Location of two simultaneously active continuous acoustic emission sources on an aluminum beam

Tadej Kosel and Igor Grabec

Introduction
A common goal of all non-destructive testing methods is to discover defects in materials. Acoustic emission analysis (AE) is a passive testing method which is used to locate and characterize defects which emit sound (McIntire and Miller, 1987). There are many ways to deduce the location of an AE source from electrical signals detected by a chain of sensors. The corresponding problems may be classied by the type of acoustic source mechanism as the location of a continuous emission source, such as that generated by a leak, or as the location of discrete emission, such as the AE from a growing crack. This paper describes a method for processing AE signals to determine the time delays between signals and thus the location of AE sources. It should be pointed out that source AE parameters, such as count, count rate, amplitude distribution, and conventional time delay measurement, become meaningless when dealing with continuous acoustic sources such as leaks. The basic information for AE source location consists of time delays between stress waves detected at different positions on a specimen. In the case of only one active AE source, time delays of continuous acoustic waves can be estimated using the crosscorrelation function of sensor signals (McIntire and Miller, 1987; Grabec, 1978; Kosel et al., 2000). In the case of two (or more) simultaneously active AE sources, this method is not applicable, since analysis of the cross-correlation function leads only to the time delay of the most powerful AE signal. Detection of simultaneously active independent AE source signals therefore requires a more sophisticated approach. The purpose of our study was to nd a suitable method for processing a mixture of two simultaneously active continuous AE signals to determine the time delays and, related to this, the coordinates of both AE sources. We found that the blind source separation (BSS) method solves this problem satisfactorily. Blind source separation is a general signal processing method consisting of recovering the contributions of different sources from a nite set of observations recorded by sensors, independent of the propagation medium and without any prior knowledge of the sources. BSS has already been successfully applied in 4

The authors Tadej Kosel and Igor Grabec are at the Faculty of Mechanical Engineering, University of Ljubljana, PO Box 394, SI-1000 Lubljana, Slovenia Keywords Aluminum, Defects, Fault analysis Abstract Acoustic emission analysis (AE) is used for characterization and location of developing defects in materials. AE sources often generate a mixture of various statistically independent signals. One difcult problem of AE analysis is the separation and characterization of signal components when the signals from various sources and the way in which the signals were mixed are unknown. Recently, blind source separation (BSS) by independent component analysis (ICA) has been used to solve these problems. The main purpose of this paper is to demonstrate the applicability of ICA to time-delay estimation of two independent continuous AE sources on an aluminum beam. It is shown that it is possible to estimate time delays by ICA, and thus to locate two independent simultaneously emitted sources. Electronic access The current issue and full text archive of this journal is available at http://www/emeraldinsight.com/0002-2667.htm

Aircraft Engineering and Aerospace Technology Volume 74 Number 1 2002 pp. 48 q MCB UP Limited ISSN 0002-2667 DOI 10.1108/00022660210414967

Two simultaneously active continuous acoustic emission sources

Aircraft Engineering and Aerospace Technology Volume 74 Number 1 2002 48

Tadej Kosel and Igor Grabec

medicine, telecommunications, image processing, etc. (Lee, 1998). However, it is also a promising method for AE analysis, because AE signals in an industrial environment are often disrupted by multiple signals from various sources. BSS could extract the specic signature of each AE source, which can be further be used for location and characterization purposes or to isolate AE sources from background noise. We conducted experiments with BSS on an aluminum beam on which two continuous AE sources were generated simultaneously by air ow.

Methods
In this section we explain two different methods for time-delay estimation of AE sources. The rst method is based on analysis of the cross-correlation function and is convenient for time-delay estimation of continuous AE sources (McIntire and Miller, 1987; Kosel et al., 2000; Ziola and Gorman, 1991). This method is only applicable for one active AE source. In the case of two (or more) simultaneously active AE sources, a different method has to be used based on blind source separation algorithms (Lee et al., 1997). For AE source location, the time delays of each signal regarding sensor pairs are required. Location of two simultaneously active AE sources was performed by an adaptive locator based on a general regression neural network (Grabec and Sachse, 1997). The crosscorrelation function (CCF) exhibits a peak when the delay parameter t compensates the time delay Dt between the sensor signals x1(t ) and x2(t ) (McIntire and Miller, 1987). The time delay is thus determined by the position of the highest peak of the CCF. The method was successfully used for time delay estimation of one continuous AE signal (Kosel et al., 2000). Multichannel Blind Source Separation has recently received increased attention due to the importance of its potential applications (Burel, 1992). It occurs in many elds of engineering and applied sciences, including processing of signals from antenna array processing, speech and geophysical data processing, noise reduction, biological system analysis, etc. It consists of recovering signals emitted by unknown sources and mixed by an unknown medium, using only several 5

