Jianguo Wu Acoustic Emission Monitoring for Ultrasonic ... · can resume immediately if necessary....

12
Jianguo Wu Department of Industrial and Systems Engineering, University of Wisconsin-Madison, 3255 Mechanical Engineering, 1513 University Avenue, Madison, WI 53706 e-mail: [email protected] Shiyu Zhou 1 Department of Industrial and Systems Engineering, University of Wisconsin-Madison, 3270 Mechanical Engineering, 1513 University Avenue, Madison, WI 53706 e-mail: [email protected] Xiaochun Li Department of Mechanical Engineering, University of Wisconsin-Madison, 1035 Mechanical Engineering, 1513 University Avenue, Madison, WI 53706 e-mail: [email protected] Acoustic Emission Monitoring for Ultrasonic Cavitation Based Dispersion Process In the manufacturing of micro/nanocomposite materials, micro/nanoparticles need to be dispersed evenly into the base materials. However, due to their high surface-to-volume ratio and high surface energy, the micro/nanoparticles tend to agglomerate and cluster together. Ultrasonic cavitation is effective to disperse micro/nanoparticles. However, works on correlating the cavitation parameters with the micro/nanoparticle dispersion are limited. This paper presents a real-time acoustic monitoring method based on cavita- tion noises to monitor the micro/nanoparticle dispersion status. In this paper, two types of cavitation noise power indices computed based on the raw cavitation noise signals are used to monitor the cavitation status. Both off-line and on-line steady state detection algorithms are developed. These algorithms can be used to determine the critical process parameters including the power of the ultrasonic sound and the dispersion time. Exten- sive experiments have been conducted to illustrate the effectiveness of the developed methods. [DOI: 10.1115/1.4024041] 1 Introduction Recently, micro/nanoparticles have attracted significant scien- tific interests, due to a wide variety of potential applications in biomedical, optical, electronic, mechanical fields, etc. They have been widely used as the additives in the fabrication of high per- formance composites [1,2]. Well dispersed micro/nanoparticles can significantly improve the mechanical properties, including toughness, stiffness, ductility, machinability of the composites [36]. Due to their high surface energy, large surface-to-volume ratio, poor wettability in the liquid, however, micro/nanoparticles tend to agglomerate and cluster together [5,7], which greatly lim- its their effectiveness. It is essential that the particles are dispersed evenly into the base materials before use. Ultrasonic cavitation is an effective method to disperse micro/ nanoparticles [5,811]. The basic idea is to shoot a beam of ultrasonic sound through the particle-liquid system. Then due to local violent pressure variations caused by ultrasonic vibrations [12], we will get a “cavitation” phenomenon, which refers to the formation, growth, oscillation, and implosive collapse of gas or vapor bubbles in liquids caused by the ultrasound. Based on the du- ration of bubbles, the cavitation is classified into two types: stable cavitation and transient cavitation [13]. For the stable cavitation, the bubbles oscillate nonlinearly around the equilibrium size. They are relatively stable and last for many cycles of the acoustic pres- sure. While for the transient cavitation, the bubbles usually oscil- late for much shorter time. They explosively grow into a cavity with a size of many times of their original sizes and then collapse violently. When the bubble collapses, it produces transient micro “hot spot” that can have temperatures of about 5000K, pressures above 1000 atms, and heating and cooling rates above 10 10 K/s, high speed liquid jets of up to 300m/s [12]. Due to these intense effects, the cavitation can effectively mix and also break particle agglomerates into well-dispersed particles in the liquid. There are several methods to detect and monitor cavitation process, including high-speed photography [14,15], laser diffrac- tion technique [16], phase-Doppler technique [16,17], acoustic attenuation method [18,19], and cavitation noise spectrum analysis technique [11,2025], etc. The cavitation noise spectrum analysis is the most popular method due to its low cost, easiness to imple- ment and its ability to capture various information of cavitation using acoustic transducers. The fundamental mechanism of acous- tic cavitation has been experimental and theoretically studied in the last several decades to interpret the cavitation noise spectrum. It is known that the cavitation noise spectrum consists of continu- ous components and various discrete frequency components [2629] close to nf =m ð Þ where f is the fundamental or driving fre- quency, and m, n are integers. These discrete components are: har- monics ( n=m ð Þ is integer), subharmonics n ¼ 1; m ¼ 2; 3; ð Þ and ultraharmonics (m > n; n=m ð Þ is noninteger). The continuous com- ponents are the broadband components (also called “white noise” [13]) that lie between the discrete components. The harmonics of the fundamental frequency are easily explained by the nonlinear characteristic of forced pulsations of bubbles [30]. However, for the other components, the origin is still under debate; many theo- ries have been proposed [25,31]. For the origin of “white noise”, there also exist different explanations. One explanation is that it originates from the shock waves produced by the collapse of bub- bles [21,32]. Using numerical simulation, Yasui et al. [33] explained that the temporal fluctuation in the number of bubbles results in the broadband noise. In other words, the transient cavita- tion results in the broadband noise. Stable cavitation does not cause the broadband noise even if it emits shock waves. All these explan- ations lead to that the broadband noise can be used as an indicator of the intensity of acoustic cavitation. Although the mechanism of cavitation has been intensively studied, the works on real-time monitoring of the ultrasonic cavi- tation based material processing is very limited. The ultrasonic power and processing time are usually chosen somewhat arbitra- rily in practice. An unnecessarily high ultrasonic power level or long processing time may result in waste of time and energy, while too low a power level or too short a processing time may lead to insufficient treatment. Some research works have been 1 Corresponding author. Contributed by the Manufacturing Engineering Division of ASME for publication in the JOURNAL OF MANUFACTURING SCIENCE AND ENGINEERING. Manuscript received July 18, 2012; final manuscript received February 12, 2013; published online May 24, 2013. Assoc. Editor: Robert Gao. Journal of Manufacturing Science and Engineering JUNE 2013, Vol. 135 / 031015-1 Copyright V C 2013 by ASME Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Transcript of Jianguo Wu Acoustic Emission Monitoring for Ultrasonic ... · can resume immediately if necessary....

Jianguo WuDepartment of Industrial and

Systems Engineering,

University of Wisconsin-Madison,

3255 Mechanical Engineering,

1513 University Avenue,

Madison, WI 53706

e-mail: [email protected]

Shiyu Zhou1

Department of Industrial and

Systems Engineering,

University of Wisconsin-Madison,

3270 Mechanical Engineering,

1513 University Avenue,

Madison, WI 53706

e-mail: [email protected]

Xiaochun LiDepartment of Mechanical Engineering,

University of Wisconsin-Madison,

1035 Mechanical Engineering,

1513 University Avenue,

Madison, WI 53706

e-mail: [email protected]

Acoustic Emission Monitoringfor Ultrasonic CavitationBased Dispersion ProcessIn the manufacturing of micro/nanocomposite materials, micro/nanoparticles need to bedispersed evenly into the base materials. However, due to their high surface-to-volumeratio and high surface energy, the micro/nanoparticles tend to agglomerate and clustertogether. Ultrasonic cavitation is effective to disperse micro/nanoparticles. However,works on correlating the cavitation parameters with the micro/nanoparticle dispersionare limited. This paper presents a real-time acoustic monitoring method based on cavita-tion noises to monitor the micro/nanoparticle dispersion status. In this paper, two typesof cavitation noise power indices computed based on the raw cavitation noise signals areused to monitor the cavitation status. Both off-line and on-line steady state detectionalgorithms are developed. These algorithms can be used to determine the critical processparameters including the power of the ultrasonic sound and the dispersion time. Exten-sive experiments have been conducted to illustrate the effectiveness of the developedmethods. [DOI: 10.1115/1.4024041]

