Mechanical Systems and Signal Processing (1997) 12(6), 783–795

Article No. pg980163

IDENTIFICATION OF THE IMPACT LOCATION ON

A PLATE USING WAVELETS
L. G  S. H
Institute A of Mechanics, University of Stuttgart, Allmandring 5 b, 70550 Stuttgart,
Germany

(Received March 1998 , accepted June 1998)

This paper presents an experimental method that has been developed for detecting
flexural waves in plates by the use of piezoelectric films. The recorded signals are analysed
by a ‘wavelet transform’ to determine arrival times of waves at different frequencies. These
arrival times are used in conjunction with an optimisation method to identify the location
of the impact (i.e. the x- and y-coordinate of the source location). Also, the time lag
between the instant of the impact and the recording of the signals together with the group
velocity of the dispersive waves are specified. Close agreement is observed between the
experimentally determined results and the exact solutions. This technique is a valuable
method for identification of source locations that may be applied to problems related to
in-service impact loading of structures, seismic data from earthquakes, and to acoustic
emission signals from propagating cracks.
7 1998 Academic Press

1. INTRODUCTION
Large delaminations in composite materials are generally attributed to impact loadings.
These delaminations can induce changes in the stiffness of a structure leading to unsafe
conditions. It is not always possible to identify the damaged region of a structure
particularly if such damage occurs in regions which are inaccessible to visual inspection.
Many non-destructive evaluation methods (NDE), however, exist for determining the size
of delaminations; these include ultrasonic inspection and lock-in thermography. A
disadvantage of some NDE methods is that the structure needs to be out of service to
perform the NDE tests. Therefore, special NDE methods have been developed for
so-called ‘structural health monitoring’, which consists of two major components: passive
sensing monitoring and active sensing monitoring. The active sensing monitoring needs
both sensors and actuators to evaluate the damage of a structure. The passive sensing
monitoring uses sensors which only ‘listen’ to signals emitted by internal or external
sources such as propagating cracks or impacts. This research deals with passive sensing
monitoring in order to determine the location of an impact. It is important to establish
the location of the impact for determining the force–time history of the impact. This
information can be compared with previously recorded measurements for the same
material and geometric characteristic. The differences can be attributed to the ‘damage’
in the structure.
Extensive studies have been conducted on the analysis of dispersive waves in structures.
Doyle [1] has presented a method for determining the location of dispersive pulses by

0888–3270/97/060783 + 13 $30.00/0 7 1997 Academic Press

784 .   . 
applying the phase information from spectral analysis. The approach in the work of Doyle,
however, is restricted only to beams. Choi and Chang [2] have suggested an impact load
identification technique using piezoelectric sensors. They used a structural model and a
response comparator for solving the inverse problem. The structural model characterises
the relation between the input load and the sensor output. The response comparator
compares the measured sensor signals with the predicted model. An extension to this work
is given by Tracy and Chang [3], who extended their work not only to beams but also
composite plates. Their paper presents a computer code which automatically identifies the
impact load and location. In this approach exact information of the geometry and material
properties of the plate are needed. Often such data are not available, due to effects such
as aging of the material and/or deformation of the plate.
Only a few studies deal with dispersive waves in structures in conjunction with the
time–frequency analysis. Newland [4] described three different methods of time–frequency
analysis for vibration problems: the short-time Fourier transform (STFT), the
Wigner–Ville distribution (WVD), and the wavelet transform (WT). Kishimoto et al. [5]
developed a tool for determining the impact location of a beam using the WT. Here again,
they restricted their analysis only to beams. Inoue et al. [6] validated the method suggested
by Kishimoto et al. experimentally by impacting a simply supported beam with a sphere.
They used strain gauges for measuring the strain caused by flexural waves.
This paper extends the approach previously developed for beams to include plates.
Advantages of the proposed techniques are the accuracy of the results and that the exact
parameters concerning material and the geometry are not needed. First, a brief
introduction and a comparison of the three different time–frequency analyses are given and
the WT is described in more detail. The application of the WT to wave propagation
problems is also demonstrated. The experimental set-up of measuring flexural waves in
plates using piezoelectric films is then documented. The signals for the flexural waves are
then used to compute the time–frequency analysis, using the WT. Finally, a method is
shown for calculating (i) the coordinates of the impact location, (ii) the time lag between
the arrival of the impactor and the time of triggering of the signal, and (iii) the dispersion
relation of the group velocity. The application of the procedure to the study of a plate
is challenging because of the non-linear relationships between the group velocity, the time
lag and the coordinates.

