Invited Paper Independent Component Analysis Defect Detection For Improved in Guided Wave Monitoring
Invited Paper Independent Component Analysis Defect Detection For Improved in Guided Wave Monitoring
Invited Paper Independent Component Analysis Defect Detection For Improved in Guided Wave Monitoring
Independent Component
Analysis Defect Detection for
Improved
in Guided
Wave Monitoring
This paper applies guided wave methods to detect damage in complex
industrial pipeline structures using independent component analysis
(ICA).
By Jacob Dobson and Peter Cawley
ABSTRACT
|
Guided wave sensors are widely used in a number of industries and have found particular application in the oil and
gas industry for the inspection of pipework. Traditionally this type of sensor was used for one-off inspections, but in recent years
there has been a move towards permanent in- stallation of the sensor. This has enabled highly repeatable readings of the same
section of pipe, potentially allowing im- provements in defect detection and classification. This paper proposes a novel approach
using independent component analysis to decompose repeat guided wave signals into consti- tuent independent components. This
separates the defect from coherent noise caused by changing environmental conditions, improving detectability. This paper
demonstrates independent component analysis applied to guided wave signals from a range of industrial inspection scenarios.
The analysis is per- formed on test data from pipe loops that have been subject to multiple temperature cycles both in undamaged
and damaged states. In addition to processing data from experimental da- maged conditions, simulated damage signals have been
added to ‘‘undamaged’’ experimental data, so enabling multiple dif- ferent damage scenarios to be investigated. The algorithm
has also been used to process guided wave signals from finite
element simulations of a pipe with distributed shallow general corrosion, within which there is a patch of severe corrosion. In all
these scenarios, the independent component analysis algorithm was able to extract the defect signal, rejecting coherent noise.
KEYWORDS
|
General corrosion in pipes; guided wave moni- toring; independent component analysis
I. INTRODUCTION AND RELATED WORK Guided wave sensors are widely used in a number of in- dustries and
have found particular application in the oil and gas industry for the inspection of pipework. This type of sensor
generates a low-frequency ultrasonic wave (G 100 kHz) that travels along the axis of the pipe using the walls as a
waveguide. The wave has full volumetric cov- erage and very low attenuation in steel, making it espe- cially suited
for long-range screening applications. It is possible to inspect a large length of pipework with only a small number
of sensors; a single sensor can routinely inspect more than 50 m of pipe from a single location [1]. This coverage
comes at the cost of lower sensitivity, with commercial guided wave systems typically sensitive to changes in cross
section of around 5% and above in a single inspection [2]. This type of sensor is therefore usually used
Manuscript received February 24, 2015; revised June 18, 2015; accepted June 23, 2015.
in conjunction with another localized, high-accuracy tech-
Date of publication September 15, 2015; date of current version July 15, 2016. This work was supported by BP plc, the U.K.
Research Centre in NDE, and the U.K. Engineering and Physical Sciences Research Council.
nique such as ultrasonic thickness gauging [3] to do follow- up inspections of the areas the guided wave sensor has The authors
are with the Department of Mechanical Engineering, Imperial College of Science, Technology and Medicine, London SW7 2AZ,
U.K. (e-mail: [email protected]).
identified as suspect.
There are several commercial guided wave systems for
Digital Object Identifier: 10.1109/JPROC.2015.2451218
pipework, and in all cases these systems use an array of This work is
licensed under a Creative Commons Attribution 3.0 License. For more information, see
http://creativecommons.org/licenses/by/3.0/
1620 Proceedings of the IEEE | Vol. 104, No. 8, August 2016
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
sensors attached to the outside of the pipe [4]–[6]. These
baseline to compare to a reading [13]. However, in an
in- sensors operate in pulse-echo mode, exciting a controlled
dustrial setting, it is often not possible to collect
multiple wave that travels through the structure. When this wave
signals under prescribed environmental and structural
reaches a change in the impedance of the structure, caused
conditions. by a change in the cross section for
example, some of the
In response to this limitation, the baseline stretch
wave will be reflected back and received by the sensor. A
technique was developed, which seeks to compensate
for trained operator can then assess the shape, amplitude, and
signal changes due to temperature, the most significant
mode of these reflections to infer the nature of the feature
environmental change for guided wave signals [12].
When that caused the reflection. Damage such as corrosion will
the temperature of a pipe changes, the dimensions of
the produce a reflection [7], [8], as will benign pipe features
structure change as does the propagation velocity of the
such as welds and supports [9]. There can also be signal
inspection wave [14]. This causes changes in the arrival
components due to the excitation and reception of un-
time of reflections and leads to imperfect cancellation
wanted modes and imperfect direction control [10]. These
during subtraction. Baseline stretch attempts to
compen- additional components introduce a background noise that
sate for this change through compression or dilation of
the sets the sensitivity of the techniqueVtypically defects
signal and can be applied globally or to windowed
portions must produce a reflection twice as large as the background
of the signal [13], [15]. noise in order to be detected.
These unwanted signals are
This compensation procedure is limited however.
The deterministic and cannot be removed through averaging;
stretch process does not perfectly compensate for
temper- for this reason, they are often known as coherent noise
ature, and its performance deteriorates as the
temperature [10]. Standards exist to guide the collection and interpre-
difference increases. This stretch procedure also ignores
tation of guided wave data [11], limiting the impact of
the influence of other environmental effects, which
have coherent noise as far as possible.
been found to be significant in guided wave systems [16].
