Professional Documents
Culture Documents
IPC2022
September 26 – 30, 2022, Calgary, Alberta, Canada
IPC2022-86934
NOT ALL DATA IS GOOD DATA: THE CHALLENGES OF USING MACHINE LEARNING
WITH ILI.
1 © 2022 by ASME
verify the generalization of a model and more accurately the standard reported features of depth, length and width.
represents future performance. Regression models have been used as far back as Gauss in his
This paper describes the development of an ILI machine modeling of the heavenly bodies [6] and are a staple of
learning model at each stage of the process and contrasts the empirically derived engineering models [7].
performance of correctly utilizing data in training and testing to As computational speeds and memory advanced, so did the
that of incorrectly utilized data. methodologies that used them. AI got its real start when
mathematician and computer scientist Alan Turning asked the
NOMENCLATURE question, “Can machines think?” [8] Symbolic AI was the first
AI Artificial Intelligence attempt at modeling how to think using predefined rules, much
AUT Automated Ultrasonic Technology as humans do with logic, mathematics and language.
DEF Internal Geometry or Deformation Symbolic AI developed considerably over the course of the
Measurements 1960s, but due to the a priori knowledge that had to be
DVP Dig Verification Program programmed into these systems they began to be supplanted by
EMAT Electromagnetic Acoustic Testing neural networks. Neural networks are a form of AI called
GPS Global Positioning System machine learning designed to learn from data that is presented to
GPU Graphical Processing Unit it. There is no need for machine learning to be programmed with
ILI In-line Inspection explicit sets of instructions and symbols.
KNN K-Nearest Neighbors Classification Model Neural networks were used on many ILI problems from
LFM Low Field Magnetic Flux Leakage classifying features to determining the severity of mechanical
MDS Multiple Dataset damage [9].Neural networks derive their name from the original
MFL Magnetic Flux Leakage idea to mimic the connections of neurons in the brain. However,
NDE Non-Destructive Examination even today these methods are grossly oversimplified when
PAUT Phased Array Ultrasonic Technology compared to the human brain. Neural networks comprise layers,
SMFL Spiral Magnetic Flux Leakage each one containing a node (see Figure 1). The node is basically
UT Ultrasonic Testing a function that takes in inputs, processes them and then sends out
outputs that may be the inputs to another layer. These nodes
1. INTRODUCTION have numerical coefficients associated with them and they
Inspection of steel pipelines with magnetic flux leakage “learn” by adjusting these coefficients to match their outputs to
(MFL) tools has been around since the 1960s. At first, the data the expected outputs they are being trained on. The data set used
from these tools was analog with coarse circumferential and for this learning step is called the training set. After being
axial resolutions measured by induction coil sensors. The sizing trained, the model outputs are compared to data it has not been
algorithms used were mostly based on the amplitude of the trained on, called the validation set.
recorded signals with a lot of analyst interpretation. Since then
things have changed considerably. During the 1990’s two
revolutions occurred in the in-line inspection (ILI) community:
advances in measurement technologies with the application of
hall sensors and ultrasonic testing (UT), and the computer
revolution that ushered in powerful desktop digital processing.
The latter continued into the 21st century with data processing
power growing exponentially and breakthroughs in artificial
intelligence (AI) including machine learning and deep learning.
Now, in the 2020s, we are seeing the impact of these
advancements over a range of ILI technologies used in pipeline
integrity [1]‒[4].
Machine learning tools are powerful but “with great power
there must also come great responsibility.” [5] There are two
regimes that must be addressed in building a reliable machine
learning algorithm: the type of machine learning methods to use
and the proper curation of truth data.
2 © 2022 by ASME
There have been many adaptations and variations on the generalize well to new examples. This is generally overcome by
regression fitting of nodes, including trees, forest, boosting. and, introducing additional training examples that more
most recently, deep learning. comprehensively represent the prediction space. The quantity of
Trees and forests use nodes but instead of assembling them training data may determine how sophisticated the model can be.
in layers they structure them in branches and leaves. Boosting A small data set may be more prone to overfitting when using
guides the learning process by combing the outputs of many deep learning techniques, whereas it may be more appropriate to
weak learners to obtain an overall strong model. Deep learning, apply random forest or gradient boosting [10].
the next generation of neural nets, exploits the ability to mine
large amounts of data using the computational power of
specifically engineered processing units including graphical
processing units (GPU). Deep learning models have enormous
amounts of nodes in many layers densely connected to each
other.
Another technique worth mentioning are support vector
machines, which belong to a category of pattern recognition
algorithms called kernels. This technique takes in an array of
features for each data point and operates on the distances
between these data points, treating the feature array as a
coordinate in a higher dimensional space. Support vectors are
(a) PIPELINE (A) METAL LOSS (b) PIPELINE (A)
very good at grouping similar data points and separating groups
DEPTH DISTRIBUTION RESIDUAL ERRORS
into classes.
3 © 2022 by ASME
to detect and tune to pipeline differences by training it on data a digital gage to the pipe with a magnetic bridging bar can ensure
across all the cases to be inspected. a more level mounting on the pipe surface, reducing errors due
to rocking or tilting of the gage. Both gages are limited, however,
3. DATA CORRELATION by the width of the point extending into the metal loss.
ILI uses various technologies to gather data on the condition Because measuring only the deepest point can be difficult
of a pipeline while traveling internally through it. Technologies and it ignores other zones in the area, aligning the appropriate
include magnetic flux leakage (MFL), ultrasonic testing (UT), ILI signal associated with the actual deepest metal loss region
electromagnetic acoustic testing (EMAT) and internal geometry may not be feasible. Since areas of corrosion often manifest
measurements (DEF). Inspections can be done with separate multiple ILI signal responses there can be a one-to-many
tools; however, their results must be aligned for comprehensive relationship of NDE-to-ILI. This can lead to correlation errors
integrity analysis. Aligning data requires matching the position between ILI and NDE by assigning an ILI training sample
and orientation of each inspection and accounting for the errors (signal) to the incorrect target NDE value (measured depth).
associated with each tools’ method of measuring axial and
circumferential orientation. Pattern matching techniques of the
datasets can be used to refine alignment, but measurement
uncertainty will always be present, which complicates utilizing
separate inspections. The error resulting from matching ILI to
ILI or NDE to ILI will be referred to as correlation error. One
way to minimize ILI to ILI correlation error is to run multiple
technologies on a single inspection tool.
When an ILI is completed using multiple datasets, a more (a) MANUAL GAGE (b) DIGITAL GAGE
comprehensive view of what is happening on the pipeline can be FIGURE 3: EXTERNAL PIT GAGES
visualized.[11] This allows for the development of more
comprehensive machine learning model inputs. By capturing
NDE data collected at precisely known locations on the pipe, an
alignment can also be made with each of the ILI datasets,
building up the prediction space and reducing the error
associated with correlation.
4 © 2022 by ASME
to map the external surface [13] and automated ultrasonic multiple dataset (MDS) tool in a joint of test pipe. Standard ILI
technology (AUT) [14] maps the internal surface. and NDE reports use boxes to identify and correlate metal loss
Figure 5 shows the use of a handheld profilometry unit in features, with NDE generally only providing information on the
the ditch and the associated digital output of the scanned area. deepest feature in the area. This is a poor utilization of the true
The pipe is prepared by removing a small area of coating and power of using profilometry data to validate ILI results, as
placing some magnetic markers for reference calibration. The evident in Figure 7, where far more comparisons could be made
unit is manually positioned and the software automatically if signal to signal matching occurred. In cases where corrosion
calculates the surface profile of the pipe at resolutions much finer is complex, for example, matching boxes can lead to
than can be achieved with pit gages. The data is stored in digital misalignments as overlapping areas can include pits and signals
formats that can easily be converted into usable formats by ILI not used in the calculation of the boxes. Using actual depth
providers and integrity consultants. Figure 6 shows an AUT measurements at each axial and circumferential position allows
setup that can provide resolution and digital outputs similar to metal loss to be mapped directly to the ILI response location.
that of laser profilometry. The features from the ILI data can then be mapped directly to the
NDE data, minimizing the correlation error, maximizing the
number of correlated points and creating a larger model
prediction space to be developed.
5 © 2022 by ASME
if any analysis adjustments may be warranted. Any applicable planning of the Plan step. This transition from Act to Plan is a
adjustments can then be implemented into the final report. More decision process made by operators, vendors and third-party
verification digs can be done after the final report, and the cycle consultants. Planning for NDE at this stage will direct future
of assessing performance and implementing applicable attention of ILI data analysis toward an operator’s most pressing
adjustments and refinements can be repeated until a stable set of integrity threats, assist in the selection of appropriate NDE
accurate and actionable information is achieved. methods and facilitate the logistics of selecting dig sites. The
Check section is modified so as not to explicitly state digs
because the integrity planning is done in the Plan stage.
Use of the continuous improvement cycle of providing
successive feedback allows digs to be augmented with historical
and previous ILI data. The key in this process is the input from
Act into Plan, which provides a means to leveraging historical
data and continually improving high-resolution ILI. This
minimizes future reliance on validation digs and provides greater
confidence in using ILI data in integrity programs.
6 © 2022 by ASME
profilometry depth data and the yellow contour lines are the pit
dimensions called by the NDE. The bottom picture is the ILI
MFL data, and the red squares are the pit dimensions called by
the ILI. Looking at the axial-circumferential data displayed in
the image, the MFL signals can be matched to the corresponding
measurements in the NDE data. The discrepancy in cluster
sizing is due to the implementation of interaction criteria.
Aligning both data sets enables a more accurate correlation of
ILI and NDE.
7 © 2022 by ASME
Figure 13 shows the KNN model retrained using field data
instead of machine pull test features. As can be seen there is a
dramatic improvement in performance, with a tighter error
distribution. However, the field data contains a population
imbalance that has far more samples in the moderate depth range.
This introduces a bias into the model as the majority of neighbors
are in the moderate depth range and highlights the importance of
having representation over the full range of geometries to be
sized.
For the KNN model, using curated field data results
improves accuracy by relying on nearest neighbors. In densely
populated regions it performs well but struggles in regions more
sparsely populated with feedback. Switching to a gradient boost
(XGB) model allows for a more adaptive fitting using an
ensemble of many small decision trees that each minimize the
residual prediction errors [18]. Figure 14 shows the result of
using the XGB model. The linearity of the unity line has been
captured and the mean error is lowered.
A special note should be made regarding Figure 14. Split 1
shows a subpopulation that significantly differs from the other
points in the validation data (blue box). This subpopulation is
clearly an outlier from the rest of the population and is not
represented well in the training data. By identifying this, a
FIGURE 12: RESULTS OF TRAINING A K NEAREST detailed review of those specific anomalies can be done and new
NEIGHBOR (KNN) MODEL ON MACHINED ANOMALIES AND model inputs can be engineered or changes in the modeling
PREDICTING DEPTHS OF FIELD CORROSION. architecture can be made to improve performance.
FIGURE 13: RESULTS OF TRAINING A K NEAREST FIGURE 14: RESULTS OF TRAINING A XGB MODEL ON
NEIGHBOR (KNN) MODEL ON FIELD DATA AND FIELD DATA AND PREDICTING THEIR DEPTHS. AN OUTLIER
PREDICTING THEIR DEPTHS. SUBPOPULATION IS HIGHLIGHTED IN BLUE.
8 © 2022 by ASME
8. CONCLUSION tree induction." Journal of Artificial Intelligence
As demonstrated in this paper, machine learning is a Research (2003): 315–354.
powerful tool, but only when carefully applied. Data used in [11] Kirkwood, M., Burden, D. and Maldonado, M., “The Art of
model development requires diligent curation, precise Looking: An In-Line Inspection Perspective,” Proceedings
correlation and accurate measurements. Subject matter experts of the 2019 Pipeline Pigging and Integrity Management
play a vital role in selecting a model, engineering its inputs and Conference, Houston, TX, USA, February 18-22, 2019, No.
checking the prediction space is well represented. Operators can 10.
