You are on page 1of 50

Application of Artificial

Intelligence in
Electromagnetic Methods
of Non-Destructive Testing
Tomasz Chady
& KETiI NDT Team
(R.Sikora, M. Caryk, P. Łopato)
Westpomeranian University of Technology
Department of Electrical and
Computer Engineering
Szczecin, POLAND
Artificial Intelligence in NDT – ANN Models for MMFECT
• A process of defects’ detection and recognition requires the use of
equipment and intelligence.
• Human interpretation is subjective, inconsistent, and often biased.
• The additional problems are caused by the insufficient quality of utilized
signals or images.
• An incorrect classification may result in rejection of a part in good
conditions or acceptance of a part with defects exceeding the limit
defined by the relevant standards.
• A proper interpretation of the achieved data is a crucial and complicated
process requiring involvement of natural or artificial intelligence.
• Despite of the development of various types of modern testing
techniques, an involvement of human senses in NDT is still very
significant and difficult to substitute.
• In this presentation application of Artificial Neural Networks for
Automatic or Assisted Defect Recognition is proposed
2
Multi-Frequency Eddy Current Non-Destructive Testing
Signal Spectrum

Windowing Averaging
e.g. Hann FFT (optional)

Noise removal - Block of digital filters


Multi-frequency 1-D low pass filter in 1-D low pass filter in
signals “frequency f domain” “position x domain”

Detrending

Spectrogram

Flaw profile
ANN or
Flaw parameters

T. Chady, M. Enokizono, R. Sikora, T. Todaka, Y. Tsuchida: “Natural Crack Recognition Using Inverse Neural Model and Multi-Frequency Eddy
Current Method”, IEEE Transactions on Magnetics, Vol. 37, No. 4, July 2001, pp. 2797-2799

3
Multi-Frequency Eddy Current Non-Destructive Testing
A massive multi-frequency excitation and spectrogram eddy current method
(MMFES) is a very effective tool of nondestructive testing. It allows to identify
defects using a frequency characteristic of defects. In the MMFES system a
complex signal containing selected harmonic components is used as an
excitation:

n
s E (t )  U i  sin2f i t  i 
i 1

4
Multi-Frequency Eddy Current Non-Destructive Testing

Signal measured from a pick-up coil of the transducer is amplified, filtered, digitized
and analyzed using discrete Fourier transform (DFT). The frequency components of
the pick-up signal measured over the defect is utilized to create a spectrogram.

5
Multi-Frequency Eddy Current Non-Destructive Testing
The spectrogram is a two-dimensional plot of the relative amplitude of the frequency
components of a signal from pick-up coil versus the sensor position. The information
about the defect can be obtained from the frequency characteristic measured at the
position XMAX corresponding to the peak value of the spectrogram SMAX

Spectrogram with depicted Frequency characteristic S(f) of the


parameters defect’s respond measured at x=XMAX
SMAX
SMAX

fMAX

x=XMAX

XMAX fMAX

T. Chady, M. Enokizono, R. Sikora, T. Todaka, Y. Tsuchida: “Natural Crack Recognition Using Inverse Neural Model and Multi-Frequency Eddy
Current Method”, IEEE Transactions on Magnetics, Vol. 37, No. 4, July 2001, pp. 2797-2799

6
Multi-Frequency Eddy Current Non-Destructive Testing

7
Multi-Frequency Eddy Current Non-Destructive Testing
y x
EC EB

EA ED

x
Spectrogram
(xmax, fmax, smax)
Single-frequency signal
2
2
1.5 magnitude s (mV) 1.5
magnitude s (mV)

1
1
0.5
0.5 0

0 200
10
-15 -10 -5 0 5 10 15 100 0
sensor position x (mm) frequency f (kHz) -10
sensor position x (mm)

8
Multi-Frequency Eddy Current Non-Destructive Testing

Parameter (mm) UNIBig UNISmall


D1 12 8.5
D2 2 1.5
L1 5 3.5
H1 10 7
H2 6 4.5
UNI-Small

Depth d [mm]
Depth d [mm]

Position x [mm] Position x [mm]