1 Introduction

Recently, micro/nanoparticles have attracted significant scien-tific interests, due to a wide variety of potential applications inbiomedical, optical, electronic, mechanical fields, etc. They havebeen widely used as the additives in the fabrication of high per-formance composites [1,2]. Well dispersed micro/nanoparticlescan significantly improve the mechanical properties, includingtoughness, stiffness, ductility, machinability of the composites[3–6]. Due to their high surface energy, large surface-to-volumeratio, poor wettability in the liquid, however, micro/nanoparticlestend to agglomerate and cluster together [5,7], which greatly lim-its their effectiveness. It is essential that the particles are dispersedevenly into the base materials before use.

Ultrasonic cavitation is an effective method to disperse micro/nanoparticles [5,8–11]. The basic idea is to shoot a beam ofultrasonic sound through the particle-liquid system. Then due tolocal violent pressure variations caused by ultrasonic vibrations[12], we will get a “cavitation” phenomenon, which refers to theformation, growth, oscillation, and implosive collapse of gas orvapor bubbles in liquids caused by the ultrasound. Based on the du-ration of bubbles, the cavitation is classified into two types: stablecavitation and transient cavitation [13]. For the stable cavitation,the bubbles oscillate nonlinearly around the equilibrium size. Theyare relatively stable and last for many cycles of the acoustic pres-sure. While for the transient cavitation, the bubbles usually oscil-late for much shorter time. They explosively grow into a cavitywith a size of many times of their original sizes and then collapseviolently. When the bubble collapses, it produces transient micro“hot spot” that can have temperatures of about 5000 K, pressuresabove 1000 atms, and heating and cooling rates above 1010 K/s,high speed liquid jets of up to 300 m/s [12]. Due to these intenseeffects, the cavitation can effectively mix and also break particleagglomerates into well-dispersed particles in the liquid.

There are several methods to detect and monitor cavitationprocess, including high-speed photography [14,15], laser diffrac-tion technique [16], phase-Doppler technique [16,17], acousticattenuation method [18,19], and cavitation noise spectrum analysistechnique [11,20–25], etc. The cavitation noise spectrum analysisis the most popular method due to its low cost, easiness to imple-ment and its ability to capture various information of cavitationusing acoustic transducers. The fundamental mechanism of acous-tic cavitation has been experimental and theoretically studied inthe last several decades to interpret the cavitation noise spectrum.It is known that the cavitation noise spectrum consists of continu-ous components and various discrete frequency components[26–29] close to nf=mð Þ where f is the fundamental or driving fre-quency, and m, n are integers. These discrete components are: har-monics ( n=mð Þ is integer), subharmonics n ¼ 1;m ¼ 2; 3;…ð Þ andultraharmonics (m > n; n=mð Þ is noninteger). The continuous com-ponents are the broadband components (also called “white noise”[13]) that lie between the discrete components. The harmonics ofthe fundamental frequency are easily explained by the nonlinearcharacteristic of forced pulsations of bubbles [30]. However, forthe other components, the origin is still under debate; many theo-ries have been proposed [25,31]. For the origin of “white noise”,there also exist different explanations. One explanation is that itoriginates from the shock waves produced by the collapse of bub-bles [21,32]. Using numerical simulation, Yasui et al. [33]explained that the temporal fluctuation in the number of bubblesresults in the broadband noise. In other words, the transient cavita-tion results in the broadband noise. Stable cavitation does not causethe broadband noise even if it emits shock waves. All these explan-ations lead to that the broadband noise can be used as an indicatorof the intensity of acoustic cavitation.

Although the mechanism of cavitation has been intensivelystudied, the works on real-time monitoring of the ultrasonic cavi-tation based material processing is very limited. The ultrasonicpower and processing time are usually chosen somewhat arbitra-rily in practice. An unnecessarily high ultrasonic power level orlong processing time may result in waste of time and energy,while too low a power level or too short a processing time maylead to insufficient treatment. Some research works have been

1Corresponding author.Contributed by the Manufacturing Engineering Division of ASME for publication

in the JOURNAL OF MANUFACTURING SCIENCE AND ENGINEERING. Manuscript receivedJuly 18, 2012; final manuscript received February 12, 2013; published online May24, 2013. Assoc. Editor: Robert Gao.

Journal of Manufacturing Science and Engineering JUNE 2013, Vol. 135 / 031015-1Copyright VC 2013 by ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

conducted to study the relationships between the ultrasonic cavita-tion parameters and processing efficiencies [10,11,23]. Althoughthese studies provided insights on how to select optimal process-ing parameters, these studies are essentially off-line studies onspecific system configuration. Thus, the results may not be appli-cable to general situations since the processing efficiency dependson many factors, such as volume, particle concentration, viscosityand temperature. Therefore an effective on-line technique to mon-itor the ultrasonic cavitation based dispersion process is of criticalimportance in engineering practices.

In this work, we developed a real-time monitoring technique tomonitor micro/nanoparticle dispersion in aqueous liquid. Thistechnique is tested in tap water with an addition of Al2O3 par-ticles. The paper is organized as follows. In Sec. 2, the experimen-tal procedure is introduced. Section 3 presents descriptive analysisof the acoustic signals collected. Several off-line and on-linesteady state detection methods are presented and compared inSec. 4. The conclusions are presented in Sec. 5.

2 Experimental Procedure

The experimental setup mainly consisted of six components:Misonic Sonicator 4000, an ultrasonic horn/probe, a glass beaker,a titanium rod, an acoustic sensor and a Tektronix DPO7354Oscilloscope, as shown in Fig. 1.

The Misonic Sonicator 4000 has an operating frequency of20 KHz and the output amplitude can be controlled by setting arange from 1 to 100% of the maximum vibration amplitude55 lm. The tip of the ultrasonic probe, made of niobium alloyC103, is 12.7 mm in diameter. It is positioned in the center ofthe beaker and the distance between the probe tip and the surfaceof the water is about 2.0 cm. The vibration and shock waves pro-duced by the ultrasonic cavitation are collected by the titaniumrod with a length of 61.72 cm and a diameter of 1.59 cm. The tita-nium rod is immersed in the water with length of 3.0 cm and witha distance of 3.0 cm to the probe tip. A MISTRAS R15S acousticsensor was coupled to the top of the titanium rod by an ultrasoniccouplant. The piezoelectric signal of the acoustic sensor wasacquired by the Tektronix DPO7345 Oscilloscope.

