Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
A Methodology to Reduce Tank Inspection Frequency for U.S. Navy Surface Ships
by Brian S. Tait
B.S. in History, May 1989, U.S. Naval Academy
M.S. in Electrical Engineering, December 1995, U.S Naval Postgraduate School
A Praxis submitted to
The Faculty of
The School of Engineering and Applied Science
of The George Washington University
in partial fulfillment of the requirements
for the degree of Doctor of Engineering
January 10, 2020
Praxis directed by
Timothy Blackburn
Professorial Lecturer of Engineering Management and Systems Engineering
Amir Etemadi
Associate Professor of Engineering and Applied Science
ii
The School of Engineering and Applied Science of The George Washington University
certifies that Brian Steven Tait has passed the Final Examination for the degree of Doctor
of Engineering as of October 14, 2019. This is the final and approved form of the Praxis.
A Methodology to Reduce Tank Inspection Frequency for U.S. Navy Surface Ships
Brian S. Tait
Praxis Research Committee:
Timothy Blackburn, Professorial Lecturer of Engineering Management and
Systems Engineering, Praxis Co-Director
Amir Etemadi, Associate Professor of Engineering and Applied Science, Praxis
Co-Director
Ebrahim Malalla, Visiting Associate Professor of Engineering and Applied
Science, Committee Member
iii
© Copyright 2019 Brian S. Tait
All rights reserved
iv
Dedication
This praxis is dedicated to the engineers of the United States Navy, past, present
and future. Your engineering disciplines are indispensable in the formation and
sustainment of naval forces.
.
v
Acknowledgements
The author wishes to thank the LORD for His sovereignty and provision over all
things, especially this praxis. Thank You for answering prayer and being an ever-present
help. The author wishes to thank his wonderful wife and family for their enduring love
and support, patience and encouragement to undertake and complete such a formidable
endeavor.
vi
Abstract of Praxis
A Methodology to Reduce Tank Inspection Frequency for US Navy Surface Ships
The United States Navy is, by far, the largest navy in the world. With over 200
vessels that have more than 15,000 shipboard tanks to maintain, an average of nearly
2,000 tank inspections has been conducted annually since 2011 (Csapo, 2019). Twenty-
five percent of the annual Navy’s maintenance budget is spent on corrosion (Parsons,
1994). This research effort introduced a novel methodology for analyzing and integrating
factors of ultra-high and high solids (UHS/HS) tank coating types, coating age, surface
preparation method, ship class, tank type and tank criticality from existing U.S. Navy
ship tank inspection databases in order to reduce required tank inspection frequencies by
25-50%. Ways to mitigate model potential overconfidence were explored. It was also
demonstrated that significant savings were accrued by a reduction of required tank
inspections. The practical applications of this proposed methodology are: 1) tank
inspection periodicities will be optimized, 2) unnecessary tank inspections will not be
conducted, saving shipyard resources (cost and schedule) and 3) future planning efforts
will be adjusted.
vii
Table of Contents
Dedication……………………….…………………………………………….………….iv
Acknowledgements………………………………………………………………………..v
Abstract of Praxis………………..…………………………………………...…………...vi
List of Figures………………………………………………………...………..................xi
List of Tables …………………………………………………………….……...……..xiii
List of Acronyms……....…………………………………………………………......…xvi
Glossary of Terms……….……………….………………………………………..……xvii
Chapter 1—Introduction………………………………………………………...….…..1
1.1 Background…………………...……………………………………………………..1
1.2 Research Motivation……………….......……………………………………………2
1.3 Problem Statement………...………………………………………………………...3
1.4 Thesis Statement…………………...………………………………………………..3
1.5 Research Objectives……………...………………………………………………….3
1.6 Research Questions and Hypotheses……………...…………………………………4
1.7 Scope of Research………………………...…………………………………………5
1.8 Research Limitations………………………………………...………………………6
1.9 Organization of Praxis……………...………………………………………………..6
Chapter 2—Literature Review….…………...……………………...…………………..7
2.1 Introduction……………………..…………………………………………………...7
2.2 Overview……………..…………………………………………...…………………7
2.3 Corrosion……..…………………………………………………...…………………8
2.3.1 Corrosion Process………..………………………………………………………8
viii
2.3.2 Corrosion Types…….…….……………………………………………………..8
2.3.3 Corrosion Phases……..………………………………………………………….9
2.3.4 Factors Affecting Corrosion……………..……………………………………....9
2.3.5 Corrosion Rates……..………………………………………………………….11
2.3.6 Corrosion Inspections…….…………………………………..………………...12
2.3.7 Corrosion Consequences………………………………………………………..12
2.4 Corrosion Prevention……………………………………………………………….13
2.5 Coatings………………………………………………………………………….…13
2.6 Corrosion Modeling…………………………………………………………….…..14
2.7 Corrosion Risk Assessment………………...……………………………………....21
2.8 Cost of Corrosion…………………………………………………………………...22
2.9 Summary and Conclusion……………………………………………...…………...23
Chapter 3—Methodology…………………...………………………………..……..….24
3.1 Introduction……………………………………………………………..…………..24
3.2 Surface Ship Tank Inspection Data………………………………………..…….….24
3.3 Database Shaping……………………………………………………………..…….25
3.4 Database Filtering……………………………………………………………….….26
3.5 H1A and H1B Database Analysis……………………………………………….….27
3.5.1 H1A Data Set……………………………………………………………………28
3.5.2 H1B Data Set…..……..……………..………….………………….……………29
3.5.3 H1A and H1B Data Partitioning…..………..….………………………..……...30
3.5.4 H1A and H1B Model Training……..……..….………………………..……….30
3.5.5 H1A and H1B Model Testing……..……..….…………………………..……...31
ix
3.5.6 Model Classification Matrix and Optimization..……………………….......….33
3.5.7 H1A Coating Age Prediction Tool……………..……………………………...37
3.5.8 H1B Structural Score Prediction Tool…………..……………………………..37
3.6 H2 Data and Data Analysis…….…..………………………………………………38
3.7 H3 Data and Data Analysis…………………………………………...……………39
3.8 Conclusion……………………………..……………………………………..……41
Chapter 4—Results……………………...……………………………………………...42
4.1 Introduction………...……………………………………………………………….42
4.2 Models’ Assumptions and Results……..…………………………………………..42
4.2.1 Database Shaping and Filtering Results….…………………………………….48
4.3 H1A Database Composition…………..……………………………………………50
4.3.1 H1A Train Database Composition………………..…………………………….53
4.3.2 H1A Test Database Composition…………………………..…………………...56
4.4 H1B Database Composition……………...………………………………………....59
4.4.1 H1B Train Database Composition…..…………………………………………..60
4.4.2 H1B Test Database .……………………..…………………………………...…62
4.5 H1A OLR Results…………………………………………………………………..65
4.5.1 H1A Model Training and Testing Results………..……………………………..66
4.5.2 H1A Model Optimization and Classification Results………………….....……..69
4.5.3 H1A Prediction Tools…………………..…………………………………….…71
4.5.3.1 H1A Coating Score Prediction Tool #1…….…………………………….…71
4.5.3.2 H1A Coating Score Prediction Tool #2……….…………………………….72
4.6 H1B OLR Results…………………………………………………………...….…..73
x
4.6.1 H1B Model Training and Testing Results……..………………………….……74
4.6.2 H1B Model Optimization…………..……………………………………….….77
4.6.3 H1B Prediction Tool………………..…………………………………...….….81
4.6.3.1 H1B Structure Score Prediction Tool………..………………………..……81
4.7 H2 Results……………………………………………………………………,.……82
4.8 H3 Results……………………………………………………….………….………83
Chapter 5—Discussion and Conclusions………………………………………..…….86
5.1 Discussion….…………………………..……………………………………..……86
5.1.1 Scope of Analysis and Framework….………..…………………………..…….87
5.1.2 H1A Discussion…….……………………..………………………………..…..87
5.2 H1B Discussion……………………………………………………………….……88
5.3 H2 Discussion………………………………………………………………..……..90
5.4 H3 Discussion………………………………………………………………..……..91
5.5 Conclusions…………………………………………………………………..…….92
5.6 Body of Knowledge…………………………………………………………..……92
5.7 Future Study…………………………………………………………………..……92
References…………………………………………………………………………..…..93
xi
List of Figures
Figure 1 – Coating Age Prediction Tool Sample……………………….…..……………37
Figure 2 – Structure Score Prediction Tool Sample………………………..……………38
Figure 3 – H1A Coating Age (Predictor Variable) Parallel Test Results……..…………45
Figure 4 – H1A Coating Type (Predictor Variable) Parallel Test Results…..……..……46
Figure 5 - H1B Coating Age (Predictor Variable) Parallel Test Results……..……….....47
Figure 6 – H1B Coating Score (Predictor Variable) Parallel Test Results……..………..48
Figure 7 – H1A Test Database Frequency Distribution (Graph)………………..……….51
Figure 8 – H1A Test Database Coating Age (PV) Frequency Distribution……..………52
Figure 9 – H1A Test Database Frequency Distribution (Graph)………………..……….54
Figure 10 – H1A Train Database Coating Age (PV) Frequency Distribution……..…….55
Figure 11 – H1A Test Database Frequency Distribution (Graph)……………….....……57
Figure 12 – H1A Test Database Coating Age (PV) Frequency Distribution………....…58
Figure 13 – H1B Train Database Frequency Distribution (Graph)……………….……..60
Figure 14 – H1B Train Database: Coating Age Frequency Distribution………..……….60
Figure 15 – H1B Train Database Frequency Data Base (Graph)………………..…..…..62
Figure 16 – H1B Train Database: Coating Age Frequency Distribution………...……....62
Figure 17 – H1B Test Data Base Frequency Distribution……………………...…….….64
Figure 18 - H1B Test Data: Coating Age Frequency Distribution……………...…….…64
Figure 19 – H1A OLR Results……………………………………………………….…..65
Figure 20 – H1A Coating Score Prediction Tool #1…………………………………..…72
Figure 21 – H1A Coating Score Prediction Tool #2…………………………………..…73
Figure 22 – H1B OLR Results ……………………………………………………..……74
xii
Figure 23 – H1B Structure Score Prediction Tool……………………………………….81
xiii
List of Tables
Table 1- Parameters Affecting Corrosion in Bulk Carriers………...………………10
Table 2 - NSTM 100 Corrosion Rates for Steels………………………….……………..11
Table 3 – Damage Inspection Techniques Matrix……………………………………….13
Table 4 – CCAMM Database Sample…………………………………………………...26
Table 5 – HA1 Database Filtering Actions………………………………………………27
Table 6 – HB1 Database Filtering Actions..…………………………………………..…27
Table 7 – H1A Database Sample Set…………………………………………………….29
Table 8 – H1B Database Sample Set…………………………………………………….30
Table 9 – Training Database Testing Sample Set………………………………….…….32
Table 10 – Test Results Sample of Testing Data Set……………………………….……33
Table 11 – H1A Design of Experiments Sample…………………………………….…..33
Table 12 – H1B Design of Experiments Sample……………………………….………..34
Table 13 – H1A Classification Matrix Example…………………………………………35
Table 14 – H1B Classification Matrix Example…………………………………………36
Table 15 – H2 Data Base Sample………………………………………………………..39
Table 16 – H3 Data Base Sample………………………………………………………..40
Table 17 – Assumption Results for H1A and H1B Models………….…………………..43
Table 18 – H1A Correlation Results………………………………………….………….44
Table 19 – H1B Correlation Results…………………………………………….……….44
Table 20 – Data Shaping and Filtering Results………………………………………….49
Table 21 - H1A Database Frequency Distribution………………………………………50
Table 22 - H1A Train Database Frequency Distribution………………...………...…….53
xiv
Table 23 – H1A Test Data Frequency Distribution………………………...……………56
Table 24 – H1B Train Database Frequency Distribution……………………..…………59
Table 25 - H1B Train Database Frequency Distribution……………………..………….61
Table 26 – H1B Test Database Frequency Data Base…………………………..……….63
Table 27 – H1A Model Training Results………………………………………..……….67
Table 28 – H1A Model Testing Results…………………………………..……………..68
Table 29 – H1A Classification Matrix…………………………………..……………….69
Table 30 - H1A Model Training Results Classification Matrix (Pre-Optimization)..…...70
Table 31 – H1A Model Training Results Classification Matrix (Post-Optimization).......70
Table 32 – H1A Testing Data Results Classification Matrix (Pre-Optimization)…....….70
Table 33 – H1A Testing Data Results Classification Matrix (Post-Optimization)….......71
Table 34 – Model H1A Classification Performance………………………………….….71
Table 35 – H1B Model Training Results…………………………………………….…..75
Table 36 – H1B Training Result Classification Matrix (Pre-Optimization)…………..…76
Table 37 – H1B Classification Matrix………………………………………………..….78
Table 38 – H1B Training Result Classification Matrix (Pre-Optimization)…………..…78
Table 39 – H1B Training Results Classification Matrix (Post-Optimization)……….….79
Table 40 – H1B Testing Data Results Classification Matrix (Pre-Optimization)…….…79
Table 41 – H1B Testing Data Results Classification Matrix (Post-Optimization)……....80
Table 42 – Model H1B Classification Performance……………………………….…….80
Table 43 – H2 Results…………………………………………………………….….…..83
Table 44 –In-Service Tank Frequency Distribution……………………………….….…84
Table 45 – Average Tank Inspection………………………..……………………….…..84
xv
Table 46 – H3 Results………………………………………………………...……….…85
Table 47 - HA1 OLR Regression Results…………………………………………….….88
Table 48 – HB1 OLR Regression Results……………………………………………….89
Table 49 – H2 Results……………………………………………………………………90
Table 50 – H3 Results…………………………………………………………….......….91
xvi
List of Acronyms
ABS American Bureau of Shipping
CPA Carrier Planning Activity
CCAMM Corrosion Control Assessment and Maintenance Manual
CCIMS Corrosion Control Information Management System
HS High Solids
NAVSEA Naval Sea Systems Command
NSTM Naval Ship Technical Manual
NRL Naval Research Laboratory
SURFMEPP Surface Maintenance Engineering Planning Program
UHS Ultra High Solids
USN United States Navy
xvii
Glossary of Terms
CPA The Carrier Planning Activity is a United States Navy
organization dedicated to the planning of aircraft carrier
maintenance.
CCAMM Corrosion Control Assessment and Maintenance Manual is
a United States Navy Maintenance manual containing tank
inspection and repair requirements.
CCIMS Corrosion Control Information Management System is a
corrosion control management system maintained by the
Surface Ship Engineering Planning Program.
HS High Solids is coating that contains 60-90% solids and 250-
150 g/L Volatile Organic Compounds.
NAVSEA Naval Sea Systems Command is a United States Navy
organization supporting the design, construction, support,
sustainment and modernization of naval vessels.
NRL Naval Research Laboratory is a United States Navy
organization dedicated to conducting research for naval
applications.
SURFMEPP Surface Maintenance Engineering Planning Program is a
United States Navy organization dedicated to planning of
surface ship maintenance.
UHS Ultra High Solids is coating that contains greater than 90%
solids and less than 150 g/L Volatile Organic Compounds.
1
Chapter 1—Introduction
1.1 Background
Vessels of the United States Navy (USN) span the globe in support of the interest of
the United States. The operational tempo of keeping the ships on mission at sea has
strained naval maintenance repair periods. Given the number of the tanks, from such a
large fleet, to be maintained, and given the reduced amount of time for maintenance, it is
incumbent upon the USN technical community to conduct the right maintenance at the
right time, so that the ships can be ready for tasking when needed.
In the life of a shipboard tank, a tank coating life begins with a fresh coat of paint
which is fully adhered and fully protects all structural members. As the coating ages,
however, the coating begins to deteriorate, and gaps in the coating’s coverage begin to
appear, which results in a loss of coating integrity. In coating deficient areas, tank
structure is exposed which begins the onset of corrosion. When this occurs, the tank
coating has reached its service life. Normally, tank coatings are replaced prior to any
severe structural corrosion. In the coating refurbishment process, the aged tank coating is
stripped away exposing bare tank structure metal. The metal surface is prepared using
standardized methodologies to ensure tight coating adhesion, and then, a fresh coating is
applied. Once the coating cures, the tank is placed back into service, and the coating
aging process begins all over. This process of blasting and painting of tanks is routine
and periodically occurs in the life of a ship. Because of the number of tanks (over
15,000), the aggregate cost of maintenance and inspection is very significant.
2
The U.S. Navy regularly inspects its tanks for coating and structural degradation. The
periodicity of the tank inspections requirements and tank inspection (both coating and
structure) grading criterion are detailed in the USN Corrosion Control Assessment and
Maintenance Manual (CCAMM) revision three dated 30 June 2015. The CCAMM is a
foundational document for this research effort. The USN has been documenting tank
inspection results for decades, and results are warehoused in the Corrosion Control
Information Management System (CCIMS) which is the tank inspection results master
databank for all surface ships.
The present work effort intends to introduce a novel methodology by analyzing and
integrating the factors impacting tank coating inspection scores, structural inspection
scores, coating type, coating age, surface preparation method, tank type, and tank
criticality from the CCIMS in order to increase current tank inspection periodicity by 25-
50% with minimal risk acceptance. The practical applications of this proposed
methodology are: 1) tank inspection periodicities will be optimized, 2) unnecessary tank
inspections will not be conducted, saving shipyard resources (cost and schedule) and 3)
future planning efforts will be adjusted.