Depth d [mm]

Depth d [mm]
UNI-Big

Position x [mm] Position x [mm]

9
Multi-Frequency Eddy Current Non-Destructive Testing

OF 40 %
Flaw length 2 mm, x - direction

OF 20 %
OF 10 %

10
Multi-Frequency Eddy Current Non-Destructive Testing

OF 40 %
Flaw length 2 mm, y - direction

OF 20 %
OF 10 %

11
Multi-Frequency Eddy Current Non-Destructive Testing

Flaw depth [%]


Flaw
length UNI-Big UNI-Small
(mm) 10 20 40 60 80 10 20 40 60 80

2 15.37 28.22 38.90 47.74 53.31 11.00 22.24 33.70 44.06 50.62

3 17.55 30.76 42.35 50.80 56.27 12.38 24.50 36.30 46.07 52.77

5 21.79 33.32 44.57 51.33 57.43 15.22 29.35 40.88 47.77 54.36

12
Multi-Frequency Eddy Current Non-Destructive Testing

Approximation algorithm

f n – given frequency

T. Chady and P. Lopato, “Flaws identification using an approximation function and artificial neural networks”, IEEE transactions on magnetics,
2007, vol. 43, no 4, pp. 1769-1772

13
Multi-Frequency Eddy Current Non-Destructive Testing
Simplified approximation Extended approximation
2.5 2.5
noisy signal noisy signal
2 filtered signal 2 filtered signal
simplified approximation extended approximation

magnitude s (mV)
magnitude s (mV)

1.5 1.5

1 1

0.5 0.5

0 0

-0.5 -0.5
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
sensor position x (mm) sensor position x (mm)

 c1b1 ( x  x1 ) 2
   ( x  x)   ( x  x) 
2 2
s( x)  a1 (b1 ( x  x1 )  1)  e
2

s ( x )  s m  k u  e e 
   a1 a 2 (b1 ( x  x1 )  1)  e
2  c1b1 ( x  x1 ) 2

 c2  x 2
 b2  e

T. Chady and P. Lopato, “Flaws identification using an approximation function and artificial neural networks”, IEEE transactions on magnetics,
2007, vol. 43, no 4, pp. 1769-1772

14
Multi-Frequency Eddy Current Non-Destructive Testing
Function Approximation for signal restoration ( x - domain )
 b0 x 2  b1  x  x1 2  b2  x  x 2  2
U ( x)  a0e  a1e  a2e

T. Chady and P. Lopato, “Flaws identification using an approximation function and artificial neural networks”, IEEE transactions on magnetics,
2007, vol. 43, no 4, pp. 1769-1772

15
Multi-Frequency Eddy Current Non-Destructive Testing
Function Approximation for signal restoration ( f - domain )

U ( f )  cf 2 e  df

Length-2mm , Crack IF 20 Length-2mm , Crack IF 60 Length-2mm , Crack IF 100


0.35 1.4 2
Approximation Approximation
0.3 1.2 Measurements 1.8 Measurements

Approximation 1.6
0.25 1
Signal value [mV]

Signal value [mV]

Signal value [mV]


Measurements
1.4
0.2 0.8
1.2
0.15 0.6
1
0.1 0.4
0.8

0.05 0.2 0.6

0 0 0.4
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Frequency [kHz] Frequency [kHz] Frequency [kHz]
Length-1mm , Crack IF 20 Length-1mm , Crack IF 60 Length-1mm , Crack IF 100
0.2 0.7 0.8
Approximation Approximation
0.18 Measurements Measurements
0.6 0.7
Approximation
0.16 Measurements
0.5 0.6
Signal value [mV]

Signal value [mV]

Signal value [mV]


0.14

0.12 0.4 0.5

0.1 0.3 0.4


0.08
0.2 0.3
0.06
0.1 0.2
0.04

0.02 0 0.1
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Frequency [kHz] Frequency [kHz] Frequency [kHz]

16
Multi-Frequency Eddy Current Non-Destructive Testing

Results of signal approximation procedure

Measured signal Approximated signal Generated signal


length 2 mm, relative depth 20% length 2 mm, relative depth 20% length 3.5 mm, relative depth 35%