The experiments were carried out in tap water of 500 mL con-tained in a standard 500 mL glass beaker. The Al2O3 particleswith a diameter of 1 lm were added to the tap water along thewall of the glass beaker before the power switch of the ultrasonicsonicator was turned on. The trigger mode was used in the oscillo-scope and the cavitation noise signal was immediately acquiredafter the ultrasonic sonicator was turned on. The memory of theoscilloscope is capable of storing 5� 108 samples. With a sam-pling rate of 1� 106 samples/second, each cycle of signal acquisi-tion lasted about 500 s. The signal can be stored to hard drivewithin about ten seconds and the next cycle of signal acquisition

can resume immediately if necessary. The ultrasonic intensity wascontrolled by setting the vibration amplitude of the probe tip inthe range of 1–100% of the maximum amplitude.

3 Descriptive Analysis of the Cavitation Noise Signal

3.1 Cavitation Noise Signal. Figure 2 shows two representa-tive cavitation noise waveforms with duration of 500 s under ultra-sonic power 40 W from pure tap water and Al2O3-particle-filledtap water, respectively. There are 12 s of pretrigger samples ineach signal. Both waveforms show three stages: (I) immediatelyafter the ultrasonic power is turned on, there appears a high peakin the waveform; (II) after the peak, the cavitation noise signalreaches the weakest and then gradually increases; (III) finally thesignal enters into steady state. The obvious difference betweenthese two waveforms is that in stage II, for tap water with Al2O3

particles, the initial cavitation noise is lower than that withoutparticles, and it increases more significantly than that withoutparticles. This phenomenon is somewhat similar to Wojs’s results[11] that for pure water, there was no significant change on thespectrum characteristics at time 0, 15, 30, 60 mins while forPAA 0.1% solution, the spectrum was moved slightly upwardsafter 60 mins.

Stage I reflects the step response of the beaker, water, sonicatorsystem excited by the change of the power status, i.e., from off toon. When the step response diminishes, the cavitation noise falls.In stage II, an increasing number of air bubbles are formed by the

Fig. 1 Experimental setup (a) and its schematic representation (b): (1) Misonix Sonicator 4000; (2) ultrasonic horn/probe; (3)standard 500 mL glass beaker; (4) titanium rod; (5) acoustic sensor; (6) Tektronix DPO7345 oscilloscope

Fig. 2 Two representative cavitation noise waveforms withultrasonic power 40 W for pure tap water and tap water with 20 gAl2O3 particles

031015-2 / Vol. 135, JUNE 2013 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

rectified diffusion process [34] and thus the intensity of the cavita-tion noise increases gradually. In this process, the dispersion ofinitial impurities and the formation of a huge amount of small airbubbles cause more cavitation nuclei (note that in the pure tapwater there are also many impurities). As for stage III, which ischaracterized as the steady state, the liquid becomes uniform andthe cavitation becomes the most intensive.

The influence of Al2O3 particles or the possible reasons thatresult in the difference between tap water with and without Al2O3

in stage II are: first, the unwettable Al2O3 particles and extra airbubbles brought by these particles in suspension absorb part of theultrasonic energy in the process of formation, growth of cavitationbubbles, and the vibration and breakage of Al2O3 agglomerates.Second, the addition of Al2O3 particles increases the ultrasonicattenuation coefficient due to the scattering and absorption effects.Allegra and Hawley [35] studied the attenuation of sound forsolid-in-liquid suspensions and the scattering coefficient wasobtained by:

as ¼1

2�k4

c R3 1

3

bc � b0

bc

� �2

þ q0 � q2q0 þ q

� �2" #

(1)

where � is the volume fraction of the suspended particles, kc is thecompressional wave number for the suspending medium, R is theradius of the suspended particles, bc is the compressibility of thesuspending medium, b0 is the thermal dilation of the suspendedparticles, q and q0 are the densities of suspending medium andsuspended particles, respectively. Equation (1) shows that thescattering coefficient is proportional to the cubic of the particle ra-dius. In the cavitation and dispersion process in stage II, the sizesof Al2O3 clusters gradually reduce, which gradually decreases thescattering coefficient. The reduction of attenuation coefficient, theincrease of cavitation nuclei caused by the breakage of Al2O3

particles, and the fast development of cavitation intensify thecavitation noise in stage II until it enters into stage III where theparticles are well dispersed and uniformly suspended. Therefore

the cavitation noise signal can be effectively used to monitor thestatus of the cavitation and dispersion.

To confirm the above analysis and statements, we conducted anexperiment. The basic idea of this experiment is to disperse theparticles with different dispersion time and then we let the mixturesit for a fixed amount of time. Then, we can compare the severityof the segregation occurred after the sitting period. A better dis-persed mixture should have less segregation. Specifically, in theexperiment, six beakers were used with each beaker containing20 g Al2O3 particles and 500 mL tap water. The first beaker wasused as the control group where there was no ultrasonic treatment.For the other five beakers, the ultrasonic processing times were34.2 s, 80 s, 180 s, 300 s and 450 s, respectively. The ultrasonicdriving power was 40 W in the experiment. Please note that fromFig. 2, we can see that after roughly 300 s of the dispersion time,the acoustic noise is in the steady state. Figure 3 shows the Al2O3

suspension immediately after the ultrasonic treatment whereAl2O3 particles are evenly distributed in the water.

Figure 4 shows the segregation between the clusters of Al2O3

particles (the white layer at the bottom) and the water (the clearlayer on the top) 26 hs after the ultrasonic treatment with differentamount of processing times. It is clear that volume of Al2O3 layerincreases significantly at first (0–34.2 s), and then expands slowly(34.2–300 s) and finally become stable (300–450 s) as the process-ing time increased. The reason for this phenomenon is that whenthe particles are dispersed, the spaces between the neighboringparticles are enlarged, and thus the volumes of the Al2O3 layerincreased. When the particles are completely dispersed, the subse-quent increase of ultrasonic processing time will result in nochange in the volume, as shown in Fig. 4 where sample 5 andsample 6 have almost the same volume for the Al2O3 layer.

To make this point clear, the volume of the Al2O3 layer as afunction of processing time is shown in Fig. 5. Clearly, the trendshown in Fig. 5 is identical to that of cavitation noise signals. Webelieve this experiment directly supports our statement that whenthe cavitation noise signals are steady, the particles are well dis-persed. Thus, by detecting when the cavitation noise signals go

Fig. 3 Al2O3 suspension immediately after the ultrasonictreatment

Fig. 4 Deposited Al2O3 particles (Al2O3 20 g, ultrasonic driving power 40 W, 26 hsafter treatment)

Fig. 5 The volume of the deposited Al2O3 particles as afunction of the processing time

Journal of Manufacturing Science and Engineering JUNE 2013, Vol. 135 / 031015-3

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

into steady state, we can determine when the dispersion is suffi-cient. To achieve this goal, we shall introduce two quantitativeindices that measure the power level of the cavitation noises.

3.2 Indices of Cavitation Noise Power. Figure 6 shows thefrequency spectrum of the cavitation noise 40 s after the ultrasonicpower is turned on for (a) 40 W and (b) 100 W ultrasonic powersin tap water with addition of 10 g Al2O3. For (c) and (d) in Fig. 6the cavitation noise spectrum is expressed in a logarithmic scale.From this figure we can clearly see the harmonics, ultraharmonics,subharmonics, and “white noise”. For ultrasonic power 100 W, allof these components, especially the “white noise” and subhar-monics, are stronger than that for power 40 W, indicating a moreviolent cavitation.

