LPIT C-Scan Enabled Characterization of MFL Signals using Artificial Neural Networks

sciencediscussionAI and Robotics

Oct 20, 2013 (3 years and 9 months ago)

80 views


Rowan:
RTD Proposal Ver 1
.0


1






LPIT C
-
Scan Enabled Characterization of MFL Signals using Artificial
Neural Networks





A Proposal Submitted To


Mr.
Munendra Tomar

ExxonMobil Research & Engineering


By


Shreekanth Mandayam, Ph. D.


College of Engineering

Rowan University

201 Mullic
a Hill Road

Glassboro, NJ 08028




Executive Summary


Rowan University will design, develop and test a system for
characterizing
magnetic flux
leakage signals using artificial neural networks
.

The networks will be trained using LPIT C
-
scan
signatures and w
ill be used to predict 3
-
D profiles of pipe
-
wall corrosion anomalies.





Rowan:
RTD Proposal Ver 1
.0


2

Project Description


The overall objective of a desired signal characterization algorithm is illustrated in Figure 1. The
matrix consisting of magnetic flux leakage (MFL) signal ampl
itudes

must be transformed
to
provide an equivalent matrix indicative of the corrosion profile


elements of this matrix have
amplitudes corresponding to the percentage of pipe
-
wall loss at a specified location. Artificial
neural networks are ideal mechan
isms for performing such matrix
-
to
-
matrix signal
transformations.

A “brute
-
force” approach employing artificial neural networks for effecting MFL signal
transformations to predict corrosion profiles is illustrated in Figure 2 (a). The artificial neural
net
work operates in three phases. During the “training” phase, the network receives inputs that
are examples of MFL signals. The corresponding desired output for each input MFL signal is
indicated


these desired outputs are corrosion profile matrices that ar
e obtained following LPIT
laser C
-
scans of the pipe
-
wall. The “synaptic weights” of the neural network are determined at
this stage.
Following a successful training algorithm, that is indicated by minimizing specified
error
-
goals, the network is ready for
the “validation” phase. MFL signals, with known corrosion
profiles, but have not been previously introduced to the network are presented for analysis. The
network predictions are compared with the desired output. The network is said to be successfully
trai
ned if it is able to perform this task without appreciable error. In the third, or “test” phase, the
network is deployed for analyzing unknown MFL signals.

There is a significant problem with this “brute
-
force” approach for training, validating and
testing

this signal characterization neural network. In order to successfully train the network so
that it is able to analyze new MFL signals, a comprehensively large training data set is required.
Network training that employs a matrix
-
matrix transformation as d
escribed is very sensitive to
small variations in signal amplitudes, thus rendering prediction results for unknown signals
questionable.

A two
-
pronged alternate strategy can be envisioned for analyzing MFL signals, especially if
LPIT C
-
scans are available
for a set of corrosion anomalies. The heart of this algorithm is the
determination of corrosion
-
profile information contained within the MFL signature that is
in
common with the same information contained in the LPIT C
-
scan. Also, one can determine
corrosi
on
-
profile information that is missing in the MFL signature, but is available in the LPIT
C
-
scan. Results obtained from a pair of artificial neural networks that can be trained to predict

Rowan:
RTD Proposal Ver 1
.0


3

these two separate sets of information can be combined to accurately

and confidently estimate
the true corrosion
-
profile. The following paragraphs describe this strategy in more detail.