1.5 5
1.5 magnitude s (mV)
magnitude s (mV)

magnitude s (mV)
4
1 1 3
0.5 0.5 2
1
0 0 0
200 200 200
10 10 10
100 0 100 0 100
-10 0
frequency f (kHz) sensor position x (mm) -10 -10
frequency f (kHz) sensor position x (mm) frequency f (kHz) sensor position x (mm)

T. Chady and P. Lopato, “Flaws identification using an approximation function and artificial neural networks”, IEEE transactions on magnetics,
2007, vol. 43, no 4, pp. 1769-1772

17
Flaws Identification - ANN Models

k
h
b
g

Low Pass Filter

approximation of the signal

approximation of the frequency


characteristic
determination of smax, fmax
parameters
determination of k, b , g
parameters

Artificial Neural Network

T. Chady and P. Lopato, “Flaws identification using an approximation function and artificial neural networks”, IEEE transactions on magnetics,
2007, vol. 43, no 4, pp. 1769-1772

18
Flaws Identification - ANN Models
Noise influence on multifrequency signal parameters
Measured SNR (dB) for outer flaws

Fluctuations of fmax caused by noise


without approximation with approximation

T. Chady and P. Lopato, “Flaws identification using an approximation function and artificial neural networks”, IEEE transactions on magnetics,
2007, vol. 43, no 4, pp. 1769-1772

19
Flaws Identification - ANN Models
Neural models for flaw identification using spectrogram parameters
Flaw depth identification error
(xmax, fmax, smax)

2
magnitude s (mV)

1.5
1
0.5
0

200
10
100 0
frequency f (kHz) -10
sensor position x (mm)

f max
h
s max

T. Chady and P. Lopato, “Flaws identification using an approximation function and artificial neural networks”, IEEE transactions on magnetics,
2007, vol. 43, no 4, pp. 1769-1772

20
Multi-Frequency Eddy Current Non-Destructive Testing

Test Sample made of INCONEL600


with Complex Flaws

21
Multi-Frequency Eddy Current Non-Destructive Testing

UNI UNI
Big Small
dl = 4 mm

20%

20%
dl = 2 mm

20%

20%
dl = 1 mm

20%

20%

22
Multi-Frequency Eddy Current Non-Destructive Testing
UNI UNI
Big Small

80%
40% 20%
20%
20%
60%
80%
2 mm

IF20%

OF60%
4 mm

20%
3 mm

40%
20%
60%

ISEM2005
23
Multi-Frequency Eddy Current Non-Destructive Testing
1 mm

2 mm
OF40%
OF40%

OF20%

24
Multi-Frequency Eddy Current Non-Destructive Testing

2 mm

4 mm
OF20
3 mm

OF40
OF20
OF60

25
Multi-Frequency Eddy Current Non-Destructive Testing

80%
40% 20%
20% 20%
60%

80%

26
Neural Models of Multi-Frequency ECT Transducers
FORWARD model was designed in order to INVERSE model enables us to determine
attain a neural model of the eddy current plate geometry from a signal generated by
probe. This model allows to predict output the probe. The input vector contains
signal for a given flaw profile. The input samples of the output signal from the
vector contains co-ordinates of the flaw probe and the net output corresponds to
profile, taken from a moving window. predicted co-ordinates of the plate
surface.
h Frequenc
arequenc f2
Signal amplitude Frequency
fn
Window Window f1
Specimen

Flaw
x x
Sensor
position

Forward Model Inverse Model

h
Signal amplitude Estimate
Current
Estimate
value for fn Specimen
estimated
value for f2
value for f1 Estimated
point
Flaw
x x
Sensor position Sensor position

