D. G. Davydova^a

Ufa State Petroleum Technological University, ul. Kosmonavtov 1, Ufa, Bashkortostan, Russia

The results of using thresholdless data recording during acousticemission testing are pre

sented. A method for processing acousticemission signals that allows the detection of acousticemis

sion sources at a signaltonoise ratio of <1 is proposed.

Keywords: acousticemission testing, thresholdless data recording, adaptive filtering

DOI: 10.1134/S1061830915050034

The urgency of using acousticemission (AE) systems for monitoring the state of equipment increases

from year to year. However, diagnosing technical devices without their decommissioning often has sub

stantial limitations. This is primarily associated with the high sensitivity of the informative parameters of

acoustic emission to noises of different natures. Testing is significantly impeded if there are multiple noise

sources: turbulent or cavitation regimes of the workingmedium motion, synthesis reactions, rubbing

units and parts, electric interferences, vibrations, and other unavoidable technological types of noise that

are not related to the operation of a diagnosed device [1]. As a consequence, an additive noise signal may

appear with an amplitude that far exceeds the signal amplitudes from AE sources that indicate the pres

ence of flaws [2]. Correspondingly, an urgent problem for actual industrial objects is the development of

a noiseimmune algorithm for identifying AE sources, which provides the reliable detection and selection

of an AE signal from a flaw in real time within a wide signaltonoise ratio (SNR) range.

The existing AEtesting systems do not allow one to solve the formulated problem because the main

noisefiltering method reduces to the detection of  an AE signal above the preset threshold amplitude

value. The presence of a threshold during data recording involves a loss in the information on a “useful”

signal whose amplitude is below the preset value. As the threshold increases, the testing quality consider

ably decreases down to the impossibility of accomplishing AE measurements. The questions that are

related to interpreting data under the conditions of the presence of highamplitude noises have been

widely developed in the fields that are related to the AE diagnostics of technological equipment. The issue

is the investigation of seismoacoustic emissions [3, 4] and the digital processing of noisy speech signals

[5, 6]. The principles that were expounded in [5] are identical to the situation of AE testing with threshold

data recording and represent a solution that has an expert nature. The threshold value is determined by the

microphone type and the noise level in the facility. The high stability of the obtained recognition results

directly depends on the threshold value. The authors of [5] pointed out that the method of setting thresh

olds is insufficiently reliable.

An alternative method for solving a noisesuppression problem that was proposed in [3, 6] consists in

the use of multichannel adaptive spatial filters in the selfadaptation mode. This allowed the filtering sys

tem to be tuned with respect to the signal–noise mixture (as in the case of AE testing when it is impossible

to record “pure” noise). As was noted by the authors of paper [3], the operation of such filters is not as

stable and efficient as that of filters that are adapted with respect to “pure noise.” The use of the proposed

approach made it possible to obtain a tool for suppressing intense technogeneous noise when analyzing

seismic data. Another promising method that was presented in [4] is a technique for predicting the seismic

activity using the nonlineardynamics apparatus. It should be noted that in this case the determining con

dition is the presence of an undistorted signal, i.e., the absence of a threshold during data recording. The

validity of this approach is associated with the fact that a system that is studied in seismoacoustics is not 

The idea of TDR was described in [7] and primarily implies the exclusion of a detector–discriminator,

which is present in all existing AE systems, from the testing scheme. The rejection of the amplitudedis

crimination principle allowed us to achieve continuous signal recording without losing the information

content. Owing to this, the TDR that underlies the scheme (Fig. 1) allows one to calculate the spectral,

correlation, and local static and dynamic characteristics without information losses.

It is important that in this case the level of noise and different types of interference may exceed the sig

nal level from a flaw. In this case, the filtering problems are solved already not by using hardware via setting

a discrimination threshold in the analog form but by using software that is based on modern methods for

filtering experimental data (the fast Fourier transform, maincomponents method, adaptivefiltering

method, etc.).

The conclusion on the necessity of the preliminary preparation of AE data was drawn on the basis of

studies of the characteristics of inputsignal spectra. In particular, the analysis of the amplitude spectra

allowed the regions that are characteristic of only the noise component and the noninformative highfre

quency signal region to be revealed. Introducing a prefiltering procedure using a bandpass filter made it

possible to increase the SNR.