In the early stages of development, guided wave sen-
Other researchers have therefore investigated baseline-
sors were typically used in an inspection configuration. A
free methods such as the time reversal method, a
technique sensor would be attached to a section of pipe, an
for defect detection based on reciprocal inspections
inspection made, and then the sensor moved to another
between sensor pairs [17]. Such methods have obtained
location. It is, however, becoming more common to
good results in the laboratory, for example on Lamb
wave permanently bond the sensor to the pipe and operate the
inspection of composite plates [17], but their
effectiveness sensor in a monitoring configuration [2]. In guided wave
on industrial data is still an open question. Others have
techniques, this trend is partly motivated by high access
pursued data-driven approaches that take a pool of
ultra- cost. Pipelines are often buried underground or at
sonic signals and look for meaningful trends within the
elevation, requiring costly digging or scaffolding equip-
data [18], [19]. These data-driven approaches have been
ment for access. In many cases, access to the pipe will also
based on established data mining techniques such as
sup- have health and safety implications. If the sensor is
port vector machines, neural networks, Singular Value
permanently installed, these access costs need only be
Decomposition (SVD), and Principal Component
Analysis incurred once, and an umbilical cord can be run to a safer,
[20]. Of these, SVD is the most developed for guided
wave more convenient location [2].
systems, with several examples in the literature of SVD
The other driver for permanent installation is the po-
being successfully applied to experimental data [16],
[21]. tential for improvements in defect detection and classifi-
In these studies, SVD was able to compensate for real
cation. Because the sensor is bonded to the pipe in a
environmental and operational variations and extract a
permanent installation, it is possible to obtain highly re-
representation of the defect with lower coherent noise
peatable readings of the same section of pipe. If we make
than the original signal. an initial inspection of the pipe
when it is in a known
Independent Component Analysis (ICA) is similar
to structural condition, we can use this information to sub-
SVD in that it seeks to extract meaningful trends in
data. tract out coherent noise from later measurements. This is
The difference between these techniques is that SVD
uses the principle of baseline subtraction where the early mea-
a decorrelation of the data to separate the information,
surement, the baseline, is subtracted from the later mea-
while ICA minimizes the mutual information between
surement, the reading [12].
groupings. ICA has been used alongside neural networks to
This technique is effective only if the change between
identify damaged and undamaged states in vibration
data baseline and reading is a change in the structural condition
from simple truss structures [22], suggesting it may be
of the structure. If there is also a change in the environ-
advantageous to use it with guided wave systems. The
mental conditions surrounding the pipe, this will introduce
purpose of this paper is to investigate the potential of
ICA additional changes that are difficult to distinguish from
for processing the signals from permanently installed
those caused by structural change. This can partly be dealt
guided wave systems. This is achieved by applying ICA
to with by collecting a large number of baselines under differ-
signals that are representative of a range of monitoring
ent environmental conditions and selecting the optimal
conditions that might be faced in industry.
Vol. 104, No. 8, August 2016 | Proceedings of the IEEE 1621
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
1622 Proceedings of the IEEE | Vol. 104, No. 8, August 2016
Section II introduces Independent Component Analy-
When FastICA is applied to guided wave data, the
rows sis and the implementation used in this paper, as well as
of S will be guided wave signals containing the
reflections the procedure for selecting the damage-sensitive feature.
from different features, while the columns of A show
the Section III then introduces the guided wave data that will
amplitude of these signals across the input data. The
col- be passed to the ICA algorithm, and Section IV shows the
umns of A therefore track the amplitude progression of
the results of this analysis. The signals analyzed in Section IV
rows of S, the components. In this paper, all columns of
A all relate to discrete echoes from features with small axial
are scaled such that the maximum value in each column
is extent so echoes are well separated. In Section V, we con-
one, with the amplitude of the rows of S adjusted
accord- sider the more challenging case of general corrosion
ingly. Note that the product of the row of S and the
growth within which there is a region of more severe
weighting function A will give the true amplitude of
that growth. Section VI presents the conclusions of the work
component of the signal. How the amplitude is
apportioned and a discussion of the potential of the technique.
between A and S is a matter of presentational clarity. It is then necessary to know which row(s) of S contain infor-
mation about defect growth. Since the defect location is II. ICA METHODOLOGY
known, we merely select the row that is most similar to the
Independent Component Analysis is a statistical technique for revealing the hidden trends and groupings that under-
lie a set of data. The technique takes a set of multidimen- sional data and transforms it into components that are as
statistically independent as possible [23], [24]. The inten- tion of using ICA on guided wave data is that after
applying the transform, one of the independent components will contain data relating to the defect, while most
coherent noise will be rejected to other independent components;
defect reflection response. However, if the method is to be used in practice with unknown defects, it will be
necessary to produce an automatic method to distinguish components relating to true defect growth from those
relating to random noise or environmental effects. A method based on k-means clustering has been proposed in the
literature [21], but so far it has only been proven on step-type defect growth. One potential alternative is change
detection using the gener- alized likelihood ratio, as reported in [28].
so giving a clearer representation of the defect. Numerous implementations of ICA are available [25], [26], but the
FastICA algorithm [23] is used in this study because it
III. DATASETS
achieves similar performance to other algorithms but at greater speed [27]. A brief overview of the FastICA method
is given here, but the interested reader can find full details in [23].
The FastICA method assumes that the input data X can be described by the model
A. Laboratory Experiment
The first set of data comes from an experiment with a 6 m length of 8-in schedule 40 piping that was temper- ature
cycled in the laboratory. The purpose of this expe- riment was to collect commercial quality guided wave signals
from a length of plain pipe experiencing known changes in environmental condition. Since temperature is X 1⁄4 AS
(1)
the main contributor to signal changes in guided wave inspection, this was the condition that was varied. With a
sufficient pool of data collected in the undamaged condi- where S is a statistically independent representation of the
tion, the second stage was to introduce a small point
defect data and A is a matrix of scalar values that relates S to X.
and see how this changes the received signal. In the
case of guided wave analysis, X is an 1⁄2m  n matrix
The pipe was heated via a resistive heating element
that of guided wave signals collected from a sensor. Each of the
was suspended in the center using sheet metal inserts.
In m rows is a guided wave response signal as a function of
this way, the element is prevented from contacting the
time (which can be converted to distance knowing the
pipe wall while the sheet metal inserts produce
negligible wave velocity), with each of the n columns containing the
reflection during the guided wave inspection. A
commer- amplitude of that signal at a certain time. The signals that
cial guided wave sensor (gPIMS unit manufactured by
are the rows of S will be the new representation we want
Guided Ultrasonics, Ltd.) was then installed on the
outside in which the defect signal and the coherent noise have
of the pipe, 2 m from one end as well as seven thermo-
been separated into different components. Independent
couples distributed across the outer surface. This
assembly component analysis allows us to find S by performing the
was then wrapped in an insulating material. A diagram
of operation S 1⁄4 WX, where W is the inverse of the mixing
this setup is given in Fig. 1. matrix A. To calculate W,
the algorithm uses the property
Using the thermocouples, the heating element and
an that the signals in S must, by definition, be maximally
appropriate feedback control loop the temperature of
the statistically independent. The algorithm iterates through
pipe was varied. The pipe was heated to 90 C, held to
possible values of W to find the matrix that minimizes
allow the temperature to equalize throughout the pipe,
mutual information between the rows of S.
and then cooled to 30 C. During this cooling
cycle, 42 measurements were made using the gPIMS sensor in approximately equal steps of time. The measure-
ments were made using an eight-cycle hanning window toneburst, center frequency 31 kHz, T(0,1) mode. This
process was repeated for six temperature cycles to give a total of 252 measurements across a range of temperatures.
An example measurement from the sensor is given in Fig. 2, where only the forward direction is shown (forward
direction indicated by arrow in Fig. 1). Note that we consider even those parts of the signal that indicate a distance
further than 4 m. These are not caused by physical features beyond the end of the pipe, but arise due to
reverberations between the pipe ends. These signal components serve as a useful model for a longer pipe with
multiple features, which is how they are used in this study.
After this initial period of temperature cycling, the insulation was removed, and a flat bottom hole was drilled
2.3 m from the gPIMS sensor. This hole had a depth of 4 mm and a diameter of 7 mm, representing a cross-sectional
change of 0.5% at its maximum extent. The insulation was then reapplied, and the same process used
Fig. 2. Example pulse-echo measurement from the laboratory experiment, showing reflected T(0,1) mode in the forward direction
(forward direction indicated by arrow in Fig. 1). Sensor located at the origin, with reflections caused by (a) end reflection, (b)
imperfect direction control, (c) ring reflection, and (d) first reverberation.
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
to collect readings from the damaged pipe between 90 C and 30 C.
B. Pipe Loop Experiment
The second set of data comes from a pipe loop installed at an industrial test site. The purpose of this experiment
was to collect commercial quality guided wave signals from a sec-
Fig. 1. Equipment for the laboratory experiment: NPS8 pipe with internal heating element and guided wave sensor (gPIMS)
permanently attached. Note the noncontacting heating element
tion of pipe with features such as welds, bends, and flanges. The other purpose of the experiment was to understand
how these signals change when defects are introduced in both through the center of the pipe. Not to scale.
plain sections of pipe and at welds, before and after bends. Experiments were performed on a section of pipework
made from NPS8 schedule 40 piping. At one end, the pipework begins with a blanked flange, connected to a straight
section of pipe 5.79 m long. A weld then joins this to another section of straight pipe that is 6.82 m long. A Guided
Ultrasonics, Ltd., gPIMS unit was permanently attached to this length of pipe, 11.17 m from the flange. This in turn
is connected by a weld to a 90 bend, beyond which there is a further section of straight pipe that extends beyond the
region of interest. The pipe is a closed system filled with a liquid whose temperature can be controlled; in this way,
the temperature of the pipe could be controlled during the experiment. This setup is shown in Fig. 3.
Initially, the temperature of the pipe loop was set to 38 C, and four measurements were made under the no-
damage condition. The measurement was made using an eight-cycle hanning window toneburst, center frequency 31
kHz, T(0,1) mode. The temperature of the loop was then ramped to 90 C and allowed to cool again to 38 C. The
purpose of this ramping was to cause thermal stress in the sensor and study the effect of sensor changes due to envi-
ronmental cycling. Three 1-in-diameter flat bottom holes were then drilled to give defects A, B, and C indicated in
Fig. 3. Defect A was drilled in a clean section of pipe and should be the easiest to detect. Defect B was drilled adja-
cent to a weld and will be more difficult to detect due to masking by the weld reflection. Defect C is adjacent to a
weld and beyond a bend, making it the most difficult to detect due to masking and the fact that bends tend to exa-
cerbate environmental effects. Initially, the holes in these three locations were drilled to a depth that gives a 0.25%
cross-sectional change. Four measurements were made using the sensor, and the loop is again ramped to 90 C and
Vol. 104, No. 8, August 2016 | Proceedings of the IEEE 1623
Fig. 3. Equipment for the pipe loop experiment: feature-rich pipe with welds and bends, filled with temperature controlled fluid
and with a guided wave sensor (gPIMS) permanently attached. Not to scale.
cooled to 38 C. The holes were enlarged to 0.5% cross- sectional change, and another four measurements made.
This process was repeated a final time for a hole with 0.75% cross-sectional loss. An example output from the
sensor is given in Fig. 4, with positive direction indicated by the arrow in Fig. 3.
C. Data Post-Processing
It is possible to apply ICA directly to the signals from the laboratory and pipe-loop experiments. However, it was
found that the ICA performed better if the signals were first partly compensated for environmental changes. For the
results presented in this paper, each signal has been compensated using a stretch algorithm. This stretch in- volves
dilation or compression of the signal in the time domain, an amplitude scaling, and a delay, the purpose being to
bring the signal as close as possible to a reference signal. This manipulation was performed in the frequency domain
and in this study was achieved using a simulated annealing optimization to optimize across the three va- riables
simultaneously. Note that other implementations are available [13], [15]. For both the laboratory experiment and
pipe loop experiment, a suitable reference signal had to be chosen to stretch to. This reference was arbitrarily chosen
as the signal with a temperature closest to the mean.
IV. RESULTS OF APPLYING ICA
A. Laboratory Experiment With Simulated Simple Step Defect
The first test of the algorithm was to see if it could identify a simple step change in a pipe undergoing temper-
ature changes. This scenario was modeled by taking data from the undamaged pipe in the laboratory experiment and
adding a signal that models the reflection from a point defect. This first test looked at three scenarios: case A that
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
1624 Proceedings of the IEEE | Vol. 104, No. 8, August 2016
uses 20 experiment signals collected in the temperature range 60 C–63 C, case B that uses 20 signals collected in the
temperature range 60 C–70 C, and case C that uses 10 signals in the temperature range 60 C–70 C and 10 signals in
the temperature range 71 C–80 C. The spe- cific temperatures for each case are given in Table 1. Note that these
temperatures do not come directly from ther- mocouple measurements, but have been inferred from the dilation or
compression between signals within a known temperature range. In all three cases, the first 10 signals were left in
the undamaged condition, while the second 10 signals had another signal added. This signal is a model
Fig. 4. Example pulse-echo measurement from the pipe loop experiment showing reflected T(0,1) mode in the forward and
of the reflection from a point defect and is simply an am- plitude scaled version of the excitation signal used by the
backward directions (forward indicated by arrow in Fig. 3). Sensor
sensor: an eight-cycle hanning window toneburst, center
located at the origin (0 m), with reflections caused by (a) flange,
frequency 31 kHz. This signal was added at a constant (b)
weld and defect B, (c) defect A, (d) weld before bend, and (e) weld after bend and defect C.
amplitude of 0.75%, thus simulating a simple step change in the structural condition of the pipe. The amplitude is
given here as a percentage of the amplitude of the largest reflector in the signal (the first end reflection). This signal
was added at a distance of 7.1 m from the transduction, remote from the major feature reflections. Note that this is
not a physically possible damage location in the 6 m pipe, but instead is a portion of the signal that arises due to
reverberation between the pipe ends.
Some of the outputs from the ICA algorithm for case A are shown in Fig. 5. Parts (a) and (b) are the signal (row
of matrix s) and weighting function (column of matrix A) for the defect signal, showing a defect signal that is clearly
separated from coherent noise and a weighting function that correctly tracks the true amplitude of the defect [dotted
line in part (b)]. Parts (c)–(f) show some of the other output signals from the algorithm. Parts (c) and (d) show the
signal and weighting for end reverberations and ring reflections, these parts having constant weighting
Table 1 Temperature Distribution of Signals Across Three Example Cases: A, B, and C
Fig. 5. Output from the ICA algorithm for a simulated step defect in the laboratory experiment (temperature case A), showing: (a)
the recovered defect signal and (b) the associated weighting function (solid line is the weighting calculated by ICA, while the
dotted line is the true weighting [i.e., the known weighting of the added defect]), (c) the major feature reflections and (d) the
associated weighting function, (e) one component of the noise introduced from imperfectly compensated environmental effects
and (f) the associated weighting function.
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
because these parts of the signal do not change across the input signals. Parts (e) and (f) show the signal and weight-
ing for one of the components of noise introduced by im- perfectly compensated environmental effects. The noise
component with the largest amplitude has been shown, but the ICA algorithm outputs a total of six noise signals for
this set of inputs. These noise signals all have similar oscillating weighting functions, but their maximum amplitude
varies from 2% in the case shown in Fig. 5(e) to a minimum of 0.4%. Although in some cases the noise components
have a larger amplitude than the defect signal, their weighting function is different to that of the defect. This
difference in weighting function may be one way of distinguishing be- tween damage and noise components.
The ICA output for temperature case B is shown in Fig. 6(a) and (b), where only the defect signal has been
shown. The other outputs of the algorithm are similar to those shown in Fig. 5 although there are now 10 noise
Vol. 104, No. 8, August 2016 | Proceedings of the IEEE 1625
Fig. 6. Example output of the ICA algorithm for simulated step defects in the laboratory experiment, showing: (a) defect signal
for temperature case B and (b) the associated weighting function, (c) the defect signal for temperature case C and (d) the
associated weighting function.
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
components due to imperfectly compensated environmen- tal variations. These noise components have a maximum
amplitude that varies from 4.0% to 0.2% across the 10 noise outputs. This increase in the number of components is
due to the greater temperature difference among the input signals, which causes an overall increase in the
background noise. These noise components all have nonmonotonic weighting functions that vary in a pattern similar
to that shown in Fig. 5(f). Similarly, the defect signal for tem- perature case B has more coherent noise than
temperature case A. This is again thought to be because there is greater temperature difference among the input
signals, and hence the algorithm has greater difficulty separating the defect signal. The ICA output for temperature
case C is shown in Fig. 6(c) and (d), where only the defect signal has been shown. As before, the other outputs of the
algorithm are similar to those shown in Fig. 5, although there are now 14 noise components in the ICA output. These
noise compo- nents have a maximum amplitude that varies from 6.0% to 0.1%, with the maximum noise occurring
at the end re- flections. Again, the weighting functions for these noise components are nonmonotonic, varying in
patterns similar to those shown in Fig. 5(f). The noise on the defect signal is also higher than temperature case A or
B, again because there is greater temperature difference on the input signals. It is likely that greater temperature
variations would lead to increased noise and deteriorating performance of the ICA, and this is now being
investigated.
B. Laboratory Experiment With Simulated Growth of a Flat Bottom Hole
The second test of the algorithm was to look at a more realistic model for defect growth. Instead of modeling a
simple step change, this test modeled the reflection from a flat bottom hole with increasing cross-sectional loss over
time, but a constant depth-to-diameter ratio. This is a model of, for example, growing localized corrosion. This
scenario was modeled by taking 20 signals from the undamaged pipe in the laboratory experiment and adding a
defect signal to the final 10. For the sake of brevity, only temperature case B is presented in this paper, where all
signals have a temperature within the range 60 C–70 C (see Table 1 for details). The added defect signals were
generated using the analytical model of Cegla et al. [29]. Using this model the reflection coefficient versus
frequency behavior of 10 different flat bottom holes was calculated. Each hole had a depth-to-diameter ratio of 1:2
and repre- sented a cross-sectional loss of between 0.1% and 1.0% in the 8-in pipe. These reflection coefficient
curves are shown in Fig. 7. To generate a signal to add to the laboratory data, the frequency spectrum of the
inspection signal (eight- cycle hanning window toneburst, 31 kHz center frequency) was multiplied by these curves.
Transforming this data back into the time domain gave a signal with the correct am- plitude and frequency content
for a reflection from each hole. Note that as the hole size increases, the frequency
1626 Proceedings of the IEEE | Vol. 104, No. 8, August 2016
Fig. 7. Reflection coefficient versus frequency behavior for a flat bottom hole. Values along each curve are the percentage
cross-sectional change represented by each curve. All holes have a constant depth-to-diameter ratio of 1:2.
content of reflection changes, so the defect signal is no longer simply an increasing amplitude of the same signal. It
is of interest to see whether ICA can still identify this hole as a single component at all hole sizes. Although this
model is for flat bottom holes in plates, the ratio of pipe radius to hole radius was large enough to justify this
approximation. This signal was again added at a distance of 7.1 m, away from major feature reflections.
The results from the ICA algorithm are given in Fig. 8, which shows the defect signal and associated weighting
function. The figure shows that the reflection from the
Fig. 8. Output from the ICA algorithm for a simulated flat bottom hole grown in stages in the laboratory experiment data,
showing: (a) defect signal and (b) the associated weighting function. Solid line is recovered weighting function, while dotted line
is the true. Signals have the temperature variation described by case B.
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
defect can be identified and the weighting function cor- rectly tracks the true weighting function [dotted line in Fig.
7(b)], albeit with some noise. The other outputs from the algorithm are similar to those shown in Fig. 5, with a total
of 10 noise outputs in this case. The maximum am- plitude of these noise components varies from 3.7% to 0.4%,
and their weighting functions vary nonmonotoni- cally in a pattern similar to Fig. 5(f).
C. Laboratory Experiment With Drilled Flat Bottom Hole
The previous tests with laboratory data have used sig- nals from the undamaged pipe with added artificial defect
signals. This is a reasonable approach given that the inter- action of guided waves with defects in pipes is well
under- stood [29]. However, there may be some subtlety that arises with a physically introduced defect that is not
captured in this previous approach. For that reason, the next test of the ICA algorithm involved data from both the
undamaged and damaged pipe in the laboratory experiments. Thirty signals were selected, 15 from a pipe in the
undamaged condition and 15 from a pipe containing a 7-mm-diameter, 4-mm- deep flat bottom hole; such a hole
represents a cross- sectional loss of 0.75% This hole was drilled 2.3 m from the sensor. These 30 signals were
collected in the temperature range 60 C–70 C, with the temperatures covering this full range. These signals were
then passed to the ICA algorithm, although note that only the first 4 m of the signals were processed to keep the
analysis physical.
The results of the ICA algorithm are given in Fig. 9, which shows: (a) the reflection from the flat bottom hole (at
a distance of 2.3 m), and (b) the associated weighting function. The algorithm has extracted the defect signal with a
fairly low level of coherent noise given the size of the
Fig. 9. Result of applying the ICA algorithm for a drilled flat bottom hole in the laboratory experiments, showing: (a) the
reflection from the flat bottom hole, beginning at 2.4 m, and (b) the associated weighting function (dotted line is true). All signals
at 60 C–70 C.
Vol. 104, No. 8, August 2016 | Proceedings of the IEEE 1627
defect and has accurately tracked the amplitude growth of the signal. Note that in this case the algorithm is trying to
extract a defect reflection of around 0.75%, whereas in one off guided wave nondestructive testing, the operator
would typically look to detect changes of 5% or greater. The other outputs of the ICA algorithm are similar to those
shown in Fig. 5, with a total of 12 noise components. These noise components have a maximum amplitude that
varies from 1.4% to 0.2%, and again the corresponding weighting func- tions vary nonmonotonically in a pattern
similar to Fig. 5(f).
D. Pipe Loop Experiment With Physical Flat Bottom Holes
The last analysis on signals with discrete echoes looks at the reflection from holes physically introduced in the
pipe loop experiment. As with the previous test, the defects in this case have been physically drilled in the pipe. The
difference compared to the previous case is that these holes have been grown in stages and have been drilled not
only in clean sections of pipe, but also close to welds and beyond bends. Fourteen signals were selected, with this
group representing four different damage conditions: four signals in the undamaged condition, four signals with
holes repre- senting a 0.25% cross-sectional change, four signals with holes representing a 0.50% cross-sectional
change, and three signals representing a 0.75% cross-sectional change. All of these signals were collected at a
temperature of 38 C, but with temperature cycling in between each da- mage condition to simulate ageing of the
sensor and hence possible changes in sensitivity or frequency response.
The analysis of this data is split into two: the data from the ‘‘forward’’ direction looking towards the bend and
the data in the ‘‘backwards’’ direction that looks towards the flange (forward direction indicated by arrow in Fig. 3).
Looking in the backwards direction first [Fig. 10(a) and (b)], the ICA algorithm extracts defects A and B, which are
in a clean section of pipe and close to the weld, respectively. Both defects occur on the same output because the
holes were drilled at the same time. All signals therefore either have neither or both defects present. Although the
defects have the same cross-sectional loss, they have different amplitudes in the output from the ICA. It is thought
that the amplitude of defect B is underestimated because of the difficulty in separating it from the reflection from the
neighboring weld. The weighting function [Fig. 10(b)] correctly tracks the known weighting function of the de- fect.
The other outputs from the ICA algorithm are the unchanging weld and flange reflections and 12 noise com-
ponents. These noise components have a maximum ampli- tude between 4% and 0.2% and weighting functions that
vary nonmonotonically in patterns similar to Fig. 5(f). In the forward direction [Fig. 10(c) and (d)], the algorithm
extracts the reflection from the drilled hole with little co- herent noise. The algorithm also correctly tracks the
weighting of the defect, albeit underestimating the ampli- tude when the defect is small. The fact that the weighting
function is poorer in the forward direction than the back- wards direction is to be expected: A defect at a weld
beyond a bend is a challenging inspection scenario, especially for this size of defect. The other outputs from the ICA
algo- rithm are the unchanging reflection from the welds before and after the bend and 12 noise components. These
noise components have a maximum amplitude between 4% and 0.2% and weighting functions that vary
nonmonotonically in a pattern similar to Fig. 5(f).
V. SIMULATED GENERAL CORROSION
A. Method
The analysis so far has focused on discrete echoes from features with small axial extent. The purpose of this
section is to consider the case of interacting echoes from distri- buted general corrosion. Specifically, we are
interested in whether the ICA algorithm is able to identify the reflection from a region of deep corrosion growing
within a patch of
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
1628 Proceedings of the IEEE | Vol. 104, No. 8, August 2016
shallower corrosion. Finite element simulations were used to study this issue given the difficulty of preparing long
sections of pipe with a prescribed surface profile.