27
Inverse Neural Models of Multi-Frequency ECT Transducers

without feedback
D D
Structures
D D

D D

D D
f1 f1
f2 f2

fn fn

D
D
D
with feedbacks

D
D
Structures

D
D f1
f1 f2
f2
D
fn
fn
D
D

D D

T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

28
Inverse Neural Models of Multi-Frequency ECT Transducers

n
f
n
f

2
f 1
2
f 1

f
f
D

D
D

D
Structure: 5 1, window length: 6, number of frequencies: 5 Structure: 5 4 1, window length: 12, number of frequencies: 5
100 100
estimation estimation
80 80
depth [%]

depth [%]
original profile original profile
60 60
40 40
20 20
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Position [mm] Position [mm]

Structure: 3 1, window length: 15, number of frequencies: 3 Structure: 4 2 1, window length: 20, number of frequencies: 5
100 100
estimation estimation
80 80

depth [%]
depth [%]

original profile original profile


60 60
40 40
20 20
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Position [mm] Position [mm]

Structure: 3 1, window length: 15, number of frequencies: 1 Structure: 6 5 1, window length: 15, number of frequencies: 5
100 100
estimation estimation
80 80
depth [%]

depth [%]

original profile original profile


60 60
40 40
20 20
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Position [mm] Position [mm]
T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

29
Inverse Neural Models of Multi-Frequency ECT Transducers

n
f
n
f

2
f 1
2
f 1

f
f

D
D

D
D

D
D
D

Structure: 5 1, window length: 10, num. of delays in feedback: 7 Structure: 5 1 1, window length: 10, num. of delays in feedback: 2
100 100
estimation estimation
80 80
depth [%]

depth [%]
original profile original profile
60 60
40 40
20 20
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Position [mm] Position [mm]

Structure: 3 1, window length: 10, num. of delays in feedback: 3 Structure: 3 3 1, window length: 10, num. of delays in feedback: 5
100 100
estimation estimation
80 80
depth [%]

original profile

depth [%]
original profile
60 60
40 40
20 20
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Position [mm] Position [mm]
Structure: 5 1, window length: 5, num. of delays in feedback: 3 Structure: 5 3 1, window length: 10, num. of delays in feedback: 15
100 100
estimation estimation
80 80
depth [%]

depth [%]

original profile original profile


60 60
40 40
20 20
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Position [mm] Position [mm]

30
Inverse Neural Models of Multi-Frequency ECT Transducers
D

D
Influence of number of frequencies
D
(ANN model with 2 hidden layers,
D

f
2
f
1 without feedbacks)
f
n

x 10
-3 Window length 5 x 10
-3 Window length 20
3 1.4
n1=2 n1=2
n1=3 n1=3
1.2
2.5 n1=5 n1=5
n1=7 n1=7
n1=9 1 n1=9
2 n1=11 n1=11
n1=13 n1=13
0.8
MSE

MSE
1.5
0.6

1
0.4

0.5
0.2

0 0
1 3 5 1 3 5
number of frequencies number of frequencies
T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

31
Inverse Neural Models of Multi-Frequency ECT Transducers
D

D
Influence of number of frequencies
D
(ANN model with 3 hidden layers,
D

f
2
f
1 without feedbacks)
f
n

Window length 5 Window length 20


-3 -4
x 10 x 10
2 6
311 311
1.8 411 411
631 5 631
1.6
651 651
1.4
4
1.2
MSE

MSE
1 3

0.8
2
0.6

0.4
1
0.2

0 0
1 3 5 1 3 5
number of frequencies number of frequencies

32
Inverse Neural Models of Multi-Frequency ECT Transducers
D

D
Influence of windows size
D
(ANN model with 2 hidden layers,
D

f
2
f
1 without feedbacks)
f
n

Number of frequencies 5 Number of frequencies 1


-3 -3
x 10 x 10
1.2 3
n1=2 n1=2
n1=3 n1=3
1 n1=5 2.5 n1=5
n1=7 n1=7
n1=9 n1=9
0.8 n1=11 2 n1=11
n1=13 n1=13
MSE

MSE
0.6 1.5

0.4 1

0.2 0.5

0 0
10 15 20 10 15 20
window length window length

33
Inverse Neural Models of Multi-Frequency ECT Transducers
D

D
Influence of windows size
D
(ANN model with 3 hidden layers,
D

f
2
f
1 without feedbacks)
f
n

Number of frequencies 5 Number of frequencies 1


-4 -3
x 10 x 10
4.5 2
311 311
4 411 1.8 411
631 631
3.5 1.6
651 651
1.4
3
1.2
2.5
MSE