The classical scheme of using an AF as a noise suppressor is based on the use of a reference (noise) sig

nal [10]. In a general case, the filter operation implies the use of two separate acousticemission transduc

ers (AETs), viz., independent information channels. In  this case, one transducer, as is shown in Fig. 2

(AET no. 2) must be far from potential flaws (AE sources) in the tested object. A signal is transmitted to a

receiver in the form of a mixture of a useful signal, which is initiated by the presence of a flaw, and a noise

component; the AF input signal is s + n0. An auxiliary transducer records the noise, n1, that is not corre

lated with the signal but directly correlates with the noise, n0. Noise n1 is a reference signal for the noise

suppressing device. The task of the AF is to transform the input signal so as to maximally approximate it

to the reference signal. Because only the noise component of the reference signal is correlated with the

input signal, an estimate of noise that is present in the input signal will be obtained at the filter output after

the adaptation process is over. As a result of noise filtering, a signal is formed. This signal is subtracted from

the input signal, thus forming a useful signal, a socalled error signal e (Fig. 3). The filtering error is min

imized via a feedback and a corresponding numerical algorithm (e.g., the leastsquares method [10]). The

AF scheme in the twochannel operating mode was realized earlier as applied to the detection of a weak

periodic AE signal in highamplitude noise [7] and was described in detail in [11].

The absence of an AE signal from a flaw in the region where an auxiliary transducer is installed is a

determining condition for the working capacity of the classical AF scheme. The latter condition is not

always fulfilled during AE diagnostics of actual objects because an uncertainty in the flaw location in the

tested object is a priori present. In this respect, the AF scheme with one input information channel is

much more important for practice, especially when constructing AE monitoring systems in real time. This

channel contains only a recorded time series consisting of an additive mixture of an AE signal and noise,

which arrives from a single AET that is positioned near the flaw–AE source in the tested object. In this

study, the singlechannel AF structure was implemented by modifying the filtering algorithm, which

allowed the “blind”adaptation scenario to be accomplished [10]. The essence of the proposed approach

consists in the setting of a time delay that is equivalent to the introduction of a virtual standard signal (Fig. 3).

The delay value is determined by the ratio of the noise and signal correlation times (see below).

To estimate the filteringalgorithm efficiency, numerical criteria are proposed that are based on the dif

ference between the statistical, spectral, and correlation properties of a signal from a flaw and stationary

noise. For this purpose, postprocessing of the AF data was performed, which consisted of four indepen

dent stages: the crosscorrelation proximate diagnostics, calculation of the correlation intervals, analysis

of time intervals between pulses, and the statistical analysis of an AE signal and noise.

that had a crack on its case [12]. Files were recorded using an LCard E2010 datacollection card and the

recording was conducted in the twochannel mode. To record the signal, noise, and their mixture, wide

band piezoelectric acoustic AE transducers of the GT200 type (Global Test, Sarov) were used. A highly

noisy input signal was created using synthesized highamplitude interference (surrogate data [13]), whose

amplitude–frequency characteristic is identical to a noise signal that was obtained during testing. Intro

ducing random mixing of the initialsignal phase and samples fully excludes the possible crosscorrelation of a

signal and noise, with the additive character of the investigated signal being preserved. Exactly this statement of

the problem is of interest from the standpoint of detecting a signal from a flaw. The SNR that was calculated

from the AEsignal and noise power within the recordingsystem passband did not exceed 0.03.

Figure 4 shows a typical result of applying the AF with one information input for detecting a nonperi

odic AE signal that is contained in highamplitude noise. As is seen in Figs. 4a and 4b, the AF efficiently

detects the signal from a flaw. In this case, the number of false (or missed) pulses is within 10%. To estimate

the filtering quality, the functions of the cross correlation, AE coherence, and the signalcorrelation coef

ficient at the filter input and output were calculated [14]. The crosscorrelation and coherence functions

determine the degree of signal similarity in the time and spectral (frequency) domains, respectively. The

correlation coefficient characterizes the degree of statistical correlation of signals and is close to unity, if

this correlation is linear (Figs. 4c, 4d). Figure 4c shows that the positions of the maxima of the crosscor

relation function on the time axis coincide with the corresponding AEsignal pulses at the AF input and

output, and the coherence function in the recorded frequency band slightly differs from unity (Fig. 4d).

In this case, the value of the Pearson correlation coefficient was found to be 0.2035. This indicates a non

linear statistical interrelation of AE signals at the AF input and output with a high degree of similarity in

the time and frequency domains. This allows one to go over to the identification of AE sources on the basis

of analyzing the spectral, correlation, and statistical properties of signals and noise at the AF output.