1.2 Research Motivation
With a fleetwide focus of reducing maintenance dollars and minimizing maintenance
schedules in order to keep ships at their primary mission, being at sea, every maintenance
task must be performed at the right maintenance period, when required, in order to avoid
unnecessary cost in both schedule and dollars as well as avoiding unacceptable risk.
Tank maintenance is typically a primary factor, for cost and schedule, during a ship’s
major shipyard maintenance period. Therefore, it makes sense to focus on this area for
3
any potential cost and schedule savings. Additionally, the USN technical community has
received several first-hand reports from waterfront tank inspectors that question the
reasoning for inspecting pristine tanks. These reports of unnecessarily conducted tank
inspections bring into question the validity of the current tank inspection periodicity
requirements. This occurrence has been more frequent since the introduction of high-
performance high solids and ultra-high solids (HS/UHS) paints into the Navy coatings
inventory in the 2000s. Conducting necessary inspections of tanks is a waste of precious
resources (time and material), especially given the numbers of tanks in the surface navy
inventory.
1.3 Problem Statement
Tanks aboard U.S. Navy surface ships are being inspected too frequently wasting
money and maintenance time.
1.4 Thesis Statement
An integrated methodology model factoring tank coatings deterioration, structural
degradation, coating type, surface preparation, tank type, and tank criticality will enable
increasing tank inspection periodicity by 25-50% with minimal risk acceptance.
1.5 Research Objectives
Given HS/UHS coatings, determine the impact on:
tank inspection coating and structural inspection scores
factors affecting coating and structural inspection scores
optimum tank inspection periodicities
Optimizing tanks inspection periodicities will balance cost and risk.
4
1.6 Research Questions and Hypotheses
The intent of the Research Questions is to explore the relationship between the
HS/UHS coatings with coating age, coating inspection score, and structure inspection
score, and to establish tank inspection periodicities within acceptable established risk
guidelines and recoup any cost saving associated with revised tank inspection
periodicities.
RQ1: What are the current coatings used in U.S. Navy class ships tanks, and what
are their inspection requirements? What are the different tank types and their
criticality towards the mission of the ship?
RQ2: What costs are associated with the current tank inspections?
RQ3: What are the coatings and structural inspection scores associated with high
solids and ultra-high solids tank coatings?
RQ4: What are the relationships between coating, coating age, inspection scores
(both coating and structural), coating type, surface preparation method, tank type,
tank pressure, ship class and tank criticality?
RQ5: What is the optimized UHS coating inspection periodicity for each type of
tank? What is the risk associated with each revised tank inspection periodicity?
RQ6: What are the savings implications of the optimized UHS tank coating
inspection periodicities (from RQ #5)?
There are four Research Hypotheses associated with this research effort:
H1A: Coating Type, Application Method, Tank Type, Ship Class, Tank
Criticality, Coating Age are predictors of Coating Condition Score.
5
H1B: Coating Condition Score, Ship Class, Tank Pressure, Coating Age, Tank
Criticality are predictors of Structural Condition Score.
H2: Tank inspection periodicities, where HS/UHS coatings are employed, can be
increased 25-50% using the coating and structure score predictions (developed from
H1A and H1B) with minimal increased risk of failure.
H3: There are significant savings when tank inspection periodicities increase by
25-50% (results from H2).
1.7 Scope of Research
The Scope of Research for this effort is broad because the extent and use of tank
coatings in the U.S. Navy is broad. This research effort involves 10 ship classes, 13
different tank types, over 15,000 individual tanks, 33 different coating types, etc. with an
overall total of about 25,000 unique inspection records spanning 18 years as documented
in the CCIMS.
For the HS/UHS coatings analysis portion of this present work, only tank coatings
used since 2001 were analyzed since the USN maintenance community began to
transition to HS/UHS coatings in 2000. Only the data records contained in the CCIMS
were utilized.
In the U.S. Navy, ships of the same design are organized into ship classes. A ship
class can have as few as six ships or as high as 78 ships in a ship class. The U.S. Navy
has more than 250 surface ships in the fleet. Navy ships classes have different missions,
and those missions affect the mix of tank types required for each ship class. Thirteen
different tank types represent the gamut of tanks in the surface fleet, from fuel to lube oil
to seawater to fresh water. Tanks have varying degrees of criticality codes associated
6
with their function, like fresh water being very critical (crew survival) to waste oil tanks
being not so critical. Some ship classes are designed disproportionally with more of one
type of tank than another ship class due to their mission. Thus, the range of tank types
from ship class to ship class is broad and varied.
1.8 Research Limitations
The research effort was limited to surface ships only. The research effort was limited
by the quantity and quality of data contained in the CCIMS database. Aircraft carriers
and submarines, although they have thousands of tanks as well, were not analyzed since
the CCIMS contains only surface ship data. Certainly, the methodology in this effort
could be applied toward aircraft carriers and submarines in a future study.
1.9 Organization of Praxis
The Praxis is organized as follows: Chapter 2 contains a Literature Review as it
relates to this research, Chapter 3 discusses the Methodology used in the analysis,
Chapter 4 contains the Results of the analysis, and Chapter 5 contains the Discussion,
Conclusion, how this effort added to the Body of Knowledge, and Future Study
Recommendation.
7
Chapter 2—Literature Review
2.1 Introduction
The intent of this praxis Literature Review is to provide the results of a literature
search for those writings that support the practical application of this research effort. In
doing this, this research effort will have an academic basis. This research effort will be
placed in the context of existing standards and previous similar analysis efforts, and this
research effort will demonstrate the uniqueness of this effort as compared with others.
Since this applied research effort is intended to solve a specific United States Navy
(USN) coating problem, the literature review, by necessity, will include a significant
number of USN references, manuals, technical standards, technical presentations,
technical papers, and prior naval coating analysis and studies. Numerous sources outside
of the USN references were included as well for completeness, rigor, and thoroughness of
review.
2.2 Overview
The U.S. Navy has long recognized the value of potential savings given newer tank
coatings (NAVSEA, 1997). This applied research effort involves an analysis of the
impact and relationship of the degradation of existing tank coatings and tank structure in
various types of tank coating aboard USN surface ships using high-performance high
solids and ultra-high solid (HS/UHS) tank coatings aboard USN surface ships in order to
determine optimized tank inspection periodicities and potential cost and schedule
savings. The scope of this literature search is limited to the practical application of this
research effort.
8
2.3 Corrosion
In this section, an overview of corrosion will be presented to include the corrosion
process, the types of corrosion, the phases of development in corrosion, the factors
influencing the development of corrosion, the rates of shipboard corrosion, and
inspection methods of corrosion.
2.3.1 Corrosion Process
The process of corrosion is a result of electrochemical reactions between materials
and substances in their environment. It involves an anode, a cathode and an electrolyte
(Gardiner and Melchers, 2003). The anode is the site of the corroding metal. The
electrolyte is the corrosive medium that enables the transfer of electrons from the anode
to the cathode. The cathode forms the electrical conductor in the cell that is not
consumed in the corrosion process (Poplar, 2013). This results in an anodic reaction
where the metal is oxidized (metal ions and free electrons released into solution), and the
cathodic reaction takes place where the electrons are reduced by an electron acceptor
(Gardiner and Melchers, 2003). Thus, a type of metal loss does occur in corrosion with
associated structural strength (Wang et al, 2003). Additionally, an oxide layer is formed,
thus helping to prevent further corrosion from occurring (American Bureau of Shipping,
2017b).
2.3.2 Corrosion Types
Corrosion is a significant factor in the marine environment. The general types of
corrosion seen in merchant ships are general (uniform), pitting (localized), stress
corrosion cracking, groove, weld metal, and microbial induced corrosion (Gardiner and
Melchers, 2001); (Rizzo et al, 2008); (American Bureau of Shipping, 2017a). General
9
corrosion is corrosion that is broadcast over a general area in a uniform fashion
(American Bureau of Shipping, 2017a). Pit corrosion is characterized as a localized
corrosion that has a pit to it (Guedes and Garbotov, 1999). Stress corrosion cracking
(SCC) is a form of localized corrosion which produces cracks in metals by simultaneous
action of a corrodent and tensile stress. (Pooplya, 2013). Lastly, microbial induced
corrosion occurs when living organisms, present in the fluid, deposit organic matter near
the metal boundary layer, and that induces corrosion (Pooplya, 2013). These types of
corrosion will occur given the proper environmental conditions.
2.3.3 Corrosion Phases
Paik et al (2003) and Gardiner and Melchers (2003) identified three phases of
corrosion developing in tanks: durability of the coating, transition (corrosion initiation),
and progress of the corrosion (corrosion of exposed metal). Coating durability is the
length of time the coating is effective. This will be discussed in detail in a subsequent
section. Transition, or corrosion initiation, is the time between the end of coating
effectiveness and the onset of measurable corrosion. Progress of corrosion phase begins
when measurable corrosion is present. In bulk carriers, the transition time has been
reported to be up to three years (Paik et al, 2003). The phases of corrosion are offered as
general background knowledge of the corrosion process experienced in tank coatings.
This present work will be examining the length of time a HS/UHS coating will last until
exposure of metal.
2.3.4 Factors Affecting Corrosion
The most significant research on the corrosion of commercial bulk carriers was
previously conducted by Guedes and Garbatov (1999), Gudze and Melchers (2008),
10
Melchers and Jiang (2006), Melchers (2006), and Gardiner (2001). They identified the
factors of the type of cargo, the time spent in ballast, the effectiveness of corrosion
protection, the environmental conditions, and the tank geometry as vital factors in the
corrosion process in bulk carriers. Gardiner and Melchers (2003) noted that factors could
be grouped into two categories: operational vs internal. Operational factors are type of
cargo, time spent in ballast, and environmental conditions (due to trade routes). Internal
conditions were due to design like geometry of tank and type of corrosion control
protection. Industry also recognized the importance of coating type and surface
preparation in the role of the corrosion process (Wheat, 2015). This complex array of
important identified factors also played a vital role in the USN corrosion research and are
factored into this praxis effort as well. In order to enhance the reliability of prediction
models for bulk carriers, Gardiner and Melchers (2003) proposed recording the
parameters listed in Table 1 for corrosion database for different shipboard operations.
Parameters affecting corrosion in bulk carriers
Table 1- Parameters Affecting Corrosion in Bulk Carriers
Operational parameters Internal (vessel-related) parameters
Cargo Type of corrosion protection
Type of ballast Member location
Operational route Corrosion Protection Systems
Cargo corrosivity
Frequency of cargo change
11
2.3.5 Corrosion Rates
Gardiner and Melchers (2003) reported that previous studies concluded that
corrosion rates were highly variable. Given the complexity of the factors affecting
corrosion, however, Gardiner (2001) identified a pattern of corrosion in bulk carriers. His
analysis provided empirically derived corrosion rates in bulk carriers (Gardiner, 2001)
from a range of 0.05 to 0.45 (exposed weather decks) mm/year depending on the
location. Likewise, the USN technical community has grouped tanks with similar
corrosion factor attributes into similar type tank groups called service codes, depending
on the content of the tank and its location. Table 2 below is the USN standards for steel
structures corrosion rates (taken from empirical data) (NAVSEA, 2017). Contents of the
tank and its location are significant to the corrosion rate per year of a particular tank. It
should also be noted the similarity in corrosion rates ranges between the two
aforementioned sources.
Table 2 - NSTM 100 Corrosion Rates for Steels
Service
Code
Type of Service Rate
Mils per Year
(mpy)
A Interior compartments, normally dry 6
B, C Interior compartments, frequently or cyclically wet
from either fresh water or seawater
15
D Seawater immersion 13
E Fresh water immersion (full) 12
F Fuels (when water is present) 13
G Contaminated fuels, oils, oily waste, waste water
(includes potentially contaminated storage tanks)
17
H CHT/VCHT 13
I Weather exposed areas or compartments 45
12
2.3.6 Corrosion Inspections
Different types of tank inspections are conducted to ascertain the condition of the
tank, its coating condition, its structural condition, and the condition of any in-place
corrosion prevention systems. They are either non-destructive or destructive type testing,
typically non-destructive. Non-destructive type testing involves a visual survey, ultra
sound in specific areas (to determine wastage), and infrared technology. Destructive
techniques involve determining if structural corrosion is present behind an apparently
intact coating surface (Martin et al, 2018). Industry and the USN use licensed inspectors
to help ensure inspection quality (American Bureau of Shipping, 2017b); (NAVSEA,
2015). American Bureau of Shipping (2017b) lists specific types of inspection types and
periodicities guidelines. U.S. Navy specific types of inspections with associated
inspections periodicities are listed in the CCAMM. Lastly, Table 3 below lists inspection
techniques that can be used while performing tank inspections to determine damage to
shipboard coating and/or structures.
2.3.7 Corrosion Consequences
The consequences of corrosion vary with the degree of severity (Rizzo et al, 2008).
Generally, corrosion in primary hull structure can lead to a significant loss of hull
strength (Wang et al, 2003). Table 3 provides a table of inspections with degradation
mechanisms (Rizzo et al, 2008).
13
Table 3 – Damage Inspection Techniques Matrix
Inspection
Technique
Damage Mechanism
Complexity in
Execution
Corrosion Fracture Mechanical
Visual X X X Low
Digital
Imaging
X X X Moderate
Pressure Test X X X Moderate
Dye Penetrant X High
Ultra Sound X X High
Magnetic
Particle
X High
2.4 Corrosion Prevention
Rizzo et al (2008) stated corrosion prevention methods were classified into two types:
active and passive. Active methods are barrier method, cathodic protection or impressed
current. Passive methods make the protective substance long-lived through design
(avoiding discontinuities or stress concentrations) or corrosion allowances (Rizzo et al
2018); (American Bureau of Shipping, 2017b); (NTSM 633 1996). Guedes and Garbatov
(2009) demonstrated that the key component to corrosion inspections is to inspect at the
right time to minimize cost.
2.5 Coatings
Coatings researched in this effort are limited to high-performance high solid and ultra-
high solid (HS/UHS) coatings that are approved for USN shipboard use. These are
14
located on the USN qualified products database (Qualified Products Database, 2019).
Thirty-three tank coatings are to be examined in this research effort. The American
Bureau of Shipping (ABS) (2017a) provides commercial recommendations for tank
coatings.
USN surface ship coating and structure coating requirements are listed in two source
documents: Naval Ship Technical Manual (NSTM) Chapter 100 (Hull Structures 2017)
and NSTM Chapter 631 (Preservation of Ships In-Service 1996). These two source
technical documents are the foundation of the Corrosion Control Assessment and
Maintenance Manual (CCAMM), the primary document for USN surface ship tank
maintenance. The CCAMM contains the requirements for tank coatings, tank inspections
periodicities, and tank risk assessments. USN technical authority over all three of these
foundational documents (NSTM 100, NSTM 631 and CCAMM) rests with Naval Sea
Systems Command (NAVSEA) Code 05P. These three documents frame any USN
surface tank technical requirements discussion regarding coatings, structure and tank
coating inspection and maintenance periodicities. This praxis research will also, by
necessity, be framed with these source documents. The American Bureau of Shipping
(2017a) contains the coatings typically used for commercial vessels.
2.6 Corrosion Modeling
The purpose of this section is to explain previous works involving corrosion modeling
and how this present work fits within the context of previous works.
Much research has been performed in order to understand and predict corrosion
progress for shipboard applications. Guedes and Garbatov (1999) proposed a non-linear
model of general corrosion wastage in a plate across all three-development phases of
15
corrosion. The conventional models of corrosion of that time assumed a constant
corrosion rate which led to a linear relationship between the materials lost versus time
(Guedes and Garbatov, 1999). Evidence of corrosion data sample reported by numerous
authors showed that a non-linear model corrosion prediction is more suitable (Guedes and
Garbatov, 1999). The significance of Guedes and Garbatov’ s research is that they
proposed a more realistic corrosion prediction model in plates across all phase of
corrosion (Guedes and Garbatov, 1999). They did not, however, factor into their work
any environmental factors (Guedes and Garbatov, 1999).
Gardiner and Melchers (2001) did, however, address the need in adding more realistic
factors in corrosion prediction modeling in bulk carriers because of the high variability of
survey thickness data. They investigated operational and internal parameters to include:
cargo, ballast, trading route, cargo corrosivity, frequency of cargo changes, and type of
corrosion protection (Gardiner and Melchers, 2001). They also proposed that these
additional parameters be factored into future corrosion modeling efforts by including
these factors into samples of future surveys of thickness measurement for inclusion in
future corrosion modeling efforts that will lead to better corrosion modeling accuracy
(Gardiner and Melchers, 2001). Factors affecting corrosion are considered in the present
work.
Gardiner and Melchers (2003) continued their work in the analysis of the factors of
immersion in seawater, exposure in enclosed atmospherics, and exposure to porous
media. Gardiner and Melchers (2003) postulated three factors impacting corrosion
modeling accuracy in bulk carriers: 1) how well the model represents corrosion
phenomena, 2) how well the model represents the uncertainty of parameters influencing
16
the corrosion phenomena, and 3) the quality of the data used to determine the empirical
corrosion model parameters. These factor into the present work. Gardiner and Melchers
(2003) and Gardiner and Melchers (2001) focused on addressing their third postulated
factor. Both Gardiner and Melchers (2001) and (2003) postulated that corrosion
prediction models for bulk carriers could be improved by the development of a corrosion
rate database that included the different operational parameters of a vessel. This database
would include data of operational factors (immersion, atmospheric, exposure to porous
media) which then could be included in corrosion modeling analysis. Melchers (2003)
then proposed a simplistic non-linear probabilistic corrosion model (for steel) that
included the aforementioned operational factors.