Two indices are used to quantitatively describe the cavitationnoise power (CNP) in this research. The first one, termed asCNP-1, is defined as the integration of cavitation noise spectrumover frequency from 0–200 KHz in a logarithmic scale to enhancethe “white noise” contribution:

CNP1 ¼ð

A fð Þdf �X

A fð ÞDf (2)

where A fð Þ is the DFT spectrum amplitude in a logarithmic scaleand f denotes the frequency. This method was developedby Frohly et al. [25] and later used by Gibson et al. [11], whoshowed that CNP-1, multiplied with time t, is directly proportionalto the ultrasound energy density obtained by the calorimetry tech-nique. The second method, termed as CNP-2, is defined as theaveraged square of the cavitation noise signal in each second,

CNP2 ¼

Xn

i¼1

U2i

n(3)

where Ui is the cavitation noise signal and n is the number of sam-ples in each second. Using Parseval’s theorem, it can be proventhat CNP-2 is proportional to the summation of the spectralenergy density (the square of the spectrum amplitude) across allfrequency components.

In Fig. 7(a), the CNP-1 is plotted as a function of time for dif-ferent amounts of Al2O3 particles with ultrasonic power 40 W.

Three stages are clearly seen in the figure, the initial burst in stageI, the increasing region in stage II, and the steady state in regionIII. The influence of particles concentration on the cavitation noisepower is significant. The suspension with more particles has lowercavitation noise power, especially in stage II. This is consistentwith what we expect since Al2O3 particles absorb and scatteracoustic energy. The more particles, the higher the ultrasonicattenuation coefficient and thus, the lower the cavitation noisepower. After the particles are completely dispersed, the scatteringeffect is almost eliminated, which can be seen from the CNP-1curves in the steady state that there is little difference among thesecurves.

Figure 7(b) shows the influence of ultrasonic power on CNP-1for tap water with 30 g Al2O3 particles. We can clearly see thatincreasing the ultrasonic power could increase the cavitation noisepower. Besides, it is faster for CNP-1 to reach steady state withhigher ultrasonic driving power. The reason is obvious thatincreasing the ultrasonic driving power could intensify the cavita-tion, especially the transient cavitation, and thus increase the cavi-tation noise power and dispersion efficiency. We can also find thatwhen the ultrasonic driving power is above 70 W, there is almostno significant change on CNP-1 curves. The possible reason isthat for the ultrasonic driving power above 70 W, the cavitation isfully developed. The corresponding curves for CNP-2 are shownin Fig. 8, from which we can find that the variance of CNP-2 big-ger than that in CNP-1. Note that we present them separately toavoid overlapping due to large noise. In Sec. 4, we will focus onthe dispersion status detection by monitoring the CNP indices.

4 Steady State Detection

From above discussion, we can see that to monitor theultrasonic cavitation based dispersion process using the acousticemission signal, it is critical to detect the steady state of the acous-tic signal. In the literatures, there exist some steady state detecttechniques. Many of these exiting techniques are developed andused in the discrete-event simulations to remove or truncate theinitialization bias [36–39]. These techniques are off-line methodsand not applicable for on-line monitoring purpose because theyrequire a large number of observations in the steady state to accu-rately estimate the truncation point. In real time monitoring, wewant to detect the steady state as soon as possible with a verylimited number of observations in the steady state. There are very

Fig. 6 Cavitation noise spectrum for tap water with 10 g Al2O3 particles at the timeof 40 s after the ultrasonic power is turned on: (a) 40 W, (b) 100 W, (c) 40 W, naturallogarithmic scale, (d) 100 W, natural logarithmic scale

031015-4 / Vol. 135, JUNE 2013 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

limited on-line steady state detection techniques. Among thesemethods, a modified variance ratio test (call it R-test here)[40,41], which was first used in chemical process control, is avery effective and well known method with low computationallycost and relative independence of system noise.

In this section, we will introduce one off-line and two on-linesteady state detection methods. The off-line method is called asexponentially weighted moving average-marginal standard errorrules (EWMA-MSER) method. This method is refined upon theexiting MSER method to make it more robust to noise. AlthoughEWMA-MSER is an off-line method, it can provide insights tothe cavitation based dispersion process and serve as a benchmarkto evaluate the performance of on-line detection algorithms.Among the two online methods introduced in this section, one isthe newly proposed nonoverlapping slope detection method

(NSDM) and one is the existing R-test method. The performanceof these two online methods will be systematically evaluated andcompared as well.

4.1 Off-Line Detection

4.1.1 EWMA-MSER Method. The MSER [42] determinesthe truncation point (steady state point in this research) thatminimizes the width of the marginal confidence interval about thetruncated sample mean (steady state mean). It outperforms otherheuristic algorithms on models that contain exponential shift bias[43] and these models are very similar to CNP signals. A laterrefinement, MSER-5 [34], was developed where the raw observa-tions are grouped into nonoverlapping batches with each batchhaving five observations and MSER is performed on these batch

Fig. 8 CNP-2 as a function of time for different ultrasonic power in tap water with 30 g Al2O3

particles

Fig. 7 The influence of particle concentration and ultrasonic power on CNP-1: (a) CNP-1 as a function of time for differentamounts of Al2O3 particles with ultrasonic power 40 W, (b) CNP-1 evolves with time for different ultrasonic power in tap waterwith 30 g Al2O3 particles

Journal of Manufacturing Science and Engineering JUNE 2013, Vol. 135 / 031015-5

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

means. It was shown that MSER-5 was better than MSER in mostcases [43]. However, MSER-5 did not work well on CNP signalsbecause using MSER-5 made the sample size very small, whichsignificantly reduced its detection accuracy.

Mathematically, the MSER method can be briefly described asfollows. Given the observations Yi : i ¼ 1; 2;…; nf g, assume thesteady state samples are Yi : i ¼ d þ 1; d þ 2;…; nf g. Then thehalf-width of the 100 1� að Þ% confidence interval for the estimateof the steady state mean is given by

CI d þ 1; nð Þ ¼za=2Sn;dffiffiffiffiffiffiffiffiffiffiffi

n� dp (4)

where za=2 is the inverse of the cumulative density function forstandard normal distribution at probability 1� a=2, and Sn;d is thestandard sample deviation given by

Sn;d ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1

n� d � 1

Xn

i¼dþ1

Yi � �Yn;d

� �2

s(5)

where

�Yn;d ¼1

ðn� dÞXn

i¼dþ1

Yi

Thus, the optimal truncation point d� minimizes the confidenceinterval and is given by

d� ¼ arg minn�d�0

CI d þ 1; nð Þð Þ

¼ arg minn�d�0

CI2 d þ 1; nð Þ� �

¼ arg minn�d�0

Xn

i¼dþ1

Yi � �Yn;d

� �2

n� dð Þ n� d � 1ð Þ

0BBB@1CCCA

(6)

Since n� d, the denominator can be simplified from n� dð Þn� d � 1ð Þ to n� dð Þ2. Thus, the monitoring statistic of this

method, denoted as “MSER”, is given as

MSER ¼ 1

n� dð Þ2Xn

i¼dþ1

Yi � �Yn;d

� �2(7)

Figure 9 is an example of MSER on CNP indices with duration of500 s, which shows that the transition time estimated by MSER isa little shorter than the true transition time. It is consistent withWhite et al.’s results [43] that MSER failed to truncate all of thebias, particularly when the process noise level is high.

