凳子的英文
Abstract —As flaw classification is norm ally m anual determ ination in ultrasonic nondestructive testing field, an automatic identification of flaw type bad on Lifted Wavelet Transform (LWT) and BP neural network (BPN) is introduced in this paper. LWT is propod to extract flaw feature from ultrasonic echo signals, ideally matched local characteristics of original signals. The co putational speed and flaw classification efficiency is incread. Then a feature library is constructed. A modified BPN is followed as a classifier, trained by the library. And then when feature is extracted from any other flaw echo, the feature eigenvector is nt to the trained BPN. The output of the BPN is the input flaw signal’s type, realizing autom atic flaw classification. For com parison, a Radial Basis Function neural network (RBFN) is tested under the sam e condition as BPN. Experim ent results prove the propod m ethod, LWT with BPN, is fit for autom atic flaw classification.
I. I NTRODUCTION
IPELINE is widely ud in transporting petroleum and natural gas, important energy materials nowadays. Pipeline safety is cloly related to public living environment and national economy [1]. D uring welding procedure and usage of pipelines, veral kinds of flaws may be formed, which are hidden dangers to trigger leakage accidents. Currently, ultrasonic phad array instrument is adopted for quick automatic detection and positioning of flaws in the pipeline girth weld. But traditional manual i
dentification method is still ud in flaw classification and human-made error is easily introduced. Since different kinds of flaw have different influences on pipeline safety, it is very important to recognize the type of tested flaw [2]. Searching for an automatic flaw classification method is a challenge but necessary to increa the recognition precision and velocity. The key of automatic flaw classification is to extract feature which reflecting defects nature from ultrasonic echo signals. There have been many rearches done in this field. Traditional methods include spectrum analysis method, time transition diffraction, and synthetic aperture focusing and split spectrum analysis [3]. Wavelet packet transform (WPT)
Manuscript received January 23, 2009. This work was supported in part by National Natural Science Foundation of China (NO. 60534050) and Tianjin Natural Science Foundation (06YFJMJC02000).
Jian Li is with the faculty of State Key Laboratory of Precision Measuring Technology & Instruments, Tianjin University, P.R.China 86-22-27401462 86-22-27402366 tjupipe@
Xianglin Zhan is with the department of Electrical Engineering, Civil Aviation University of China, P.R.China xlzhan@
Shijiu Jin is with the faculty of State Key Laboratory of Precision Measuring Technology & Instrument
s, Tianjin University, P.R.China shjjin@).
is a common feature extraction method. However, as the wavelet basis function is constructed by Fourier Transform, involving mass convolution computation and occupying vast inner memory, its computational speed is limited [4]. Modern signal processing methods are ud in flaw feature extraction and pattern recognition is ud in automatic flaw classification [5]. Performances of RBFN and back-propagation neural network (BPN) are compared while experiment results showed that RBFN has better classification accuracy and speed [6]. The studies are mainly bad on traditional ultrasonic testing signals and there is still no mature method in flaw qualitative analysis. In this paper, a new method combined Lifted Wavelet Transform (LWT) with modified BPN is studied. The cond part of this article describes the principle of feature extraction method bad on LWT, the third part introduces application of BPN as the classifier to realize automatic flaw classification before moving into experiment experimental configuration and results in the fourth part. Finally, future rearch is pointed out.lead
In this paper, LWT is propod to extract flaw feature, becau it has many merits, such as symmetrical, compactly supported, decaying oscillation with impulsive signature and constructing wavelet according to the signal characteristics. A modified BP neural network is ud as classifier, re
alizing automatic flaw classification. Then, experimental configuration and results are introduced in detail and experiment results validate that LWT is more fit for flaw classification of ultrasonic nondestructive testing..
II. F EATURE EXTRACTION METHOD
L WT utilizes a classical 2-channel filter bank as a framework for multiresolution analysis [7]. It is not necessarily translates and dilates of one mother wavelet as wavelet transform (WT) or wavelet packet transform (WPT) does. Through lifting scheme, the wavelet basis can be constructed in time domain completely, greatly improving the computation speed. In addition, vanishing moments can be incread by lifting scheme according to engineering requirements, and the shape of the wavelet can be getting smoother. This makes the wavelet ideally match the analysis signal and be propitious to feature extraction.
