Wavelet analysis of acoustic emission signals from thermal barrier coatings

Available online at www.sciencedirect.com mClRNCC

Science Press

d

DIRECT.

Transactions of Nonferrous Metals Society of China

Trans. Nonferrous Met. SOC.China 16(2006) s270-s275 www.csu.edu.cru'ysxb/

Wavelet analysis of acoustic emission signals from thermal barrier coatings YANG Li(&

IF!),

ZHOU Yi-chun(M%%)

Key Laboratory for Advanced Materials and Rheological Properties of Ministry of Education, Faculty of Material Photoelectron Physics, Xiangtan University, Xiantan 41 1 105, China Received 10 April 2006; accepted 25 April 2006 Abstract: The wavelet transform is applied to the analysis of acoustic emission signals collected during tensile test of the Zr02-8% Y 2 0 3 (YSZ) thermal barrier coatings (TBCs). The acoustic emission signals are de-noised using the Daubechies discrete wavelets, and then decomposed into different wavelet levels using the programs developed by the authors. Each level is examined for its specific frequency range. The ratio of energy in different levels to the total energy gives information on the failure modes (coating micro-failures and substrate micro-failures) associated with TBCs system. Key words: wavelet transform; acoustic emission; tensile test; thermal barrier coatings; failure modes

1 Introduction Acoustic emission(AE) technique for non-destructive evaluation of material failure has been widely used to study the fracture behavior of thermal barrier coatings[ 1-41. The fundamental characteristics of acoustic signals, such as the attenuation, event count, ring-down count, peak amplitude distribution and energy count are investigated to describe the initiation and propagation of cracks in thermal barrier coatings. In the fracture process of the TBCs, the number of AE events is in a good agreement with the number of micro-failures[5], the amplitude or energy count are commonly used to estimate the deformation degree[ 1-41, and the waveform is occasionally introduced to analyze the failure mode of TBCs[l]. In fact, the amplitudes attenuate greatly with the increment of the propagation length and the distance of sensor, moreover, an AE signal may be the result of multi-failure, while the frequencies of AE signals from the certain failure modes are almost unchanged. This proves that the frequency analysis is an effective way in processing AE signals. Another point motivating the development of frequency analysis is the stochastic noise mixed in AE signals. The signals obtained during the recording of AE events and storing their waveforms are usually not recorded in a way to apply more sophisticated analysis tools. The so-called raw data have to be processed in a

way to extract detailed information about the source terms of fracture analysis. To do so, all influences not related to source information have to be eliminated. The data processing can include several steps. One can be the application of a band-pass filter to cut-off frequencies out of the frequency range of interest as well as suppressing alias effects. There is also a demand to eliminate artifacts caused by the sensor transfer functions (sensor frequency response). A more effective way to filter the void information is the use of frequency analysis by reducing the noise mixed up with signal parts related to the fracture process and enhancing the signal-to-noise ratio. In frequencies analysis to AE signals, one method is the fast-Fourier transform (FFT) to identify various failure modes, etc. [6-81. However, the noise information can not be eliminated because of its randomicity, and the time information of failure modes can not be obtained. A more sophisticated way of denoising AE signal and discriminating failure modes is using wavelets. Wavelet techniques have been suggested to enhance the signal-to-noise ratio[9]. Wavelets are mathematical functions that cut up data into different frequency components, and then reconstruct the useful components by thresholding policy[ 10-151. In this study, the tensile test and AE monitoring are performed on TBCs. The AE signals are de-noised using wavelet technique, and the de-noised

Foundation item: Project(10525211, 5053 1060) supported by the National Natural Science Foundation of China. Corresponding author: ZHOU Yi-chun; Tel: +86-732-8293568; Fax: +86-732-8292468; E-mail: zhouyc@xtu,edu.cn

YANG Li, et al/Trans. Nonferrous Met. SOC.China 16(2006)

AE signals are decomposed into different frequency levels by wavelet to discriminate the failure modes by the ratio of energy at different levels to total energy.

s211

resolution at jth order, respectively. They are can be represented as

cjs(f)= C C j , & h ( l l Z- 2k) k

2 Wavelet transform A wavelet ~ ( t is) an oscillating function of short duration, temporarily localized around the center t-1/ 2 , Its spectrum concentrates in a bilateral band O
I,?(CO~*

q,k(f)=2"2V(Yt-k) P o , k E z

Cj,kWj,k(t) j

s(O=Cdt) c~(t)=c,s(t)+D,s(t)