Wang et al (2003) conducted a survey type study of multiple smaller databases for
several types of tanks and several different types of ships. Wang et al (2003) concluded
through his analysis that: 1) corrosion wastage exhibits high variability, 2) corrosion
exhibits an increasing trend with the passage of time, and 3) corrosion wastage has an
influence on the hull girder strength. Interestingly, Gardiner and Melchers (2001) made a
similar observation of corrosion being highly variable. This research effort relates to the
work of Wang et al in that the scope will be broad like their research effort, but this
research will be of greater scope with more factors and more data points. Additionally,
this research effort will yield different research products, specifically, tank inspection
periodicities as a research product, instead of general observations that were his research
conclusions.
Paik et al (2003) developed a more complex corrosion modeling methodology for
predicting corrosion wastage as a function of time in primary structure locations (34
17
different locations in all) of single and double-hull tankers, floating storage and off load
units, and floating production storage and off load units. Approximately, 34,000 data
points were analyzed for all three phases of corrosion (Paik et al, 2003). Paik et al
concluded (2003) the following: the average annual rate of corrosion followed a Weibull
distribution but became more normal in distribution when in the severe corrosion range,
the annualized corrosion rate varied from location to location, corrosion depth could be
accurately predicted, and study findings could be used in future designs for a more
accurate corrosion allowance provision. Lastly, (Paik et al 2003) warned, for statistical
analysis of past data, to not apply corrosion models beyond that justified by the
characteristics of the database underpinning them.
In an effort to get quality inspection data for the development of a mathematical
model of coating deterioration, Melchers and Jaing (2006) developed a questionnaire that
was sent to a broad audience of tank coating surveyors. Melchers and Jaing (2006)
concluded that owners of the vessels were more pessimistic in surveying for corrosion
than expert surveyors and that good data input is vital for the development of accurate
coating deterioration modeling.
To address uncertainties in data collection and environmental complexities, Noor et al
(2007) developed and demonstrated the application of the probability distributions in
modelling the time-dependent growth of corrosion pits in vessels’ seawater ballast tanks.
In particular, Noor et al (2007) utilized the Weibull probability distribution functions to
address the high variability of corrosion wastages in seawater ballast tanks. Noor et al’s
(2007) new model demonstrated that it can be used to predict the likely variation of
corrosion depth at any point of time without having to estimate the corrosion growth rate
18
for each single defect. This simplified the corrosion modelling process, thus, making
available more data more fully useful for prediction purposes (Noor et al, 2007).
Noor et al (2013) similarly to Noor et al (2007) utilized Weibull analysis to predict
corrosion rate in structure pitting in salt water ballast tanks in bulk carriers with limited
success. Noor et al (2013) concluded the necessity of quality data in order for a more
accurate structural degradation model.
Gudze and Melchers (2008) developed a corrosion model for the prediction of
corrosion loss in the seawater ballast tanks of naval vessels. The model incorporated their
previous models (Guedes and Melchers, 2001 and 2003) which factored immersion and
atmospheric corrosions. Gudze and Melchers (2008) included the impact of the
environmental condition, namely seawater temperature, and naval operation of ballast
tanks, to model corrosion loss in this unique application. Gudze and Melchers (2008)
incorporated these two factors into their corrosion prediction model for naval units.
Guedes Soares and Garbatov (2009) produced a tank maintenance and repair model.
They conducted a probabilistic maintenance and repair analysis of two different types of
tanker deck plates subjected to general corrosion. Guedes and Garbatov (2009)
addressed the question of when to conduct maintenance and repair on corroded deck
plates. Additionally, they added a risk component in their study in the form of corrosion
tolerance. Their analysis was a Weibull analysis, which is a common methodology for
deterioration analysis (Ivosevic et al, 2017). Their analysis differs from the present work
in that the present work will be looking at more factors as well as utilizing a different
type of analysis methodology to optimize USN tank coating inspection periodicities as
previously mentioned. Guedes and Garbatov (2009) also factored repair cost in their
19
optimization which this present work will not. Guedes and Garbatove (2009) used the
term corrosion tolerance which will be utilized in this present work in presenting tank
inspection options with associated risk.
Ivosevic et al (2017) provided a probabilistic corrosion model for inner bottom plates
for fuel oil carriers. Ivosevic et al (2017) showed a logistic distribution or normal
distribution to be appropriate for the corrosion estimation model for inner bottom plates
of fuel oil carriers. Previous research efforts did not address fuel tank corrosion modeling
(Ivosevic, 2017)
In the area of tank coatings, beginning in the 1990s to the 2000s, the Naval Research
Laboratory (NRL) led the USN research effort to investigate tanks’ UHS coatings
durability for possible requirements changes in the CCAMM tank coating inspection
requirements (Lucas, 2000). Lucas (2000) proposed a USN comprehensive tank corrosion
monitoring approach. Slebodnick, (2008) concluded that UHS coating systems offered
attractive advantages and potential for a 20-year coating service life. He also reported the
critical nature that surface preparation and quality assurance has upon coating service
life. His findings were not predictive in nature and were based only on a coating failure
analysis of several limited scope case studies, and his study was not a comprehensive,
across the board, statistical regression type analysis that is proposed in this research.
The USN technical authority for tank coatings, based on previous NRL research, also
concluded that HS coatings provided promise of reaching 20 years, but more data and
more analysis were needed in order to make a definitive conclusion (Ingle, 2011). Ingle
(2011) presented NRL research that graphically compared limited tank inspection coating
results from pre-HS coatings (N=255) vs HS coatings (N=133). The graphical
20
representation demonstrated an improved longevity performance of HS coatings. Again,
no predictive, probabilistic, statistical analysis was offered to support his conclusion. By
2017, however, the USN tank coating technical community began advertising a 15-20
years UHS coating life on most types of tanks that were using the improved UHS tank
coatings (NAVSEA, 2017b), although the tank inspection periodicities requirements in
the CCAMM remained unchanged. Interestingly, the major shipbuilders of the world
(Korea, China and Japan) were long using an epoxy type HS/UHS coating for their
ballast tanks that they determined to have an expected coating life of 15 years (Eliasson
et al, 2007).
Another USN analysis effort to address extending coating inspection periodicities for
USN surface ships, given the usage of high-performance UHS coatings, was conducted
by Surface Maintenance Engineering Planning Program (SURFMEPP) (SURFMEPP,
2015). SURFMEPP (2015) conducted a limited scope Weibull type coatings analysis and
coating failure analysis. The analysis report and results, however, were not peer reviewed
nor were they accepted by the USN coatings technical warrant holder (Carrier Planning
Activity, 2016). Recently, however, the SURFMEPP methodology underwent a peer
review and was found to be faulty (Kent, 2019) as well. The SURFMEPP analysis,
although well intended, missed the mark in academic rigor and methodology, and thus, it
had to be discredited. Currently, another USN maintenance community effort, apart from
this research effort, is underway to determine tank coatings periodicities, probably using
Weibull analysis techniques.
The USN tank corrosion community recognizes that coating age, tank type, type of
coating, and surface preparation play a factor in coating deterioration (NAVSEA, 2015).
21
The relationship between coating deterioration, structural deterioration and risk
management is characterized and recognized in the CCAMM as well. The underpinnings
of the relationships of these three factors, however, is established by the expert
judgement of USN coatings and structural technical authorities, not statistical analysis.
Another aspect in the present work is to establish a regression equation between coating
and structural deteriorations, given the existing USN tank inspections data results in the
Corrosion Control Information Management System (CCIMS). This will help better
define the relationship and understanding of coating and structural degradation rates in
the USN technical community.
While these previous efforts used mathematics to predict corrosion rate, this research
effort will be utilizing a different analysis technique for predicting coating and structural
degradations than what was previously employed. This research effort will be utilizing
ordinal logistic regression since the tank inspection results in the CCIMS database are
ordinal in nature. A mathematical model will be built to predict structural corrosion
scores, given coating age, tank type, surface preparation, and tank pressure.
2.7 Corrosion Risk Assessment
In the area of risk, risk management, and risk-based inspections, Balkey et al (1990)
provides general guidelines for industrial applications. For tank applications, however,
Ifeuze and Tobins (2013) explored the role of the corrosion engineer through risk-based
inspections. In a case study, they concluded the role of risk-based inspections were valid
in determining tank inspection periodicities. Geis (2002) demonstrated the use of a risk
management tool in conducting tank inspections and maintenance. Muhammet Gul et al
(2017) explored the use of fuzzy logic in ballast tank maintenance to determine risk for
22
better safety operations. Rizzo et al (2007) considered risk-based decision procedures for
remedial actions to mitigate age related corrosion effects. USN tank risk guidance and
maintenance procedures to mitigate risk are found in the CCAMM. This research effort
will be utilizing the risk guidance found in the CCAMM. Additionally, this research
effort will employ a risk-based inspection methodology, but the scope will be much
greater than Ifeuze and Tobins (2013). Like Geis (2002), this research effort will be
utilizing a risk management tool which is in CCAMM. The risk assessment approach
used will be much simpler than Muhammet (2017) in that it will involve using a basic
lookup table in the CCAMM vice using analytical methodology. Overall, it will be
utilizing a risk-based methodology similar to others but employing the methodology to a
specifically broad issue of USN tank inspection periodicities.
2.8 Cost of Corrosion
The cost of corrosion is well documented (Parson et al, 2011); (Ingle, 2011b);
(NAVSEA, 2016). Twenty-five percent of the annual USN maintenance budget is
devoted to corrosion control (Parsons, 1994). The merit and potential savings of using
HS/UHS is and has been recognized by industry and the USN (Stein, 2015); (NAVSEA,
1997); (Ingle, 2011b); (NAVSEA 2016). Cost savings figures in the research effort are
taken directly from the USN surface navy maintenance planning organization,
SURFMEPP.
This research effort will not include:
Any tanks associated with aircraft carrier or submarines.
Any repair cost associated after a tank inspection is conducted.
23
2.9 Summary and Conclusion
Given the complex factors involved with shipboard corrosion and the complex
operational nature of USN surface ships, an academically sound and rigorous statistical
methodology is needed and required to determine and set tank coatings inspection
periodicities with associated risk. Academia, industry and the USN technical community
have identified factors affecting corrosion and have applied statistical methodologies to
predict corrosion behavior. Previous research efforts have predominately utilized Weibull
model to predict corrosion behavior (Paik et al, 2003), (Noor et al, 2007), (Noor et al,
2013), (Garbatov and Guedes, 2009). Previous research efforts have focused on many
aspects of corrosion and on trying to predict corrosion (Gardiner and Melcher, 2001),
(Gardiner and Melcher, 2003), Melcher (2003), Wang et al (2003). Given the existing
USN surface ship tank inspection data, this present work intends to conduct a much
broader analysis. Through the use of ordinal logistic regression, this present study seeks
to understand the relationship between tank coatings, surface preparation type, type of
tank, ship class, tank criticality, and coating and structural degradation in order determine
tank inspection periodicities and associated risk. This work intends to yield significant
savings for the United States Navy and for U.S. taxpayers.
24
Chapter 3—Methodology
3.1 Introduction
The intent of this chapter is to give the reader a detailed overview of how this research
was conducted and to discuss the process and procedures, step by step, from raw data to
accepting or rejecting the hypothesis.
3.2 Surface Ship Tank Inspection Data
Corrosion Control Information Management System (CCIMS) contains the United
States Navy (USN) tank inspection data repository. Tank inspectors record tank
inspection results into the CCIMS database after conducting an inspection. Tanks are
assessed using the scoring criterion provided in the Corrosion Control Assessment and
Maintenance Manual (CCAMM) (NAVSEA, 2015). Whether a coating condition
inspection or structural condition inspection, each tank inspection is given inspection
scores (ordinal data) upon inspection completion. These scores are recorded, along with
other pertinent data, in CCIMS. In the 2000s, the USN tank maintenance community
decided to consolidate tank inspection data into a single location. In 2014, the USN tank
community added a tank structural score to its inspection regiment in order to better
assess and document tank conditions (NAVSEA, 2015). Today, the CCIMS database, an
Excel spreadsheet, has over 50,000 tank inspection records, some dating back to the
1970s. A single CCIMS data record can provide up to 115 unique pieces of information
surrounding a tank inspection. Each tank inspection record contains information like:
date opened, date closed, date of inspection, the ship, tank unique identification number,
tank type, tank volume, tank area, overall coating condition, structural condition,
25
inspector’s name, coating type (if recoated), surface preparation (used), why the tank was
inspected, etc. Very specific information is also potentially available such as: square
footage and type of corrosion, repair activity doing the work, contracting organization,
etc. The database shows a mixed degree of completed fields in the records. For this
research effort, through the filtering process of record collecting, records with incomplete
fields were systematically excluded from this study.
The key fields for this present work are: Inspection Date, Ship Class, Tank Type,
Coating Type, Surface Preparation Method (for coating), Coating Inspection Score, and
Structural Inspection Score.
3.3 Database Shaping
By using the formula features in Excel, a column was created to track the time (in
years) between tank inspections for each unique tank by taking the date from one
inspection to the next inspection. When a tank was refreshed and renewed with a fresh
coat of paint, then, the coating score for that tank was reset from high coating condition
score (3 or 4) to a low coating score (1). When the tank coating condition was reset, then,
another column was added to track the age of that particular tank coating from inception
at a fresh coat of paint to subsequent inspection (with its coating inspection score) to a
following inspection (with its coating inspection score) to eventually when it was blasted
away and a new coat was applied. With this methodology, the service life of a tank
coating condition, for a particular coating type, could be tracked through the years, and
the degradation of its coating condition could be documented, monitored and tracked.
Additionally, a column was added to identify those tanks that are pressurized tanks
because the present work included an examination of the impact of pressure upon coating
26
and structural degradation. Lastly, a column was added to include the tank criticality code
taken from the Table 6-2 of the CCAMM. By USN maintenance doctrine, each tank is
assigned a tank criticality code according its mission essentialness. Tanks with a
criticality code of three are considered very mission essential while tanks with a
criticality code of one are low in mission essentiality. Table 4 displays the sample of data
base shaping.
Table 4 – CCAMM Database Sample
3.4 Database Filtering
Once the coating age, tank criticality, and tank pressure columns were set, then the
database was filtered to support the analysis of H1A and H1B. Table 5 displays actions
taken with the CCIMS database to filter the data set for H1A. Table 6 displays actions
taken with the CCIMS database to clean, shape and focus the data set for H1B. These
filtering actions formed the foundation database for H1A and H1B models.
SHIP
CLASSTANK # TANK TYPE INSPECTION #
Tank Pressure
(1 -Yes, 2,
No)
Tank Severity (1-
low, 3 high)INSP DATE
TIME BTW
INSPCOATING AGE
Coating
Score
9 665 CHT 1 No 3 16-Jan-71 0.0 0.0 1
9 665 CHT 2 No 3 12-Jun-98 27.4 27.4 4
9 666 SALT WATER BALLAST 1 No 3 16-Jan-71 0.0 0.0 1
9 666 SALT WATER BALLAST 2 No 3 12-Jun-98 27.4 27.4 2
9 666 SALT WATER BALLAST 3 No 3 03-Jul-03 5.1 32.5 4
9 667 SALT WATER BALLAST 1 No 3 16-Jan-71 0.0 0.0 1
9 667 SALT WATER BALLAST 2 No 3 03-Jul-03 32.5 32.5 4
9 668 FUEL OIL 1 No 2 16-Jan-71 0.0 0.0 1
9 668 FUEL OIL 2 No 2 03-Jul-03 32.5 32.5 1
9 669 SALT WATER BALLAST 1 No 3 16-Jan-71 0.0 0.0 1
9 669 SALT WATER BALLAST 2 No 3 12-Jun-98 27.4 27.4 3
9 669 SALT WATER BALLAST 3 No 3 03-Jul-03 5.1 32.5 4
9 670 SALT WATER BALLAST 1 No 3 16-Jan-71 0.0 0.0 1
9 670 SALT WATER BALLAST 2 No 3 12-Jun-98 27.4 27.4 2
9 670 SALT WATER BALLAST 3 No 3 03-Jul-03 5.1 32.5 4
27
Table 5 – HA1 Database Filtering Actions
Filter Action Reason Result
1. Records of Decommissioned
ships excluded
Present work studying
active fleet only
Only active ships
considered
2. Records prior to 2000
excluded
High solids/Ultra high-
solids (HS/UHS) not
implemented prior to 2000
Records of legacy paint
inspections excluded
3. Only Tanks renewed with
fresh paint coat or recoat after
2000 were included
UHS/HS were in use after
2000
Records examined in
this present work will
only be UHS/HS
coatings
4. Only Records with fields:
Coating Type, Coating Age,
Ship Class, Tank Type,
Coating Type, Surface Prep,
Coating Score were included
These factors were being
studied.
Simplified the database
field to this present
study; easier to compute
Table 6 – HB1 Database Filtering Actions
Filter Action Reason Result
1. Records of Decommissioned
ships excluded
Present work studying
active fleet only
Only active ships
considered
2. Records with fields:
Coating Score, Coating Age,
Ship Class, Tank Type,
Structural Score, Tank
Severity and Tank Pressure
were included
These factors were being
studied.