In order to make the method more robust, we propose to useexponentially weighted moving average (EWMA) to smooth outshort-term fluctuations without impacting on the long-term trendsand then perform MSER on the filtered samples. We call thismethod as EWMA-MSER. Specifically, for the observations

Yi : i ¼ 1; 2;…; nf g, the smoothed samples are given by

Fig. 9 Illustration of MSER for CNP signals (power 40 W, Al2O3 30 g)

Fig. 10 An example of MSER-EWMA on CNP signals (40 W, 30 g Al2O3)

031015-6 / Vol. 135, JUNE 2013 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Yf ;i ¼ kYi þ 1� kð ÞYf ;i�1 (8)

where k is a parameter such that 0 < k � 1. A small k puts a lightweight on the recent observations and more noises are smoothedout. However, too low k will delay the detection when the processenters into steady state. Here, we choose k ¼ 0:5 and 0.1 forCNP-1 and CNP-2, respectively, and the detected transition timesfor all power levels are quite consistent with visually examinedvalues. Figure 10 shows an example of MSER-EWMA.

4.1.2 Results and Discussion on Off-Line Steady StateDetection for Acoustic Signals. Figure 11 shows the MSER andEWMA-MSER determined transition times as a function of ultra-sonic powers for both CNP-1 and CNP-2. There is no significantdifference between the transition times of CNP-1 and CNP-2using the same detection method, indicating that both signals canbe used to monitor the dispersion status. We can also find that thedetection results of EWMA-MSER are larger than those byMSER, for the reason that MSER-EWMA has successfullyreduced the influence of noise and more accurately detected thetransition times than MSER did. We choose the time instancedetermined by EWMA-MSER method as the benchmark in thefollowing work.

Figure 12 shows the dispersion time (transition time) as a func-tion of Al2O3 concentration. We can see that when the particleconcentration is small �20 g=500 mLð Þ, there is no significantchange on the dispersion efficiency. This result is consistent withGibson et al.’s finding [11] that changing particle concentrationhad relatively little effect (<5%) on the ability of ultrasound to

break particles. The reason is that the suspended Al2O3 particlescould act as cavitation nuclei and enhance the cavitation process.Increasing particle concentration could increase the acousticenergy loss due to attenuation effects. On the other hand, itcan also increase the cavitation nuclei, which improves thedispersion process. When the particle concentration is high>20 g=500 mLð Þ, the acoustic attenuation effects overwhelm the

influence of cavitation nuclei and therefore the dispersion timeneeded to break Al2O3 particles is significantly increased by add-ing more Al2O3 particles.

Figure 13 shows the mean CNP-1 calculated by averagingthe CNP-1 indice (Fig. 7(b)) in the transient state (as marked inFig. 9) under different ultrasonic power levels. There are obviousthree regions. For the ultrasonic power less than 50 W, the meanCNP-1 grows slowly by increasing the ultrasonic power. The cavi-tation type under this ultrasonic power level may be mainly stablecavitation. For ultrasonic power from 50 W to 70 W, there appearsa fast mean CNP-1 increasing region, which is caused by the onsetof transient cavitation. For ultrasonic power above 70 W, themean CNP-1 reaches the maximum level and the subsequentincrease of the ultrasonic power will not result in any significantchanges. We can treat the cavitation in this region as the fullydeveloped transient cavitation. Also in this region, the dispersionefficiency is almost unchanged, as shown in Fig. 11. Thereforeultrasonic power 70 W can be considered as the optimal cavitationparameter in this experiment.

4.2 On-Line Steady State Detection

4.2.1 Description of NSDM and R-Test Methods. The algo-rithm of NSDM is fairly simple and easy to implement. In thismethod, an ordinary least square (OLS) linear regression over anonoverlapping moving data window with m samples of the CNPsignal is performed until the fitted line is “flat” and continuously“flat” for D consecutive windows. Suppose the detected startingpoint of the stead steady is bTs, the estimated slope is Si for ith win-dow and the slope threshold is Sc, then

bTs ¼ m N þ D� 1ð Þ (9)

where

N ¼ arg min i bSiþk

������ ��� < Sc; k ¼ 0; 1;…;D� 1n o

An existing on-line method, the variance ratio test [44], is alsoan effective method to detect steady state. In this method, the var-iance of a moving data window is calculated in two differentways: (1) mean squared deviation from the average V1ð Þ and (2)

Fig. 11 MSER and MSER-EWMA detected transition times asfunctions of ultrasonic power for CNP-1 and CNP-2 (30 g Al2O3)

Fig. 12 The influence of Al2O3 concentration on dispersiontime estimated by MSER-EWMA on CNP-1 (40 W)

Fig. 13 Mean CNP-1 in the transient state as a function of ultra-sonic power (30 g Al2O3)

Journal of Manufacturing Science and Engineering JUNE 2013, Vol. 135 / 031015-7

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

mean squared differences of successive data V2ð Þ. In the non-steady or transient state, the first variance will be larger than thesecond variance and the ratio V1=V2 is larger than 1. In the steadystate, this ratio is expected to approach 1. In the test, the nullhypothesis (steady state) will be rejected until the ratio is belowa threshold. In order to reduce the computational cost and datastorage, Cao and Rhinehart [40,41] used an recursive method toestimate the variances S2

1;i and S22;i:

Yf ;i ¼ k1Yi þ 1� k1ð ÞYf ;i�1

V2f ;i ¼ k2 Yi � Yf ;i�1

� �2þ 1� k2ð ÞV2f ;i�1

d2f ;i ¼ k3 Yi � Yi�1ð Þ2þ 1� k3ð Þd2

f ;i�1

S22;i ¼ 2� k1ð ÞV2

f ;i=2

S22;i ¼ d2

f ;i=2

(10)

Here k1, k2, and k3 are the parameters with 0 < kj � 1j ¼ 1; 2; 3ð Þ. The ratio is given by

Ri ¼S2

1;i

S22;i

¼2� k1ð ÞV2

f ;i

d2f ;i

(11)

Similary, suppose the ratio threshold is Rc, then bTs is expressed by

bTs ¼ arg min i Ri < Rcjð Þ (12)

There is a trade-off between rapid tracking of the process andseparating the probability density function of R between thesteady state and the nonsteady state in the selection of the parame-ters k1, k2, and k3. In general, small parameters can reduce theinfluences of noise on estimating the variances and lead to biggerseparation in the probability distribution of R of the steady stateand the nonsteady state. However, small parameters may delaythe detection. Cao [41] provided some settings of parameters andtheir detection performance in different situations. Interested read-ers may refer to their paper for more details.