(1)

The wavelet transform can be considered a correlation between the signal under study s(t) and the wavelets. For a given a signal s(t), it is possible to represent it by the series: s(t)=

Thus, the wavelet decomposition of a given signal s ( t ) , is then expressed as:

(2)

s(t)= c , s ( t ) + z D j s ( t ) . (11) Let E f s ( t ) and Dps(t) be ttie'accumulative energy at low frequency of Jih order and that at high frequency ofjth order at any time t during the test, respectively. Namely

k

where Cj,k is the wavelet coefficients. Assuring that the set { W / , k ( f ) , k E Z } is the orthonormal basis of W,. Let 4(f) is the scale function the wavelet ~ ( t )and , the set { 4j,k(t),k E Z} (where 4j,j.k(f) are obtained by expanding and translating of 4(t)) is the orthonomal basis of 5. According to the scale and wavelet function, a low-pass filter and a high-pass filter construct, and their response function h(k) and g(k) are defined as

h(k)=< %(O, VJ-I,W)>

(3 1

g(k)=< 4 , ( d , 4 r , k ( d >

(4)

r=l

With the total energy given by J

Es(t)= E:s(f)

+ C( E f s ( f ) r j=l

To be more precise, the ratio of energy in different levels to the total energy is more important in discriminating how the energy is displayed in different levels. Let

A space J$, is decomposed a lower resolution space 5 plus a detail space W,, W, is the orthogonal complement of the spaces V, and VJ-], and V, c V,-' , thus

v,.,=v, fB W,

(5)

The orthogonal projection of a signal on is decomposed as the sum of projections on 5 and W,, which can be obtained by the convolution between the orthogonal projection on 5 and low-pass filter, high-pass filter, respectively. Thus, the coefficient cJ,kand d,,k can be expressed as 'J,k

='

J

fkl =

-2k)cJ-1

rml

(6)

rn 'J,&

=dl f k l =

g ( m-

2kk,-,Lml

(7)

m

h(m-2k) and g(m-2k) are the new response functions reconstructed by inserting j-1 zeros between each near two coefficients of h(k) and g(k), respectively. Defining C,s(t) and DJs(f)are the lower and higher

By tracing the values defined by the level energy distribution coefficient rEF and r E f , the energy variation of each fracture mode is expected to be evaluated. Although there are many types of wavelets, we restrict ourselves in this study to the Daubechies[161, which is good enough for the purpose of engineering applications.

3 Experimental 3.1 specimen preparation

The TBCs used for this research project were provided by Beijing Institute of Aeronautical Materials.

s272

YANG Li, et aVTrans. Nonferrous Met. SOC.China 16(2006)

First, Ni based superalloy DZ125 which consisted of Engation/% multiple elements Cr~.9C0.1C~10W7M02A15.2Ti~.9Ta0.8Hfl.sBo.ol~Ni was used as substrate, its geometry is 5 mm in width, 3 mm in thickness and 60 mm in length with a gage length of 40 mm. Next, a NiCrAlY bond coat about 50 pm thick was deposited on one surface of the beam substrate by vacuum plasma-spray technology using Sulzer-Metco systems (where the powder feed rate is 45 g/min). Last, electron beam physical vapor deposition technique applied to manufacture Zr02 stabilized with 7%-8%Y203 ceramic coat in a total thickness of approximate of 10 pm on the top of bond coat, during this deposition, the power of electric gun is 15 kW, and Time/us the surface temperature of the gun is 1 173 K. Fig.1 Relationship between AE events and loading time

c

C

I

3 6

Y

b)

w d

3.2 Tensile test monitored by acoustic emission The tensile test was carried out to understand the microscopic failure behavior of thermal barrier coatings using AE technique. The experiment was carried out under load control at a rate of. Two AE sensors with resonance frequency of 600 kHz were used. The AE sensors were attached firmly to the surface of substrate by an elastic cord and sealed with butter a thin film of butter to ensure maximum ultrasonic signal transmission. AE signals traveled through TBCs, were received by sensors, were filtered and magnified by DB16 AE analyzer, and then recorded automatically by a computer.

4 Result and discussion 4.1 Micro-failure mechanism and AE characteristic Fig.1 shows typical stress-strain curves and the relationship between the load increment and the AE events. The deformation behavior and corresponding AE response of TBCs can be described in three different steps 1) All layers of 'TBCs system are hlly elastic, and no AE signals are observed because no irreversible deformation behavior occurs. 2) The YSZ layer deform plastically, but the substrate and the bond coat deform elastically, AE events increase gradually because the YSZ ceramic is so brittle that it begins to fracture in the plastic region. 3) All layers of TBCs deform plastically till ultimate fracture occurs, AE event count increases rapidly compared with region 2) due to the occurrence of micro-failures in each layers of TBCs, for example, the fractureb of YSZ coat, the delemination between YSZ coat and bond coat, as well as the plastic deformation and fracture of the substrate and those of the bond coat. Fig.2 shows the cross-section of the TBCs specimen at the end of stage I and stage 11, respectively. No

Fig2 Cross-section of TBCs specimen at end of stage I (a) and stage I1 (b)

cracks has been found in YSZ layer in Fig.2(a), but serious fractures in YSZ layer have been found and some of these has penetrated into the bond coat in Fig.2(b). The small magnification is applied for Fig.2(b) in order to exhibit the panorama of TBCs. In addition, the great mass of YSZ layer has spalled after the ultimate fracture of TBCs, thus the cross-section is absent. 4.2 Denoising of AE signals AE data are usually disturbed by noises of different frequency ranges. On the one hand, AE data contain a lot of high-frequency noise, mainly caused by the measurement equipment (preamplifier etc.) and the surrounding. On the other hand, due to the testing process itself, a low-frequent signal caused by the testing device (loading machine) often superimposes AE signals additionally. When a signal is transformed into the wavelet domain, the noise in the signal often corresponds

YANG Li, et aVTrans. Nonferrous Met. SOC.China 16(2006)

to the coefficient whose absolute value is relatively small. This noise can be excluded easily because of its small absolute coefficient compared with the threshold. Once the threshold is obtained, one can apply soft thresholding policy: for each wavelet coefficient c ~, ,the~ soft thresholded value is calculated as c,,,=sgn(c,)(lcj,kl-th)

if

3th

(16)

q k = O if I cj,k 1
(17)

It can be observed that, when a testing device disturbance occurs, the absolute values of wavelet coefficients at the final scale level would become very large and exceed the threshold more, Extracting this coefficient, disturbance due to the testing device is erased. The computations were performed with Matlab for Windows version 7.0[ 171, using our own programs. Each of the wavelet function has different characteristic. The wavelet filter which is optimal for a given signal, is not necessarily the best for another type of signal. Improper choice of filter can cause distortions and artefacts in the reconstructed signal. By comparing the original and reconstructed signals, such as, signal-to-noise ratio, neatness without white noise, the wavelet function DB7 with level 7 was used in AE signal denoising. A filtering example of AE signal for TBCs is shown in Fig.3. Specifically, soft thresholding with universal threshold and the last level of decomposition was restored for all of the signals. 250 I

1

I

loob

460

860 1200 Time/ PS

I600

Fig3 Original AE signal and denoised AE signal using wavelet technique

4.3 Signal classification using energy analysis Base on the theoretical ground of the work discussed in section 2, the denoised AE signals are decomposed into six different levels. The decomposed AE signals in each level and corresponding frequency spectrum is shown in Fig.4. Each level represents a specific

s273

frequency range from 0-156.25 to 2 500-5 000 kHz. Visually, none of these is similar to white noise, and none of these exhibits abrupt change. The features of these waveforms, such as, duration time, amplitude and frequency spectrum are useful to illustrate the characteristics of AE signal and distinguish the different AE signals from various possible failure modes in TBCs. Given the AE signal that decomposed in Fig.4, it is seen that the duration of each level shows no obvious difference; the amplitudes distributing is that it is higher in D2, D3 and D4 than in the rest levels; corresponding energy intensities are high in these three levels with the maximum in D3. The results show that the signal can be distinctly identified with its character concentrate in D2 to D4. Information shown in Fig. 4 can give immediate access to AE signal and are always convincing. But due to the large number of AE events, analyzing by this way is very time consuming. Note that the ratio of energy defined as Eqn.( 15), which includes not only the time information such as duration and amplitude but also the frequency information, is a favorable index to discriminate the AE signals. The automatic application of wavelet algorithms for distinguishing is simple as long as the AE signals can be classified in a certain way, i.e. typical models for AE signals exist. In this work, one hundred AE signals from TBCs during tensile test are chosen and analyzed using wavelet transform. Three typical models are found and the distributions of their energy ratios at different levels are shown in Fig.5. From the figure, it can be seen that the energy is concentrated on scale D4, D3 and D2 (with a frequency range from 312.5-625 kHz, 625-1 250 kHz and1 250-2 500 kHz) as the same result shown in Fig.4. The histograms in Fig.5 also reveal that the biggest value of energy ratio for each type exceeds 0.4, such as, 0.58 at scale 2 for model 1, 0.52 at scale 3 for model 2 and 0.68 at scale 4 for model 3. Thereafter, a given signal can be discriminated by comparing the biggest value of energy ratio with 0.4 and confirming which level of the biggest value occurs in. Specially, for some signals can’t be regarded as any one of the existed modes, such as, the values of energy ratio at all levels are small than 0.4, the biggest one exists not in D2-D4 but in D 1 or C5, discriminating should be practiced again to find another possible model. In case of our work, no else model can be established because of small quality of these cases. The accumulative AE event counts of each model at different stage of tensile test are illustrated in Table 1. Among, stage I is absent in that no AE signal recorded, stage I1 is divided into two period at its