2. TIME–FREQUENCY-ANALYSIS
In this section an introduction to three different methods of time–frequency analysis,
namely, the STFT, the WVD and the WT, are given and their special features as well as
advantages and disadvantages are discussed.

2.1. -   ()

The STFT uses a constant time window, and this short time window of the original
signal is transformed into the frequency domain. The window is then shifted to a new
position and the transform is repeated with a different part of the signal. Mathematically,
the STFT can be expressed as a windowed Fourier transform [7] and is a function of the
frequency v and the position b:

g
a
1
F(v, b) = f(t)g(t − b) e−ivt dt. (1)
2p −a
      785
This is the Fourier transform of function f(t) windowed by g(t) for all b. Hence one can
obtain a time–frequency plot of the entire signal. The disadvantage of the STFT is that
it is not possible to obtain a high resolution in the time domain and in the frequency
domain. The reason for this drawback is that the selected time window has a constant size
for all frequencies. Many signals require a more flexible approach where the window size
can be varied to determine more accurately either the time or the frequency.

2.2. –  ()

An alternative method is based on the concept of a correlation function. An
autocorrelation is a correlation of the signal with itself and gives an energy density which
allows determination of some properties of the signal. The WVD of a function f(t) is a
combination of the autocorrelation function and the Fourier transform defined as

g 0 10 1
a
t t −ivt
W(v, t) = f t+ f t− e dt, (2)
−a
2 2

where W is quadratic in f. The complex conjugate f is introduced for the purpose of

generalising the analysis to complex signals. The advantage of the WVD is the facility for
obtaining, simultaneously, high resolution in the time and frequency domains. However,
the WVD has three fundamental shortcomings: firstly, the procedure is computing
intensive; second, W(v, t) is not always positive, which is unrealistic from a physical point
of view; and finally, W(v, t) introduces cross-terms when signals with multifrequency
components are analysed [8]. These drawbacks are not present in the time–frequency
analysis scheme which uses the wavelet transform.

2.3.   ()

The continuous wavelet transform of an arbitrary function f(t) is defined by (see e.g.
Strang and Nguyen [7]):

g 0 1
a
1 t−b
FC (a, b) = f(t)C dt, (3)
za −a
a

where

0 1
1 t−b
Ca,b (t) = C (4)
za a

are the wavelet functions. The position variable b is the same as in the STFT. The difference
consists of using a scale variable a in the WT instead of a frequency variable in the STFT.
This scaling is a primary characteristic of the wavelet analysis. In this paper, the relation
between the scale variable a and the frequency v is v = v0 /a, where v0 is a positive
constant. The function C(t) is termed as the analysing function. It may be considered as
a window function both in the time and frequency domain. The size of the time window
is controlled by the translation, while the length of the frequency band is controlled by
the dilation. It is possible to change the window size either in the time or in the frequency
domain by controlling dilation and translation. This property of the WT is called
multiresolution [9]. As has been suggested by Kishimoto et al. [5], the Gabor function is
786 .   . 
used as the analysing wavelet because it provides a small window in the time as well as
in the frequency domain. The Gabor function is defined by

Cg (t) =
1
zp
4 X v0
g
exp − 0
2$
(v /g)2 2
t + iv0 t .
% (5)

Following [5], the positive constants g and v0 are set as g = pz2/ln 2 and v0 = 2p,
respectively. The Gabor function can be assumed as a complex sinusoidal function with
a Gaussian window as shown in Fig. 1.
Using Parseval’s theorem [7] the WT can be rearranged as

g
a
za
FC (a, b) = f
(v) eibvC
(av) dv, (6)
2p −a

where

g
a

f
(v) = f(t) e−ivt dt (7)
−a

is the Fourier transform of the function f(t). The Gabor function in the frequency domain
can be expressed by

C
g (v) =
2p
zp
4 X g
v0
exp −
$
(g/v0 )2
2
(v − v0 )2 .
% (8)

3. WAVELET ANALYSIS OF WAVES

It is important to have a knowledge of the relationship between wave propagation and
the wavelet analysis in order to determine the arrival times of the flexural waves. In general,
a wave propagating in a plane can be specified by the wavelength l, and the angle between
the x1 axes and the propagation direction, U. Hence, one can obtain the apparent
wavelengths in the x1 and x2 directions defined by l1 = l/cos U and l2 = l/sin U,
respectively (Fig. 2).

