Академический Документы
Профессиональный Документы
Культура Документы
Abstract
One problem of the acoustic emission testing is a great number of noises affecting the
diagnosis results. Electric noises, electromagnetic interference, background acoustic noise,
rubbing noises are far from the full list of noises present under measurements. At the high level of
noises the operator has to increase the recording threshold of the acoustic emission impulses
through reducing the testing sensitivity. False determination of AE parameters can result in
wrong location and false determination of the AE source danger level. To improve the noise
immunity of the acoustic emission system, the data filtering algorithms are to be used.
The main filtration algorithms applied in the A-Line system (Interunis company) are
frequency filtering (FIR and IIR filter, filters with an arbitrary transfer function)
wavelet filters for removing an additive white noise
suppression of impulse distortion
statistical filtration on the basis of principal-component analysis
The traditional frequency filtering is applied when the signal and noise lie in the different
frequency ranges. The possibility to define an arbitrary shape of the filter transfer characteristic
allows for carrying out a deconvolution operation or inverse filtering, that results in compensation
of distortions contributed into the signal by the measuring path.
The median-type filtering is applied for suppressing the impulse noises. The impulse noises
are widespread: various electric and electromagnetic noises have an impulse waveform. The
median filtering can also help in filtering of narrowband noises; in this case the signal spectrum
undergoes the median filtering. The spectrum of such signal has a well-defined single peak which
is removed during the median filtering of the signal spectrum.
For filtering of additive random noises, the A-Line 32D system employs the wavelet-
thresholding algorithm. The acoustic emission signal is represented as a wavelet-decomposition,
as a set of detailing and approximation coefficients, which correspond to different scale values.
Application of the threshold algorithm is equivalent to an adaptive smoothing. When the detailing
coefficients readings are put to zero, which correspond to the signal high-frequency components,
the signal low-frequency filtering is carried out.
One of key questions is the threshold correct selection. When selecting the threshold optimum
value, the risk function equal to mathematical expectation of the norm of difference of functions
before thresholding and function after thresholding is minimized. The criterion Stein's unbiased
risk estimation (SURE) is the most effective and widespread
On of the most intelligent filtration method is a statistical filtration on the basis of main-
component method. In this method the AE data represents as a continuous time series. The
principal-component analysis allows to divide noises and useful AE signals on the basis of the
difference in the they arrival time distribution.
Realization of various algorithms of AE data filtering allows for improving the noise
immunity of the acoustic emission system as a whole. Elimination of stochastic noise from the
acoustic emission signals allows for determining more precisely AE parameters, which can form
the basis for authentical determination of the acoustic emission source characteristics.
Impulse noise filtering
Application of the frequency filtering for removing the impulse noises does not result in an
acceptable solution, because the noise-like impulse signals have wide spectrum overlapped with
spectrum of acoustic emission signal. The median filtering is applied for suppressing the impulse
noises. A one-dimensional median filter is a sliding window including an odd quantity of signal
readings. The central reading is replaced with the median of all readings in the window. When
selecting the filtering window that exceeds duration of impulse components of noises, the last
will be suppressed, while energy of the acoustic emission signal will decrease insignificantly.
m
mV
mV
V
a. b. c.
Fig.1. Some types of noise signals a) electromagnetic noise; b) electric noise; c)
narrowband harmonic noise
The median filtering can also help in filtering of narrowband amplitude-modulated noises
(Fig. 1c); in this case the signal spectrum undergoes the median filtering. The spectrum of such
signal has a well-defined single peak which is removed during the median filtering of the signal
spectrum (1). Energy of the narrowband quasiharmonic noise decreases by hundreds of times
during the spectrum median filtering, while energy of the acoustic emission signal, which has no
defined single peaks in the spectrum, varies insignificantly.
2πi
1 N − jn
s ( n) = ∑ median ( F ( Ω )) e N
(1)
2π i =0
Upon allocation of the desired frequency range in the signal and elimination of the impulse
noises, the problem of additive white noise elimination from realizations of the acoustic emission
signals remains unsolved. For filtering of additive random noises, the A-Line 32D system
employs the wavelet-thresholding algorithm.
where ϕ(n) – basic function of sequence of orthogonal embedded subspaces, while ψ(n) – its
orthogonal complement.
Representation as a wavelet-decomposition is organic (natural) for a multicomponent signal
of acoustic emission, which is a package of different wave modes. Characteristics of separate
modes of the wave package are described by the detailing and approximation coefficients.
Fig.2 shows the algorithm diagram. Upon calculation of the wavelet-decomposition, a
threshold restriction of the detailing coefficients is performed followed by a signal reverse
recovery [4].