The finite element model consisted of a 3.0-m-long section of pipe, 168-mm outer diameter, and a wall thick- ness
of 7 mm (NPS 6 schedule 40). This model was dis- cretized into tetrahedral elements with a characteristic length of
2 mm using the Netgen algorithm [30]. The ele- ments were then given the properties of steel except for the first 0.3
m that was set up as an absorbing region following Petit et al. [31]. This absorbing region attenuates all waves
entering it to negligible amplitude. The inner surface of the pipe from 0.7 to 2.7 m was then adjusted to follow a
given corrosion profile. The corrosion profile was generated separately as a grid 2.0 m by 0.484 m (inner
circumference of the pipe) broken into 2-mm elements. Each point in the grid was given a value taken from a
Gaussian distribution with desired mean and given standard deviation. An analysis of industrial corrosion patches
indicates that the Gaussian distribution is a suitable model for distributed general corrosion [32]. This surface was
then multiplied by a 3D-Gaussian to give the surface a specific correlation length in both directions, and all heights
greater than zero are set to zero (no increase in pipe thickness). The nodes of the pipe model were then adjusted to
match this corrosion profile, with smoothing applied in the radial direction to reduce mesh distortion. Further mesh
smoothing was applied using the mesh smoothing function in Gmsh [33]. A uniform tangential force was then
applied to a ring of nodes on the outside surface of the pipe, 0.4 m from one end. This changing force generated a
T(0,1) mode, eight-
Fig. 10. Results of applying ICA to pipe loop data showing: (a) defect A and B, (b) the associated weighting function, (c) defect
C, and (d) the associated weighting function. All signals collected at 38 C.
cycle hanning window toneburst, center frequency 31 kHz. This model was then stepped in time using POGO, a
finite element solver based on the GPU [34]. This software is two orders of magnitude faster than similar
commercial pack- ages and enables the current work [34]. The T(0,1) mode reflection from the corrosion patch was
then measured at the location where the T(0,1) mode was generated. This model is shown in Fig. 11.
B. Dataset
Initially, 10 random corrosion surfaces were generated having a correlation length of 5 mm and a root mean
square depth of 0.1 mm. These shallow surfaces were used to generate a low level of background noise to model the
low levels of coherent noise observed when inspecting a nominally clean section of pipe. A further 10 surfaces were
Fig. 11. Geometry of the model used in general corrosion simulations. Not to scale.
then generated to model a patch of general corrosion gradually increasing in depth and severity.
The first surface was a Gaussian random surface with a correlation length of 5 mm and a root mean square depth
of 0.15 mm. However, a small region of this surface was given a root mean square depth of 0.3 mm. This region
extended from 0.95 to 1.05 m along the axis of the pipe and around half of the circumference, with Gaussian
smooth- ing applied at the interface with the rest of the corrosion surface to prevent discontinuities.
To generate the second surface, a perturbing surface was then added to this original surface. The perturbing
surface had an increased correlation length and root mean square depth and again had a small region from 0.95 to
1.05 m where the corrosion was deeper. This perturbing surface used a random number grid that was a 50:50
weighting of the previous random number grid and a new random number grid. This process of taking the previous
surface and perturbing it was continued nine times to give a total of 10 random surfaces with increasing mean depth
and correlation length. The correlation length of the perturbing surface was increased from 5 to 50 mm, and the
mean depth was increased from 0.1 to 0.5 mm across the nine surfaces. This gave the final corrosion surface shown
in Fig. 12, where the shallow general corrosion had a maximum depth of 1.5 mm, while the deep corrosion pits had
a maximum depth of 3 mm. Note that these are absolute limits on the depth; outlier regions deeper than this have
been set to the maximum depth. Although convoluted, this process of perturbing previous surfaces was preferable to
generating completely new random surfaces for each increment. This is because a perturbation more closely
resembles the physi- cal corrosion process. In the real corrosion of pipework, each new corrosion surface must
necessarily be an incre- ment of the corrosion surface that preceded it.
The amplitude of T(0,1) reflection from the general corrosion surface in Fig. 12 is shown as the solid black line
in Fig. 13. This reflection amplitude is shown alongside the cross-sectional loss represented by the corrosion. As ex-
Fig. 12. Thickness map of the final state of the simulated corrosion patch.
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
pected, there is broad agreement between the two varia- bles, although the agreement will not be perfect due to the
complex constructive and destructive interference between overlapping waves. This figure shows clearly the
increased reflection from the more severely corroded section of pipe. Note that this figure does not show the
reflection from the end of the pipe (2.6 m from the monitoring point) that dominates the received signal.
C. Results of Applying ICA Algorithm
Some example outputs from the ICA algorithm are given in Fig. 14, where the input to the ICA algorithm is the
first 3.0 m of the signals from the model. Fig. 14(a) and (b) shows the signal and weighting function for the severely
corroded section of pipe. This signal contains a major re- flection from the severely corroded region and is the only
signal of this type in the ICA output. This is a very encour- aging result, showing that ICA tracks the severe
corrosion, keeping it in single component even though the corrosion is growing at more than one point and its shape
is changing. Fig. 14(c) and (d) shows the signal and weighting function for the pipe end reflection. This weighting
function cor- rectly shows a decrease in the amplitude of the reflection across the final 10 signals as the corrosion
grows. This happens because the increase in corrosion causes a reduc- tion in the T(0,1) mode reaching the pipe end.
Fig. 14(e) and (f) shows the signal and weighting function for one of the signal components due to the shallow
general corro- sion. Note that the algorithm separates the reflection from the shallow corrosion into 14 components,
with the largest amplitude component shown in Fig. 14. The other com- ponents have a maximum amplitude from
2.4% to 0.6%, and these components constructively and destructively in- terfere to give the total reflection seen in
Fig. 13.