Figure 1. Gabor function in the time (a) and in the frequency domains (b). ——, Real; · · · · , imaginary.
      787

Figure 2. Wave propagation in a plane.

Assuming two waves travelling in a plane in the same direction with the same amplitude
and with the frequencies v1 and v2 , respectively, this can be expressed mathematically as
u(x1 , x2 , t) = ei(v1 t − k11 x1 − k12 x2 ) + ei(v2 t − k21 x1 − k22 x2 ), (9)
where kji are wave numbers corresponding to the frequency vj and to the coordinate xi .
In general, the wavenumber is defined as k = 2p/l. Applying this definition leads to the
wavenumbers kj1 = kj cos U and kj2 = kj sin U and therefore the wave equation can be
transformed as
u(x1 , x2 , t) = ei(v1 t − k1 cos Ux1 − k1 sin Ux2 ) + ei(v2 t − k2 cos Ux1 − k2 sin Ux2 ). (10)
Defining kc = (k1 + k2 )/2, vc = (v1 + v2 )/2, Dk = (k1 − k2 )/2 and Dvc = (v1 − v2 )/2, the
wave equation can be rewritten as
u(x1 , x2 , t) = 2 cos (Dvt − Dk cos Ux1 − Dk sin Ux2 ) ei(vc t − kc cos Ux1 − kc sin Ux2 ). (11)
The phase velocity cp and the group velocity cg of flexural waves are given by
cp = vc /kc and cg = Dv/Dk, (12)
respectively, with the assumption that Dk is sufficiently small [10].
By discrete Fourier transform (DFT) and discrete wavelet transform (DWT) of
u(x1 , x2 , t), the following expression is obtained:
F(x1 , x2 , a, b) = za [ei(v1 b − k1 cos Ux1 − k1 sin Ux2 )C
(a v1 )
+ ei(v2 b − k2 cos Ux1 − k2 sin Ux2 )C
(a v2 )]. (13)
Multiplying the complex conjugate of F(x1 , x2 , a, b) with F(x1 , x2 , a, b) leads to the
magnitude of the wavelet transform under the assumption that C
(v) is real: i.e.

=F(x1 , x2 , a, b)= = za [C
(av1 )2 + C
(av2 )2
1
+ 2C
(av1 )C
(av2 ) cos (2Dvb − 2Dk cos Ux1 − 2Dk sin Ux2 )]2.
(14)
Two different cases have to be considered now: where Dk is large and where Dk is small.
788 .   . 
3.1.  
If Dk is large, the two frequencies v1 and v2 are unequal or distinctly different. The
Gabor wavelet has its maximum value when v = v0 /a. Hence, the peaks of the Gabor
wavelets are at different locations and therefore C
(av1 )C
(av2 ) vanishes. The magnitude
of the wavelet transform is then obtained by
1
=F(x1 , x2 , a, b)= 1 za [C
(av1 )2 + C
(av2 )2]2. (15)
It is easily seen that the maximum of equation (15) does not depend on b. It can be
concluded that the peaks of the wavelet transform, which are located at a1 = v0 /v1 and
a2 = v0 /v2 , represent stationary vibration if the two wave numbers are dissimilar.

3.2.  

In the second case Dk is considered to be small. Then k1 1 k2 , v1 1 v2 and
C
(av1 ) 1 C
(av2 ) 1 C
(avc ), the magnitude of the wavelet transform may be expressed as
1
=F(x1 , x2 , a, b)= 1 z2a =C
(avc )=[1 + 2 cos (2Dvb − 2Dk cos Ux1 − 2Dk sin Ux2 )]2. (16)
If
2Dvb − 2Dk cos Ux1 − 2Dk sin Ux2 = 0 and a = v0 /vc , (17)
the magnitude of the wavelet transform has its maximum value. This is the case when
Dk 1
b= (cos Ux1 + sin Ux2 ) = (cos U x1 + sin Ux2 ). (18)
Dv cg
The peak of the magnitude of the WT indicates the arrival time of the flexural wave with
the group velocity cg and the wave number kc .

