on a Generalized S-Transform and Module Time-Frequency Matrix. Sensors. 2016; 16(6):941. https://doi.org/10.3390/s16060941

Chicago/Turabian Style

Liu, Yushun, Wenjun Zhou, Pengfei Li, Shuai Yang, and Yan Tian. 2016. "An Ultrahigh Frequency Partial Discharge Signal De-Noising Method Based on a Generalized S-Transform and Module Time-Frequency Matrix" Sensors 16, no. 6: 941. https://doi.org/10.3390/s16060941

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop