Development of AE Monitoring System with Noise Reduction Function by Spectral Subtraction

(1)

Development of AE Monitoring System with Noise Reduction Function

by Spectral Subtraction

Takuma Matsuo and Hideo Cho

Department of Mechanical Engineering, Faculty of Science and Engineering, Aoyama Gakuin University, Sagamihara 229-8558, Japan

In order to monitor acoustic emission (AE) in a noisy environment, an AE monitoring system with a real-time noise reduction function using spectral subtraction (SS) was developed. First, the improvement in the S/N ratio and distortion of the waveform after the process were compare to those in¾-ﬁlter and wavelet shrinkage methods. The improvement in the S/N ratio of the waveform processed by SS constantly increased and was observed to be independent of the S/N ratio of the wave before processing. The distortion of the waveform processed by SS was less than the distortions of the waves processed using¾-ﬁlter and by wavelet shrinkage. Next, the effect of two parameters in the SS process®frame number (fn) and over-subtraction factor (¡)®on the noise reduction performance was studied. The S/N ratio of the signals processed by the SS technique improved with decreasingfnand with increasing¡. However, the processed waveform was distorted when¡was large. It is necessary to set the value offnto 16 or less and¡to 5 or less so that SS showed an advantage to reduce noise with low waveform distortion.

Cylindrical wave AE signals produced by the Hsu-Nielsen source (pencil lead breaking) were monitored with the developed system in an environment with artiﬁcial noise. The developed system was able to acquire AE signals with a sampling frequency of up to 25 MHz. The noise was reduced and AE signals with S/N ratio of 0 dB before the process could be detected. The S/N ratio of the AE signal was improved by approximately 10 dB using the SS technique. [doi:10.2320/matertrans.I-M2011855]

(Received November 9, 2009; Accepted October 29, 2011; Published January 18, 2012)

Keywords: acoustic emission, spectral subtraction, noise reduction, signal processing, real-time monitoring

1. Introduction

An acoustic emission (AE) technique has been utilized for

a condition monitoring of structures1,2) _{and machines.}3)

However, it is a difﬁculty to monitor AE signals in a noisy

environment where vibration or electromagnetic noise exist

as environmental noise. A frequency ﬁlter is generally used

to reduce background noise, however, a frequency ﬁlter is

unable to considerably reduce wide frequency band noise or noise due to variation in frequency with time. Noise is not just a problem associated with AE monitoring but a common

problem in the ﬁeld of sound engineering, medical

electronics and geophysics explorations that involve signal

detection.4)In order to overcome this problem, various signal

processing methods were proposed. Domoho proposed a

de-noising technique involving a discrete wavelet transform.5)

Arakawa et al. proposed a nonlinear ﬁlter called ¾-ﬁlter.6)

Lim and Oppenheim proposed a Wienerﬁlter that utilized a

mean square error.7)_{Boll proposed spectral subtraction (SS),}

a technique for reducing noise on the basis of the frequency

spectrum.8)_{SS is a computationally ef}_ﬁ_{cient technique and is}

widely used for noise reduction of speech signals.9,10)_{It has}

also been used for noise reduction of AE signals.11)_{We also}

applied SS for AE signals and embedded it in a real-time

noise reduction system by using SS.12) However, noise

reduction performance has not been compared to other techniques. Moreover, the previous system has a dead time of 30 ms.

In this study, we developed an AE monitoring system with a real-time noise reduction function using the SS technique with no dead-time. First, we compared the noise reduction performance between SS and other techniques for the inherent background noise in the monitoring system and the sensor. Then, we monitored the cylindrical wave AE

signals produced by the HsuNielsen source (pencil lead

breaking) in a noisy environment with a wide frequency band.

2. Spectral Subtraction

The AE signal x(t) is the combined waveform of the

original signals(t) and noise signaln(t), detected by the AE

monitoring system. The AE signal x(t) and its frequency

spectrumX(f) are expressed by

xðtÞ ¼sðtÞ þnðtÞ ð1Þ XðfÞ ¼SðfÞ þNðfÞ ð2Þ

whereX(f),S(f), andN(f) represent the frequency spectra of

x(t),s(t) andn(t), respectively. The spectrum of the original

signal can be estimated by subtracting the frequency spectrum

of the noise signal, N(f), from X(f). Figure 1 shows a

schematic diagram of the SS process used in this study.

(1) The detected signalx(t) was divided intofnequal parts.

Here,fntakes an arbitrary number. In Fig. 1, fnwas 4.

In this study, fn was 4, 8, 16, 32, and 64. The data

length in one frame depends on the frame number. (2) A fast Fourier transform (FFT) was performed for all

frames individually.

(3) As N(f) should be determined before the noise

reduction process, the signal in the ﬁrst frame was

assumed to be the noise component.

(4) The phases of each divided signal (ª1(f),ª2(f),+) was

calculated.

(5) The frequency spectrum of the original signal S(f) is

given by

SðfÞ ¼ XðfÞ ¡NðfÞ (XðfÞ>¡NðfÞ)

0 (otherwise)

ð3Þ

where¡is an over-subtraction factor. In this paper, the

value of¡ is varied from 1 to 8.

Special Issue on APCNDT 2009

(2)

(6) After an inverse fast Fourier transform (IFFT) was performed for the obtained frequency spectra of the original signal, divided signals were combined in the same order of the detected signal.

3. Noise Reduction by Spectral Subtraction and Comparison of Noise Reduction Performance

[image:2.595.108.490.70.309.2]

3.1 Noise reduction by spectral subtraction

Figure 2 shows the experimental setup for estimation of noise reduction performance in AE monitoring. Two AE transducers (PAC, PICO center frequency: 450 kHz) were mounted on a type-304 stainless steel plate with 1 mm thickness. One transducer was used for the excitation of an

artiﬁcial AE signal and another transducer was used as a

sensor for detecting the signal. AE was excited by applying 1 cycle of sine voltage at a frequency of 450 kHz to the excitation transducer which was placed 120 mm away from

the detection transducer. The input voltage Vin for the

excitation transducer was varied from 70 mV to 2.0 V in

order to change the S/N ratio of the AE signal. The detected

AE signal was ampliﬁed by 40 dB with a pre-ampliﬁer and

was fed to a digitizer. The sampling interval and data points were set to 40 ns and 4096 points, respectively. Figure 3 shows a comparison between waveforms before and after the noise reduction or SS process. The panels on the left of Fig. 3 show the waveforms before noise reduction, and panels on the right show waveforms in Fig. 3 show

waveforms after noise reduction by SS with fn=4 and

¡=2. In this case, AE propagates as a Lamb wave. The

detected waveform forVin=2.0 V (a) is an ideal Lamb wave

composed of a strong A0mode followed by a weak S0mode.

The waveform after SS (b) displays a reduced background

noise, and the S/N ratio of the waveform increased by

10.4 dB. The waveforms (b) and (c) were processed into

waveforms (e) and (f), respectively. The S/N ratios of

waveforms (e) and (f) were thereby improved by 13.2 and 10.2 dB, respectively.

3.2 Improvement in S/N ratio and waveform distortion by SS process

The relationship between the S/N ratios of the signal

before and after SS was studied and the characteristics of noise reduction by SS were compared to the characteristics of

noise reduction using the¾-ﬁlter (EF)6)_{and wavelet shrinkage}

(WS).5) _{The parameter of SS were} _fn₌_{8 and} _¡₌_{2. The}

parameters of EF were ¾0=10, ¾1=3.0 mV and ¾2=

1.5 mV. In WS, Daubechies wavelet (N=10) was adopted

as the mother wavelet. Figure 4 shows a comparison between the waveforms before noise reduction (a) and the waveforms

processed using spectral subtraction (b), ¾-ﬁlter (c), and

wavelet shrinkage (d), respectively. We can observe that the

background noise was reduced and the S/N ratio improved in

all the processes. Figure 5 shows a change in the S/N ratio

before and after processing by the aforementioned three

techniques. The improvement in S/N ratio of the waves

processed by both SS and EF is approximately constant.

Thus, the improvement in S/N ratio by these processes is

not dependent on the amplitude of the noise components.

0 - 0.005

0 0.005

N(k)

θx

(k)

0

- 0.005 0 0.005

x₁(t)

X₁(f) X₃(f) X₄(f)

X

₂

(f)

-

α

N(k)

X

₃

(f)

-

α

N(k)

X

₄

(f)

-

α

N(k)

X₂(f)

s(t)

0 - 0.005 0 0.005

x(t)

x₂(t) x₃(t) x₄(t)

1.Dividing

x(t)

2.FFT

x

n

(t)

X

n

(f)

3. Noise estimation

N(f)

=

X

1

(f)

4. Phase

calculation

5. Subraction

S

₂

(f)

=

X

₂

(f)

–

α

N(f)

6. IFFT

and combing

Fig. 1 Procedure of spectral subtraction.

Digitizer

Sensor

(Center Frequency , 450 f/kHz)

Excitation for AE

(Center Frequency, 450 f/kHz)

Type 304

Stainless steel plate

Pre-Amp (40 (dB))

450 f/kHz 1 Cycle sine wave Vin= 2 V/V – 70 V/mV

Function Generator (AE Generator)

[image:2.595.305.547.348.469.2]

(3)

However, the S/N ratio did not improve after WS for the

waveform with an S/N ratio lower than 5.1 dB. In the WS

method, the original signal was distinguished from noise by obtaining decomposed signals by a wavelet transform and removing the noise on the basis of a threshold. Therefore, when the amplitude of the amplitude of original signal is less than that of noise, the WS technique does not work.

Figure 6 shows the correlation between the waveforms processed by the three types of noise reduction techniques and the waveform shown in Fig. 3(a). Here, low correlation

signiﬁes that the process caused a distortion in the waveform.

Although all processed signals tend to be distorted in the

signal with lower S/N ratio, correlation coefﬁcient for the

wave processed by the SS technique was higher than the

correlation coefﬁcients for the waves processed by the

other two processes. Distortion by EF may be caused by

inappropriate parameters. In AE monitoring, it is difﬁcult

to choose the appropriate parameters because the amplitude

and/or frequency characteristics of the noise component vary

with time.15) _{Noise reduction by the SS technique did not}

depend on the S/N ratio, hense, SS was considered to be a

convenient technique for AE monitoring.

3.3 Effect of frame number

Figure 7 shows waveforms processed by SS withfnof 4,

16 and 64, and ¡of 2 for the waveform shown in Fig. 3(c)

(Vin=70 mV). In order to evaluate the effect of the frame

number on noise reduction, fn was changed from 4 to 64,

while ¡ was kept at 2. The S/N ratio increased with

decreasing frame number. For anfnof 64, 16, and 4, the S/N

ratio was 10.9, 9.8, and 4.9 dB, respectively. Figure 8 shows

the changes in the S/N ratio and correlation coefﬁcient for

the waveform processed by SS [shown in Fig. 3(a)] as a

function of fn. Improvement of S/N ratio decreased with

increasing fn. This is because the difference of data length

in one frame. As the frequency resolution of Fourier

transform depends on the data length of the waveform,13)

the frequency spectrum of noise can be determined with

high resolution by using a long data length or a small fn.

The correlation coefﬁcient is maintained at more than 0.92

for any frame number. There was no advantage of noise reduction performance in SS over EF and WS by using

the value of fnto 32 or more because improved S/N ratio

rapidly decreased. Thus, a smaller fn is suitable for AE

monitoring.

Time, t/μs

0 40 80

Output,

V/

mV

50

-50

V

_in

=2.0

V

/V

0

S/N : 27.0(dB)

(a)

0 40 80

50

-50 0

S/N : 37.4(dB)

(d)

0 40 80

5.0

-5.0

V

_in

=0.2

V

/V

0

0 40 80

5.0

-5.0 0

S/N : 5.1(dB) S/N :18.3(dB)

V

_in

=0.07

V

/V

0 40 80

2.0

-2.0 0

0 40 80

2.0

-2.0 0

S/N : 0.7(dB) S/N :10.9(dB)

(c)

(f)

(b)

(e)

Before SS

After SS

S

₀

A

₀

Time, t/μs

Time, t/μs Time, t/μs

Output,

V/

mV

Output,

V/

mV

Output,

V/

mV

Output,

V/

mV

Output,

V/

mV

[image:3.595.136.459.66.484.2]

(4)

Some researchers have reported that an artifact signal called musical noise or pre-echo, appeared just before the AE

signal when a small fn was used.14,15) _{This artifact signal}

obstructs the analysis of the AE waveform, especially AE

source characterization16) and/or source location.17) This

artifact signal tends to appear when the signal amplitude changes rapidly and it is prevented by using a short sampling

interval.4,14,15)

3.4 Effect of over-subtraction factor

Figure 9 shows waveforms with Vin=70 mV [shown in

Fig. 3(c)] after they are processed with the SS technique

using¡=1, 4, and 8. In this case,fnwasﬁxed at 8. The S/N

ratio of the waveform was improved to 10.1, 19.5 and

24.7 dB as ¡ increased from 1 to 4 to 8. However, the

maximum amplitude of the processed signals decreased. The SS technique does not reduce only the noise signal but also a

part of the original signal when ¡ is too high. Figure 10

shows a change in the improvement of the S/N ratio of the

waveform before and after processing by SS the correlation

coefﬁcient as functions of ¡. A trade-off was observed

between the S/N ratio improvement and waveform distortion

against the over-subtraction factor. Thus,¡is required to be

determined by a high noise reduction performance or low waveform distortion of AE signals. However, correlation

coefﬁcient is suitable more than 0.9 to get an advantage to EF

and WS in Fig. 6. To be higher correlation coefﬁcient more

than 0.9,¡should be 5 or less.

Time,

t

/

0

5.0

-5.0

0

40

80

0

40

80

0

40

80

(c)

ε

-filter

(d) Wavelet

shrinkage

Time,

t

/

μ

s

Time,

t

/

μ

s

Time,

t

/

μ

s

Output,

V

/mV

Output,

V

/mV

Output,

V

/mV

Fig. 4 Comparison of waveforms before noise reduction (a) and waveformsﬁltered by spectral subtraction (b),¾-ﬁlter (c) and wavelet shrinkage (d).

ε-filter(EF)

Wavelet Shrinkage(WS)

S/N ratio before process, (dB)

0 10 20 30

Correlation coef

ficient

1.0

0.25 0.5 0.75

SS EF WS

Spectral Subtraction

Fig. 6 Comparison of correlation coefficients for waveformsfiltered by spectral subtraction, e-filter, and wavelet shrinkage.

0 20 30

S/N ratio after process, (dB)

0 70

-10

10 30

Spectral Subtraction

ε-filter Wavelet Shrinkage

SS EF WS

S/N ratio before process, (dB)

[image:4.595.109.490.71.367.2] [image:4.595.329.521.418.571.2] [image:4.595.74.266.425.570.2]

S/N :4.9 (dB)

fn

=4

fn

=16

fn

=64

Time,

t

/

μ

s

Output,

V/

mV

Time,

t

/

μ

s

Output,

V/

mV

Time,

t

/

μ

s

Output,

V/

mV

Fig. 7 Waveformsﬁltered by spectral subtraction withfn=4, 16, 64 and over-subtraction factor¡=2.

frame number, fn

Improved S/N ratio by SS, (dB)

70 13.0

6.0 10.0

10 20 30 40 50 60 0

Correlation coef

ficient

1.0

0.9

0.8

S/N :24.7 (dB)

α

= 1

α

= 8

α

= 4

Time,

t

/

μ

s

Output,

V/

mV

Time,

t

/

μ

s

Output,

V/

mV

Time,

t

/

μ

s

Output,

V/

mV

Fig. 9 Waveforms ﬁltered by spectral subtraction with frame number

fn=8 and over-subtraction factor¡=1, 4, 8.

Over-subtraction factor, α

Improved S/N ratio by SS, (dB)

25

Correlation coef

ficient

0 2 4 6 8

0 10 20

[image:5.595.69.511.65.521.2]

0.8 0.9 1

[image:5.595.330.525.71.521.2] [image:5.595.72.267.586.713.2] [image:5.595.325.526.587.713.2]

(6)

4. Developed AE Monitoring System with Noise Reduction Function and Its Application in AE Monitoring

4.1 AE monitoring system with noise reduction function by spectral subtraction

Figure 11 shows the block diagram of the developed AE monitoring system with real-time noise reduction by SS. Signals detected by an AE sensor were temporarily stored

in the ﬁrst memory in the A/D board (ﬁrst buffer). When

ﬁrst buffer was completely ﬁlled with the data, the stored

data were transferred to the main memory in the PC, and signals were buffered in the second memory in the board simultaneously. As signals were alternately buffered in the two memories, the transferred signals were processed by SS simultaneously. When the maximum amplitude of the processed signal exceeded the predetermined threshold level, the processed data were saved in the hard disk. To process the entire transient data during AE monitoring without dead time, parallel computing was performed in this system. This developed system can acquire AE signals with a sampling frequency of up to 25 MHz (40 ns sampling intervals) with no dead-time. This sampling interval is

sufﬁciently small to prevent the generation of the signal by

SS process.

4.2 Monitoring of cylindical wave AE in noisy environ-ment

Figure 12 shows the experimental setup for monitoring

cylindrical wave AE signals using the developed system. An AE sensor was set on a steel pipe with 6000 mm in length, 114.3 mm in diameter, and 6 mm in thickness. An adhesive tape (JIS Z1901, vinyl chloride tape of 0.4 mm thickness)

was wrapped over 3000 mm long on the pipe. An artiﬁcial

noise was generated by a PZT transmitter placed on one edge of the pipe, and a white noise electronic signal with a peak-to-peak voltage of 5 V was applied to the transmitter for noise

generation. AE signals were excited by the HsuNielsen

source (pencil lead breaking) at a distance (L) of 13 m from

the AE sensor. The frame number fnwas set to 4 and

over-subtraction factor¡was set to 4.

Figure 13 shows unprocessed waveforms [(a)(c)] and

waveforms processed by the SS technique [(d)(f)]. The

attenuation of amplitude of the AE wave increased with increasing L due to the absorption of the wave energy by the

adhesive tape. In particular, the S/N ratio of the unprocessed

wave (c) for propagated distance of 3.0 m was approximately 0 dB, and it could not be distinguished from noise. The AE

signal could be identiﬁed in the waveform (f) after waveform

(c) was processed and its S/N ratio was improved to 12.8 dB.

The S/N ratio of other processed AE signals were also

improved by approximately 10 dB. Thus, we found that the developed system is suitable for AE monitoring in a noisy environment.

5. Conclusion

An AE monitoring system with real-time noise reduction by spectral subtraction (SS) process was developed and demonstrated for cylindrical wave AE monitoring. The results are summarized below:

(1) Three types of noise reduction techniques, wavelet

shrinkage (WS), ¾-ﬁlter (EF) and SS were compared.

For an environment with a high background noise,

the correlation coefﬁcient in the case of SS was higher

than that in other processes.