4.2.2 Performance Evaluation and Comparison. To evaluateand compare the performance of steady state detection algorithms,it is natural to use the bias in the detection as the evaluation met-rics. Thus, in this research, we define a criterion named theexpected detection bias (EDB) as

EDB ¼ EjbTs � T0j (13)

where bTs and T0 are the starting point of the steady state detectedby the algorithms and the underlying true value, respectively. Forthe cavitation based dispersion, an early detection, i.e., bTs < T0,will lead to insufficient dispersion and bad quality product. Thus,it is critical to also evaluate the probability of early detection.Toward this goal, we define another criterion named false alarmrate (FAR),

FAR ¼ PrðbTs < T0Þ (14)

to quantitatively evaluate it.For NSDM, these criteria (EDB and FAR) can be derived as

follows. Given the observations Yi ¼ ym i�1ð Þþ1; ym i�1ð Þþ2;…; ymi

� �for the ith data window and the time index Ti ¼ tm i�1ð Þþ1;

�tm i�1ð Þþ2;…; tmiÞ, where m is the window size, the OLS estimatorof the slope is

Si ¼

Xm

k¼1

tm i�1ð Þþk � �Ti

� �ym i�1ð Þþk

Xm

k¼1

tm i�1ð Þþk � �Ti

� �2

where �Ti is the mean value of the time index. Suppose theobservation noises follow independent and identically distributed nor-

mal distribution, �m i�1ð Þþk N 0;r2ð Þ and ym i�1ð Þþk¼ f tm i�1ð Þþk

� �þ�m i�1ð Þþk where f tm i�1ð Þþk

� �is the expected value, thenbSiN li;r

2i

� �with

li ¼

Xm

k¼1

tm i�1ð Þþk � �Ti

� �f tm i�1ð Þþk

� �Xm

k¼1

tm i�1ð Þþk � �Ti

� �2

(15)

r2i ¼

r2Xm

k¼1

tm i�1ð Þþk � �Ti

� �2

(16)

Define ai as the probability that the absolute value of the slopeof the ith data window is below the slope threshold Sc, then

ai ¼ Pr jSi

�� < Sc

� �¼ / Sc � lið Þ=rið Þ � / �Sc � lið Þ=rið Þ (17)

Define the probability mass function (PMF) Pn as the probabil-ity of receiving the steady state alarm after monitoring thenth; nþ 1ð Þth;…; nþ D� 1ð Þth nonoverlapping moving win-dows (total D windows, suppose we stop the monitoring processimmediately after we receive the steady state alarm). Let a0¼ 0,then

Pn ¼ Pr N ¼ nð Þ ¼ Pr N ¼ njN > n� 1�Dð ÞPr N > n� 1�Dð Þ

¼1� an�1ð Þan…anþD�1½ 1�

Xn�1�D

i¼1

Pi

!; for n > 1þD

1� an�1ð Þan…anþD�1; for n � 1þD

8>><>>:9>>=>>;

(18)

FAR can be calculated by

FAR ¼Xn0

i¼1

Pi (19)

Here, n0 ¼ T0=m� Dþ 1b c, the largest index of the data win-dow where the following D� 1 data windows are before thesteady state transition time T0. EDB is expressed as

EDB ¼ E Ts � T0

�� �� ¼X1n¼1

Pn m nþ D� 1ð Þ � T0j j (20)

Although, there are infinite terms in the expression above, itconverges very fast due to the rapid convergence of Pn and weonly need to sum up a small number of terms to calculate it.Clearly, bTs is required to be as close as possible to T0 and thus thesmaller the EDB, the higher the detection accuracy.

For the R-test method, it is very difficult to get the analyticalexpression for these evaluation criteria due to the complexity ofthe algorithm. Thus, Monte Carlo simulations have to be used tocompute them. In the simulation, we will need to simulate the sig-nal with noise many times and then apply the detection algorithmto the simulated signals. Finally the detection results will be aver-aged to obtain the values of EDB and FAR.

To compute EDB and FAR, we need to know the underlyingtrue value of the starting point of the steady state. Thus, we need toassume an underlying function to describe the changes of the sig-nal. Here, we select the exponential bias function as the underlyingfunction. This function was used by Cash et al. [36] and White etal. [43] as a generic function to assess off-line heuristic algorithms.

031015-8 / Vol. 135, JUNE 2013 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Furthermore, the behavior of this function is quite similar to ourCNP signals. The mathematical form of the function is given as

f ið Þ ¼C 1� ea i�1ð Þ� �

; i � T0 að ÞC 1� ea T0�1ð Þ� �

; i > T0 að Þ

((21)

where T0 að Þ is the smallest integer i where the derivative of

C 1� ea i�1ð Þ� �is less than C� 10�4. C ¼ 0:7 is chosen to match

the CNP signals. The time series are generated by Yi ¼ f ið Þ þ �i

where �i N 0;r2ð Þ and r ¼ 0:04. Eight values of a and the cor-responding T0, as shown in Table 1, were chosen to study theinfluence of signal changing rate on the detection accuracy. Onerepresentative signal generated with T0 ¼ 461 is shown in Fig. 14.

Figure 15 shows the EDB and FAR as functions of detectionthreshold for NSDM and R-test. EDB and FAR of NSDM weredirectly calculated by Eqs. (20) and (21), respectively. For R-test,computer simulations were performed where computer experi-ments were repeated for 30,000 times for each set of detectionparameters and signal parameters. For NSDM, m ¼ 50;D ¼ 2 andfor R-test, k1 ¼ 0:05, k2 ¼ 0:05; k3 ¼ 0:08. From Fig. 15, we canfind that:

(1) For both NSDM and R-test, as we increase the detectionthreshold, EDB decreases rapidly at first, and then gradu-ally increases. FAR is always nondecreasing when thethreshold is increased. The optimal threshold should be thevalue that has low FAR and also low detection bias.

(2) Both methods perform better on signals with fast changingrate (large a) than on slow changing signals in most cases.For these rapidly changing signals, NSDM and R-test havelow detection bias and false alarm rate. Besides, the detec-tion bias is more stable under different detection thresholds.In this situation, the R-test is better than NSDM due tolower computational cost and data storage.

(3) For the signals with low changing rate, the bias and falsealarm rate of R-test are more sensitive to the change ofdetection threshold than NSDM. An optimal detectionthreshold for one signal may work badly on the othersignals.

Table 1 Bias function parameters

Model # 1 2 3 4 5 6 7 8

a 0.01 0.012 0.015 0.019 0.026 0.039 0.07 0.1T0 461 339 334 276 214 153 94 69

Fig. 14 An example of generated signal (a 5 0.01, T0 5 461)

Fig. 15 The expected detection bias and false alarm rate as functions of detection threshold for NSDMand R-test (NSDM: m 5 50, D 5 2; R-test: k1 5 0.05, k2 5 0.05, k3 5 0.08)

Journal of Manufacturing Science and Engineering JUNE 2013, Vol. 135 / 031015-9

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

(4) For R-test, the optimal detection threshold Rc for differentsignals varies significantly. Two or even more sets of detec-tion parameters are required to make R-test work well onall signals with different To. NSDM outperforms R-test onsignals with large range of changing rate in terms of easi-ness in selection of detection parameters and the stability ofdetection accuracy.

4.2.3 Results of Online Steady State Detection on CNPIndices. The detection parameters (shown in Table 2) for NSDMand R-test were selected by minimizing the difference between

the detected starting points of the steady state with EWMA-MSER detected results.

