PHASED ARRAYS
1. Introduction
The ultrasonic phased array (UPA) technique has rapidly become
a new method in the industrial non-destructive testing (NDT) eld
for quality assessment and materials characterisation by providing
the distinct advantages of electronic beam-steering, sector scanning
and dynamic focusing
[1]
. In particular, it has been successfully
applied to automatic defect detection in long-distance pipeline girth
welds
[2]
.
Although many UPA systems have been developed with
capabilities for automatic data acquisition and display of echo
signals, the operation of flaw classification still follows traditional
manual approaches. Typically, the testing technician determines
the type of detected flaw signal according to the echo’s envelope
and certain features, such as echo shape and amplitude level
[3]
. For
example, gas pores produce low-amplitude signals and the echoes
produced by cracks are very directional
[4]
. The operator selects
images and tries to determine the presence of flaws according
to his own observation of visual images. Some researchers have
combined the shape features in an implicit way and derived the
diagnosis
[3]
. As a result of its manual nature, the inspection results
largely depend on the operator’s experience and competence, and
are inherently error-prone.
Flaw classification in ultrasonic phased array testing is one of
the most important, not yet completely solved problems. Knowing
the type of flaw present in the material can give crucial information
for identifying whether the tested material is safe for future
applications and for predicting the remaining life of the tested
structure. To cope with the problems of manual classification, there
is a strong motivation to automate the flaw classification process.
In fact, the automation of flaw classification is a pattern
recognition problem. Significant research has been conducted in
this field. In general, an automatic flaw classification system has
three stages: (1) signal preprocessing; (2) extracting useful features;
and (3) designing a classifier to realise automatic classification
[5]
.
In the first stage, the Gabor filter
[4]
, Fourier transform (FT)
[6,7]
,
Hilbert-Huang transform
[8]
, wavelet transform (WT)
[3,5,9-12]
and the
lifted wavelet transform (LWT)
[13]
have all been used for signal
preprocessing. They were either used to suppress noise or applied to
decompose the input signal. In
[5,6]
, experimental results illustrated
that the WT is a multi-resolution analysis technique and is therefore
more appropriate for analysing non-stationary ultrasonic NDE
signals than the FT
[5]
. The advantage that all operations are carried
out in the spatial domain and are in-line computations motivates
us to choose the LWT as the signal preprocessing method.
In
[14-16]
, it was shown that the LWT has the outstanding characteristics
of high computational efficiency, less memory requirement and
easy implementation by scalable hardware, which is appropriate
for online automatic flaw qualitative analysis.
As for the second stage, there are generally three types of
features: ‘directly obtained’ features
[3,17,18]
, statistical features
[9-11]
and coefficient features of the non-linear transformation
[4,5,10,12,18]
.
For example, the first type includes the time of flight and amplitude
of the A-scan image in the time domain, the second type includes
energy and kurtosis, and the third type includes wavelet coefficients.
Furthermore, the methods of principal component analysis
(PCA)
[4,8]
, the k-means algorithm (KM)
[8,18]
and the PCA-KM
method
[18]
have been used to reduce the size of feature vectors.
Though these features are commonly used, the inherent
characteristics of defect echoes are not considered here. In
general, defects exhibit non-linearity and the interaction between
ultrasonic beams and defects is an extremely non-linear process.
This motivated us to investigate the fractal properties of ultrasonic
echoes. In
[13]
, the fractal dimensions of the defect echo signals were
calculated and the relationship between fractal features and flaw
types was quantitatively established. But in
[13]
, the scaleless range
in which the fractal dimension exists was determined by manual
observation. In this paper, an adaptive subsection search method
based on the average residual sum of squares (RSS) is applied
to determine the scaleless range, improving the computational
precision of the fractal dimension.
For the final stage, various classifiers have been proposed.
Presently, many artificial neural networks (ANNs)
[3,11,15,18-20]
have
been widely used for automatic flaw classification. Moreover,
clustering analysis, including, in particular, the fuzzy c-means
clustering algorithm
[4]
or the k-means clustering algorithm
[7]
, is
another potential solution to automatic classification. In
[11]
, the
k-nearest neighbour method, ANN and the Bayesian statistical
method were compared. In
[19]
, a hybrid neural-fuzzy classifier was
employed to classify weld defects based on the ultrasonic time-
of-flight diffraction (TOFD) technique. However, these approaches
are mainly used for offline analysis or semi-automatic inspection
because ANNs and clustering algorithms are conducted under
the premise that the training samples tend to be infinite. But in
the real-life inspection of pipeline girth welds, the difficulty of
obtaining samples in large quantities acts as a negative factor in
the classification process
[17]
. Classifiers with high classification
precision and high recognition velocity, especially those suitable
for small numbers of samples, should be studied. Therefore, the
support vector machine (SVM) method is tried here to solve this
problem.
The signal processing and pattern recognition techniques used
in the work reviewed above have tremendous potential in ultrasonic
NDT. The present paper is a sequel to
[14-16,20-23]
in automatic flaw
inspection and classification. An UPA system inspects a pipeline
girth weld block and collects the flaw signals. The flaw signals are
decomposed by the LWT. Next, the fractal dimension is computed
and optimal features are selected. A support vector machine (SVM)
is designed to realise automatic classification.
The paper is arranged as follows. Section 2 describes the
experimental set-up. The signal analysis method, LWT, is
Jian Li is with the State Key Laboratory of Precision Measuring Technology
& Instruments, Tianjin University, China.
Xianglin Zhan is with the College of Aeronautical Automation, Civil
Aviation University of China, China.
Shijiu Jin is with the State Key Laboratory of Precision Measuring
Technology & Instruments, Tianjin University, China.
An automatic aw classication method for ultrasonic
phased array inspection of pipeline girth welds
Jian Li, Xianglin Zhan and Shijiu Jin
Submitted 29.08.12
Accepted 27.03.13
DOI: 10.1784/insi.2012.55.6.308
308 Insight Vol 55 No 6 June 2013