4. EXPERIMENTAL SET-UP AND PROCEDURE

The experimental set-up consists of a suspended plate which is impacted by a pendulum
at a specified location (x0 and y0 ). The mode of support of the plate is not critical here
since the considered time window is relatively short (3 ms) relative to the travel time of
the incident wave. Therefore, the dominant part of the signal consists of the initial flexural
waves which are propagating from the location of the impact to the sensors. The side length
of the square plate is 2 m and the thickness is h = 10 mm. The material of the plate as
well as the material of the tip of the pendulum is steel (Young’s modulus
E = 210 000 N/mm2, Poisson’s ratio n = 0.3, mass density r = 7860 kg/m3).
In the experiment, four piezoelectric films (thickness 110 mm, diameter 15 mm) are used
and they are located at each corner of a square as shown in Fig. 3. The advantage of the
piezoelectric film is that this sensor has the property of multidirectional sensitivity as
opposed to strain gauges and they can be used without any strain amplifiers. The
oscilloscope which records the signals is triggered by one of the sensor signals and the data
are sampled at 1 ms. Figure 4 shows the measured strain history for each sensor. It is clear
that reflections from the edges of the plate are present in the data. The start of the trace
is the classic form of a flexural wave generated by a broadband source, the
highest-frequency components arriving first, followed by a long sequence of slow,
lower-frequency components. The higher-frequency components reflect from the plate
edges before the lower-frequency components arrive.
It has been established that the strain in the plate is proportional to the output voltage
of the piezoelectric film. In this research, the exact knowledge of the proportionality factor
between the output voltage and the strain is not required, since there is no need to establish
      789
the absolute magnitude of the wavelet transform. As a result, the strain can be given in
an unfactored, dimensionless form.

5. WAVELET TRANSFORM OF STRAIN HISTORY SIGNALS

The signal processing procedure can proceed once the experimental data are saved on
a workstation. It is important to window the signals before applying the Fourier transform.
In this research, a Hanning window gave best results. However, the maximum of the
Hanning window is at the half of the input sequence length of the signal whereas the
interesting part of signal might not be at the same place as the maximum of the Hanning
window. Therefore by using a common Hanning window the interesting part of the signal
may be reduced in amplitude whereas the less interesting part of the signal may be
unchanged. To avoid this problem, zeros are added at the beginning or at the end of the
signal in order to ‘shift’ the interesting part of the signal to the corresponding maximum
range of the Hanning window.
The signals are then transformed into the frequency domain using the fast Fourier
transform (FFT) algorithm. The discrete form of equation (6) leads to the wavelet domain
of the four signals of the piezoelectric films. The parameters used in this transform are
a = 2m/4 and b = nDt, (19)
where m and n are integers. Newland [8] suggested the use of the wavelet transform starting
in the time domain, transforming into the frequency domain and finally transforming into
the wavelet domain. This is preferred as opposed to a direct transformation of the time
domain signal into the wavelet domain as shown in equation (3). The primary advantage
is the large decrease in computational time. Figure 5 is the three-dimensional plot of the
magnitude of the wavelet level of sensor 1. The maximum of the plot indicates the arrival
time of the flexural wave. Figure 6 shows the contour plot of the time–frequency analysis
using the WT for each sensor signal. As described in the previous section, the

Figure 3. Experimental set-up (dimensions in mm).

790 .   . 

Figure 4. Strain histories of sensors 1 to 4 [(a)–(d), respectively].

higher-frequency components reflect from the plate edges before the lower frequencies have
arrived. The wavelet analysis resolves the signal in time as well as in frequency and
therefore, the early reflected waves containing the high-frequency components are
separated from the incident waves containing the lower-frequency components.
Furthermore, the incident waves are larger in amplitude than the reflected ones and the
reflected waves need not be considered further in the procedure.

6
Magnitude (∞ 10–3)

0
2.5
2 –11
1.5 –12
Tim 1 –13
e 0.5 –14
2 π /ω
(ms )
–15
) 0 –16 log 2 (
Figure 5. Three-dimensional plot of the wavelet transform.
      791

Figure 6. Contour plots of the time–frequency analysis including the sensors location.