(2) The improvement in the S/N ratio and waveform

distortion by SS decreased with increasing frame

number. In contrast, improvement in S/N ratio

in-creased with increasing¡, and a trade-off relation was

observed between S/N ratio improvement and

wave-form distortion against the over-subtraction factor. SS had an advantage of noise reduction performance and

6000

2000 3000

6.0

1

14.3

Adhesive tape for corrosion protection (JIS Z 1901)

L

Pre-Amp(40dB) Developed AE monitoring system

(CPU:Intel CORE i7, A/D board:Alazartech ATS660)

Steel pipe

500

Hsu-Nielsen source (Pencil lead breaking) Function

Generator

Transmitter for noise (KGK, K-98)

AE sensor (PAC, PICO)

[image:6.595.55.284.439.580.2]

Unit, L/mm

Fig. 12 Experimental setup for monitoring cylindrical wave produced by Hsu-Nielsen source in noisy environment.

Digitize data

Receive data

Spectral Subtraction AE

sensor

Save Signal

A/D board Noise Reduction Program

Developed AE Monitoring System

Check data

Over the threshold

Under the threshold

1st Buffer

2nd Buffer Store data

Send data Clear data

Check all data

[image:6.595.107.480.624.765.2]

(7)

suppress of waveform distortion by using the value of

fnto 16 or less and¡to 5 or less.

(3) An AE monitoring system with noise reduction function was developed. The developed system can acquire AE signals with a sampling frequency of up to 25 MHz with no dead-time.

(4) Cylindrical wave AE signals produced by the Hsu–

Nielsen source were monitored in an environment

containing wide frequency band noise. The S/N ratio of

the AE signal was improved by approximately 10 dB using the SS technique.

REFERENCES

1) B. Muravin, G. Muravin, G. Kravertz and T. Duchin: Progr. Acoustic Emission13(2006) 183193.

2) T. Shiotani, X. Lio and H. Haya: J. Acoustic Emission24(2007) 205 214.

3) C. Scheer, W. Reimche and F. W. Bach: J. Acoustic Emission25(2008) 331340.

4) S. V. Vaseghi: Advanced Digital Signal Processing and Noise

Reduction, (Wiley, England, 1996).

5) D. L. Domoho:IEEE Trans. Inf. Theory41(1995) 613627.

6) H. Watase, K. Arakawa and Y. Arakara: IEICE Trans.E83-A(2000) 283290.

7) J. S. Lim and A. V. Oppenheim:IEEE Trans. ASSP26(1978) 197210.

8) S. F. Boll:IEEE Trans. ASSP27(1979) 113120.

9) R. M. Udrea, N. Vizireanu, S. Ciochina and S. Halunga:Signal Process. 88(2008) 12991303.

10) H. Saruwatari, S. Kajita, K. Takeda and F. Itakura: IEICE Trans. FundamentalsE83-A(2000) 866876.

11) M. Wevers, L. Rippert, J.-M. Papy and S. Van Huffel:NDT&E Int.39 (2006) 229235.

12) T. Matsuo and H. Cho: J. Japan Society for Non-Destructive Inspection

58(2009) 549555 (in Japanese).

13) M. Takemoto, H. Cho and H. Suzuki: J. Acoustic Emission24(2006) 1221.

14) M. Okazaki, T. Kunimoto and T. Kobayashi: Proc. the 2004 IEEE Int. Conf. on Acoustics (ICASSP, 2004) pp. 805808.

15) Z. Goh, K.-C. Tan and B. T. G. Tan: IEEE Trans. Speech Audio Process.6(1998) 287292.

L = 3.0

(a)

(b)

(c)

(d)

(e)

(f)

Time,

t

/ms

Output,

V/

mV

Time,

t

/ms

Output,

V/

mV

Time,

t

/ms

Output,

V/

mV

Time,

t

/ms

Output,

V/

mV

Time,

t

/ms

Output,

V/

mV

Time,

t

/ms

Output,

V/

[image:7.595.124.474.71.448.2]

mV

[doi:10.2320/

627.

210.

120.

1303.

235.

IEEE Trans. Speech AudioProcess.

110.