An illustration of R-test on CNP signals is shown in Fig. 16.Figure 17 shows the transition time as a function of ultrasonic

power detected by EWMA-MSER, NSDM, and R-test. It shouldbe noted that the changing rate for CNP-1 before the steady stateis lower than CNP-2. CNP-1 increases rapidly to a level close tothe steady state level at first, and then drifts slowly into the steadystate, while CNP-2 increases with a relatively constant and highchanging rate.

The difference between CNP-1 and CNP-2 leads to differentperformance of the detection methods. R-test works better onCNP-2 signals than on CNP-1 signals, which is consistent withthe simulation results that for high changing rate, R-test performswell with only one threshold while for signals with low changingrate, it is hard to find a threshold that works for all signals. NSDMworks well on both CNP-1 and CNP-2 signals.

The above results show that R-test and CNP-2 are the optimalchoice in our current experiments. R-test is less influenced by thenoise. In addition, R-test requires less calculation and data storagethan NSDM. In the real-world application, however, the signalsmay show wide range of changing rate, where NSDM may be

Table 2 Detection parameters for NSDM and R-test

NSDM R-test

Sc

CNP m D Power � 50W Power > 50W k1 k2 k3 Rc

CNP-1 20 2 1� 10�4 3� 10�4 0.1 0.05 0.05 2CNP-2 30 2 2� 10�3 4� 10�3 0.05 0.05 0.08 2

Fig. 16 Illustration of NSDM (a) and R-test (b) for CNP signals (40 W, Al2O3 30 g)

Fig. 17 Transition time detected by MSER-EWMA, NSDM and R-test (Al2O3 30 g)

031015-10 / Vol. 135, JUNE 2013 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

preferred. CNP-2 is better than CNP-1 for detection purpose sinceit has larger changing rate than CNP-1. Besides, CNP-2 is compu-tationally less expensive to calculate than CNP-1 since the latterrequires Fourier transform.

5 Conclusion

In the present work we have proposed a method based on thecavitation noise to monitor the particle dispersion process. Thecavitation noise signals and their spectrum are analyzed anddiscussed in details. The cavitation noise signals are divided intothree stages. The first stage corresponds to the step response of thecavitation system. The second stage is the most important stagewhich characterizes the evolving of the cavitation and thedispersion process. The third stage is the steady state in which theparticles are dispersed well. The Al2O3 particles can reducethe strength of the cavitation noise by increasing the acousticattenuation characterized as absorption and scattering of theacoustic wave. The attenuation effect is reduced as the particlesare well dispersed. These characteristics of the cavitation noisecan be used to monitor the dispersion status.

Two quantitative indices (CNP-1 and CNP-2) are chosen tocapture the evolution of the cavitation noise and CNP-2 is betterin terms of computational cost and detection accuracy. The off-line method MSER and its modification EWMA-MSER areused to identify the dispersion steady state. The proposedEWMA-MSER works quite well and its detection results are usedas the benchmark to develop and evaluate the on-line detectionmethods. Two online methods, NSDM and R-test are applied andsystematically compared. In the comparison, we proposed to usethe expected detection bias and the false alarm probability toquantitatively evaluate the performance of these two detectionmethods. We further derived the analytical expressions for thesequantities for the proposed nonoverlapping slope detectionmethod. With these expressions, we can easily calculate the aver-age run length, expected detection bias and false alarm rate for agiven signal. We also obtained these quantities for R-test usingnumerical methods. Both methods work well on signals with highchanging rate and R-test outperforms NSDM in terms of computa-tional costs and data storage. For signals with large range chang-ing rate, NSDM outperforms R-test in terms of easiness inselection parameters in the algorithm and the stability of detectionaccuracy.

The results of this research provide useful guidelines forestablishing a real-time process monitoring and control schemefor ultrasonic cavitation based dispersion processes, which is acritical process in the manufacturing of many composite materi-als. Admittedly, we only tested this monitoring technique inAl2O3-water system, but it is generally believed that the trend andcharacteristics obtained from Al2O3-water based experiment willhold for other particle-liquid system that can be easily cavitated.Unfortunately, limited by our experimental setup at this stage, wehave to leave this work to the future.

Acknowledgment

The authors would like to thank the editor and the referees fortheir valuable comments and suggestions. The financial support ofthis work is provided by National Science Foundation with GrantNo. 0926084 and the Technology Innovation Program of NationalInstitute of Standards and Technology.

References[1] He, F., Han, Q., and Jackson, M. J., 2008, “Nanoparticulate Reinforced

Metal Matrix Nanocomposites—A Review,” Int. J. Nanoparticles, 1(4), pp.301–309.

[2] Bakshi, S., Lahiri, D., and Agarwal, A., 2010, “Carbon Nanotube ReinforcedMetal Matrix Composites—A Review,” Int. Mater. Rev., 55(1), pp. 41–64.

[3] Arsenault, R., Wang, L., and Feng, C., 1991, “Strengthening of CompositesDue to Microstructural Changes in the Matrix,” Acta Metall. Mater., 39(1), pp.47–57.

[4] Nardone, V., and Prewo, K., 1986, “On the Strength of Discontinuous Sili-con Carbide Reinforced Aluminum Composites,” Scr. Metall., 20(1), pp.43–48.

[5] Yang, Y., Lan, J., and Li, X., 2004, “Study on Bulk Aluminum Matrix Nano-Composite Fabricated by Ultrasonic Dispersion of Nano-Sized Sic Particles inMolten Aluminum Alloy,” Mater. Sci. Eng., A, 380(1), pp. 378–383.

[6] Cao, G., Konishi, H., and Li, X., 2008, “Mechanical Properties and Microstruc-ture of Mg/SiC Nanocomposites Fabricated by Ultrasonic Cavitation BasedNanomanufacturing,” ASME J. Manuf. Sci. Eng., 130(3), p. 031105.

[7] Gedanken, A., 2007, “Doping Nanoparticles Into Polymers and Ceramics UsingUltrasound Radiation,” Ultrason. Sonochem., 14(4), pp. 418–430.

[8] Yang, Y., and Li, X., 2007, “Ultrasonic Cavitation-Based Nanomanufacturingof Bulk Aluminum Matrix Nanocomposites,” ASME J. Manuf. Sci. Eng.,129(2), pp. 252–255.

[9] Li, X., Yang, Y., and Weiss, D., 2007, “Ultrasonic Cavitation Based Dispersionof Nanoparticles in Aluminum Melts for Solidification Processing of BulkAluminum Matrix Nanocomposite: Theoretical Study, Fabrication andCharacterization,” AFS Trans., 115, Paper No. 07-133(02), p. 249.

[10] Bittmann, B., Haupert, F., and Schlarb, A. K., 2009, “Ultrasonic Dispersion ofInorganic Nanoparticles in Epoxy Resin,” Ultrason. Sonochem., 16(5), pp.622–628.

[11] Gibson, J. H., Hon, H., Farnood, R., Droppo, I. G., and Seto, P., 2009, “Effectsof Ultrasound on Suspended Particles in Municipal Wastewater,” WaterResearch, 43(8), pp. 2251–2259.

[12] Suslick, K. S., Didenko, Y., Fang, M. M., Hyeon, T., Kolbeck, K. J., McNa-mara, W. B., Mdleleni, M. M., and Wong, M., 1999, “Acoustic Cavitation andIts Chemical Consequences,” Philos. Trans. R. Soc. London, Ser. A, 357(1751),pp. 335–353.