aid in the development by utilizing NDE techniques that [12] McNealy, R., McCann, R., Van Hook, M., Stiff, A. and
maximize the number of measurements per dig and provide Kania, R., “In-Line Inspection Performance III, Effect of In-
detailed information on anomaly geometry. Ditch Errors in Determining ILI Performance,” Proceedings
of the 8th International Pipeline Conference, IPC2010-
ACKNOWLEDGEMENTS 31269, pp. 469-473, Calgary, Alberta, Sept 27 – Oct 1, 2010,
Tod Barker, Jeremy Clark, Josh Yunik, Robert Coleman, https://doi.org/10.1115/IPC2010-31269
Chris Goller, Jed Ludlow, James Simek, Andrew Moore, Callie [13] Tomar, M., Fingerhut, M., and Yu, D., “Qualification of ILI
Lamb and Eric South. Performance in Accordance with API 1163 and the Potential
Impact for Management of Pipeline Integrity,” Proceedings
REFERENCES of the IPC2008, 7th International Pipeline Conference,
[1] Burden, D., Dalfonso, P. and Belanger, A., “The Current IPC2008-64469, pp. 693-703, Calgary, Alberta, Sept 29-
Progeny of In-Line Inspection Machine Learning,” Oct3, 2008, https://doi.org/10.1115/IPC200864469
Proceedings of the 2020 Pipeline Pigging and Integrity [14] Feole, R. A., “Automation of Fast Ultrasonic Technique for
Management Conference, Houston, TX, USA, February 17- Pipeline Weld Examinations,” PRCI, 1998, PR-250-9612,
21, 2020, No. 54. Cat. No. L51777.
[2] Conrad, B., Kissel, M., Tse, V., Barlow, J., Chen, J., and [15] American Petroleum Institute. (April 2019). Managing
Kania, R., “Considerations for validating an ILI technology: System Integrity for Hazardous Liquid Pipelines. API
controlled implementation of a new MFL product,” Standard 1160, Section 4.1
Proceedings of the 2020 Pipeline Pigging Management [16] Lamb, C. and Dalfonso, P., “How ILI Feature Grouping Can
Conference, Houston, TX, USA, February 17-21, 2020, No. Impact Apparent Accuracy from NDE,” Proceedings of the
19 2022 Pipeline Pigging and Integrity Management
[3] Chakraborty, I. and Vyvial, B., “Using deep learning to Conference, Houston, TX, USA, Feuruary 2-40 2022, No.
identify the severity of pipeline dents,” Proceedings of the 43.
2020 Pipeline Pigging and Integrity Management [17] Goldberger, J., Rowels, S., Hinton, G. and Salakhutdnov,
Conference, Houston, TX, USA, February 17-21, 2020, No. R., “Neighbourhood Components Analysis,” Advances in
70 Neural Information Processing Systems, Vol. 17, May 2005,
[4] Hurd, G., Elliott, J., Farnie, S., and Sutherland, J., pp 513-520.
“Advancing In-line Inspection technology and pipeline risk [18] Friedman, J. H. 1999, "Greedy Function Approximation: A
management through advanced analytics of big data,” Gradient Boosting Machine" (PDF), IMS 1999 Reitz
Proceedings of the 2020 Pipeline Pigging and Integrity Lecture.
Management Conference, Houston, TX, USA, February 17-
21, 2020, No. 13
[5] Lee, S., 1962, Amazing Fantasy, #15, Marvel.
[6] Gauss, Carl F., 1809, Theory of the Motion of the Heavenly
Bodies Moving about the Sun in Conic Sections, Little,
Brown and Company, Boston, Doi: 10.5962/bhl.title.19023
[7] Box, G., and Draper, N., 1987, Empirical Model-Building
and Response Surfaces, Wiley, New York.
[8] Turing, A., 1950, “Computing Machinery and Intelligence,”
MIND a Quarterly Review of Psychology and Philosophy,
pp 433-460
[9] Bubenik, T. A., Nestleroth, J.B., Davis, R. J., Crouch, A.,
Upda, S., and Afzal, A. K., 2000, “In-line Inspection
Technologies for Mechanical Damage and SCC in
Pipelines: Final Report,” US DOT, OPS, Report No.
DTRS56-96-C-0010.
[10] Weiss, Gary M., and Foster Provost. "Learning when
training data are costly: the effect of class distribution on
9 © 2022 by ASME