YANG Li, et al/Trans. Nonferrous Met. SOC.China 16(2006)

s274

Q

0

'-

AE wavdonn in different wavelet levels

3

-0O' -208 0

1000

500

1500

2000

@

Corresponding frequency spectra IOXlW 5 0

X 0

l 0.1

W

0.2

W

0.3

0.4

d 0.5

C

W

. 9 ; $ 2 x 10"

cnn

E

I X 106

0

0

500

1000

I500

2000

0

0.1

0.2

0.3

0.4

0.5

r 400 4 x 101, 2x 10"

'i; -408

0

500

IOOO

1500

2000

lox 107 5 x 107 0

0

500

IOOO

1500

I

0

2000

0.1

0.2

0.3

0.4

0.5

0.1

0.2

0.3

0.4

0.5

I

0

:::1 18; $ 0

0

500

0

500

1000

1500

2000

0

0.1

0.2

0.3

0.4

0.5

1000

1500

2000

0

0.1

0.2

0.3

0.4

0.5

Time/ us

b requency/MHZ

Fig. 4 AE waveforms and corresponding frequency spectra in different wavelet levels

Table 1 Percentage of each model for AE signals of TBCs at different deformation stages type Event count Model 1 Model 2 Model 3 Other

Stage I 43 19(44.2%) 16(37.2%) 1(2.3%) 7(16.3%)

Stage I1 143 56(39.1%) 67(46.9%) 9(6.3%) 11(7.7%)

approximate midpoint, named as stage I1 (F) and stage I1 (L), respectively. According to Table 1, AE signals of model 1 and model 2 are active during the whole tensile test, with a total percentage of 73%. It can also be seen that the percentage of model 2 and model 3 keeps rising, but that of model 1 keeps declining with the deformation of TBCs increasing. Table 1 also displays that a decrease trend in AE count of the others with the increase of the TBCs deformation. Considering the AE characteristics and the micro-failing process discussed in section 4.1, AE signals are mainly caused by the plastic deformation and fracture occurred in YSZ layer in stage I1 , but those in stage I11 are induced by the micro-fractures of both YSZ coat and substrate. As we know, the frequencies of AE signals resulted from the p1asti.c deformation are commonly higher than those caused by the fracture of a certain material[l8]. These results show that AE signals induced by micro-failures of

Stage I11 192 24( 12.5%) 97(50.5%) 62(32.3%) 9(4.7%)

Total 378 99(26.3%) 180(47.6%) 72( 19.0%) 27(7.1%)

YSZ layer are mainly of model 1 or model 2, associated with the plastic deformation and micro-fractures respectively; most AE signals caused by the micro-failures in substrate belong to model 2 or model 3. The rapid increase of AE signals in model 3 indicates the fracture of substrate give the rise to AE signals with low frequencies consisting. In addition, it can be deduced that the model of AE signals associated with the micro-failures in the interface between the YSZ layer and the bond coat, like delemination, is one of the three models because no other model has be found. The characteristics of AE signals induced by the same micro-failure type from the same source may be different from each other due to different micro-mechanism. In terms of the cracking behavior of YSZ layer, the difference may be caused by the different crack direction, crack initiation or propagation. The firther discrimination should be discussed in future work