Simplified the database
field to this present
study; easier to compute
3.5 H1A and H1B Database Analysis
For this work, two different analyses were performed. The first considered the factors
(coating type, coating age, surface preparation, ship class, tank type, tank severity)
affecting coating score (degradation). The second considered the factors (coating score,
28
coating age, ship class, tank type, tank pressure and tank severity) affecting structural
score (degradation). Coating assessment score is valued at 1 to 4, 1 being pristine coating,
4 being severely corroded. Structural assessment score is valued at 1 to 5, 1 being in new
or near condition while 5 is severely degraded. The CCAMM has the complete grading
criterion for each.
Both response variables are ordinal, so an ordinal logistic regression (OLR)
methodology was chosen as the analytical methodology. Because of the type of response
variables, other regression analysis methods like ordinary least squares regression were
ruled out because the model assumptions would be violated by response variables being
ordinal. Ordinal Logistic Regression model assumptions for this work were confirmed.
Minitab 18 was chosen as the software to conduct the analysis. Predictive variables
that were labeled in text (all predictive variables except coating age) were assigned a
numeric value by using the date recode function in the Minitab drop down menu.
3.5.1 H1A Data Set
H1A: Coating Type, Application Method, Tank Type, Ship Class, Tank
Criticality, Coating Age are predictors of Coating Condition Score.
Thus, the data set for H1A was reshaped by recoding and prepared for analysis.
Table 7 shows a sample set of the recoded database for the H1A (coating degradation
analysis). Note from left to right the six independent variables followed by the single
dependent variable on the far right.
29
Table 7 – H1A Database Sample Set
COATING AGE (Years)
Recoded COATING
Type
Recoded Surface Prep
Recoded TANK TYPE
SHIP CLASS
RC Tank Criticality
Coating Score
3.98 8 8 7 1 3 1
4.04 8 7 7 1 3 1
2.28 8 7 11 1 1 1
4.04 8 7 7 1 3 1
4.76 28 8 7 1 3 2
4.76 28 8 7 1 3 2
4.04 8 7 7 1 3 1
9.67 28 8 7 1 3 3
3.86 8 7 7 1 3 1
3.5.2 H1B Data Set
H1B: Coating Condition Score, Ship Class, Tank Pressure, Coating Age, Tank
Criticality are predictors of Structural Condition Score.
Additionally, the data set for H1B was reshaped by recoding and prepared for
analysis. Table 8 shows a sample set of the recoded data base for the H1B (structural
degradation analysis). Note from left to right the six independent variables followed by
the single dependent variable on the far right.
30
Table 8 – H1B Database Sample Set
COATING AGE
Coating Score
Recoded Tank Type
RC SHIP CLASS
RC Tank Pressure
Tank Severity
Structural Score
32.61 3 10 8 1 3 5
32.61 3 3 8 1 2 5
15.18 4 10 8 2 3 5
32.61 4 11 8 1 1 3
33.43 4 1 8 1 1 5
10.56 2 2 1 1 3 1
3.13 1 6 1 1 2 1
3.13 1 6 1 1 2 1
10.99 2 9 1 1 3 1
3.5.3 H1A and H1B Data Partitioning
H1A and H1B data sets were each partitioned into a training data set and a testing
data set through the random function in Excel which assigned values between 0 and 1 to
each record file. Those values 0.80 and above were grouped into a testing data set. Those
values less than 0.80 were grouped in the training data set. Thus, each hypothesis had a
training data set and a corresponding testing data set.
3.5.4 H1A and H1B Model Training
Ordinal logistic analysis was performed for each training data set using Minitab.
The logit function was selected for each. Regression results were captured in an output
file. The model initially included all predictive factors. Based on the model’s initial
regression results, least significant factors were removed one at time, and the model
regression was conducted again. Regression results again were analyzed to determine
factors’ significance. Model results included the regression analysis output (response
information, logistic regression table, test of all slopes equal to zero, goodness-of-fit test,
31
and measures of association), coefficient of the estimation equation, and the event
probabilities. Each training result was tested for significance, goodness-of-fit, and
measures of association. A regression equation for reach model was derived as well and
subsequently used in model testing.
3.5.5 H1A and H1B Model Testing
For H1A and H1B, the OLR output event probabilities for each regression output
were inserted into each respective training database. The training data were subsequently
tested for accuracy using the following methodology. Using the regression output
response variable values and coefficients, logit values were calculated using the logit
function equation for each categorical outcome.
Calculated logit = B0 + B1X1+B2X2+…BkXk (3.1)
B0 is the categorical coefficient, and B1 and X1 are the predictor variable coefficient
and predictor value, respectively and so forth for other predictive variables.
Then, an event probability was calculated for each categorical outcome using the
corresponding calculated logit value input to the regression equation.
Calc 1 Prob =e(Calc Logit 1)/(1+e(Calc Logit 1)) (3.2)
Subsequent categorical outcome probabilities were calculated as follows:
Calc 2 Prob =e(Calc Logit 2)/(1+e(Calc Logit 2))-Calc 1 Prob (3.3)
Calc 3 Prob =e(Calc Logit 3)/(1+e(Calc Logit 3))-(Calc 1 Prob)-(Calc 2 Prob) (3.4)
Calc 4 Prob =1-(Calc 1 Prob)-(Calc 2 Prob)-(Clac 3 Prob) (3.5)
32
Table 9 – Training Database Testing Sample Set
The highest value of the calculated probability outcomes for each inspection record
was chosen as the predicted inspection score. See the record circled on the far right
portion of Table 9. The probability of score 1 is almost 86% percent, so coating score 1
is predicted. Predicted value was compared with actual coating score inspection value,
and the results were registered in an added column at the right of the training database.
This methodology was completed for each training data set to determine preliminary
model accuracy.
In order to test the trained models, a similar methodology was followed except the
output probabilities were not entered into the test database. Using the training data
regression output predictors’ values and coefficients, logit values were calculated using
the logit function equation for each categorical outcome, like before, except for the
testing data. The highest value of the calculated probability outcomes for each testing
data inspection record was chosen as the predicted inspection score. Predicted value was
compared with actual coating score inspection value, and the results were registered in an
added column at the right of the training database. This methodology was completed for
each testing data set to determine model accuracy. Table 10 shows a sample of this
testing of a testing data set.
Coating
Age
Coating
Type
Surface
Prep
Method
Tank
Type
Ship
Class
Tank
Criticalit
y
Coating
Inspection
Score
EPROB
1
EPROB
2
EPROB
3
EPROB
4
Calc 1
Logit
Calc 2
Logit
Calc 3
Logit
Calc 4
Logit
Calc 1
Prob
Calc 2
Prob
Calc 3
Prob
Calc 4
ProbPredict
Accurate
?
3.98 8 8 7 1 3 1 0.857 0.112 0.029 0.002 1.786 3.406 6.206 -22.09 0.856 0.111 0.03 0.002 1 yes
0.43 22 7 9 1 3 1 0.991 0.007 0.002 1E-04 4.67 6.29 9.09 -19.21 0.991 0.007 0.002 1E-04 1 yes
4.04 8 7 7 1 3 1 0.984 0.013 0.003 2E-04 4.082 5.702 8.502 -19.8 0.983 0.013 0.003 2E-04 1 yes
2.28 8 7 11 1 1 1 0.995 0.004 8E-04 5E-05 5.364 6.984 9.784 -18.52 0.995 0.004 9E-04 6E-05 1 yes
4.04 8 7 7 1 3 1 0.984 0.013 0.003 2E-04 4.082 5.702 8.502 -19.8 0.983 0.013 0.003 2E-04 1 yes
4.76 28 8 7 1 3 2 0.661 0.249 0.084 0.006 0.662 2.282 5.082 -23.22 0.66 0.248 0.086 0.006 1 no
4.76 28 8 7 1 3 2 0.661 0.249 0.084 0.006 0.662 2.282 5.082 -23.22 0.66 0.248 0.086 0.006 1 no
= B0 + B1X1+B2X2+…BkXk = e(Calc Logit 1)/(1+e(Calc Logit 1))
33
Table 10 – Test Results Sample of Testing Data Set
The testing of the trained models for H1A and H1B was performed in this manner.
3.5.6 Model Classification Matrix and Optimization
A classification matrix for each model was built using the training data model
results to determine how well each model classified. Models were then optimized to
address model overconfidence and to reduce the number of false positives (Type 1 errors)
while maintaining nearly the same model accuracy.
Minitab’s Design of Experiments was utilized to determine the required combinations
needed, given the number of factors (response variables logit coefficients) factors and
levels for each factor. Tables 11 and 12 are samples of the Design of Experiments factors
used for H1A and H1B. For H1A, the four responsive variable outputs (coating scores 1-
4) represented the four factors in the Design of Experiments.
Table 11 – H1A Design of Experiments Sample
COATING
AGE
Coating
Type
Surface
Prep
TANK
TYPE
SHIP
CLASS
Tank
Criticality
Coating
Insp
Score
EPROB1
from
Minitab
EPROB
2 from
Minitab
EPROB3
from
Minitab
EPROB4
from
Minitab
Calc 1
Logit
Calc 2
Logit
Calc 3
Logit
Calc 4
Logit
Calc 1
Prob
Calc 2
Prob
Calc 3
Prob
Calc 4
ProbPredict
Accurate
?
0.347222 18 8 3 1 2 1 No values 3.4 5.02 7.82 -20.48 0.968 0.026 0.006 4E-04 1 yes
3.863889 8 7 7 1 3 1 4.147 5.767 8.567 -19.73 0.984 0.012 0.003 2E-04 1 yes
0.405556 8 7 13 1 2 1 5.458 7.078 9.878 -18.42 0.996 0.003 8E-04 5E-05 1 yes
3.863889 8 7 7 1 3 1 No values 4.147 5.767 8.567 -19.73 0.984 0.012 0.003 2E-04 1 yes
4.041667 8 7 7 1 3 1 4.082 5.702 8.502 -19.8 0.983 0.013 0.003 2E-04 1 yes
4.041667 8 7 7 1 3 1 4.082 5.702 8.502 -19.8 0.983 0.013 0.003 2E-04 1 yes
0.072222 20 8 2 1 3 1 No values 3.028 4.648 7.448 -20.85 0.954 0.037 0.009 6E-04 1 yes
4.130556 8 7 6 1 2 1 4.57 6.19 8.99 -19.31 0.99 0.008 0.002 1E-04 1 yes
StdOrder RunOrder PtType BlocksCS Coeff
1
CS Coeff
2
CS Coeff
3
CS Coeff
4
1 1 1 1 22.33 24 26.8 -1.5
2 2 1 1 22.33 24 26.8 -1
3 3 1 1 22.33 24 26.8 -0.5
4 4 1 1 22.33 24 26.8 0
5 5 1 1 22.33 24 26.8 0.5
6 6 1 1 22.33 24 26.8 1
7 7 1 1 22.33 24 26.8 1.5
34
Table 12 – H1B Design of Experiments Sample
For H1A, seven levels were chosen for each factor. The range of the levels was
centered around the response variable coefficients from the OLR in order to examine the
results region around the OLR response variable coefficients. For example, the H1A OLR
response coefficient for Coating Score 1 was 23.88. CS1 (coating score) coefficient levels
were 25.33, 24.88, 24.33, 23.88, 23.33, 22.88 and 22.33. For H1A, there were 2,401
response variable logit coefficient combinations examined (47) (number of factors raised
to the number of levels).
For H1B, five factors were chosen since there are five response variable coefficients.
Five levels were chosen for each factor. Similar to H1A, the range of levels were
centered around the response variable coefficient from OLR. For H1B, there were 3,125
response variable logit coefficient combinations examined (55) (number of factors raised
to the number of levels).
Visual basic for application (VBA) code in Excel was used to automate the input of
each combination of response coefficients to execute the logit function (Equation 3.1), to
calculate probabilities (Equation 3.2), to generate a classification matrix and to populate
an overall model results spreadsheet, given all combinations of factors. Criteria for the
selection of the optimized factors were based on overall model performance balanced
StdOrder RunOrder PtType BlocksCS Coeff
1
CS Coeff
2
CS Coeff
3
CS Coeff
4
CS Coeff
5
1056 1 1 1 2.3 3.55 3.19 2.76 -1
1192 2 1 1 2.3 4.05 3.19 3.76 -0.5
190 3 1 1 1.8 2.55 3.19 3.26 1
1905 4 1 1 3.3 2.05 2.69 2.26 1
1041 5 1 1 2.3 3.55 2.69 3.76 -1
1838 6 1 1 2.8 4.05 3.69 3.26 0
2370 7 1 1 3.3 3.55 4.19 3.76 1
35
with minimizing the risk to the most mission critical tanks (tank criticality 3) from Type
1 errors from overconfident models. Model classification performance measures
included: overall accuracy, error rate average, precision average, recall average, false
positive rate (FPR) average, Cohen's kappa coefficient (k) and Mathews correlation
coefficient (MCC).
For both models, classification matrices were then obtained by using the optimized
regression coefficients. Additionally, classification matrices for each model were also
built using model testing data results to confirm intended model optimizations. Tables 13
and 14 are classification matrix examples for H1A and H1B, respectively.
Table 13 – H1A Classification Matrix Example
For Table 13, H1A predicted coating scores were classified against actual observed
inspection results. True positives (predicted value matched actual values) are shaded in
the gray diagonal. TP1 represents all tank coating inspection values predicted as coating
score 1 which when inspected were actually inspected as coating score 1. The other
model results classified in error are located on either side of the gray diagonal. The error
1 2 3 4
1 TP1 EE12 EE13 EE14
2 EE21 TP2 EE23 EE24
3 EE31 EE32 TP3 EE34
4 EE41 EE42 EE43 TP4
Predicted
H1A Model
Classification Matrix
Observed
36
cell annotation EERC means error on row R, column C. For example, EE13 (highlighted in
Table 13) represents all tank coating inspection values predicted as coating score 1, but
when inspected, the tank inspection results actually indicated a coating score of 3. Table
14 represents the H1B classification matrix. The notation in Table 14 is the same as that
which is used in Table 13. It should be noted that Table 14 has an additional row and
column since H1B has five categorical responses vice four in H1A.
Thus, for each model, the classification matrix tool was used to determine the number
of false positives (type 1 errors), false negatives (type 2 errors), true positives, and true
negatives for each categorical outcome. Model optimization focused on minimizing
errors in the pink regions of Tables 13 and 14 while maintaining overall model
performance. This pink region represented areas where model performance was
overconfident; that is, predictions were optimistic, while actual results were worse than
predicted. For example, the highlighted cell EE13 in Table 13 represents tanks predicted
to be a coating score 1 but were actually a coating score 3 when inspected. This is a false
positive situation. Type 1 errors represent the worst case scenario for this research work.
Table 14 – H1B Classification Matrix Example
1 2 3 4 5
1 TP1 EE12 EE13 EE14 EE15
2 EE21 TP2 EE23 EE24 EE25
3 EE31 EE32 TP3 EE34 EE35
4 EE41 EE42 EE43 TP4 EE45
5 EE51 EE52 EE53 EE54 TP5
H1B Model Observed
Classification Matrix
Predicted
37
3.5.7 H1A Coating Age Prediction Tool
As a product of the regression analysis of the coating score training data, by using
the regression output predictors’ values and coefficients, the coating age can be predicted
if other independent variables are known, and the desired output coating score is known.
This can be accomplished through manipulating the logit function and the regression
equation. Figure 1 shows a sample for the Coating Age Prediction Tool.
Figure 1 – Coating Age Prediction Tool Sample
3.5.8 H1B Structural Score Prediction Tool
As a product of the regression analysis of the coating score training data, by using
the regression output predictors’ values and coefficients, the structural score can be
predicted if other independent variables are known. The methodology for developing this
Predict age, given
Inputs below and Prob
entered to the right,
Enter Prob (0.XX) 0.9
AGE
Predicted
(years)
20.20
Input Factors projected minimal years to Coating Score 3
Coating Type (1-34) 27
App Method (1-8) 7
Tank Type (1-13) 10 COEFF Score
Ship Class (1-14) 11 23.8872 1.0000
Severity (1,2,3) 3 25.5316 2.0000
Coating Score (1-4) 3 28.3489 3.0000
0.0000 4.0000
COEFF 28.3489
CA -0.3626
APL -0.0422
SP -2.3174
TT -0.0660
SC 0.0503
TS -0.4537
38
tool is the same as the previous tool, manipulation of the logit function and the regression
equation. Figure 2 shows a sample for the Structural Score Prediction Tool.
Figure 2 – Structure Score Prediction Tool Sample
3.6 H2 Data and Data Analysis
A sample of the H2 database is below in Table 15. H2 is:
H2: Tank inspection periodicities, where HS/UHS coatings are employed, can be
increased 25-50% using the coating and structure score predictions (developed from
H1A and H1B) with minimal increased risk of failure.
Factors
Enter
Values
Below
Predicted
Structural
Score
Coating Age (yrs) 15 1
Coating Insp Score (1-4) 3
Tank Type (1-13) 10
Ship Class (1-10) 10
Tank Pressure (1-2) 2
Tank Criticality (1 -3) 3
1.991
3.052
3.193
3.254
05
-0.0202 CA
-0.2260 CS
-0.0487 TT
-0.0938 SC
0.7490 TP
-0.5257 TC
39
Table 15 – H2 Database Sample
The methodology used to prove H2 is systematically explained by explaining the
columns in Table 15. Each tank type is listed. Tank criticality code is listed for each tank.