observations of the mixtures. The only assumptions made are the linearity of the mixing system and the statistical independence of original signals. BSS methods may be classied in several ways. One possible classication that can be made depends on whether the mixtures are instantaneous or convolutive (Charkani and Deville, 1999a,b). Convolutive mixtures correspond to a mixing system with time dependent memory. They represent a more general case than instantaneous mixtures, and they have in particular acoustic applications. Recently, the principle of independent component analysis (ICA) was applied in BSS, and it was found to be a simple and powerful tool (Hyvarinen and Oja, 2000). This study deals with the separation of two convolutively mixed independent continuous acoustic emission signals by ICA. The mixing and ltering processes of the unknown input signals sj(t ) may have different mathematical or physical backgrounds, depending on specic applications. In this paper, we focus mainly on the simplest cases with n signals xi(t ) linearly composed in n unknown statistically independent, zero mean source signals sj(t ). The composition is expressed in matrix notation as (Lee, 1998) x = A*s, where * is convolutive multiplication, x x1 t; . . .; xn tT is the vector of sensor signals, s s1 t; . . .; sn tT is the vector of source signals and A is an unknown full rank nxn mixing matrix whose elements are FIR lters. We assume that only vector x is available. The goal of ICA is to nd a matrix W, by which vector x can be transmitted into the source signals u = W*x. Matrix W is simply the inverse of A. However, when noise corrupts the signals, matrix W must be found by optimal statistical treatment of the inverse problem. The optimal matrix W can be estimated by a feed-forward neural network operating in the frequency domain. A learning algorithm with Amaris natural gradient can be written as (Amari, 1998): ~ ~ u Wzx ~ 1 ~ ~ ~ ~ Wt 1 Wt aDWt hDWt 2 1; 2 ~ ~ DW I 2 yzu H W; ~ ~ y tanh?u itanh?u; ~ ~ ~ 3 4

Two simultaneously active continuous acoustic emission sources

Aircraft Engineering and Aerospace Technology Volume 74 Number 1 2002 48

Tadej Kosel and Igor Grabec

where a is the learning rate, h is the moment of learning, I is the identity matrix and the tilde , represents a frequency domain. The location problem is usually solved by various triangulation techniques based on mathematical analysis of ultrasonic ray trajectories. This analysis is rather cumbersome and cannot be readily performed if the form of the specimen is geometrically complicated. In order to avoid this difculty, several empirical approaches based on learning from examples have already been proposed (Grabec and Sachse, 1997). An instrument that is able to learn from examples is called an adaptive locator. It replaces information obtained by analysis of sonic ray trajectories with information provided by simulated AE events on the specimen being tested. The standard calibration procedure, which has to be performed in any event, could be used to train the adaptive locator. In our case, the adaptive locator was used to locate two independent continuous AE sources from time delays.

Experiments
We performed the experiment with two independent continuous AE sources on an aluminum band of dimensions 4000 40 5mm3. Reections at the end of the band were reduced by wrapping the ends in putty. The testing area was on the longitudinal axis in the middle of the band, were 23 holes of diameter 2mm and mutual separation 100mm were prepared. Two AE sensors were mounted 100mm away from the terminal holes, that 2.4m from each other. The origin of the coordinate system was in the middle of the band and the testing area extended from 2 1.1m to +1.1m. AE signals were excited by two independent air jets owing through the holes. The source position was arbitrarily selected at +100mm and +800mm. Air jets were formed by two nozzles with hole diameter of 1mm and pressure 7 bar. The experimental set-up consisted of the test specimen, two AE sensors (pinducers), two AE actuators (air jets), two pre-ampliers, two ampliers, a digital oscilloscope and a computer. We performed three experiments: (1) time-delay estimation using a cross-correlation of two AE signals that were not simultaneously active; (2) timedelay estimation using a cross-correlation 6

function of two AE signals which were simultaneously active and (3) time-delay estimation of AE signals using ICA. Location of sources, based on time delays, by the adaptive locator was performed in all three cases. In the rst experiment only one air jet was activated for a particular measurement. In the second experiment both air jets were activated. Sensor signals were linear convolutive mixtures of two independent continuous AE sources. The auto- R11, R22 and cross-correlation functions R12, R21 were calculated from sensor signals. Only one time delay of two signals can be estimated from the highest peak in both cross-correlation functions, regardless of the number of independent AE sources on the test specimen. This means that a cross-correlation function cannot be used for automatic time-delay estimation of multiple AE signals on the test specimen. The cross-correlation function exhibits various peaks which belong to various sources, but it is nearly impossible to relate these peaks to corresponding coordinates of AE sources. In the third experiment the ICA algorithm was used to solve this problem satisfactorily. The ICA algorithm results in demixing FIR lters which extract the independent source signals from sensory signals. By inverting the demixing lters W we obtain mixing lters A. In the case of two independent AE sources and two sensors, the components of A are four FIR mixing lters, as shown in Figure 1. There are two direct a11, a22 and two cross mixing lters a12, a21. The rst index of the lter represents the number of the sensor, while the second index represents the number of the source. The position of the highest peak of the cross FIR lters determine the time delay between two signals from two sensors. If we subtract the position of the highest peak of a direct mixing FIR lter a11 from the position of the highest peak of cross lter a21 we obtain the time delay of rst independent AE source, since each of the highest peaks in the FIR lters belongs to different independent AE signals and sensors.

