Multi-sensor data fusion using geometric transformations for the nondestructive evaluation of gas transmission pipelines by PJ Kulick Graduate Advisor: Dr. Shreekanth Mandayam MS Final Oral Presentation August 29, 2003, 3:00 PM
Dec 31, 2015
Multi-sensor data fusion using geometric transformations for the nondestructive evaluation
of gas transmission pipelines
byPJ Kulick
Graduate Advisor: Dr. Shreekanth Mandayam
MS Final Oral PresentationAugust 29, 2003, 3:00 PM
Outline
• Introduction
• Objectives and Scope of Thesis
• Background
• Approach
• Implementation Results
• Conclusions
Gas Transmission Pipelines
Sleeve
Weld
Corrosion
SCC
T-section
Valve
• 280,000 miles• 24 - 36 inch dia.
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
In-Line Inspection
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Nondestructive Evaluation (NDE)
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Gas Transmission Pipeline Indications
• Benign– T-sections
– Welds
– Valves
– Taps
– Straps
– Sleeves
– Transitions
• Anomalies– Stress
Corrosion Cracking
– Pitting
– Arching
– Mechanical Damage
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
NDE using Multiple Inspection Modalities
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Data Fusion
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Data Fusion
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Objectives of This Thesis
• Develop data fusion techniques for the extraction of redundant and complementary information
• Validate techniques using simulated canonical images
• Validate techniques using laboratory NDE signals
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Expected Contributions
• A data fusion algorithm with the ability to identify redundant and complementary information present in multiple combinations of pairs of NDE data sets.
i. e. (MFL-UT, MFL-Thermal, UT-Thermal)
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Ultrasonic Testing
Thermal Imaging
Acoustic Emission
Test PlatformsDigital Signal/Image
Processing
Data Fusion
AdvancedVisualization
Virtual Reality
This research work is sponsored by:• US Department of Energy• National Science Foundation• ExxonMobil
Nondestructive Evaluation of Gas Pipelines0.0” 0.2”
0.4” 0.6”
Artificial Neural Networks
1
1
1
x1
x2
x3
y1
y2
wij
wjk
wkl
InputLayer
Hidden Layers
OutputLayer
Magnetic Imaging
Previous Work in Data Fusion
• Mathematical Theory– Probability Theory
• Bayes’ Theorum
– Possibility Theory• Fuzzy logic
– Belief Theory• Dempster Shafer
– “Improved” DS Theories• Transferable Belief Model
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Previous Work in Data Fusion
• Mathematical Transforms– Discrete Fourier Transform (DFT)– Discrete Cosine Transform (DCT)– Wavelet based transforms
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Geometric Transformations
• Spatial TransformationOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
),( yxf yxg ˆ,ˆ
Geometric Transformations
• Gray-level InterpolationOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Approach
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
GeometricTransformation
Feature x1
Feature x2
Redundant/ Complementary
Information
g2(x2) Θ g1-1(x1, x2) = h
homomorphic operator
OBJECT
Approach
• Redundant Data Extraction
Train RBF (homomorphic operator +)
g1(x1, x2) = g2(x2) – h1
RBFNeural Network
x1
x2
x2 – h1
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Approach
• Redundant Data Extraction Test RBF
h1 = x2 – g1(x1, x2)
RBFNeural Network
x1
x2
h1∑-
+
x2 – h1
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Canonical Image Results
Simulation 1OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
x1x2
Redundant Complementary
• 6 Images
• 4 Training
• 2 Test
• 20 x 20 pixels
• 20 x 20 DCT sent into network in vector form
Canonical Image Results
Simulation 1: Training Data ResultsOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Canonical Image Results
Simulation 1: Test Data ResultsOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Canonical Image Results
Simulation 2OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
x1x2
Redundant Complementary
• 6 Images
• 4 Training
• 2 Test
• 20 x 20 pixels
• 20 x 20 DCT fed into network in vector form
Canonical Image Results
Simulation 2: Training Data ResultsOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Canonical Image Results
Simulation 2: Training Data ResultsOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Canonical Image Results
Simulation 2: Test Data ResultsOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Experimental Setup
• Test Specimen SuiteOUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Experimental Setup: MFL
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Clamp Pipesection
Hallprobe
Probemount
Currentleads
Experimental Setup:Tangential MFL Images
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Experimental Setup: UT
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Ultrasound transducers Concretetestspecimen
Immersiontank
Linearactuators forscanning
Scanner controller & stepper motors
PC fordata acquisition & processing
Experimental Setup:UT Time of Flight (TOF) Images
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Experimental Setup: Thermal
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Experimental Setup:Thermal Phase Images
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
What is Redundant and Complementary Information?