MSE
1
2
0.8
1.5
0.6
1 0.4

0.5 0.2

0 0
5 10 15 20 5 10 15 20
window length window length

34
Inverse Neural Models of Multi-Frequency ECT Transducers
Influence of delays in feedback
D

Number of frequencies: 5 D

Window length: 10
D

Structure: 531 f
2
f
1

-3 f
x 10 n

3 D

2.8

2.6
MSE

2.4

2.2

1.8
0 2 4 6 8 10 12 14 16 18 20
number of delays in feedback

T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

35
Multi-Frequency Eddy Current Non-Destructive Testing

20% 80%

4 mm 3 mm

20%

40% 60%

2 mm 5 mm

60% 10%

5 mm

36
Inverse Neural Models of Multi-Frequency ECT Transducers

Estimation of complex flaw’s profiles


D

D
Window length 10, Structure: 6 3 1
f
f 1
2
f
n

T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

37
Inverse Neural Models of Multi-Frequency ECT Transducers

Estimation of complex flaw’s profiles


D
D
D

Window length 10, Structure: 5 3 1


D
f
f 1
2
f
n
D
D

T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

38
Inverse Neural Models of Multi-Frequency ECT Transducers
Hidden
fN TD
layer 1 Output
Single output
TD
TD
TD neuron h
TD
model
40 TD
TD x
Spectrogram del
40 TD
ays
del
40
ays
del Flaw profile
ays
TD Hidden
TD layer 2
TD
TD

Hidden Output
layer 1 neurons
Multiple output

fN TD
TD
TD h
TD
model

TD
40 TD
TD
Spectrogram del
40 TD
ays
del
40
ays
del
ays
TD
x
TD
TD Hidden Flaw profile
TD
layer 2

T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

39
Inverse Neural Models of Multi-Frequency ECT Transducers
D - OF 40%
Flaw F, Length 7 mm Flaw E, Length 2 mm E - OF 20%
F - OF 80%

3 f , 15-5

Flaw F, Length 1 mm Flaw D, Length 2 mm Flaw E, Length 5 mm

T. Chady, M. Caryk: “Reconstruction of Flaw Profiles Using Neural Networks and Multi-Frequency Eddy Current System”, Review of Quantitative
Nondestructive Evaluation, Eds.: D. O. Thompson, D. E. Chimenti, Vol. 24, American Institute of Physics, CP 760, New York, 2005, pp. 820-827

40
ANN Inverse Models – 3D Flaws Identification

Neural identification procedure


Artificial Neural Network Reconstructed geometry
Measured signal (inverse model) of the flaw

ANN

T. Chady, P. Łopato: “Identification of 3D Conductivity Distribution Using Eddy Current Technique and Artificial Neural Networks”, Proceedings of 13th Biennial
IEEE Conference on Electromagnetic Field Computation, Athens, Greece, May 11–15, 2008, PE5 – 2, pp. 471.

41
ANN Inverse Models – 3D Flaws Identification

Neural inverse models

Feed-forward Neural Network (FFNN) Feed-back Neural Network (FBNN)


45 45
30 30

1
5x5

3x3
1

T. Chady, P. Łopato: “Identification of 3D Conductivity Distribution Using Eddy Current Technique and Artificial Neural Networks”, Proceedings of 13th Biennial
IEEE Conference on Electromagnetic Field Computation, Athens, Greece, May 11–15, 2008, PE5 – 2, pp. 471.