A. Principle of lifting scheme
LWT was propod by Sweldens and Daubechies. It does not necessarily translate and dilate of one mother wavelet. The analysis signal is firstly splitted into even quence and odd quence, which is called lazy wavelet. Then the odd
An Automatic Flaw Classification Method of Ultrasonic
Nondestructive Testing for Pipeline Girth Welds
Jian Li, Xianglin Zhan, Shijiu Jin
P
Proceedings of the 2009 IEEE
International Conference on Information and Automation
June 22 -25, 2009, Zhuhai/Macau, China
quence numbers are predicted by the even quence through a prediction operator, which is the lifting scheme. And then the deviation between prediction result and the real value of the odd quence is ud to repair the even quence by an updating operator. This procedure is called dual lifting scheme, the inver procedure of lifting scheme.
alignmentFrom another point of view, as any discrete wavelet transform can be decompod into ladder struct
ures, traditional wavelet filters can be prented by polypha matrix )(z P . It has been proved that any polypha matrix of traditional wavelet transform can be factored into lifting steps by Euclidean algorithm [4]:
,1)(0110)(1/100)(1∏=¿
简繁体字转换
¾½¯®»¼º«¬ª»¼º«¬ª»¼º«¬ª=m i i i z p z u K K
vsa
z P (1)
where
)(z u i and )(z p i are Laurent polynomials,
k k
小女孩英文名i k i z u z u −¦=)(,k
k
i k i z p z p −¦=)(,K is a non-zero
constant. From Equation (1), it can be referred that the polypha matrix can be reprented by Laurent polynomials, whereas )(z u i is the prediction operator and )(z p i is the updating operator. Therefore, the reprentation format of polypha matrix is the same as the lifting scheme.
On the basis of (1), the process of forward transform of LWT ud for signal decomposition can be expresd as Fig. 1. Here, 4-tap orthonormal filter with two vanishing moments db4 is factored and the forward SGWT is:
,3212)1(l l lrax
x x d
−=+ (2)
,4/)23(4/3)1(1
)1(2)1(+−++=l l
l l d d
x s
(3)
,)1(1
)1()2(−+=l l l
s d d
(4)
,2/)13()1(l
l s s += (5) .2/)13()2(l l d d −= (6)
B. Feature extraction algorithm bad on LWT If input flaw signal is suppod as
},...,,{110−=M x x x x , M is signal length. Bad on the
above analysis, flaw signal is decompod by the forward lifting wavelet transform as: --First, splitting:
;20l l x s = ;120+=l l x d )1(,...,1,02−=M l (7)
--Second, lifting and dual lifting:
;11¦−−−−=k
i k l i k i l i l s p d d ,1
1¦−−−+=k
i k l i k i l i l d u s s (8)
--Third, scale transform:
;m l l Kd d = ./K s s m l l = (9)
Then, smooth approximate signal 01(1)
2{,,...,}M s s s s −=
and discrete detail signal 01(1)2
{,,...,.}M d
d d d
−= are
achieved. After transferred pret levels bad on lifting scheme above, original signal is decompod into veral frequency bands. And then, the energy feature of each frequency band is extracted as:
.2
¦=k
jk lj x E (10)
where jk x is the amplitude value of discrete signal in each frequency band. And the flaw feature n T is constructed and normalized as:
].[],...,,[1
20
10¦−===l j lj
lj
lj l l n
E
E E E E T (11)
The above feature extraction on LWT can be concluded as Fig. 2. Becau only 4 types flaws are most common in girth welds of pipelines, N echoes of the 4 types flaws are collected by ultrasonic phad array detector. A feature library extracted from N flaw echoes is then constructed,
being ud to train the following BP neural network.
Fig. 1. Process of forward LWT
III. A UTOMATIC F LAW C LASSIFICATION W ITH M ODIFIED B PN A modified BPN is ud as a classifier to realize automatic flaw classification becau BPN has strong fault tolerance property. There are 3 layers of BPN. The node number of input layer is the element number of a feature. As 4 types flaw are to be identified, there are two nodes at the output layer and two binary numbers are ud to reprent flaw types. The node number of middle layer is computed with an empirical equation:
»¼»
«¬
«+=∞+22int O I
m N N N (12)
Where ¬¼x +∞
int indicates x is round-off number to positive
infinity. I N ,O N and m N are the node number of input layer, output layer and middle layer parately. The train error object is t to 0.0001. Levenberg-Marquart algorithm is train
function. Since output values of BP network are normally fractions, floating around 0 and 1, Iterative Self-Organizing
2012年英语二答案Data Analysis Techniques (ISODATA) is applied to process the output values. The algorithm is as follows:
--First, determining the control parameters and the initial values. Setting initial clustering number C , i.e. flaw types; tting initial clustering center, which is reprented by the binary numbers.
--Second, defining distance between two samples as Euclidean distance:
.)(),(2¦−=
−=i
i i
y x
y x y x D (13)
veral造句According to (13), the output value of the neural network is merged to the most possible type. --Third, determining criterion function for evaluating the clustering result. Here, error sum of squares is applied as the
criterion function:
,11
¦¦==−=C
i N j ij e j
m x J
(14)
Fig. 2. Flow chart of flaw feature extraction method bad on LWT
where
j N is the sample number of j cluster. The cluster center of each type is updated according to (14).