• We have defined this as follows:OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Defect Profile
Method 1 NDE Signature
Method 2 NDE Signature
Redundant Information
Complementary Information
Experimental Setup:Tangential MFL Images
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Experimental Setup:UT Time of Flight (TOF) Images
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Experimental Setup:Thermal Phase Images
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Data Fusion Trials
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Trial 1 UT-MFL UT-Thermal MFL-Thermal
Trial 2 UT-MFL UT-Thermal MFL-Thermal
Trial 3 UT-MFL UT-Thermal MFL-Thermal
Data Fusion Trials
• Trial #1OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
UT and MFL Data Fusion ResultsTrial 1:
UT and MFL Data Fusion ResultsTrial 1:
Data Fusion Trials
• Trial #2OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
UT and MFL Data Fusion ResultsTrial 2:
UT and MFL Data Fusion ResultsTrial 2:
Data Fusion Trials
• Trial #3OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
UT and MFL Data Fusion ResultsTrial 3:
UT and MFL Data Fusion ResultsTrial 3:
Data Fusion Trials
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Trial 1 UT-MFL UT-Thermal MFL-Thermal
Trial 2 UT-MFL UT-Thermal MFL-Thermal
Trial 3 UT-MFL UT-Thermal MFL-Thermal
Data Fusion Trials
• Trial #1OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
UT and Thermal Data Fusion ResultsTrial 1:
UT and Thermal Data Fusion ResultsTrial 1:
Data Fusion Trials
• Trial #2OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
UT and Thermal Data Fusion ResultsTrial 2:
UT and Thermal Data Fusion ResultsTrial 2:
Data Fusion Trials
• Trial #3OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
UT and Thermal Data Fusion ResultsTrial 3:
UT and Thermal Data Fusion ResultsTrial 3:
Data Fusion Trials
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Trial 1 UT-MFL UT-Thermal MFL-Thermal
Trial 2 UT-MFL UT-Thermal MFL-Thermal
Trial 3 UT-MFL UT-Thermal MFL-Thermal
Data Fusion Trials
• Trial #1OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
MFL and Thermal Data Fusion ResultsTrial 1:
MFL and Thermal Data Fusion ResultsTrial 1:
Data Fusion Trials
• Trial #2OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
MFL and Thermal Data Fusion ResultsTrial 2:
MFL and Thermal Data Fusion ResultsTrial 2:
Data Fusion Trials
• Trial #3OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
MFL and Thermal Data Fusion ResultsTrial 3:
MFL and Thermal Data Fusion ResultsTrial 3:
Accomplishments
• Development of a generalized technique for fusing data from two distinct observations of the same object
• Design of an algorithm that can extract redundant and complementary information from two distinct observations of the same object
• Validation using simulated canonical images• Validation using lab data representative of the
NDE of gas transmission pipelines
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Conclusions
• Algorithm is sufficiently general – does not specify which features are redundant or complementary
• Efficacy has been demonstrated by defining the redundancy and complementarity of two NDE images by correlating defect signature pixels with the location, size and shape of the defect
• Definition and approach are extremely accurate in all instances of training data and sufficiently accurate in all instances of test data
• Information presented to the neural network is distinct; the matrices manipulated are non-singular
• The errors that occur during certain instances of training and testing illustrate the need for a large, more diverse data set
• Data fusion of UT/MFL proved better then data fusion of UT/Thermal or MFL/Thermal
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Directions for Future Work
• Enhancement of training and test data• Explore variety of image preprocessing
techniques• Investigate various definitions of redundant
and complementary information• Test technique’s robustness with noisy real-
world NDE signals• Adapt algorithm for heterogenous datasets
OUTLINE
Introduction
Objectives/ Scope
Background
Approach
Implementation Results
Conclusions
Acknowledgements
• U.S. Department of Energy, "A Data Fusion System for the NondestructiveEvaluation of Non-Piggable Pipes," DE-FC26-02NT41648
• ExxonMobil, "Development of an Acoustic Emission Test Platform with a Biaxial Stress Loading System," PERF 95-11
• Joseph Oagaro