The current inspection requirement periodicity for each tank (from the CCAMM) is
listed. From the H1A age prediction tool, the minimum years is derived for coating
score 3 (which is considered failure in the CCAMM). From the H1B structural score
prediction tool, the predicted structural score is given for a coating score of 3. The risk
assessment is derived using CCAMM table 6-1 given inputs (coating score 3, predicted
structure score (previous column), and tank severity (previous column)). Overall number
of inspections required in reaching minimum years (Min years/Req Inspection
Periodicity) is listed.
Net inspections equals overall inspections minus one. Percent savings is net savings
divided by overall inspections times one hundred. This methodology was used to
determine any inspection savings.
3.7 H3 Data and Data Analysis
A sample of the H3 database is below in Table 16. H3 is:
H3: There are significant savings when tank inspection periodicities increase by
25-50% (results from H2).
Tank
Type
Tank Criticality
Code
Coating
Inspection
Requirement
Periodicity
(years) per Navy
Corrosion
Manual
Minimum Years
to reach Coating
Score 3, using
H1A Coating Age
Model, (given:
coating #28, sp #
7 & tank
criticality code)
for ALL classes
with 95%
probability
Projected Structural
Score, using H1B
Structural Model
prediction tool, given
Minimum Coating Age
(years), Coating Score
#3 and Tank Criticality
Risk Assumed given
Coating Score 3 &
Projected Structure
Score per Navy
Corrosion Manual Risk
Matrix (Table 6-1)
Overall #
coating
inspections
required
during
Minimum Years
is reached (Min
Yrs/Req Insp
Periodicity)
Net Inspection
savings (Overall
Insp # less 1), if
periodicity
extended to
projected
minimal years to
Coating Score 3
Percent Insp
Savings by
extending
periodicity of
inspections (for
20 year
inspection
timeframe)
1 1 10 20.8 1 Low 2 1 50.0
2 3 4 18.1 1 Low 5 4 80.0
3 2 3 19.2 1 Low 6 5 83.3
4 3 6 17.7 1 Low 3 2 66.7
40
Table 16 – H3 Database Sample
The methodology used to prove H3 is systematically explained by explaining the
columns in Table 16. Each tank is listed. Average schedule cost per inspection for each
tank is listed (provided by Surface Maintenance Engineering Planning Program
(SURFMEPP). Average material cost per for each tank is listed (provided by
SURFMEPP). Net inspections savings (from H2) is listed. Total average schedule cost
savings in man-days (per tank) is calculated (net savings times schedule cost per
inspection). Total average material savings (per tank) is calculated (net savings times
average material cost per inspection). Total number of each type of tank in the surface
navy is listed. Aggregate schedule cost savings is listed (total average savings per tank
times the total number of that type of tank). Aggregate material cost savings is listed
(total average material savings per tank times the total number of that type of tank). The
aggregates are summed at the bottom. This is the methodology used to determine the
merit of H3.
Tank
Type
Average
Schedule Cost
per Inspection
(mandays)
Average
Material Cost
per Inspection
($ dollars)
Net Inspection
savings (from H2)
Total Ave Schedule
Cost Savings in
Mandays (per tank
type)
[Net Inspection
Savings X (Schedule
Cost/Insp)]
Total Ave Material
Cost Savings in dollars
($) (per tank type)
[Net Inspection
Savings X (Ave Mat
Cost/Insp)]
Total # of each
type of tank in
the surface
fleet
For 20 year
inspection
timeframe,
Aggregate
Schedule Cost
Savings (in
Mandays) for
each type of
tank across the
fleet (Total
Sched
Savings/per
tanks X # of
tank type in
fleet)
For 20 year
inspection
timeframe,
Aggregate
Material Cost
Savings in $
(dollars) for
each type of
tank across the
fleet (Total
Material
Savings/per
tank X # of tank
type in fleet)
1 16.55 $15,719.57 1 16.55 15719.57 344.0 5,694.0 $5,407,531.22
2 11.81 $5,966.33 4 47.23 23865.31 340.0 16,057.2 $8,114,206.52
3 11.24 $5,605.15 5 56.21 28025.73 1476.0 82,959.8 $41,365,979.90
4 20.63 $60,787.06 2 41.26 121574.12 107.0 4,415.1 $13,008,430.84
41
3.8 Conclusion
The CCIMS database was shaped, filtered and recoded. The recoded databases were
partitioned into training and testing databases. Ordinal logistic regression was principally
used to determine the merit of H1A and H1B. Both models were optimized and tested.
Predictions tools for both H1A and H1B were used as input to the development of the H2
database. H2 analysis involved simple estimating. The H2 output provided an input to H3
which was another simple estimating analysis.
42
Chapter 4—Results
4.1 Introduction
The purpose of this chapter is to present the results of the methodology presented in
the previous chapter. Hypothesis results are presented in separate sections. This work’s
hypotheses are:
H1A: Coating Type, Application Method, Tank Type, Ship Class, Tank
Criticality, Coating Age are predictors of Coating Condition Score.
H1B: Coating Condition Score, Ship Class, Tank Pressure, Coating Age, Tank
Criticality are predictors of Structural Condition Score.
H2: Tank inspection periodicities, where HS/UHS coatings are employed, can be
increased 25-50% using the coating and structure score predictions (developed from
H1A and H1B) with minimal increased risk of failure.
H3: There are significant savings when tank inspection periodicities increase by
25-50% (results from H2).
4.2 Models’ Assumptions and Results
Ordinal logistic regression model assumptions and results are presented in Table 17.
43
Table 17 – Assumption Results for H1A and H1B Models
Assumption Description Response Assumption met?
#1 Response Variable should
be measured at the ordinal
levels
Models have ordinal
Response Variables
Yes
#2 One or more predictor
variables that are
continuous, ordinal or
categorical must be used.
Models have multiple
Independent
Variables that meet
criterion
Yes
#3 There is no
multicollinearity amongst
Predictor Variables.
Correlation Matrices
results provided in
Tables 18 and 19.
Yes
#4 Proportional Odds Test of Parallelism
conducted. Figures
3-6 are provided.
Yes
Results from Tables 18 and 19 indicate no multicollinearity amongst predictor
variables. For both tables, correlation results indicated that there was enough evidence to
determine that the results were valid and that correlation coefficients were significantly
less than one, which indicates weak association amongst predictor variables and reduces
the risk of multicollinearity.
44
Table 18 – H1A Correlation Results
H1A Predictor Variables
COATING AGE
COATING Type
Surface Preparation
TANK TYPE
SHIP CLASS
Tank Criticality
COATING AGE 1
COATING Type -0.07286 1
Surface Preparation 0.261756 -0.05097 1
TANK TYPE -0.01481 0.066063 -
0.03222271 1
SHIP CLASS 0.007996 0.022524 -
0.02387041 0.22416 1
Tank Criticality -0.07854 0.159009 -0.005555 0.07628 0.0444 1
Table 19 – H1B Correlation Results
H1B Predictor Variables
COATING AGE
Coating Score
Tank Type
SHIP CLASS
Tank Pressure
Tank Criticality
COATING AGE 1
Coating Score 0.379248 1
Tank Type -0.01678 -0.02201 1
SHIP CLASS 0.020805 -0.1099 0.113923 1
Tank Pressure 0.022272 0.137278 0.091712 -0.15794 1
Tank Criticality -0.10938 0.124292 -0.0329 0.047648 0.56908 1
The proportional odds assumption, also called the assumption of parallel lines,
maintains that the impact of a predictor variable on the response variable is uniform over
all categories of response variable (UCLA, 2019). H1A and H1B have six predictor
variables each for a total of twelve predictor variables. All twelve predictor variables
45
passed the proportional odds test in demonstrating variable consistency across categorical
outcomes by virtue of having parallel lines. Figures 3 and 4 are two of the six parallel test
results for H1A. Figure 3 is an H1A continuous predictor variable and Figure 4 is H1A a
discrete predictor variable. Figures 5 and 6 are two of the six parallel test results for H1B.
Figure 5 is an H1B continuous predictor variable and Figure 6 is an H1B discrete
predictor variable. The proportional odds assumption with the ordinal logistic model was
not violated for either H1A or H1B models.
Figure 3 – H1A Coating Age (Predictor Variable) Parallel Test Results
46
Figure 4 – H1A Coating Type (Predictor Variable) Parallel Test Results
47
Figure 5 – H1B Coating Age (Predictor Variable) Parallel Test Results
48
Figure 6 – H1B Coating Score (Predictor Variable) Parallel Test Results
4.2.1 Database Shaping and Filtering Results
Table 20 shows the data shaping and filtering for H1A and H1B. Filtering of surface
ships tank inspection records began with removing decommissioned ships records which
resulted in 54,738 records remaining. Because this research effort involves analysis of
49
HS/UHS tank coatings, tanks that were coated prior to 2000 were filtered out (Table 20
filtering actions 2 and 3). Lastly, the remaining 24,311 records were filtered according to
variables associated with H1A and H1B resulting in 4,250 and 2,397 records, respectively.
Partitioning of H1A and H1B datasets represents an 80/20 split into training and test sets.
Table 20 – Data Shaping and Filtering Results
Filter Action Reason Results
Tank Inspection Record
Description
Filter Action Taken Resultant
Records
Remaining
1. All Records Minus Decommissioned vessels
records
54,736
2. All Active Vessels Tank
Inspections (54,736)
Minus Records prior to the year
2000
42,099
3. All Active Vessels Tank
Inspections since 2000 (42,099)
Include only reset tank inspection
results and any subsequent
inspections since the year 2000
24,311
3a. Tank Inspection Records
for Tanks with HS/UHS
Coatings (24,311)
Include only records that have
Hypothesis 1A analysis predictor
and response variables
4,250
3b. Hypothesis 1A Database
(4,250)
Partition data into train and test
Data Sets
3,357/893
(train/test)
4. (from #2 above) All Active
Vessels Tank Inspections
(54,736)
Include only records that have
Hypothesis 1B analysis predictor
and response variables
2,397
4a. Hypothesis 1B Database
(2,397)
Partition data into train and test
Data Sets
1,943/454
(train/test)
50
4.3 H1A Database Composition
Table 21 shows the frequency distribution for six categorical variables in the H1A
database. For example, Coating Type 1 has four records, and Tank Type 1 has 146
records.
Table 21 – H1A Database Frequency Distribution
Figure 7 shows the frequency distribution for each categorical variable in H1A in
graphical form. Figure 8 shows the coating age histogram for H1A. Given these two
figures, it is observed that the H1A dataset is not evenly distributed which means that
Coating
Type (PV)
Surface Preparation
Method (PV)
Tank Type
(PV)
Ship Class
(PV)
Tank
Criticality
(PV)
Coating
Score
(RV)
1 4 7 146 569 393 3949
2 4 10 300 1814 945 192
3 15 2 384 0 2912 99
4 93 4 43 0 10
5 4 1 60 2
6 58 21 283 0
7 24 3584 1125 9
8 1064 621 2 729
9 124 575 52
10 17 809 175
11 35 247 435
12 1 12 319
13 9 264 0
14 1 146
15 92
16 2
17 310
18 349
19 165
20 385
21 89
22 122
23 2
24 70
25 5
26 10
27 22
28 1141
29 1
30 5
31 2
32 2
33 23
Frequency Count All Data Set
Predictor and Response Variables Count
Variable
Number
51
training and testing data sets should have similar distributions. Additionally, future model
performance should reflect the impact of given distributions. It should be noted that
uneven distributions do not violate the use of the proportional odds model in ordinal
logistic regression methodology (McCullagh, 1980).
Figure 7 – H1A Database Frequency Distribution (Graph)
0
500
1000
1500
2000
2500
3000
3500
4000
4500
0 5 10 15 20 25 30 35
H1A DATA BASE COMPOSTION (VARIABLES VS FREQUENCY COUNT)
Coating Type (PV) Surface Preparation Method (PV)
Tank Type (PV) Ship Class (PV)
Tank Criticalilty (PV) Coating Score (RV)
52
Fig
ure
8 –
H1A
Dat
abas
e
Coat
ing A
ge
(PV
)
Fre
quen
cy D
istr
ibuti
on
53
4.3.1 H1A Train Database Composition
Through the 80/20 split partitioning of H1A database, the train database (80%) was
formed. Table 22 represents the frequency distribution for the six categorical variables in
the H1A train database. It is read in the same fashion as Table 21. Inspection of Tables
21 and 22 shows similar uneven distributions. The H1A training database (Table 22) is a
fair representation of the H1A database (Table 21).
Table 22 – H1A Train Database Frequency Distribution
Coating
Type (PV)
Surface Preparation
Method (PV)
Tank Type
(PV)
Ship Class
(PV)
Tank
Criticality
(PV)
Coating
Score
(RV)
1 3 5 107 451 290 3120
2 3 6 231 1436 762 154
3 12 2 307 0 2305 76
4 70 4 36 0 7
5 3 1 46 2
6 45 14 233 0
7 21 2839 903 5
8 850 486 1 573
9 104 452 43
10 13 637 148
11 28 183 330
12 1 10 260
13 7 211 0
14 1 109
15 76
16 2
17 245
18 284
19 133
20 310
21 66
22 88
23 2
24 59
25 5
26 9
27 17
28 877
29 0
30 4
31 2
32 2
33 15
Frequency Count Train Data Set
Variable
Number
Predictor and Response Variables Count
54
Figure 9 shows the frequency distribution for each categorical variable in the H1A
train database in graphical form. Figure 10 shows the coating age histogram for H1A
train database. Given these two figures, it is observed that the H1A train database is an
accurate representation of the overall H1A database, comparing Figures 7 through 10.
Figure 9 - H1A Train Database Frequency Distribution (Graph)
0
500
1000
1500
2000
2500
3000
3500
0 5 1 0 1 5 2 0 2 5 3 0 3 5
HYP 1A TRAIN DATA BASE (VARIABLES VS FREQUENCY COUNT)
Coating Type (PV) Surface Preparation (PV) Tank Type (PV)
Ship Class (PV) Tank Criticality (PV) Coating Score (RV)
55
Fig
ure
10 –
H1A
Tra
in
Dat
abas
e C
oat
ing A
ge
(PV
) F
requen
cy
Dis
trib
uti
on
56
4.3.2 H1A Test Database Composition
Through the 80/20 split partitioning of H1A database, the test database (20%) was
formed. Table 23 represents the frequency distribution for the six categorical variables in
the H1A test database. It is read in the same manner as previous similar distribution
tables. Inspection of Tables 23 and 22 show similar distributions. That is, the H1A
testing database (Table 23) is a fair representation of the H1A training database (Table
22).
Table 23 - H1A Test Data Frequency Distribution
Coating
Type (PV)
Surface Preparation
Method (PV)
Tank Type
(PV)
Ship Class
(PV)
Tank
Criticality
(PV)
Coating
Score
(RV)
1 1 2 39 118 103 829
2 1 4 69 378 183 38
3 3 0 77 0 607 23
4 23 0 7 0 3
5 1 0 14 0
6 13 7 50 0
7 3 745 222 4
8 214 135 1 156
9 20 123 9
10 4 172 27
11 7 64 105
12 0 2 59
13 2 53 0
14 0 37
15 16
16 0
17 65
18 65
19 32
20 75
21 23
22 34
23 0
24 11
25 0
26 1
27 5
28 264
29 1
30 1
31 0
32 0
33 8
Frequency Count Test Data Set
Variable
Number
Predictor and Response Variables Count
57
Figure 11 shows the frequency distribution for each categorical variable in the H1A
test database in graphical form. Figure 12 shows the coating age histogram for H1A test
database. Given these two figures, it is observed that the H1A test database is an accurate
representation of the H1A train database, comparing Figures 9 through 12.
Figure 11 – H1A Test Database Frequency Distribution (Graph)
0
100
200
300
400
500
600
700
800
900
0 5 1 0 1 5 2 0 2 5 3 0 3 5
H1A TEST DATA BASE (VARIABLES VS FREQUENCY COUNT)
Coating Type (PV) Surface Preparation Method (PV)
Tank Type (PV) Ship Class (PV)
Tank Criticalilty (PV) Coating Score (RV)
58
Fig
ure
12 –
H1A
Tes
t D
atab
ase
Coat
ing A
ge
(PV
) F
requen
cy D
istr
ibuti
on
59
4.4 H1B Database Composition
The methodology for obtaining the H1B database is detailed in section 4.2.1 of this
chapter and previously in Chapter 3 Methodology. Table 24 shows the frequency
distribution for the categorical variables in the H1B database. Like the H1A database, the
H1B database is unevenly distributed. Therefore, the same implications previously stated
for the H1A database apply to H1B database.
Table 24 - H1B Database Frequency Distribution
Figure 13 shows the frequency distribution for each categorical variable in H1B in
graphical form. Figure 14 shows the coating age histogram for H1B.
Coating
Score (PV)Tank Type (PV)
Ship Class
(PV)
Tank
Pressure
(PV)
Tank
Criticality
(PV)
Structure
Score
(RV)1 928 48 192 1656 300 1718
2 613 99 1104 741 829 112
3 535 278 3 1268 53
4 321 9 15 22
5 44 334 492
6 428 2
7 563 346
8 6 224
9 201 81
10 352 96
11 252
12 12
13 105
Frequency Count All Data Set
Predictor and Response Variables Count
Variable
Number
60
Figure 13 – H1B Database Frequency Distribution (Graph)
Figure 14 – H1B Database: Coating Age Frequency Distribution
4.4.1 H1B Train Database Composition
Through the 80/20 split partitioning of H1B database, the train database (80%) was
formed. Table 25 shows the frequency distribution for categorical variables in the H1B
61
train database. Inspection of Tables 25 and 24 shows similar uneven distributions. The
H1B training database (Table 25) is a fair representation of the H1B database (Table 24).