Results
The results of time-delay estimation of two continuous independent AE sources are shown in Figure 2. Three experiments were

Two simultaneously active continuous acoustic emission sources

Aircraft Engineering and Aerospace Technology Volume 74 Number 1 2002 48

Tadej Kosel and Igor Grabec

Figure 1 Mixing lters obtained by ICA of sensory signals

done. In the rst experiment, the time delay was estimated by a cross-correlation function of two AE sources which were not active simultaneously as marked by O. Locations of these two sources estimated by the adaptive locator were +181mm and +784mm. The second experiment was performed with both AE sources active simultaneously. Time delays were also estimated by a crosscorrelation function. The highest peak position corresponds to the source location marked by . Estimated position by adaptive locator of this source was +784mm. The third experiment was performed using ICA for time-delay estimation and location by adaptive locator. The result is marked by A. Estimated positions of these two sources were +179mm and +784mm respectively. If we compare the locations of both independent AE sources estimated by the rst experiment and by the third experiment, we nd a good correspondence. If we compare estimated AE source coordinates with actual coordinates, which were +100mm and +800mm respectively, we observe a slight disagreement due to experimental error. Experimental error is about 3 percent regarding the distance between sensors. Absolute error in this case is 79mm and 16mm respectively. The results also depend on the number and distribution of prototype sources marked by , which are essential for operator of the locator. If the number of prototypes is increased, location error is reduced. In our case the prototype sources were distributed along the beam from 2 1.1m to +1.1m separated by 0.1m, so that systematic error of the locator was set to an acceptable level. More comprehensive
.

experiments with BSS and corresponding analysis are presented in Kosel (2000).

Discussion and conclusions

The cross-correlation function is applicable to time-delay estimation only in the case of one active AE source. The goal of our research is to develop a new method to estimate time delays between AE signals in the case of multiple simultaneously active continuous AE sources. We have shown that, for this purpose, independent components analysis is an applicable option. It represents a new approach to processing of AE data and further expands the applicability of AE analysis in the eld of non-destructive testing. In an
Figure 2 Results of location of two continuous independent AE sources. Symbols: AAE sources obtained by ICA; Oestimated AE sources obtained by cross-correlation function in two steps, when just one of two AE sources was active at time of measurement; estimated AE sources obtained by cross-correlation function when two AE sources were active simultaneously; .prototype AE sources required for location using adaptive locator; distribution of actual sources

Two simultaneously active continuous acoustic emission sources

Aircraft Engineering and Aerospace Technology Volume 74 Number 1 2002 48

Tadej Kosel and Igor Grabec

industrial environment multiple sources are usually active simultaneously and often represent environmental disturbances. The corresponding complex signals are not directly applicable to characterization of particular sources. However, separation of contributions by ICA analysis in fact represents a kind of ltering, increasing the applicability of ltered signals to characterization of sources in complex environments.

References
Amari, S.-I. (1998) Natural gradient works efciently in learning, Neural Computation, Vol. 10, pp. 251-76. Burel, G. (1992) Blind separation of sources: a nonlinear algorithm, Neural Networks, Vol. 5, pp. 937-47. Charkani, N. and Deville, Y. (1999a) Self-adaptive separation of convolutively mixed signals with a recursive structure. Part I: Stability analysis and optimization of asymptotic behaviour, Signal Processing, Vol. 73, pp. 225-54. Charkani, N. and Deville, Y. (1999b) Self-adaptive separation of convolutively mixed signals with a recursive structure. Part II: Theoretical extensions and application to synthetic and real signals, Signal Processing, Vol. 75, pp. 117-40.

Grabec, I. (1978) Application of correlation techniques for localization of acoustic emission sources, Ultrasonics, Vol. 16 No. 5, pp. 111-15. Grabec, I. and Sachse, W. (1997), Synergetics of Measurement, Prediction and Control, SpringerVerlag, Berlin. Hyvarinen, A. and Oja, E. (2000) Independent component analysis: algorithms and applications, Neural Networks, Vol. 13, pp. 411-30. Kosel, T. (2000), Blind separation of independent acoustic emission sources, PhD Thesis, Faculty of Mechanical Engineering, University of Ljubljana. Kosel, T., Grabec, I. and Muzic,, P. (2000) Location of continuous acoustic emission sources generated by air ow, Ultrasonics, Vol. 38 No. 1 8, pp. 924-6. Lee, T.-W. (1998), Independent Component Analysis, Theory and Applications, Kluwer Academic Publishers, Boston, MA. Lee, T.-W., Bell, A.J. and Lambert, R. (1997) Blind separation of convolved and delayed sources, Advances in Neural Information Processing Systems, Vol. 9, pp. 758-64. McIntire, P. and Miller, R.K. (1987), Acoustic Emission Testing, Vol. 5 of Non destructive Testing Handbook, 2nd, American Society for Non-destructive Testing, Philadelphia, USA. Ziola, S.M. and Gorman, M.R. (1991) Source location in thin plates using cross-correlation, J. Acoust. Soc. Am., Vol. 90 No. 5, pp. 2551-6.