The crosscorrelation proximate diagnostics of data that were obtained after the AF procedure can be

performed upon the introduction of an additional independent AET that records a noisy AE signal. The

postprocessing involves the calculation of the crosscorrelation function (CCF) between useful signals at

the outputs of the twostage AF and their corresponding noise components. Figure 5 shows the results of

calculating the CCF. It is seen that the deterministic AE signals that have a common source in the tested

object are strongly correlated (Fig. 5a), while a correlation is virtually absent between random noise com

ponents that undergo additional phase distortions (Fig. 5b). Thus, an increase in the cross correlation in

the output channels of the adaptive detector may serve as an important diagnostic indication of the occur

rence of nucleating flaws in the tested object.

The value of the correlation interval is a numerical estimate of the degree of signal correlation:

where p is the probability of the occurrence of n pulses within the time interval.

In this case, the distribution function of time intervals between pulses obeys an exponential law. The

distribution function, f(t), which is the probability that the current time interval, t, is within a preset inter

val, has the following form for the Poisson stream:

It is known that the variance  and the mean value is  [14].

Correspondingly, the ratio  for a random Poisson process (stationary noise) is equal to unity.

The presence of an AE signal determines the occurrence of a deterministic component in a random

stream, and, as a result, the distribution function of time intervals between pulses in a signal deviates from

an exponential law [16]. Thus, changes in the characteristics of the distribution function, in particular, a

deviation of the distribution function of time intervals between pulses from an exponential law, is a crite

rion for the presence of a flawproduced signal. The dimensionless parameter α that characterizes the

degree of departure of the distribution function of time intervals between pulses from an exponential law

is calculated from the formula

It is known that the amplitude distribution in stationary random noise has a normal (Gaussian) distri

bution [15]. The occurrence of a deterministic component that is related to a source of a signal from a flaw

leads to deviations from a normal law. In order to estimate the statistical characteristics of an AE signal

and noise and to determine the degree of deviation from the normal distribution, the coefficients of asym

metry γ1 (5) and kurtosis γ2 (6) are introduced, since they are independent of the AEpulse stream rate:

where μ4 is the fourth central moment and σ is the standard deviation.

The results of calculating the coefficients of asymmetry γ1 and kurtosis γ2 are presented in Figs. 8a and

8b, respectively. It was revealed that an interval of delay values exists in which the coefficients for AE sig

nals substantially exceed the corresponding values of the coefficients for noise. No dependence of the

asymmetry and kurtosis coefficients for noise on the delay value exists, whereas such a dependence is

present for an AE signal. Thus, the parameters γ1 and γ2 can be used to identify a signal from a flaw and

the noise component. Taking the data into account (Fig. 6), this allows determination of the effective

interval of delay values at the AF input at which the difference of the statistical and correlation properties

of the signal and noise is maximized. Exactly these delay values were used during the filtering of experi

mental data (Fig. 4a).

The above-described method for identifying AE sources was used when performing engineering diag

nostics of hotpiping compensators at an operating oilprocessing plant. It is significant that in this case

The simultaneous use of the threshold and thresholdless AE testing systems made it possible to imple

ment the correct detection of AE signals, determine the coordinates, and evaluate the degree of hazard of

sources under the conditions of highamplitude  nonstationary operational noise and interference.

A detailed analysis of the AEtesting results in the presence of transient noise and the possibility of con

structing TDRbased monitoring systems is the subject of further investigation.

(1) It was shown that the application of the TDR method makes it possible to perform efficient AE test

ing under the conditions of highamplitude noise, when the use of existing threshold AEtesting systems

is impossible.

(2) To analyze the TDR AEtesting data under operating conditions in the presence of highamplitude

noise, it was proposed to use an adaptivefiltering algorithm with one input information channel (the

“blind”adaptation mechanism). It was established that the use of the presented filtering technique allows

the reliable detection of an AE signal from a flaw against a highamplitude noise background at an SNR

that is much lower than unity.