6. DETERMINING THE IMPACT LOCATION AND THE GROUP VELOCITY OF

FLEXURAL WAVES
In this section, the impact location and the group velocity of flexural waves are
determined from the time–frequency analysis. As has been shown in a previous section,
the maximum of the magnitude of the wavelet transform indicates the arrival time of the
wave with the group velocity cg for small Dk. In order to obtain the arrival times, the
maximum of the magnitude of the WT has to be determined for each frequency scale. It
is evident that the incident waves are larger in amplitude than the reflected waves. This
fact is used to obtain the maximum of the incident waves. It is concluded that the
maximum of the wavelet transform denotes the arrival time of the incident wave. Figure
7 shows the location of the maxima in the time–frequency plot. It is now evident that the
arrival time of small values of log2 (2p/v) (high frequencies) is smaller than that of large
values of log2 (2p/v) (low frequencies).
Once the arrival times are known, one has to determine the coordinates of the location
of the impact, the time lag and the group velocity. The distances between sensor i and the
impact location is li . The corresponding arrival times are given by bi and the unknown time
lag is b0 . The unknown coordinates of the impact are defined as x and y, respectively, while
the side length l of the square configuration of the sensors is known. The relations between
these magnitudes can be expressed as follows:

l1 = zx 2 + y 2 (20)

l2 = zx 2 + (l − y)2 (21)
792 .   . 
l3 = z(l − x)2 + (l − y)2 (22)

l4 = z(l − x)2 + y 2 (23)

bi (a) = b0 (a) + li /cg (a) for i = 1, 2, 3, 4. (24)

This is a system of eight equations for the eight unknowns b0 , cg , x, y, l1 , l2 , l3 and l4 . By

rearranging one obtains the following system of four non-linear equations:

zx 2 + y 2
b1 (a) = b0 (a) +
cg (a)

zx 2 + (l − y)2
b2 (a) = b0 (a) +
cg (a)

z(l − x)2 + (l − y)2

b3 (a) = b0 (a) +
cg (a)

z(l − x)2 + y 2
b4 (a) = b0 (a) + . (25)
cg (a)

In practice, it is not necessary to solve this system of equations analytically. However, an

optimisation method can be applied to solve this problem. The adopted method is

Figure 7. Arrival times of the flexural waves of sensors 1 to 4 [(a)–(d), respectively].

      793
described in more detail elsewhere [11]. Briefly, the set of equations (25) has to be
transformed into the form
f(x) = 0, (26)
where x is the vector of the four unknown components x = (x, y, b0 , cg ). This procedure
is based on the non-linear least squares optimisation using the Gauss–Newton method
which is defined [11] as
min >J(xj )dj − f(xj )>22, (27)
where J(xj ) is the Jacobian matrix of f(xj ) and dj is the search direction. The search
direction dj is obtained for each major iteration, j, and can be used as part of a line search
strategy to ensure that at each iteration expression in equation (27) decreases.
For this optimisation method an initial value x0 is necessary. One can obtain the initial
value by a rough estimate of the group velocity and compute analytically the initial values
for the coordinates and the time lag. The estimation of the group velocity is obtained by
implementation of an analytical relation between the group velocity and the material data
and geometry of the plate. Classical plate theory taking into account the shear deformation
leads to the phase velocity of the flexural wave