VI. CONCLUSION In this paper, we have discussed the use of Independent Component Analysis for improved
defect detection in guided
Vol. 104, No. 8, August 2016 | Proceedings of the IEEE 1629
Fig. 13. Overlay of cross-sectional loss and reflection coefficient for the corrosion patch. Note that the corroded surface starts at a
distance of 0.3 m from the monitoring point.
wave monitoring. Through a combination of the traditional baseline stretch approach and the novel use of ICA, we
are able to extract a representation of the defect that has lower
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
1630 Proceedings of the IEEE | Vol. 104, No. 8, August 2016
coherent noise than the original signal. The algorithm can also be used to accurately track the amplitude history of
the defect and hence understand its growth pattern. This is in the context of changing environmental conditions.
This paper has looked at experimental data from both a plain section of pipe and a pipe loop containing welds,
bends, and flanges. Data were collected from these pipes while they were in their undamaged and damaged con-
dition and undergoing temperature variations. Initially model defect signals were added to the data from the un-
damaged plain pipe to simulate step changes and gradual defect growth. The ICA algorithm was able to separate
these defect signals from coherent noise in the context of changes in temperature and changing frequency content of
the defect signal. The ICA algorithm was then applied to data from the undamaged and damaged pipes, and again
the algorithm was able to separate the defect signal from co- herent noise. This was for physically introduced defects
in clean sections of pipe, at welds, and at welds beyond bends. This paper also used finite element simulations to
inves- tigate generally corroded sections of pipe with a patch of severe corrosion. It was found that the ICA
algorithm cor- rectly separated the reflection due to the severe corrosion and recognized it as a single evolving
defect. This is in the context of the corrosion undergoing a complex growth pattern where its shape and amplitude is
changing.
It therefore appears that ICA is a promising method for the automated processing of guided wave data. The next
stage is to look at ways of automating identification of the component containing information about the defect. The
ICA algorithm outputs several components, some of which are noise and some of which contain useful information.
A method must be developed for automatically identifying the components containing useful information, and work
is in progress on this issue. One promising approach is to use the weighting functions of the components; in all the
cases reported here, the noise components have strongly nonmonotonic weighting functions, while the defect sig-
nals have near-monotonic weighting functions. The ICA technique is also only one of several possible techniques
for processing guided wave monitoring data. A quantitative comparison of the different techniques would indicate
when each method is most appropriate. Such a comparison is in preparation and will be reported in the literature
shortly. h
Acknowledgment Fig. 14. Results of applying ICA to the
general corrosion signals, showing: (a) the reflection from the severely corroded region and (b) the associated weighting function,
(c) the end reflection and (d) the
The authors are grateful to the UK Engineering and Physical Sciences Research Council for funding an Engi-
associated weighting function, and (e) one of the components of the
neering Doctorate studentship for J. Dobson, and to BP for
general shallow corrosion and (f) the associated weighting function.
supporting the project and providing access to the test loop. The authors are also grateful to Dr. A. Galvagni for
collecting test data, and Guided Ultrasonics, Ltd., for pro- viding the equipment and assistance in performing the
laboratory experiments. Supporting data is available on request: please contact [email protected].
Dobson and Cawley : Independent Component Analysis for Improved Defect Detection in Guided Wave Monitoring
REFERENCES
[14] R. Weaver and O. Lobkis, ‘‘Temperature
[24] A. Hyvarinen, J. Karhunen, and E. Oja,
[1] M. Lowe and P. Cawley, ‘‘Long range guided
wave inspection usageVCurrent commercial capabilities and research directions, Imperial College London, Tech. Rep. [Online].
Available: http://www3.imperial.ac.uk/pls/ portallive/docs/1/55745699.PDF [2] P. Cawley, F. Cegla, and A. Galvagni, ‘‘Guided
waves for NDT and permanently-installed monitoring,’’ INSIGHT, vol. 54, pp. 594–601, 2012. [3] P. Cawley, F. Cegla, and M.
Stone, ‘‘Corrosion
monitoring strategiesVChoice between area and point measurements,’’ J. Nondestructive Eval., vol. 32, pp. 156–163, Jun. 2013.
[4] Guided Ultrasonics, Ltd. [Online]. Available:
http://www.guided-ultrasonics.com [5] Plant Integrity, Ltd., ‘‘Teletest focus,’’
Jan. 2015. [Online]. Available: http://www. plantintegrity.com/teletest/teletest-focus/. [6] Guided Wave Analysis LLC. [Online].
Available: http://www.gwanalysis.com
dependence of diffuse phase fields,’’ Ultrasonics, vol. 38, pp. 491–494, Mar. 2000. [15] T. Clarke, F. Simonetti, and P. Cawley,
‘‘Guided wave health monitoring of complex structures by sparse array systems: Influence of temperature changes on
performance,’’ J. Sound Vibration, vol. 329, pp. 2306–2322, Jun. 2010. [16] C. Liu, J. Harley, M. Berges, D. Greve, and
I. Oppenheim, ‘‘Robust ultrasonic damage detection under complex environmental conditions using singular value
decomposition,’’ Ultrasonics, vol. 58, pp. 75–86, Apr. 2015. [17] H. Park, H. Sohn, K. Law, and C. Farrar,
‘‘Time reversal active sensing for health monitoring of a composite plate,’’ J. Sound Vibration, vol. 302, pp. 50–66, Apr. 2007.
[18] Y. Lu and J. Michaels, ‘‘Feature extraction and sensor fusion for ultrasonic structural health monitoring under changing [7]
R. Carandente and P. Cawley, ‘‘The scattering
environmental conditions,’’ IEEE Sensors J., of the fundamental torsional mode
from
vol. 9, no. 11, pp. 1462–1471, Nov. 2009. axi-symmetric defects with varying depth
profiles in pipes,’’ J. Acoust. Soc. Amer., vol. 127, pp. 3440–3448, Jul. 2010.
Independent Component Analysis. New York, NY, USA: Wiley Interscience, 2001. [25] A. Delrome and S. Makeig, ‘‘EEGLAB:
An
open source toolbox for analysis of single analysis EEG dynamics including independent component analysis,’’ J. Neurosci.