600 500
300 250
U, mV
U,mV
0 0
-300 -250
-600 -500
0 500 1000 1500 2000 0 500 1000 1500 2000
t, mcs t, mcs
a b
Fig.3 Wavelet thresholding’s results: a) initial AE signal b) signal with threshold noise reduction
The noises similar in waveform to acoustic emission signals are of particular complexity for
filtering. These are the noises generated by various mechanical reasons - knocks, knocks of
foreign objects, and precipitations. The noises of this kind also occur at heating a testing object
and at various hydrodynamic phenomena, for example, resulting from cavitation.
Such noises neither in waveform, nor in spectrum do not differ from acoustic emission signals.
Thus, neither of the abovementioned methods is suitable for their filtering. The fundamental
difference of the noise process and the plastic deformation process is a distinct regularity of
generation of impulses characterizing one or another process. Times of recording of impulses
relevant to the noise action are spread quasiuniformly (as in case of knocks or precipitations) or
in accordance with the Poisson's law (in case of cavitation or heating effect). Impulses
characterizing an active source of acoustic emission (defect) follow the complicated distribution
law, which parameters vary as defect develops. The defect impulses are recorded, as a rule, far
rarely, and times of their recording bring a disharmony in the form of a certain trend to the
distribution law of the background noise impulses.
To identify processes with a different nature of periodicity, it is possible to employ one of
methods of time series analysis –method of principal components. The method of principal
components allows for dividing the acoustic emission signal represented as the time series, into
several elements (components) – periodic components, chance variations and a trend. In the
problem of interpretation of acoustic emission signals, the periodic components characterize, as a
rule, deterministic noises, chance variations – random noises, and the trend characterizes the
acoustic emission source activity.
Signal preprocessing
When applying methods of time series analysis, long implementations of acoustic emission
signals are to be logged. Signal duration shall be estimated in tens of minutes; in this case, the
volume of information recorded will be tens of gigabytes. The analysis of such time series is not
efficient.
To reduce the data volume, it is recommended to operate not the signal itself, but its envelope.
Such replacement does not result in diminution of algorithm accuracy, since the signal spectrum,
in this case, is not informative. The envelope characterizes a low-frequency portion of the signal,
thus it can be effectively compressed through wavelet-transformation. The data compression ratio
depends on the contraction ratio of frequency range. If under detection of the signal envelope the
value of upper frequency of spectrum is decreased by 8 times, the envelope duration can be
reduced in comparison with the initial signal length by 28 times without significant losses of
information.
Noisy acoustic emission signal is showed on Fig.4. The noise source is friction, which is
artifically modeled with help of abrasive paper. Besides noise we can observe impulse AE signal
generated by Hsu-Nielsen source. Fig 4b shows the signals envelope compressed in thousand
times.
600
400
200
U, mV
-200
AE impulse
-400
Periodic noise
-600
0 4000000 8000000 12000000 16000000
n, samples
a.
18000
16000
14000
12000
10000
U, mV
8000
6000
4000
2000
0
0 200 400 600 800 1000 1200 1400 1600
n, samples
b.
Fig.4. Noisy AE signal and its envelope
Signal processing
The method of principal components is a method of data multidimensional analysis, but for the
time series analysis, the scientists from the St.-Petersburg State University designed its special
modification – singular spectrum analysis (or Caterpillar method). The given method suggests a
transformation of the univariate time series (caterpillar) to the set of time series which represent a
great number of fragments of the original signal (caterpillar links), cut with a sliding time
window (5). The time window duration shall be selected so, that all processes (the noise
processes and plastic deformation process) are able to become apparent.
Xi=(si-1,…,si+L-2), (5)
where L is window length.
If the parameters are selected correctly, as a result of application of the method of principal
components the signal is divided into several components, which characterize the various
processes generated by various sources, both the noise sources and the acoustic emission sources.
To divide the time series into components, the covariance matrix V is calculated (6), and its
eigen values and eigen vectors are calculated (7); under expansion of the covariance matrix in
terms of eigen components, a set of principal components of time implementation is obtained.
V = (1/k)XTX (6)
V = PLPT, (7)
where P – orthogonal matrix of eigenvectors.
The eigenvectors by themselves are impulse components of the filters yi (8) detecting various
components of the process.
XP = Y = (y1,y2,...,yM) (8)
In particular, among the principal components one may select a trend, slowly varying
components, periodic and impulse components. Expansion of the acoustic emission signal into
principal components is informative, because it allows for getting an insight of the processes
occurring in the testing object. Thus, for example, presence of the increasing trend in the signal
can give evidence of a progressive defect, while the periodic component characterizes a noise.
On fig.5 obvious example is shown. The signal has been spread out to the main components,
and noise periodic components, have been excluded from the further consideration
8000
7000
6000
5000
U, mV
4000
3000
2000
1000
0
0 200 400 600 800 1000 1200 1400 1600
n, samples
Conclusion
The paper is concerned with various methods of filtering of the acoustic emission data. The
algorithm of impulse noise filtering by means of the median filter is described. The efficiency of
wavelet thresholding for removal of white stochastic noise is shown. And also the method of
noise removal is described by means of the time series analysis using the method of principal
components.
References