42
ANN Inverse Models – 3D Flaws Identification
X-ray ECT Crack map FFNN FBNN

43
ANN Forward Model – Support problem
High frequency signal Low frequency signal dependent
dependent mostly on the flaw mostly on the fastener

• In case of the steam generator tubes made of Inconel600 and external magnetic fasteners the
signals obtained at a low frequency are strongly correlated with the unwanted signal, caused by
fasteners.
• The high frequency components are affected mainly by the tubes discontinuities.
• Unfortunately, the very different frequency signals could not be subtracted directly because the
shape of the signal caused by the fasteners is not retained.
• In order to overcome this problem the authors propose to use a specially designed dynamic
neural network, which is capable to reconstruct a “high frequency” signals from the measured
low frequency component.
• Such a technique enables us to detect shallow flaws signals buried in high-level background
noises.
44
ANN Forward Model – Support problem
Spectrogram for: a) tube with flaw; b) tube with fastener; c) tube with flaw and fastener

View of the tube with


the fastener and flaw
Fastener made
Thickness
of steel S45C 1.27 mm
22.23 mm

Sensor
Flaw

Tube made of
Width INCONEL600

45
ANN Forward Model – Support problem

 The lowest frequency component (S1L) of the spectrogram caused by the


fastener and flaw was used to estimate the whole spectrogram for the
support element only (ES).
 Signal S1L is provided the input of the neural network model having 40
input units through the tapped delay line.
 Number of output neurons corresponds to the number of frequency
components used to create the spectrogram.
 The network is working in the following way:
❖ signal S1L is provided to the taped delay line,
❖ output of the taped delay line (windowed part of the input signal – 40
samples) is transferred to the input units of the network,
❖ the network calculates a SINGLE sample of all of the output signals,
❖ the input signal is shifted by the taped delay line and next window is
presented to the network,
❖ the process is repeated until to the last sample of the input signal.

46
ANN Forward Model – Support problem

Spectrogram obtained Lowest frequency Dynamic neural Estimation of the


for the flaw (OF60%) component of the network signal caused only by
and the fastener spectrogram “S1” (with moving the fastener
window)
U (mV) 1.5 Input
1 TD

0.5 TD

-0.5
Estimation of the
-20 -10 0 10 20
X (mm)
flaw (OF60%) signal
X (mm)
S1 S1L TD

X (mm)
ES -
40 delays hidden 16
neurons output
neurons S1 + EF

X (mm)

The lowest frequency component (S1L) of the spectrogram caused by the fastener and flaw
was used to estimate the whole spectrogram for the support element only (ES).
Signal S1L is provided the input of the neural network model having 40 input units through the
tapped delay line. Number of output neurons corresponds to the number of frequency
components used to create the spectrogram.

47
ANN Forward Model – Support problem

Fastener + OF20% Fastener + OF40% Fastener + OF60% Fastener + OF80%


Measured
mixed signal
caused by
fastener and
flaws (S1)

Numerical
estimation of
signal caused
by fastener
(ES) 10
mm
OF20%* OF40%* OF60%* OF80%*

Numerical
estimation of S - E
1 S
signal caused
by flaws

OF20% OF40% OF60% OF80%

Measured
signal
caused by
flaws

48
ANN Forward Model – Support problem
NO SUPPORT  = 200,  = 1.3 107 NO SUPPORT  = 200,  = 1.3 107

In case of single-frequency method the In case of ”true multi-frequency” method one


selected quasi optimal frequency does not could select optimal frequency and therefore
guarantee best detectability in case of obtain always best detectability. The peak
presence of support elements. amplitude is even higher than without support.

49
ANN Forward Model – Support problem
450
Flaw signals OF80*
400

350 Estimation of OF80


flaw signals
300
SMAX ( V)

250

200 OF60*

150 OF60

100 OF40*

50 OF40
OF20*
OF20
0
60 70 80 90 100 110 120 130 140
FMAX (Hz)
 The estimated flaw signals are compared with the measurements performed on the same
specimen but without fastener. One could observe the difference between these results.
This is caused by the nonlinear effect. The signal resulting from two flaws, or from flaws
and background noise is not completely identical with the linear combination (vectorial
sum) of the two signals considered separately.
 The parameters of the spectrograms (SMAX, FMAX) were also calculated and presented in the
form of calibration curves. It is visible that the curve obtained from estimated flaw signals
are shifted in comparison with the original calibration curve, but it is still possible to
evaluate the flaw depth.
50

You might also like