YANG Li. et al/Trans. Nonferrous Met. SOC.China 16(2006)

s275

References 0.6 MA X Q, CHO S, TAKEMOTO M. Acoustic emission source analysis of plasma sprayed thermal barrier coatings during four-point bend tests[J]. Surface and Coating Technology, 2001, 139: 56-62. KUCUK K A, BERNDT C C, SENTURK U, LIMA R S. Influence of plasma spray parameters on mechanical properties of yttria stabilized zirconia coatings( I1 ): Acoustic emission response[J]. Materials Science and Engineering A, 2000,284A: 41-50. MA X Q, TAKEMOTO M. Quantitative acoustic emission analysis of plasma sprayed thermal barrier coatings subjected to thermal shock tests[J]. Materials Science and Engineering A, 2001, 38A:

0.5 -

2

6 0.4 -

P)

(r

.-*

0

0.3 -

62

0.2 0.1 -

0

I

DI

D2

D3

D4

D5

C5

0.1 0.6 x 0.5

2 P)

6 0.4 (r

0

.o 0.3

2 0.2 0.1 0

Fig.5 Three typical models of AE signals

5 Conclusions 1)AE event count changes with the deformation of TBCs. Hence, AE technique is an effective way to monitor the micro-failing of TBCs. 2) The wavelet transform can be a usefbl way to analyze AE signals by decomposing the signal into different wavelet levels. The ratio of energy at different wavelet component is a reliable and convenient index to different modes of AE signals. 3) AE signals of TBCs submitted to tensile test can be classified into three types associated with micro-failures in YSZ layer and substrate.

101-110.

FU L, KHOR K A, NG H W, TEO T N. Nan-destructive evaluation of plasma sprayed functionally graded thermal barrier coatings[J]. Surface and Coatings Technology, 2000, 130: 233-239. ZHOU Y C, HASHIDA T. Thermal fatigue failure induced by delamination in thermal barrier coating[J]. International Journal of Fatigue, 2002, 24: 407-41 7. CASEY N F, WHITE H, TAYLOR J L Frequency analysis of the signals generated by the failure of constituent wires of wire rope[J]. NDT International, 1985, 18(6): 339-344. RAMIREZ-JIMENEZ C R, PAPADAKIS N, REYNOLDS N, CAN T H, PURNMELL P PHARAOH M. Identification of failure modes in glass/polypropylene composites by means of the primary frequency content of the acoustic emission event[J]. Composites Science and Technology, 2004,64( 12): 18 19- 1827. WOODWARD B. Identification of acoustic emission source mechanisms by energy spectrum spectrum analysis[J]. Ultrasonics, I976,14(6): 249-255. GROSSE C U, FMCK F, KURZ J H, REMHARDT H W. Improvements of AE technique using wavelet algorithms, coherence functions and automatic data analysis[J]. Construction and Building Materials, 2004, 18(3): 203-213. PASTI L, WALCZAK B, MMASSART D L, RESCHOGLIAN P. Optimization of signal denoising in discrete wavelet transform[J]. Chemometrics and Intelligent Laboratory Systems, 1999,48: 2 1-34. QI G, BARHORST A, HASHEMI J, KAMALA G Discrete wavelet decomposition of acoustic emission signals from carbon-fiber-reinforced composites[J]. Composites Science and Technology, 1997, 57: 389-403. VELAYUDHAM A, KRISHNAMURTHY R, SOUNDARAPANDIAN T. Acoustic emission based drill condition monitoring during drilling of glass/pyenolic polymeric composite using wavelet packet transform[J]. Materials Science and Engineering A, 2005,412: 141-145. ADELMO R, FERREIRA D S. Wavelet denoising with evolutionary algorithms[J]. Digital Signal Processing, 2005, 15(4): 382-399. YAN R Q, GAO R X. An efficient approach to machine health diagnosis based on harmonic wavelet packet transform[J]. Robotics and Computer-Integrated Manufacturing, 2005,2 l(4-5): 29 1-30 I. WU J D, CHEN J C. Continuous wavelet transform technique for fault signal diagnosis of internal combustion engines[J]. NDT&E International, 2006, 39: 304-3 1 1. DAUBECHUES I. Orthonormal bases of compactly supported wavelets, communication on pure and applied mathematics[J]. XLI,1998,906-916. HU C H, ZHANG B J, XIA J, ZHANG W. System Analysis and Design[M]. Xidian University Press, 1999. YUAN Z M, MA K Y, HE 2 Y. Acoustic emission technique and its application[M]. Beijing: Machinery Industry Press, 1985. (Edited by CHEN Ai-bua)