$ %
−1
Eh 2 Eh 2
cp2 = k 1+ k (28)
12(1 − n )r
2
12k(1 − n 2)r

where n is Poisson’s ratio, h is the thickness of the plate, E is Young’s modulus, r is the
density of the material and k denotes the wave number [10]. The constant k denotes the
ratio of the Rayleigh wave velocity cR to the shear wave velocity cs and can be
approximately calculated using Poisson’s ratio n as
0.87 + 1.12n
k= . (29)
1+n
It is interesting to notice that the shear deformation accounts almost entirely for the
discrepancy between classical plate theory and the three-dimensional theory over the whole
wavelength spectrum.
The group velocity is related to the phase velocity by
dcp
cg = cp + k . (30)
dk
By incorporating this information in the system of four non-linear equations one reduces
the system to two non-linear equations for x and y by eliminating the time lag b0 : i.e.
zx 2 + (l − y)2 zx 2 + y 2
b2 (a) = b1 (a) + −
cg (a) cg (a)
z(l − x)2 + y 2 zx 2 + y 2
b4 (a) = b1 (a) + − . (31)
cg (a) cg (a)
In principle, this can be solved analytically. However, the results are not given here because
the resulting equations are still complicated.
By using the initial values on the optimisation method one can calculate the coordinates
of the impact location, as well as the time lag and the group velocity of the dispersive
waves. Figure 8(a) and (b) represent the values of x and y for different frequency scales,
respectively. A comparison of these values with the actual values of the impact location
794 .   . 

Figure 8. (a) and (b) Identified (w) and actual (——) x and y coordinates, respectively, of the impact location;
(c) identified (w) and actual (——) time lag; and (d) theoretical (——) and identified (w) group velocity.

indicates that all values of the coordinates of x and y which are obtained by the proposed
approach are accurate to within 10%. However, averaging the identified values of x and
y over the considered frequency scales by arithmetic mean leads to a coordinate in the
x-direction of 0.6092 m, and 0.4989 m in the y-direction. Compared to the actual values
(x0 = 0.6 m, y0 = 0.5 m) these are relatively accurate results. Figure 8(c) and (d) show the
time lag and the group velocity for different frequency scales. The identified group velocity
is compared with the analytically described group velocity. These velocities correspond
well in the lower-frequency range (larger log2 (2p/v)). It is clear, that the identified group
velocity in the higher-frequency range [smaller log2 (2p/v)] is lower than the analytically
obtained one, but their shapes are similar.

7. CONCLUSION
This paper develops an experimental methodology in order to identify the location of
impacts on elastic plates. The technique developed is based on principles of propagation
of dispersive waves in isotropic plates. Piezoelectric films are used to measure the strain
which is caused by flexural waves. The signals are then processed by the wavelet transform.
The maxima of the wavelet magnitude are used to calculate the arrival times of the waves.
The coordinates of the impact location, the time lag and the group velocity of the flexural
waves are then obtained by solving a system of four non-linear equations. A first estimate
of the group velocity of the dispersive waves has to be assumed in order to provide the
required initial values for the solution of the system of non-linear equations.
      795
It is possible to extend this procedure to quasi-isotropic materials such as composites.
The advantages of the proposed method are the accuracy with which the coordinates of
the impact can be estimated. Furthermore, no exact knowledge of the material properties
is required for the analysis. A rough estimate of the material parameters is sufficient, since
an optimisation algorithm is used to solve the non-linear equations.

REFERENCES
1. J. F. D 1987 Experimental Mechanics 27, 229–233. An experimental method for determining
the location and time of initiation of an unknown dispersing pulse.
2. K. C and F.-K. C 1996 Journal of AIAA 34, 136–142. Identification of impact force
and location using distributed sensors.
3. M. T and F.-K. C 1996 SPIE 2717, 231–236. Identifying impact load in composite
plates based on distributed piezoelectric sensor measurements.
4. D. E. N 1994 Journal of Vibration and Acoustics 116, 409–416. Wavelet analysis of
vibration: part I: theory.
5. K. K, H. I, M. H and T. S 1995 Journal of Applied Mechanics 62,
841–846. Time frequency analysis of dispersive waves by means of wavelet.
6. H. I, K. K and T. S 1996 Experimental Mechanics 36, 212–217.
Experimental wavelet analysis of flexural waves in beams.
7. G. S and T. N 1996 Wavelets and Filter Banks. Wellesley: Wellesley-Cambridge
Press.
8. D. E. N 1997 Proceedings of DETC’97 VIB-4135 , 1–12. Practical signal analysis: do
wavelets make any difference?
9. Y. W and R. D 1996 Mechanical Systems and Signal Processing 10, 29–53. Feature extraction
and assessment using wavelet packets for monitoring of machining processes.
10. R. D. M 1951 Journal of Applied Mechanics 18, 31–38. Influence of rotatory inertia and
shear on flexural motions of isotropic, elastic plates.
11. A G 1992 MATLAB Optimization Toolbox. Natick: MathWorks Inc.