Methods, vol. 134, pp. 9–21, 2004. [26] P. Hojen-Sorenson, O. Winther, and
L. K. Hansen, ‘‘Mean field approaches to Independent Component Analysis,’’ Neural Comput., vol. 134, pp. 889–918, Jan. 2002.
[27] D. Langlois, S. Chartier, and D. Gosselein,
‘‘An introduction to Independent Component Analysis: InfoMax and FastICA algorithms,’’ Tutorials Quantitative Methods
Psychol., vol. 6, pp. 31–38, 2010. [28] A. Galvagni and P. Cawley, ‘‘Reliable
identification of damage growth using guided wave SHM systems,’’ in Proc. 7th Eur. Workshop Structural Health Monitoring,
Nante, France, 2014, pp. 1720–1727. [29] F. Cegla, A. Rohde, and M. Veidt, ‘‘Analytical
prediction and experimental measurement for mode conversion and scattering of plate [19] C. Haynes, M. Todd, E. Flynn, and
waves at non-symmetric circular blind holes A. Croxford, ‘‘Statistically-based
damage
in isotropic plates,’’ Wave Motion, vol. 45, detection in geometrically-complex
structures
pp. 162–177, May 2008. [8] A.
Lovstad and P. Cawley, ‘‘The reflection
using ultrasonic interrogation,’’ Structural of the fundamental torsional guided wave
Health Monitoring, vol. 12, pp. 141–152, from multiple circular holes in pipes,’’
NDT
Mar. 2013. & E Int., vol. 44, pp. 553–562, Nov. 2011.
[20] D. Tibaduiza, L. Mujica, and J. Rodellar,
‘‘Damage classification in structural health monitoring using principal component analysis and self-organizing maps,’’ Structural
Control Health Monitoring, vol. 20, no. 10, pp. 1303–1316, Oct. 2013. [21] C. Liu et al., ‘‘Singular value decomposition
for novelty detection in ultrasonic pipe monitoring,’’ in Proc. Sensors Smart Structures Technol. Civil, Mech., Aerosp. Syst., J. P.
Lynch, C.-B. Yun, and K.-W. Wang, Eds., San Diego, CA, USA, 2013, vol. 8692. [22] C. Zang, M. Friswell, and M. Imregun,
‘‘Structural damage detection using independent component analysis,’’ Structural Health Monitoring, vol. 3, no. 1, pp. 69–83,
Mar. 2004. [23] A. Hyvarinen, ‘‘Fast and robust fixed-point
algorithms for independent component analysis,’’ IEEE Trans. Neural Netw., vol. 10, no. 3, pp. 626–634, May 1999.
Vol. 104, No. 8, August 2016 | Proceedings of the IEEE 1631
[30] J. Schoberl, ‘‘Netgen, an advancing front 2D/3D-mesh generator based on abstract rules,’’ Comput. Visual. Sci., vol. 1, pp.
41–52, 1997. [9] A. Galvagni and P. Cawley, ‘‘The reflection of guided waves from simple supports in pipes,’’ J. Acoust. Soc.
Amer., vol. 129, pp. 1869–1880, 2011.
[31] J. Pettit, A. Walker, P. Cawley, and M. Lowe,
‘‘A stiffness reduction method for efficient absorption of waves at boundaries for use in commercial finite element codes,’’ [10]
D. Thompson and D. Chimenti, Eds., Practical
Ultrasonics, pp. 1868–1879, 2014.
Long Range Guided Wave InspectionVManaging Complexity, vol. 657. College Park, MD, USA: American Institute of Physics,
2003.
[32] M. Stone, ‘‘Wall thickness distributions
for steels in corrosive environments and determination of suitable statistical analysis [11] D. Alleyne and A. Demma, ‘‘Guided
waves:
methods,’’ in Proc. 4th Eur.-Amer.
Workshop Standardization and certification,’’ in Proc.
Rel. NDE, 2009, pp. 1–23. 6th Int.
Conf. Certification Stand. NDT, Jun. 2011, pp. 27–32.
[33] C. Geuzaine and J. F. Remacle, ‘‘Gmsh: A three-dimensional finite element [12] A. Croxford, P. Wilcox, and B.
Drinkwater,
mesh generator with built-in pre- and
‘‘Strategies for guided-wave structural health
post-processing facilities,’’ Int. J.
Numer. monitoring,’’ Proc. Royal Soc. A, vol. 463,
Methods Eng., vol. 79, pp. 1309–1331,
2009. pp. 2961–2981, Aug. 2007.
[34] P. Huthwaite, ‘‘Accelerated finite
element [13] Y. Lu and J. Michaels, ‘‘A methodology for
elastodynamic simulations using the
GPU,’’ structural health monitoring with diffuse
J. Comput. Phys., vol. 257, pt. A, pp.
687–707, ultrasonic waves,’’ Ultrasonics, vol. 43,
2014. pp. 717–731, Oct. 2005.
ABOUT THE AUTHORS
Jacob Dobson received the M.Eng. degree in mechanical engineering from Imperial College London, London, U.K., in 2012, and
is currently pursuing the engineering doctorate in non-destructive evaluation at the UK Research Centre in NDE (RCNDE),
Imperial College London.
His research interests are in guided waves for structural health monitoring and advanced signal processing techniques thereof.
Peter Cawley received the B.Sc. and Ph.D. degrees in mechanical engineering from the University of Bristol, Bristol, U.K., in
1975 and 1979, respectively.
He worked in industry from 1979 to 1981 and then joined the Mecha- nical Engineering Department, Imperial College
London, London, U.K.,
initially as a Lecturer and then successively Senior Lecturer, Reader, and Professor. He is now head of the Imperial College
Mechanical Engineering Department and leads the NDE research group; he is also the principal investigator of the UK Research
Centre for NDE (RCNDE) that has its head office at Imperial College. He has published over 170 refereed journal papers and a
similar number of conference papers in this field and holds four current patents. He is a director of two spin-out companies set up
to exploit technology developed in his research group (Guided Ultrasonics, Ltd. and Permasense, Ltd., both of which supply
inspection and monitoring equipment to the petrochemical and other industries), and he is a consultant to a variety of industries.
Prof. Cawley is a Fellow of the Royal Academy of Engineering and the Royal Society.