(3) It is shown that the optimal AF tuning parameters can be selected and the filteroperation effi

ciency can be evaluated on the basis of calculations of the signal and noise correlation, statistical, and

probabilistic properties. It was concluded that the optimal value of the time delay is determined during

comparative analysis of the dependences of the correlation intervals and highorder central moments of

filter output signals. On the basis of the estimate of the degree of distortion of a filtered AE signal relative

to the initial signal, which is described by the values of the crosscorrelation function (in the time domain)

and the coherence function (in the frequency domain), it was revealed that the used filter correctly recon

structs the shape and time positions of AE pulses.

(4) Quantitative criteria that are used as diagnostic indications of the presence of an AE signal from a

flaw were established: the crosscorrelation coefficient in the output stages of the twochannel adaptive

detector, the correlationinterval value, the asymmetry and kurtosis coefficients, and the degree of devia

tion of the distribution function of time intervals between pulses from an exponential law. The above

parameters can serve as the basis for constructing the system for making decisions on the identification of

flaws during AE testing of the engineering state of equipment under the conditions of highamplitude

noise.

3. Chebotareva, I.Ya., Structure and dynamics of a geomedium in noise seismic fields, the methods and experi

mental results, Ezhegod. RAO, 2011, no. 12, pp. 147–156.

4. Novikov, A.V., Application of nonlineardynamics methods for studying the relationship between the microseismic emission and the pore pressure, in Sovremennye problemy fundamental’nykh i prikladnykh nauk—aerofizika 

i kosmicheskie issledovaniya: Sbornik trudov 49i nauchnoi konferentsii MFTI (Modern Problems of Basic and

Applied Sciences—Aerophysics and Space Investigations: Proc. of 49th Scientific Conf. of Moscow Physicote

chnical Institute), Moscow, 2006, vol. III.

5. Karpov, O., Gabovich, A.G., Marchenko, B.G., Khoroshko, V.A., and Shcherbak, L.N., Komp’yuternye tekh

nologii raspoznavaniya rechevykh signalov (Computer Technologies for Recognizing Speech Signals), Kiev,

2005.

6. Chuchupal, V.Ya., Chichagov, A.S., and Makovkin, K.A., Tsifrovaya fil’tratsiya zashumlennykh rechevykh sig

nalov (Digital Filtering of Noisy Speech Signals), Vychislitel’nyi tsentr RAN: Moscow, 1998.

7. Aksel’rod, E.G., Davydova, D.G., and Kuz’min, A.N., A noiseimmune method for detecting useful signals in

systems for acousticemission monitoring of hazardous industrial objects, Zh. Tekhnadzor,  2013, no. 5(78),

pp. 86–89.

8. Diniz, P.S.R., Adaptive Filtering: Algorithms and Practical Implementation, Springer, 2013.

9. Haupt, H., Ein auf der Schallemissionsanalyse basierendes Verfahren zur Risserkennung in Umformprozessen:

Diss., pp. 14–194, Uni Paderborn, 2003.

10. Uidrou, B. and Stirnz, S.,  Adaptivnaya obrabotka signalov (Adaptive Signal Processing), Moscow: Radio i

svyaz', 1989.

11. Shaibakov, R.A., Davydova, D.G., Kuz’min, A.N., Abdrakhmanov, N.Kh., and Markov, A.G., A noiseimmune method of acousticemission monitoring of reservoirs, Neftegaz. Delo: Elektron. Nauch. Zh., 2013, no. 4, 

pp. 448–464.

12. Kuz’min, A.N., Zhukov, A.V., and Lukin, V.A., Experimental modeling of destruction processes on defective

metal structures, V Mire Nerazrush. Kontr.: Ezhekvart. Zhurn. Obozr., 2011, no. 1(51), pp. 13–21.

13. Kantz, H. and Schreiber, T., Nonlinear Time Series Analysis, Cambridge: University Press, 2003.

14. Tikhonov, V.I., Statisticheskaya radiotekhnika (Statistical Radio Engineering), Moscow: Radio i svyaz’, 1982.

15. Baranov, V.M., Gritsenko, A.I., and Karasevich, A.M., Akusticheskaya diagnostika i kontrol' na predpriyatiyakh toplivnoenergeticheskogo kompleksa (Acoustic Diagnostics and Inspection at Plants of Fuel and Energy Com 

plex), Moscow: Nauka, 1998.

16. Braginskii, A.P., Vinogradov, A.Yu., and Leksovskii, A.M., Acousticemission amplitudefrequency analysis of the deformation kinetics of amorphous metallic glasses, Pis’ma Zh. Tekh. Fiz., 1986, vol. 12, no. 18, pp. 1111–1114.