Table 25 – H1B Train Database Frequency Distribution
Figure 15 shows the frequency distribution for each categorical variable in the
H1B train database in graphical form. Figure 16 shows the coating age histogram for
H1B train database. Comparing Figures 13 through 16, it is observed that the H1B train
database is a fair presentation of the H1B database.
Coating
Score (PV)Tank Type (PV)
Ship Class
(PV)
Tank
Pressure
(PV)
Tank
Criticality
(PV)
Structure
Score
(RV)1 748 37 156 1333 233 1401
2 498 81 904 610 663 88
3 438 221 3 1047 44
4 259 9 13 20
5 36 278 390
6 349 2
7 468 279
8 5 179
9 168 54
10 285 75
11 196
12 10
13 78
Predictor and Response Variables Count
Variable
Number
Frequency Count Train Data Set
62
Figure 15 – H1B Train Database Frequency Data Base (Graph)
Figure 16 – H1B Train Database: Coating Age Frequency Distribution
4.4.2 H1B Test Database
Through the 80/20 split partitioning of H1B database, the test database (20%) was
formed. Table 26 shows the frequency distribution for the categorical variables in the
63
H1B test database. Inspection of Tables 26 and 25 shows similar uneven distributions.
The H1B testing database (Table 26) is a fair representation of the H1B training database
(Table 25).
Table 26 – H1B Test Database Frequency Data Base
Figure 17 shows the frequency distribution for each categorical variable in the H1B
test database in graphical form. Figure 18 shows the coating age histogram for H1B test
database. Given these two figures, it is observed that the H1B test database is not evenly
distributed as well and is also an accurate representation of the train H1B database,
comparing Figures 15 through 18.
Coating
Score (PV)Tank Type (PV)
Ship Class
(PV)
Tank
Pressure
(PV)
Tank
Criticality
(PV)
Structure
Score
(RV)1 180 11 36 323 67 317
2 115 18 200 131 166 24
3 97 57 0 221 9
4 62 0 2 2
5 8 56 102
6 79 0
7 95 67
8 1 45
9 33 27
10 67 21
11 56
12 2
13 27
Predictor and Response Variables Count
Frequency Count Test Data Set
Variable
Number
64
Figure 17 – H1B Test Data Base Frequency Distribution
Figure 18 – H1B Test Data: Coating Age Frequency Distribution
65
4.5 H1A OLR Results
Initial OLR results from a regression analysis with all predictor variables (coating age,
coating type, surface preparation method, tank type, ship class, tank criticality and tank
pressure) indicated that the predictor variable tank pressure was not significant (p > 0.05).
Tank pressure was removed as a predictor variable from the analysis, and the regression
was performed again. Figure 19 displays the final regression analysis results of H1A
using the training data set.
Figure 19 – H1A OLR Results
66
H1A OLR results indicated statistical significance for predictive variables (p-values
less than 0.05). Test of All Slope equal to Zero is good (p-value less the 0.05). Goodness
of Fit is satisfactory since P =1.0, and Measures of Association between variables is
strong, since Somers’D and Goodman-Kruskal Gamma are 0.89 (near 1.0). Overall, the
H1A OLR results indicate a good fit between data and the model.
4.5.1 H1A Model Training and Testing Results
To determine accuracy of each trained model, each dataset was tested using the
following methodology. As also described in section 3.5.4 5 of Chapter 3 Methodology,
regression output coefficients were inputted into the logit equation (3.1), which in turn
provided an input to the regression equation (3.2), which yielded output probabilities for
each categorical outcome. Table 27 and Table 28 are the results of testing the H1A
model with the training and testing datasets. Both results indicate a high degree of
accuracy (93%) for this model.
67
Co
atin
g
Age
Co
atin
g
Typ
e
Surf
ace
Pre
p
Me
tho
d
Tan
k
Typ
e
Ship
Cla
ss
Tan
k
Cri
tica
lit
y
Co
atin
g
Insp
ect
ion
Sco
re
EPR
OB
1
EPR
OB
2
EPR
OB
3
EPR
OB
4
Cal
c 1
Logi
t
Cal
c 2
Logi
t
Cal
c 3
Logi
t
Cal
c 4
Logi
t
Cal
c 1
Pro
b
Cal
c 2
Pro
b
Cal
c 3
Pro
b
Cal
c 4
Pro
bP
red
ict
Acc
ura
te
?
3.9
88
87
13
10
.85
70
.11
20
.02
90
.00
21
.78
63
.40
66
.20
6-2
2.0
90
.85
60
.11
10
.03
0.0
02
1ye
s
0.4
32
27
91
31
0.9
91
0.0
07
0.0
02
1E-
04
4.6
76
.29
9.0
9-1
9.2
10
.99
10
.00
70
.00
21
E-0
41
yes
4.0
48
77
13
10
.98
40
.01
30
.00
32
E-0
44
.08
25
.70
28
.50
2-1
9.8
0.9
83
0.0
13
0.0
03
2E-
04
1ye
s
2.2
88
71
11
11
0.9
95
0.0
04
8E-
04
5E-
05
5.3
64
6.9
84
9.7
84
-18
.52
0.9
95
0.0
04
9E-
04
6E-
05
1ye
s
4.0
48
77
13
10
.98
40
.01
30
.00
32
E-0
44
.08
25
.70
28
.50
2-1
9.8
0.9
83
0.0
13
0.0
03
2E-
04
1ye
s
4.7
62
88
71
32
0.6
61
0.2
49
0.0
84
0.0
06
0.6
62
2.2
82
5.0
82
-23
.22
0.6
60
.24
80
.08
60
.00
61
no
Ye
s8
34
4.7
62
88
71
32
0.6
61
0.2
49
0.0
84
0.0
06
0.6
62
2.2
82
5.0
82
-23
.22
0.6
60
.24
80
.08
60
.00
61
no
No
59
4.0
48
77
13
10
.98
40
.01
30
.00
32
E-0
44
.08
25
.70
28
.50
2-1
9.8
0.9
83
0.0
13
0.0
03
2E-
04
1ye
sSu
m8
93
3.8
68
77
13
10
.98
50
.01
20
.00
32
E-0
4
4.1
47
5.7
67
8.5
67
-19
.73
0.9
84
0.0
12
0.0
03
2E-
04
1ye
s%
Acc
ura
cy0
.93
4
4.1
38
77
13
10
.98
30
.01
40
.00
32
E-0
44
.05
5.6
78
.47
-19
.83
0.9
83
0.0
14
0.0
03
2E-
04
1ye
s
0.4
18
77
13
10
.99
60
.00
48
E-0
45
E-0
55
.40
17
.02
19
.82
1-1
8.4
80
.99
60
.00
48
E-0
45
E-0
51
yes
3.8
68
77
13
10
.98
50
.01
20
.00
32
E-0
44
.14
75
.76
78
.56
7-1
9.7
30
.98
40
.01
20
.00
32
E-0
41
yes
3.8
18
77
13
10
.98
50
.01
20
.00
32
E-0
44
.16
85
.78
88
.58
8-1
9.7
10
.98
50
.01
20
.00
32
E-0
41
yes
0.4
18
77
13
10
.99
60
.00
48
E-0
45
E-0
55
.40
17
.02
19
.82
1-1
8.4
80
.99
60
.00
48
E-0
45
E-0
51
yes
3.9
88
77
13
10
.98
40
.01
30
.00
32
E-0
44
.10
45
.72
48
.52
4-1
9.7
80
.98
40
.01
30
.00
32
E-0
41
yes
0.4
18
71
11
11
0.9
98
0.0
02
4E-
04
3E-
05
6.0
44
7.6
64
10
.46
-17
.84
0.9
98
0.0
02
4E-
04
3E-
05
1ye
s2
3.8
81
0.0
72
08
21
31
0.9
54
0.0
37
0.0
09
6E-
04
3.0
28
4.6
48
7.4
48
-20
.85
0.9
54
0.0
37
0.0
09
6E-
04
1ye
s2
5.5
2
3.8
68
77
13
10
.98
50
.01
20
.00
32
E-0
44
.14
75
.76
78
.56
7-1
9.7
30
.98
40
.01
20
.00
32
E-0
41
yes
28
.33
0.4
18
77
13
10
.99
60
.00
48
E-0
45
E-0
55
.40
17
.02
19
.82
1-1
8.4
80
.99
60
.00
48
E-0
45
E-0
51
yes
04
3.8
68
77
13
10
.98
50
.01
20
.00
32
E-0
44
.14
75
.76
78
.56
7-1
9.7
30
.98
40
.01
20
.00
32
E-0
41
yes
-0.3
62
58
CA
0.4
18
77
13
10
.99
60
.00
48
E-0
45
E-0
55
.40
17
.02
19
.82
1-1
8.4
80
.99
60
.00
48
E-0
45
E-0
51
yes
-0.0
42
19
AP
L
9.6
72
88
71
33
0.2
47
0.3
83
0.3
36
0.0
34
-1.1
19
0.5
01
3.3
01
-25
0.2
46
0.3
76
0.3
42
0.0
36
2n
o-2
.31
73
9SP
3.8
68
77
13
10
.98
50
.01
20
.00
32
E-0
44
.14
75
.76
78
.56
7-1
9.7
30
.98
40
.01
20
.00
32
E-0
41
yes
-0.0
66
03
TT
0.4
18
72
13
10
.99
70
.00
36
E-0
44
E-0
55
.73
17
.35
11
0.1
5-1
8.1
50
.99
70
.00
36
E-0
44
E-0
51
yes
0.0
50
33
2SC
0.4
18
71
31
21
0.9
96
0.0
03
8E-
04
5E-
05
5.4
58
7.0
78
9.8
78
-18
.42
0.9
96
0.0
03
8E-
04
5E-
05
1ye
s-0
.45
36
7TS
Tab
le 2
7 -
H1A
Model
Tra
inin
g R
esult
s
68
CO
ATI
NG
AG
E
Co
atin
g
Typ
e
Surf
ace
Pre
p
TAN
K
TYP
E
SHIP
CLA
SS
Tan
k
Cri
tica
lity
Co
atin
g
Insp
Sco
re
EPR
OB
1
fro
m
Min
itab
EPR
OB
2 f
rom
Min
itab
EPR
OB
3
fro
m
Min
itab
EPR
OB
4
fro
m
Min
itab
Cal
c 1
Logi
t
Cal
c 2
Logi
t
Cal
c 3
Logi
t
Cal
c 4
Logi
t
Cal
c 1
Pro
b
Cal
c 2
Pro
b
Cal
c 3
Pro
b
Cal
c 4
Pro
bP
red
ict
Acc
ura
te
?
Fro
m
Mo
de
l
Fro
m
Mo
de
l
0.3
47
22
21
88
31
21
No
val
ue
s3
.45
.02
7.8
2-2
0.4
80
.96
80
.02
60
.00
64
E-0
41
yes
3.8
63
88
98
77
13
14
.14
75
.76
78
.56
7-1
9.7
30
.98
40
.01
20
.00
32
E-0
41
yes
Sco
re 1
? Y
es
0.4
05
55
68
71
31
21
5.4
58
7.0
78
9.8
78
-18
.42
0.9
96
0.0
03
8E-
04
5E-
05
1ye
s
3.8
63
88
98
77
13
1N
o v
alu
es
4.1
47
5.7
67
8.5
67
-19
.73
0.9
84
0.0
12
0.0
03
2E-
04
1ye
sSc
ore
1?
No
4.0
41
66
78
77
13
14
.08
25
.70
28
.50
2-1
9.8
0.9
83
0.0
13
0.0
03
2E-
04
1ye
s
4.0
41
66
78
77
13
14
.08
25
.70
28
.50
2-1
9.8
0.9
83
0.0
13
0.0
03
2E-
04
1ye
sY
es
83
2
0.0
72
22
22
08
21
31
No
val
ue
s3
.02
84
.64
87
.44
8-2
0.8
50
.95
40
.03
70
.00
96
E-0
41
yes
No
61
4.1
30
55
68
76
12
14
.57
6.1
98
.99
-19
.31
0.9
90
.00
80
.00
21
E-0
41
yes
Sum
89
3
0.4
11
11
11
78
13
12
12
.75
94
.37
97
.17
9-2
1.1
20
.94
0.0
47
0.0
12
8E-
04
1ye
s%
Acc
ura
cy0
.93
16
9
0.1
58
33
31
87
91
31
4.9
37
6.5
57
9.3
57
-18
.94
0.9
93
0.0
06
0.0
01
9E-
05
1ye
s
0.4
52
18
21
31
No
val
ue
s2
.84
94
.46
97
.26
9-2
1.0
30
.94
50
.04
30
.01
17
E-0
41
yes
0.5
22
22
22
18
13
12
12
.55
4.1
76
.97
-21
.33
0.9
28
0.0
57
0.0
14
9E-
04
1ye
s
7.3
72
22
21
77
10
11
31
2.8
4.4
27
.22
-21
.08
0.9
43
0.0
45
0.0
11
7E-
04
1ye
s
3.7
97
22
21
77
10
11
31
4.0
97
5.7
17
8.5
17
-19
.78
0.9
84
0.0
13
0.0
03
2E-
04
1ye
s
0.9
86
11
11
77
10
11
31
5.1
16
6.7
36
9.5
36
-18
.76
0.9
94
0.0
05
0.0
01
7E-
05
1ye
s
6.0
58
33
31
57
71
31
3.0
56
4.6
76
7.4
76
-20
.82
0.9
55
0.0
36
0.0
09
6E-
04
1ye
s2
3.8
81
0.1
86
11
11
17
21
31
5.6
84
7.3
04
10
.1-1
8.2
0.9
97
0.0
03
6E-
04
4E-
05
1ye
s2
5.5
2
0.1
86
11
11
17
91
31
5.2
22
6.8
42
9.6
42
-18
.66
0.9
95
0.0
04
0.0
01
6E-
05
1ye
s2
8.3
3
2.9
69
44
41
57
71
31
4.1
76
5.7
96
8.5
96
-19
.70
.98
50
.01
20
.00
32
E-0
41
yes
04
2.9
69
44
41
57
71
31
4.1
76
5.7
96
8.5
96
-19
.70
.98
50
.01
20
.00
32
E-0
41
yes
-0.3
62
6C
A
6.0
58
33
31
57
71
31
3.0
56
4.6
76
7.4
76
-20
.82
0.9
55
0.0
36
0.0
09
6E-
04
1ye
s-0
.04
22
AP
L
0.1
86
11
11
17
21
31
5.6
84
7.3
04
10
.1-1
8.2
0.9
97
0.0
03
6E-
04
4E-
05
1ye
s-2
.31
74
SP
5.4
94
44
41
17
12
12
13
.55
35
.17
37
.97
3-2
0.3
30
.97
20
.02
20
.00
53
E-0
41
yes
-0.0
66
TT
8.2
77
77
81
57
71
31
2.2
51
3.8
71
6.6
71
-21
.63
0.9
05
0.0
75
0.0
19
0.0
01
1ye
s0
.05
03
3SC
0.1
94
44
41
87
71
31
5.0
56
6.6
76
9.4
76
-18
.82
0.9
94
0.0
05
0.0
01
8E-
05
1ye
s-0
.45
37
TS
Tab
le 2
8 –
H1A
Model
Tes
ting R
esult
s
69
4.5.2 H1A Model Optimization and Classification Results
Using the output testing results of the training data (Table 27), a classification
matrix was developed to better understand the classifier accuracy of the H1A model.
Table 29 represents the general classification matrix for H1A. Table 30 is the
classification matrix of the H1A trained model, given the coating score coefficients of
23.88, 25.5, 28.3 and 0 from the OLR regression results. Table 31 represents H1A
classification results using the optimized coating score coefficients of 22.33, 24, 29.3 and
0. The impact, from optimization, upon the worst type of false positives (FP - Type 1
errors), that is, to cells EE14 and EE24 for mission criticality tanks (#3), should be noted,
as FPs go to zero. Tables 30 and 33 indicate a positive outcome through the use of
optimized coating score coefficients. Additionally, Table 34 displays the impact of model
H1A optimization upon classification matrix performance parameters. Overall, the
optimization of H1A produces the desired results of decreasing false positives while
maintaining overall model performance.
