POSSIBILITY OF APPLICATION OF ARTIFICIAL NEURAL NETWORK FOR RECOGNIZING OF ACOUSTIC-EMISSION EVENTS

apricotpigletΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

61 εμφανίσεις

POSSIBILITY OF APPLICATION
OF ARTIFICIAL NEURAL
NETWORK FOR RECOGNIZING OF
ACOUSTIC
-
EMISSION EVENTS

Pirumov
А
.
1

, Chvertko Ye.
1
, Por G.
2

, Dobjan T.

2

1

National Technical University of Ukraine "KPI", Kyiv, Ukraine

2

College

of

Dunaújváros
,
Dunaújváros
, Hungary


Advantages and features of AE
monitoring


allows to detect and identify developing
defects;


allows to control the entire object using a
small number of stationary sensors on its
surface



background noise


no database with the results of systematic
studies


the need of operator

Typical information parameters for
the AE signals analysis include:

EXPERIMENTS

Two series of experiments:



Source of Hsu
-
Nielsen (graphite rod
breaking)





Concentrated metal bullet impact

EXPERIMENTS

NEURAL NETWORK

S
1
×
1

a
1

Input

II
ndist

II

IW
1,1

LW
2,1


C


Competitive

layer

Linear

layer

P

R
×
1

S
1
×
R

S
1
×
1

R

S
1

S
2
×
1

S
2
×
S
1

S
2

n
2

a
2

=
y

S
2
×
1

n
i
1

=

II

i
IW
1
,
1



P

II

a
1

=

compet

(
n
1
)

a
i
2

=

purelin

(
LW
2
,
1

a
1
)

n
1

NEURAL NETWORK

Selection of
network type
and structure

Neural network training

Neural network
application

Data

Data
processing
results

Adjusting of
weight
coefficients

Data

Training
method and
parameters

Goals
vector

NEURAL NETWORK

START

Initializing the neural network
:

net = newlvq (minmax (P), 12, [.625 .375], 0.1);

Generation of training vectors using the experimental
data (vector P)

Generation of goals vector and transformation into a
matrix
:

Тс = [1 1 1 1 2 2 2 …]

T=full(ind2vec(Tc))

Specifying parameters of training
:

net
.
trainParam
.
epochs
=2000;

Neural network training
:

net

=
train

(
net
,
P
,
T
)

END

RESULTS OF THE FIRST TEST

Results of identification of experimental data which weren’t used
during network training

APPLICATION TO TENSILE TEST

APPLICATION TO TENSILE TEST

CONCLUSIONS


Application of artificial neural networks
makes it possible to identify AE events
sources with an accuracy of at least 75 %,
giving wide opportunities for automation
of the AE control


The preliminary experiments have shown
the possibility of determining the state of
the material under mechanical loads by
analyzing the AE signals using artificial
neural networks