[13] Neppiras, E. A., 1980, “Acoustic Cavitation,” Phys. Rep., 61(3), pp. 159–251.[14] Hentschel, W., and Lauterborn, W., 1984, “New Speed Record in Long Series

Holographic Cinematography,” Appl. Opt., 23(19), pp. 3263–3265.[15] Lauterborn, W., and Hentschel, W., 1985, “Cavitation Bubble Dynamics Stud-

ied by High Speed Photography and Holography: Part One,” Ultrasonics, 23(6),pp. 260–268.

[16] Burdin, F., Tsochatzidis, N., Guiraud, P., Wilhelm, A., and Delmas, H., 1999,“Characterisation of the Acoustic Cavitation Cloud by Two Laser Techniques,”Ultrason. Sonochem., 6(1–2), pp. 43–51.

[17] Tsochatzidis, N., Guiraud, P., Wilhelm, A., and Delmas, H., 2001,“Determination of Velocity, Size and Concentration of Ultrasonic CavitationBubbles by the Phase-Doppler Technique,” Chem. Eng. Sci., 56(5), pp.1831–1840.

[18] Nishi, R., 1975, “The Scattering and Absorption of Sound Waves by a Gas Bub-ble in a Viscous Liquid,” Acustica, 33(2), pp. 65–74.

[19] Duraiswami, R., Prabhukumar, S., and Chahine, G. L., 1998, “Bubble CountingUsing an Inverse Acoustic Scattering Method,” J. Acoust. Soc. Am., 104, pp.2699–2717.

[20] Vaughan, P., 1968, “Investigation of Acoustic Cavitation Thresholds by Obser-vation of the First Subharmonic,” J. Sound Vib., 7(2), pp. 236–246.

[21] Neppiras, E., 1969, “Subharmonic and Other Low-Frequency Emission FromBubbles in Sound-Irradiated Liquids,” J. Acoust. Soc. Am., 46, pp. 587–601.

[22] Avvaru, B., and Pandit, A. B., 2009, “Oscillating Bubble Concentration and ItsSize Distribution Using Acoustic Emission Spectra,” Ultrason. Sonochem.,16(1), pp. 105–115.

[23] Puga, H., Barbosa, J., Gabriel, J., Seabra, E., Ribeiro, S., and Prokic, M., 2011,“Evaluation of Ultrasonic Aluminium Degassing by Piezoelectric Sensor,” J.Mater. Process. Technol., 211, pp. 1026–1033.

[24] Labouret, S., Frohly, J., and Rivart, F., 2006, “Evolution of an 1 MHz Ultra-sonic Cavitation Bubble Field in a Chopped Irradiation Mode,” Ultrason. Sono-chem., 13(4), pp. 287–294.

[25] Frohly, J., Labouret, S., Bruneel, C., Looten-Baquet, I., and Torguet, R., 2000,“Ultrasonic Cavitation Monitoring by Acoustic Noise Power Measurement,”J. Acoust. Soc. Am., 108, pp. 2012–2020.

[26] Hodnett, M., Chow, R., and Zeqiri, B., 2004, “High-Frequency AcousticEmissions Generated by a 20 kHz Sonochemical Horn Processor DetectedUsing a Novel Broadband Acoustic Sensor: A Preliminary Study,” Ultrason.Sonochem., 11(6), pp. 441–454.

[27] Ilyichev, V., Koretz, V., and Melnikov, N., 1989, “Spectral Characteristics ofAcoustic Cavitation,” Ultrasonics, 27(6), pp. 357–361.

[28] Lauterborn, W., 1976, “Numerical Investigation of Nonlinear Oscillations ofGas Bubbles in Liquids,” J. Acoust. Soc. Am., 59, pp. 283–293.

[29] Prosperetti, A., Crum, L. A., and Commander, K. W., 1988, “Nonlinear BubbleDynamics,” J. Acoust. Soc. Am., 83, pp. 502–514.

[30] Young, F. R., 1989, Cavitation, McGraw-Hill, New York.[31] Guth, W., 1956, “Nichtlineare Schwingungen von Luftblasen in Wasser,”

Acustica, 6, pp. 532–538.[32] Coleman, A., Choi, M., Saunders, J., and Leighton, T., 1992, “Acoustic

Emission and Sonoluminescence Due to Cavitation at the Beam Focus of anElectrohydraulic Shock Wave Lithotripter,” Ultrasound Med. Biol., 18(3), pp.267–281.

[33] Yasui, K., Tuziuti, T., Kozuka, T., and Towata, A., 2010, “Origin of the Broad-Band Noise in Acoustic Cavitation,” Proceedings of the 20th International Con-gress on Acoustics (ICA), Sydney, Australia, Aug. 23–27, pp. 23–27.

[34] Spratt, S., 1998, Heuristics for the Startup Problem, M.S. thesis, Department ofSystems Engineering, University of Virginia, Charlottesville, VA.

[35] Allegra, J. R., and Hawley, S. A., 1972, “Attenuation of Sound in Suspensionsand Emulsions: Theory and Experiments,” J. Acoust. Soc. Am., 51(5B), pp.1545–1564.

Journal of Manufacturing Science and Engineering JUNE 2013, Vol. 135 / 031015-11

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

[36] Cash, C. R., Nelson, B. L., Dippold, D. G., Long, J. M., and Pollard, W. P.,1992, “Evaluation of Tests for Initial-Condition Bias,” Proceedings of the 24thConference on Winter Simulation, ACM, pp. 577–585.

[37] Goldsman, D., Schruben, L. W., and Swain, J. J., 1994, “Tests for TransientMeans in Simulated Time Series,” Naval Research Logistics (NRL), 41(2), pp.171–187.

[38] Robinson, S., 2002, “A Statistical Process Control Approach for Estimating theWarm-Up Period,” Proceedings of the 2002 Winter Simulation Conference,IEEE, Vol. 1, pp. 439–446.

[39] Franklin, W. W., 2009, “The Theoretical Foundation of the MSER Algorithm,”Ph.D. thesis, Department of Systems and Information Engineering, Universityof Virginia, Charlottesville, VA.

[40] Cao, S., and Rhinehart, R. R., 1995, “An Efficient Method for On-Line Identifi-cation of Steady State,” J. Process Control, 5(6), pp. 363–374.

[41] Cao, S., 1996, “Statistically Based Techniques for Process ControlApplications,” Ph.D. thesis, Chemical Engineering, Texas Tech University,Lubbock, TX.

[42] White, K. P., Jr., 1997, “An Effective Truncation Heuristic for Bias Reductionin Simulation Output,” Simulation, 69(6), pp. 323–334.

[43] White, K. P., Jr., Cobb, M. J., and Spratt, S. C., 2000, “A Comparisonof Five Steady-State Truncation Heuristics for Simulation,” Proceedings ofthe 2000 Winter Simulation Conference, IEEE, Vol. 1, pp. 755–760.

[44] Crow, E. L., Davis, F. A., and Maxfield, M. W., 1960, Statistics Manual: WithExamples Taken From Ordinance Development, Dover, New York.

031015-12 / Vol. 135, JUNE 2013 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 09/04/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use