Table 29 – H1A Classification Matrix
1 2 3 4
1 TP1 EE12 EE13 EE14
2 EE21 TP2 EE23 EE24
3 EE31 EE32 TP3 EE34
4 EE41 EE42 EE43 TP4
Predicted
H1A Model
Classification Matrix
Observed
70
Table 30 – H1A Model Training Results Classification Matrix (Pre-Optimization)
Table 31 – H1A Model Training Results Classification Matrix (Post-Optimization)
Table 32 – H1A Testing Data Results Classification Matrix (Pre-Optimization)
Model
Accuracy93.40%
1 2 3 4
1 3090 110 42 3 EE14 3
2 23 31 23 2 EE24 1
3 7 13 11 2
4 0 0 0 0
23.88 25.5 28.3 0
Tank Criticality (#3)
FP (Type 1 Errors)
Predicted
HA1A Model Classification Matrix (Pre-Optimization)
Observed
Coating Score Coefficients
Model
Accuracy91.70%
1 2 3 4
1 2968 54 16 0 EE14 0
2 61 23 6 0 EE24 0
3 91 77 54 7
4 0 0 0 0
22.33 24 29.3 0
Tank Criticality (#3)
FP (Type 1 Errors)
Predicted
HA1A Model Classification Matrix (Post-Optimization)
Observed
Coating Score Coefficients
Model
Accuracy93.20%
1 2 3 4
1 822 27 16 3 EE14 2
2 6 9 6 0 EE24 0
3 1 2 1 0
4 0 0 0 0
23.88 25.5 28.3 0Coating Score Coefficients
Tank Criticality (#3)
FP (Type 1 Errors)
HA1A Model Classification Matrix (Pre-Optimization)
Observed
Predicted
71
Table 33 – H1A Testing Data Results Classification Matrix (Post-Optimization)
Table 34 – Model H1A Classification Performance
4.5.3 Hypothesis 1A Prediction Tools
4.5.3.1 Hypothesis 1A Coating Score Prediction Tool #1
Through the use of the logit equation (3.1) and regression equation (3.2), for given
predictor variables, a predicted coating score can be ascertained. Figure 20 displays this
and represents the H1A coating score prediction tool. Users input the variables of coating
Model
Accuracy91.60%
1 2 3 4
1 799 17 3 1 EE14 0
2 12 2 3 0 EE24 0
3 18 19 17 2
4 0 0 0 0
22.33 24 29.3 0Coating Score Coefficients
Tank Criticality (#3)
FP (Type 1 Errors)
HA1A Model Classification Matrix (Post-Optimization)
Observed
Predicted
72
age, coating type, surface preparation method, tank type, ship class and tank type. The
automated spreadsheet yields a predicted coating score in the yellow field.
Figure 20 - H1A Coating Score Prediction Tool #1
4.5.3.2 H1A Coating Score Prediction Tool #2
A second coating prediction tool can be made by keeping predictor variables
constant and varying time and recording the response variable (coating score). Figure 21
represents this technique and is the H1A coating score prediction tool. It is intended to be
a laminated sheet for waterfront use to give tank planners a visual aid in the predicted
coating life of specified tanks.
Enter
Values
Below
Predicted
Coating Score
COATING AGE
(yrs) 12.002
COATING Type
(1-33) 28.00
Surface
Preparation (1-8) 7.00
Tank Type (1-14) 10.00
Ship Class (1-13) 10.00
Tank Criticality
(1 -3) 3.00
22.38 1Calc 1
Logit -0.89198
Calc 1
Prob 0.290702
25.5 2Calc 2
Logit 2.228022
Calc 2
Prob 0.612036
28.3 3Calc 3
Logit 5.028022
Calc 3
Prob 0.090753
0 4Calc 4
Logit -23.272
Calc 4
Prob 0.006509
-0.362584584 CA
-0.042188684 APL
-2.317391272 SP
-0.066025501 TT
0.050331542 SC
-0.453667123 TS
Coefficients
73
Figure 21 – H1A Coating Score Prediction Tool #2
4.6 H1B OLR Results
H1B OLR results shown in Figure 22. H1B predictor variables are coating age,
coating score, tank type, ship class, tank criticality and tank pressure. H1B OLR results
indicated statistical significance for predictive variables (p-values less than 0.05). Test of
All Slope equal to Zero is good (p-value less the 0.05). Goodness of Fit is satisfactory
since P =1.0 and Measures of Association between variables is moderate. Overall, the
H1B OLR results indicate a moderate fit between data and the model.
Coating Type 28, Surface Preparation 7, Tank Type 10, Ship Class 10 and
Tank Criticality 3. Every blue dot represents a quarterly inspection result.
UUnde
74
Figure 22 – H1B OLR Results
4.6.1 H1B Model Training and Testing Results
Model H1B was testing was accomplished in the same way as model H1A (see
section 4.5.1). Tables 35 and 36 display the H1B Model training and testing accuracy
results. Results indicate a modest degree of accuracy (70-72%).
75
COAT
ING
AGE
Coat
ing
Scor
e
Tank
Type
SHIP
CLAS
S
Tank
Pres
s
Tank
Criti
calit
y
Stru
ctur
e Sc
ore
EPRO
B
1 fro
m
Min
ita b
EPRO
B
2 fro
m
Min
ita b
EPRO
B
3 fro
m
Min
ita b
EPRO
B
4 fro
m
Min
ita b
EPRO
B
5 fro
m
Min
ita b
Calc
1
Logi
t
Calc
2
Logi
t
Calc
3
Logi
t
Calc
4
Logi
t
Calc
5
Logi
t
Calc
1
Prob
Calc
2
Prob
Calc
3
Prob
Calc
4
Prob
Calc
5
Prob
Pred
ictAc
cura
te
?
15.1
7778
310
82
35
0.62
20.
057
0.02
90.
013
0.27
70.
499
0.74
90.
892
0.95
8-2
.301
0.62
20.
057
0.03
0.01
30.
277
1no
15.0
1667
310
81
32
0.43
90.
063
0.03
50.
016
0.44
7-0
.246
0.00
40.
146
0.21
2-3
.046
0.43
90.
062
0.03
50.
016
0.44
75
no
8.50
5556
110
81
35
0.58
40.
060.
031
0.01
40.
311
0.33
70.
587
0.72
90.
796
-2.4
630.
584
0.05
90.
032
0.01
40.
311
1no
10.5
52
108
13
50.
518
0.06
30.
033
0.01
60.
371
0.07
0.32
0.46
20.
528
-2.7
30.
517
0.06
20.
034
0.01
60.
371
1no
15.1
7778
310
81
35
0.43
80.
063
0.03
50.
016
0.44
8-0
.25
5E-0
40.
143
0.20
9-3
.05
0.43
80.
062
0.03
50.
016
0.44
85
yes
12.8
0556
310
82
35
0.63
40.
057
0.02
90.
013
0.26
80.
547
0.79
70.
939
1.00
6-2
.253
0.63
40.
056
0.03
0.01
30.
268
1no
Yes
1404
32.6
1389
310
81
35
0.35
40.
060.
034
0.01
60.
536
-0.6
02-0
.352
-0.2
1-0
.143
-3.4
020.
354
0.05
90.
035
0.01
60.
536
5ye
sNo
539
15.1
7778
310
81
32
0.43
80.
063
0.03
50.
016
0.44
8-0
.25
5E-0
40.
143
0.20
9-3
.05
0.43
80.
062
0.03
50.
016
0.44
85
noSu
m19
43
8.50
5556
210
82
35
0.70
30.
050.
025
0.01
10.
211
0.86
1.11
1.25
21.
319
-1.9
40.
703
0.04
90.
026
0.01
10.
211
1no
%Ac
c0.
723
32.6
1389
33
81
25
0.56
60.
061
0.03
20.
015
0.32
70.
265
0.51
50.
657
0.72
3-2
.535
0.56
60.
060.
033
0.01
50.
327
1no
8.50
5556
210
82
35
0.70
30.
050.
025
0.01
10.
211
0.86
1.11
1.25
21.
319
-1.9
40.
703
0.04
90.
026
0.01
10.
211
1no
8.50
5556
210
82
35
0.70
30.
050.
025
0.01
10.
211
0.86
1.11
1.25
21.
319
-1.9
40.
703
0.04
90.
026
0.01
10.
211
1no
15.1
7778
410
82
35
0.56
80.
061
0.03
20.
015
0.32
50.
273
0.52
30.
666
0.73
2-2
.527
0.56
80.
060.
033
0.01
50.
325
1no
8.50
5556
110
82
35
0.74
80.
045
0.02
20.
010.
176
1.08
61.
336
1.47
81.
545
-1.7
140.
748
0.04
40.
022
0.01
0.17
61
no
32.6
1389
43
81
25
0.51
0.06
30.
033
0.01
60.
378
0.03
90.
289
0.43
10.
497
-2.7
610.
510.
062
0.03
40.
016
0.37
81
no
11.3
6389
26
81
25
0.68
40.
052
0.02
60.
012
0.22
60.
774
1.02
41.
166
1.23
2-2
.026
0.68
40.
051
0.02
70.
012
0.22
61
no2.
81
10.5
52
108
23
50.
694
0.05
10.
025
0.01
20.
218
0.81
91.
069
1.21
11.
277
-1.9
810.
694
0.05
0.02
60.
012
0.21
81
no3.
052
12.8
0556
111
81
15
0.77
80.
041
0.02
0.00
90.
153
1.25
31.
503
1.64
51.
711
-1.5
470.
778
0.04
0.02
0.00
90.
153
1no
3.19
3
15.9
9167
16
81
25
0.71
20.
049
0.02
40.
011
0.20
30.
906
1.15
61.
299
1.36
5-1
.894
0.71
20.
048
0.02
50.
011
0.20
31
no3.
264
15.6
3611
26
81
25
0.66
60.
054
0.02
70.
012
0.24
10.
688
0.93
81.
081.
146
-2.1
120.
665
0.05
30.
028
0.01
20.
241
1no
05
15.1
7778
410
82
31
0.56
80.
061
0.03
20.
015
0.32
50.
273
0.52
30.
666
0.73
2-2
.527
0.56
80.
060.
033
0.01
50.
325
1ye
s-0
.02
CA
15.1
7778
310
81
35
0.43
80.
063
0.03
50.
016
0.44
8-0
.25
5E-0
40.
143
0.20
9-3
.05
0.43
80.
062
0.03
50.
016
0.44
85
yes
-0.2
26CS
19.3
2778
411
81
15
0.60
90.
058
0.03
0.01
40.
289
0.44
30.
693
0.83
50.
902
-2.3
570.
609
0.05
80.
031
0.01
40.
289
1no
-0.0
49TT
32.6
1389
311
81
11
0.59
90.
059
0.03
0.01
40.
297
0.40
10.
651
0.79
30.
859
-2.3
990.
599
0.05
80.
031
0.01
40.
297
1ye
s-0
.094
SC12
.805
562
108
23
50.
684
0.05
20.
026
0.01
20.
226
0.77
31.
023
1.16
51.
232
-2.0
270.
684
0.05
10.
027
0.01
20.
226
1no
0.74
9TP
15.1
7778
311
81
15
0.68
0.05
20.
026
0.01
20.
229
0.75
31.
003
1.14
51.
212
-2.0
470.
680.
052
0.02
70.
012
0.22
91
no-0
.526
TS
Tab
le 3
5 –
H1B
Model
Tra
inin
g R
esult
s
76
CO
ATI
NG
AG
E
Co
atin
g
Sco
re
Tan
k
Type
SHIP
CLA
SS
Tan
k
Pre
ssu
re
Tan
k
Cri
tica
l
ity
Stru
ctu
re
Sco
re
EPR
OB
1
fro
m
Min
itab
EPR
OB
2 fr
om
Min
itab
EPR
OB
3
fro
m
Min
itab
EPR
OB
4
fro
m
Min
itab
EPR
OB
5
fro
m
Min
itab
Cal
c 1
Logi
t
Cal
c 2
Logi
t
Cal
c 3
Logi
t
Cal
c 4
Logi
t
Cal
c 5
Logi
t
Cal
c 1
Pro
b
Cal
c 2
Pro
b
Cal
c 3
Pro
b
Cal
c 4
Pro
b
Cal
c 5
Pro
bP
red
ict
Acc
ura
te?
8.51
11
08
13
50.
338
0.59
10.
729
0.79
6-2
.463
0.58
40.
060.
031
0.01
40.
311
1n
o
32.6
13
10
81
35
-0.6
01-0
.348
-0.2
1-0
.143
-3.4
020.
354
0.06
0.03
40.
016
0.53
65
yes
32.6
13
38
12
50.
265
0.51
80.
657
0.72
3-2
.535
0.56
60.
061
0.03
20.
015
0.32
71
no
15.1
84
10
82
35
0.27
40.
527
0.66
60.
732
-2.5
270.
568
0.06
10.
032
0.01
50.
325
1n
o
32.6
14
11
81
13
0.17
50.
429
0.56
70.
633
-2.6
250.
544
0.06
20.
033
0.01
50.
347
1n
o
12.8
11
10
81
35
0.25
10.
504
0.64
20.
709
-2.5
50.
562
0.06
10.
032
0.01
50.
331
no
Yes
32
0
33.4
34
18
11
50.
646
0.89
91.
037
1.10
4-2
.155
0.65
60.
055
0.02
80.
013
0.24
91
no
No
13
4
8.51
13
81
25
1.20
41.
457
1.59
61.
662
-1.5
960.
769
0.04
20.
020.
009
0.15
91
no
Sum
45
4
10.5
62
21
13
11.
116
1.36
91.
508
1.57
4-1
.684
0.75
30.
044
0.02
10.
010.
172
1ye
s%
Acc
ura
cy
0.70
5
3.13
16
11
21
1.82
32.
076
2.21
52.
281
-0.9
770.
861
0.02
80.
013
0.00
60.
093
1ye
s
3.13
16
11
21
1.82
32.
076
2.21
52.
281
-0.9
770.
861
0.02
80.
013
0.00
60.
093
1ye
s
10.9
92
91
13
10.
766
1.02
1.15
81.
224
-2.0
340.
683
0.05
20.
026
0.01
20.
227
1ye
s
29.1
42
31
12
11.
218
1.47
11.
609
1.67
6-1
.583
0.77
20.
042
0.02
0.00
90.
158
1ye
s
27.0
52
91
13
10.
442
0.69
50.
834
0.9
-2.3
580.
609
0.05
80.
030.
014
0.28
91
yes
29.1
42
11
11
11.
841
2.09
42.
232
2.29
9-0
.96
0.86
30.
027
0.01
30.
006
0.09
11
yes
20.8
64
71
23
10.
962
1.21
51.
353
1.42
-1.8
390.
723
0.04
80.
023
0.01
10.
195
1ye
s2.
81
29.1
44
71
23
10.
794
1.04
81.
186
1.25
2-2
.006
0.68
90.
052
0.02
60.
012
0.22
21
yes
3.05
42
29.1
44
71
23
10.
794
1.04
81.
186
1.25
2-2
.006
0.68
90.
052
0.02
60.
012
0.22
21
yes
3.19
23
29.1
44
71
23
10.
794
1.04
81.
186
1.25
2-2
.006
0.68
90.
052
0.02
60.
012
0.22
21
yes
3.25
84
4.45
11
11
11
52.
079
2.33
22.
472.
537
-0.7
220.
889
0.02
30.
011
0.00
50.
073
1n
o0
5
8.74
13
51
25
1.48
11.
734
1.87
21.
939
-1.3
20.
815
0.03
50.
017
0.00
70.
126
1n
o-0
.02
CA
8.74
13
51
21
1.48
11.
734
1.87
21.
939
-1.3
20.
815
0.03
50.
017
0.00
70.
126
1ye
s-0
.226
CS
6.28
21
05
23
51.
187
1.44
1.57
91.
645
-1.6
130.
766
0.04
20.
021
0.00
90.
162
1n
o-0
.049
TT
8.83
11
08
23
51.
081.
334
1.47
21.
538
-1.7
20.
747
0.04
50.
022
0.01
0.17
71
no
-0.0
94SC
8.32
21
08
23
50.
864
1.11
81.
256
1.32
2-1
.936
0.70
40.
050.
025
0.01
10.
211
no
0.74
9TP
28.0
32
10
82
35
0.46
60.
720.
858
0.92
4-2
.334
0.61
50.
058
0.03
0.01
40.
284
1n
o-0
.526
TS
Tab
le 3
6 –
H1B
Model
Tes
ting R
esult
s
77
4.6.2 H1B Model Optimization
In tank coating inspections, having an over optimistic model, that is, a model that
predicts a better coating score than actually observed when inspected (false positive),
yields the riskiest situation. Given this, in order to reduce risk, optimization involves
adjusting model classification to reduce false positives (Type 1 errors) without overly
increasing Type 2 errors.
Table 37 represents the general classification matrix for H1B. Table 38 is the
classification matrix, given the OLR coating score coefficients of 2.8, 3.05, 3.19, 3.26
and 0 from the regression results. Table 38 represents H1B classification results using
the optimized coating score coefficients of 1.8, 2.55, 3.69, 4.26, and 0. The impact upon
worst type of false positives (FP - Type 1 errors), that is, cells EE14, EE15, and EE25, for
mission criticality tanks (#3) should be noted. Tables 38 and 41 indicate a modest
outcome through the use of optimized coating score coefficients. Additionally, Table 42
displays the impact of model H1B optimization upon classification matrix performance
parameters. Overall, the optimization of H1B produces modest results of decreasing false
positives while maintaining overall model performance.
78
Table 37 – H1B Classification Matrix
Table 38 – H1B Training Result Classification Matrix (Pre-Optimization)
79
Table 39 – H1B Training Results Classification Matrix (Post-Optimization)
Table 40 – H1B Testing Data Results Classification Matrix (Pre-Optimization)
80
Table 41 – H1B Testing Data Results Classification Matrix (Post-Optimization)
Table 42 – Model H1B Classification Performance
81
4.6.3 H1B Prediction Tool
4.6.3.1 H1B Structure Score Prediction Tool
Through the use of the logit equation (3.1) and regression equation (3.2), for given
predictor variables, a predicted structural score can be ascertained. Figure 23 represents
this methodology and is the H1B coating score prediction tool. Users input the variables
of coating age, coating inspection score, tank type, ship class, tank pressure and tank
criticality. The automated spreadsheet yields a predicted structural score in the yellow
field. Given the modest accuracy of this tool, it is recommended that structural score of 3
be treated as structural score of 5 (worst case) in order to reduce potential risk from
model over optimization.