--Four, computing similarity parameter σthat indicating similarity level of every sample in a cluster to the cluster center bad on (13). At the same time, average similarity parameter δ of total samples in a cluster to its cluster center is computed, according to which splitting and the splitting policy are determined.
--Five, computing similarity level of current cluster centers and confirming merging and merging policy.openkore
--Six, if the last iteration comes, the program should be ended; if any parameter is needed to be revid, turning to step one. Otherwi, iterative computation is continued.
--Seven, when iteration computation is ended, the center of each cluster is changed. Computing the distance of the new centers to the initial cluster centers bad on (13). The one with minimum distance is merged to the initial clusters.
IV. E XPERIMENT
A. Experiment Configuration
A t of ultrasonic phad array detection and automatic flaw classification system is designed. The system is mainly made up of the host computer and the flaw detector with two ultrasonic phad array transducers. A big diameter pipeline girth weld block is tested, who groove format is CRC and wall thickness is 14.6mm. The testing configuration is shown in Fig. 3.
As root incomplete penetration, lack of fusion, center line crack and gas porosity are four main natur
al flaw types in pipeline girth weld, only four types of artificial defects are machined in the block, which are notch , flat-bottom hole, pylome and round hole to simulate the above natural flaws correspondingly. The binary outputs of the artificial neural network are as followed: (0, 0) is notch, simulating incomplete penetration; (0,1) is flat-bottom hole, simulating lack of fusion; (1,0) is through hole, simulating longitudinal crack; (1,1) is round hole, simulating gas hole. The flaw
structures are prented from Fig. 4 to Fig. 7.
Fig. 3.Testing configuration with an ultrasonic phad array system
In Fig. 4, the notch is 1.5mm in length along weld’s axial
and 0.92mm in width, 1mm in depth, with 37 degree angle to
the horizontal direction. In Fig. 5, the flat-bottom hole is
Fig. 4 Structure of a notch
Fig. 5 Structure of a flat-bottom hole
Fig. 6 Structure of a through hole
Fig. 7 Structure of a round hole
located at groove fusion line, who suface is a round ofφ2 with 45 degree angle to the axis. According to ASTM E-1961-98, flat-bottom hole can be located at hot welding zone and filling zone, to simulate lack of fusions locating at different zones. In Fig. 6, the through hole penetrates the weld along the axis and its surface is a round of φ2. In Fig. 7, the round hole is of φ1.5, to simulate the gas hole.
B.Feature Extraction Results Bad on LWT
As flaw signals of same type have similar energy features, a certain flaw signal is randomly chon in each type of signals. The energy feature is displayed in Table 1.
From Table 1, it is clearly en that energies of flaw signals are largely focud on the low-frequency bands while little energy is distributed among high-frequency bands. Especially, 99.8% energy of round-hole flaw signal is focud on the low-frequency bands, with no difference among 3 low-frequency bands. Compared with others, energy feature of through hole is distributed disperdly, who 20% energy is distributed in high-frequency bands. The main difference between notch and flat-bottom hole is the energy of high-frequency detail signal: the former is incread as the lifted decomposition procesd; though the latter is incread too, it keeps nearly unchanged at frequency T cd2 and T cd3. The above analysis indicates feature extraction bad on LWT algorithm can effectively distinguish the energy feature of flaw signals.
C.Automatic Flaw Classification Results of BPN
44 flaw echoes are collected by the system, among which 16 echoes are train samples to construct BP network, 28 echoes are test samples. Levenberg-Marquardt algorithm is applied to train the netw
ork. The pre-established error is achieved after 10 cycles, as Fig. 8 shows.
Fig. 8 Training result of the modified BPN
above automatic classification flow process, BP output result is shown in Fig. 9, who four corners reprent 4 flaw types. The square means the actual recognition result whereas the dot is the ideal recognition object. If the square and the dot are overlapping, the classification result is correct. Thus, the classification accuracy can be computed. Here, it is 82.14%. It took 24.06 conds to finish recognizing the input flaw type. Particularly, it can be found in Fig. 9 that the recognition precision of round hole is 100%. This is maybe LWT is fit for processing signals of the round hole, which have more distinct characteristics than other types of flaw signals.
For comparison, Radial Basis Function neural network (RBFN), for its outstanding computation speed, is tested on the same samples too.
D.Automatic Flaw Classification Results of RBFN and Discussion
There are also 3 layers of RBFN. Hidden layer nodes are normally compod of radial function, such as Gauss kernel function. The output layer nodes are simple linear functions. The error convergence curve of RBFN is shown as Fig. 10 and the recognition result is displayed in Fig. 11. The regnition time is 1.43 conds.
Fig. 9 BPN output result