Figure 23 – H1B Structure Score Prediction Tool
Factors
Enter
Values
Below
Predicted
Structural
Score
Coating Age (yrs) 15 1
Coating Insp Score (1-4) 3
Tank Type (1-13) 10
Ship Class (1-10) 10
Tank Pressure (1-2) 2
Tank Criticality (1 -3) 3
1.991
Calc 1
Logit-0.4944
Calc 1
Prob0.3788
3.052
Calc 2
Logit0.5656
Calc 2
Prob0.2589
3.193
Calc 3
Logit0.7056
Calc 3
Prob0.0317
3.254
Calc 4
Logit0.7656
Calc 4
Prob0.0131
05
Calc 5
Logit-0.7482
Calc 5
Prob0.3174
-0.0202 CA
-0.2260 CS
-0.0487 TT
-0.0938 SC
0.7490 TP
-0.5257 TC
82
4.7 H2 Results
H2 results (Table 43) are derived from the results of H1A and H1B. The H1A coating
score prediction tool was used to determine the time in years (column 4 – Table 43) each
type of tank (column one – Table 43) would reach coating score 3, tank coating service
life according to current United States Navy (USN) maintenance standards. To ascertain
the tank structural condition at the predicted age when the coating score reaches 3, the
H1B structural prediction tool was used. The predicted coating score (H1A resultant) and
predicted structural score (H1B resultant) along with tank criticality code (column two –
Table 43) were used to derive risk using the maintenance and risk matrix Table 6-1 in the
Corrosion Control Assessment and Maintenance Manual (CCAMM). The net savings of
inspections resulted in the difference between predicted and current tank inspection
periodicities. The last column of Table 43 shows the significant savings (in percentage)
of tank inspection periodicities, over a 20 year period, by the extension of current tank
inspection periodicities. The inspection periodicities savings range from 50 to 83.3%. The
H2 inspection periodicities savings provide an input to the H3 cost savings model.
83
4.8 H3 Results
The net inspection savings results from H2 provided an input into the H3 cost model.
Other inputs into the H3 cost model included the numbers of each type of tank in the
USN surface fleet (Table 43), and the average tank inspection cost (material and man-
days) associated with each type of tank (Table 44). H3 results (Table 45) display these
Tan
k
Typ
e
Tan
k Cr
itic
alit
y
Cod
e
Coat
ing
Insp
ecti
on
Req
uir
emen
t
Per
iod
icit
y
(yea
rs)
per
Nav
y
Corr
osi
on
Man
ual
Min
imu
m Y
ears
to r
each
Co
atin
g
Sco
re 3
, usi
ng
H1A
Co
atin
g A
ge
Mo
del
, (g
iven
:
coat
ing
#28,
sp
#
7 &
tan
k
crit
ical
ity
cod
e)
for
ALL
cla
sses
wit
h 9
5%
pro
bab
ility
Pro
ject
ed S
tru
ctu
ral
Sco
re,
usi
ng
H1B
Stru
ctu
ral M
od
el
pre
dic
tio
n t
oo
l, g
iven
Min
imu
m C
oat
ing
Age
(yea
rs),
Co
atin
g Sc
ore
#3 a
nd
Tan
k Cr
itic
alit
y
Ris
k A
ssu
med
giv
en
Coat
ing
Sco
re 3
&
Pro
ject
ed S
tru
ctu
re
Sco
re p
er N
avy
Corr
osi
on
Man
ual
Ris
k
Mat
rix
(Tab
le 6
-1)
Ove
rall
#
coat
ing
insp
ecti
on
s
req
uir
ed
du
rin
g
Min
imu
m Y
ears
is r
each
ed (
Min
Yrs/
Req
Insp
Per
iod
icit
y)
Net
Insp
ecti
on
savi
ngs
(O
vera
ll
Insp
# l
ess
1), i
f
per
iod
icit
y
exte
nd
ed t
o
pro
ject
ed
min
imal
yea
rs t
o
Coat
ing
Sco
re 3
Per
cen
t In
sp
Savi
ngs
by
exte
nd
ing
per
iod
icit
y o
f
insp
ecti
on
s (f
or
20 y
ear
insp
ecti
on
tim
efra
me)
11
1020
.81
Low
21
50.0
23
418
.11
Low
54
80.0
32
317
.91
Low
65
83.3
43
617
.71
Low
32
66.7
53
617
.51
Low
32
66.7
62
1018
.61
Low
21
50.0
73
617
.11
Low
32
66.7
82
1018
.24
1Lo
w2
150
.0
93
616
.81
Low
32
66.7
103
616
.61
Low
32
66.7
111
619
1Lo
w3
266
.7
122
317
.51
Low
65
83.3
132
617
.31
Low
32
66.7
Tab
le 4
3 –
H2 R
esult
s
84
inputs as well the projected savings (calculated through estimating) over a 20 year period.
H3 results (Table 44) show a significance cost savings (in net present value) over a 20
year period of 582,420 man days and $890,242,549. The average annual savings (in net
present value) is projected to be 29,121 man-days and $44,512,127.
Table 44 – In-Service Tank
Frequency Distribution Table 45 – Average Tank Inspection
Tank Type # In-Service% of Total
In-Service
1 344 2.14
2 340 2.12
3 1476 9.19
4 107 0.67
5 199 1.24
6 3369 20.97
7 2941 18.31
8 42 0.26
9 643 4.00
10 2018 12.56
11 3878 24.14
12 156 0.97
13 552 3.44
Total 16065 100.00
Tank
Type
Average
Material
Cost ($)
Average
Manday
Cost
1 $15,719.57 16.55
2 $5,966.33 11.81
3 $5,605.15 11.24
4 $60,787.06 20.63
5 $17,648.65 14.91
6 $16,768.53 16.82
7 $48,178.56 25.04
8 $4,849.24 10.15
9 $16,260.25 15.80
10 $82,081.53 26.69
11 $14,981.51 14.86
12 $5,625.79 11.14
13 $2,230.67 10.00
Average Tank Inspection Cost
85
Tan
k
Typ
e
Ave
rage
Sch
edu
le C
ost
per
Insp
ecti
on
(man
day
s)
Ave
rage
Mat
eria
l Co
st
per
Insp
ecti
on
($ d
olla
rs)
Net
Insp
ecti
on
savi
ngs
(fr
om
H2)
Tota
l Ave
Sch
edu
le
Cost
Sav
ings
in
Man
day
s (p
er t
ank
typ
e)
[Net
Insp
ecti
on
Savi
ngs
X (
Sch
edu
le
Cost
/In
sp)]
Tota
l Ave
Mat
eria
l
Cost
Sav
ings
in d
olla
rs
($)
(per
tan
k ty
pe)
[Net
Insp
ecti
on
Savi
ngs
X (
Ave
Mat
Cost
/In
sp)]
Tota
l # o
f ea
ch
typ
e o
f ta
nk
in
the
surf
ace
flee
t
For
20 y
ear
insp
ecti
on
tim
efra
me,
Agg
rega
te
Sch
edu
le C
ost
Savi
ngs
(in
Man
day
s) f
or
each
typ
e o
f
tan
k ac
ross
th
e
flee
t (T
ota
l
Sch
ed
Savi
ngs
/per
tan
ks X
# o
f
tan
k ty
pe
in
flee
t)
For
20 y
ear
insp
ecti
on
tim
efra
me,
Agg
rega
te
Mat
eria
l Co
st
Savi
ngs
in $
(do
llars
) fo
r
each
typ
e o
f
tan
k ac
ross
th
e
flee
t (T
ota
l
Mat
eria
l
Savi
ngs
/per
tan
k X
# o
f ta
nk
typ
e in
fle
et)
116
.55
$15,
719.
571
16.5
515
719.
5734
4.0
5,69
4.0
$5,4
07,5
31.2
2
211
.81
$5,9
66.3
34
47.2
323
865.
3134
0.0
16,0
57.2
$8
,114
,206
.52
311
.24
$5,6
05.1
55
56.2
128
025.
7314
76.0
82,9
59.8
$4
1,36
5,97
9.90
420
.63
$60,
787.
062
41.2
612
1574
.12
107.
04,
415.
1
$1
3,00
8,43
0.84
514
.91
$17,
648.
652
29.8
235
297.
3019
9.0
5,93
4.0
$7,0
24,1
62.7
0
616
.82
$16,
768.
531
16.8
216
768.
5333
69.0
56,6
51.0
$5
6,49
3,16
4.65
725
.04
$48,
178.
562
50.0
896
357.
1229
41.0
147,
292.
0
$2
83,3
86,3
01.3
4
810
.15
$4,8
49.2
41
10.1
548
49.2
442
.042
6.5
$2
03,6
68.2
7
915
.80
$16,
260.
252
31.5
932
520.
5164
3.0
20,3
12.7
$2
0,91
0,68
7.60
1026
.69
$82,
081.
532
53.3
716
4163
.06
2018
.010
7,71
0.1
$331
,281
,048
.34
1114
.86
$14,
981.
512
29.7
229
963.
0238
78.0
115,
235.
6
$1
16,1
96,5
96.1
2
1211
.14
$5,6
25.7
95
55.7
228
128.
9315
6.0
8,69
2.1
$4,3
88,1
13.4
6
1310
.00
$2,2
30.6
72
20.0
044
61.3
455
2.0
11,0
40.0
$2
,462
,658
.23
582,
420.
0
$8
90,2
42,5
49.2
0
Tab
le 4
6 –
H3 R
esult
s
86
Chapter 5—Discussion and Conclusions
5.1 Discussion
This work’s hypotheses are:
H1A: Coating Type, Application Method, Tank Type, Ship Class, Tank
Criticality, Coating Age are predictors of Coating Condition Score.
H1B: Coating Condition Score, Ship Class, Tank Pressure, Coating Age, Tank
Criticality are predictors of Structural Condition Score.
H2: Tank inspection periodicities, where HS/UHS coatings are employed, can be
increased 25-50% using the coating and structure score predictions (developed from
H1A and H1B) with minimal increased risk of failure.
H3: There are significant savings when tank inspection periodicities increase by
25-50% (results from H2).
The results indicate that the use of ultra-high solids/high solids (UHS/HS) coatings
aboard United States Navy (USN) surface ships, along with other stated factors like
surface preparation method, have extended tank coating service life well beyond current
USN surface ship tank inspection periodicities. Thus, tank inspection periodicities can be
extended by 25-50% with minimal risk acceptance. Therefore, this confirms the problem
statement that unnecessary tank inspections are being conducted, incurring unnecessary
cost. The results of this work are vital to optimizing USN surface ship tank inspections
requirements, while the integrated methodology presented offers a framework for
analysis of the USN Corrosion Control Assessment and Maintenance Manual (CCAMM)
database.
87
5.1.1 Scope of Analysis and Framework
This work focused on in-service USN surface ship tank inspection data contained
in the CCAMM database. The quality and quantity of the data determined the extent of
the analysis performed. Conclusions were not drawn that could not or would not be
supported by the data and the chosen analysis methodology. The chosen analysis
methodology model’s assumptions were tested and verified. For H1A, coating inspection
scores relationship to UHS/HS coatings (along with other factors) were examined. For
H1B, structural inspection scores relationship to coating inspection scores (along with
other factors) were also examined. H2 examined the outputs of H1A and H1B to
ascertain tank inspection periodicities savings with associated risk characterization. H3
calculated and accrued savings, given revised tank inspection periodicities from H2.
5.1.2 H1A Discussion
The H1A database was partitioned into training and test sets. Frequency
distributions of variables indicate fair representation amongst the two different data sets.
The H1A model results from the Ordinal Logistics Regression (OLR) regression table
indicated that coating age, coating type, surface preparation method, tank criticality and
ship class were statistically significant factors (p-value < 0.5). Additionally, OLR
goodness of fit results, Pearson and Deviance Methods, equal to one indicates the model
fit the data set well. Also, from the OLR analysis results, the measures of association
between response variable and predictor probabilities had strong correlation (0.89).
Testing of the H1A model, with partitioned test data, indicated a strong degree of model
accuracy (93 percent). The H1A model results were optimized to reduce false positives
(Type 1 error) while maintaining overall model classification performance. Then with
88
the results listed in Table 46, based on HA1 data set OLR results, the null hypothesis
(H1Ao – coating age, coatings type, surface preparation method, tank type, ship class and
tank criticality are not predictors of coating inspection score) can be rejected, and the
present research H1A can be adopted.
Table 47 – HA1 OLR Regression Results
With this optimized and tested model, several practical predictive tools were also
developed: 1) coating score prediction tool and 2) coating score prediction graphs.
These tools are intended to be applied in waterfront tank planning and applications.
Lastly, the H1A coating age prediction tool was integrated into the H2 analysis in order
to predict when tank coatings would reach coating condition 3, the USN threshold for
coating renewal, and thus, aide in determining an optimized inspection periodicity.
5.2 H1B Discussion
The H1B database was partitioned into training and test sets. Frequency distributions
of variables indicate fair representation amongst the two different data sets. For H1B, the
results of the OLR regression table indicated that coating age, coating inspection score,
tank type, ship class, tank pressure and tank criticality were statistically significant
Ha
Confirmed
Predictor P Lower Upper
Coating Age 0.000 0.7 0.67 0.73 Yes
Coating Type 0.001 0.96 0.94 0.98 Yes
Surface Preparation
Method0.000 0.1 0.07 0.14 Yes
Tank Type 0.031 0.94 0.88 0.99 Yes
Ship Class 0.018 1.05 1.01 1.1 Yes
Tank Criticality 0.002 0.64 0.48 0.84 Yes
Odds
Ratio
95% CI
H1A Ordinal Logistic Regression Results
89
factors (p-value < 0.5). Ordinal Logistics Regression goodness of fit results indicated
that the Deviance Method equaled one, indicating the model fit the data set well.
Additionally, from the OLR analysis results, the measures of association between
response variable and predictor probabilities had moderate correlation (0.31). Testing of
the H1B model, with partitioned test data, indicated a moderate degree of model accuracy
(70 percent). The H1B model optimization results, to reduce false positives (Type 1
errors) while maintaining overall model classification performance, were moderately
successful. Then with the results listed in Table 47, based on HB1 data set OLR results,
the null hypothesis (H1Bo – coating age, coating inspection score, tank type, ship class,
tank pressure and tank criticality are not predictors of structural inspection score) can be
rejected, and the present research H1B can be adopted.
Table 48 - HB1 OLR Regression Results
With this optimized and tested model, a structure score predictive tool was developed
as well. The H1B structural score prediction tool integrated into the H2 analysis, as
well, in order to aid in characterizing risk by predicting and assessing the structural score
when optimizing tank inspection periodicities.
Ha
Confirmed
Predictor P Lower Upper
COATING AGE 0.003 0.98 0.97 0.99 Yes
Coating Score 0.000 0.8 0.72 0.88 Yes
Tank Type 0.007 0.95 0.92 0.99 Yes
Ship Class 0.000 0.91 0.88 0.95 Yes
Tank Pressure 0.000 2.11 1.61 2.77 Yes
Tank Criticality 0.000 0.59 0.49 0.71 Yes
95% CIOdds
Ratio
H1B Ordinal Logistic Regression Results
90
5.3 H2 Discussion
Previously developed H1A and H1B tools along with a risk assessment from the
maintenance decision matrix in CCAMM (Table 6-1) were utilized to assess optimal tank
inspection periodicities. Tank inspection periodicities for each tank type were developed
based on the predicted time (in years) it would take a coating to reach coating condition
3, the coating condition that requires coating replacement. The difference between
current tank inspection periodicities and predicted tank inspection periodicities (using
H1A) provided a net inspection periodicities savings.
With the H2 revised results listed in Table 48, the null hypothesis (H2o – Tank
inspection periodicities, where UHS coatings are employed, cannot be increased 25-50%
using the coating and structure score predictions, from H1A and H1B, with minimal
increased risk of failure) can be rejected.
Table 49 – H2 Results
91
5.4 H3 Discussion
The results from H2 provided an input for H3. The cost savings were derived over a
twenty-year period since the maximum projected coating life was 20 years. This allowed
sufficient time passage for cost saving accrual in the most limiting case (longest projected
longevity - 20 years). All savings were represented in net present value (2019).
Aggregate savings were the sum of the individual tank types savings. Individual tank
savings were derived from the net tank inspection savings for each tank type (from H2)
multiplied by the average cost data (material and labor) for an inspection for each type of
tank. Then with the results listed in Table 49, the null hypothesis (H2o – There are no
significant savings when tank inspection periodicities increase by 25-50% (results from
H2)) can be rejected, and the present research H3 can be adopted.
Table 50 – H3 Results
Tank
Type
Average
Schedule Cost
per Inspection
(mandays)
Average
Material Cost
per Inspection
($ dollars)
Net Inspection
savings (from H2)
Total Ave Schedule
Cost Savings in
Mandays (per tank
type)
[Net Inspection
Savings X (Schedule
Cost/Insp)]
Total Ave Material
Cost Savings in dollars
($) (per tank type)
[Net Inspection
Savings X (Ave Mat
Cost/Insp)]
Total # of each
type of tank in
the surface
fleet
For 20 year
inspection
timeframe,
Aggregate
Schedule Cost
Savings (in
Mandays) for
each type of
tank across the
fleet (Total
Sched
Savings/per
tanks X # of
tank type in
fleet)
For 20 year
inspection
timeframe,
Aggregate
Material Cost
Savings in $
(dollars) for
each type of
tank across the
fleet (Total
Material
Savings/per
tank X # of tank
type in fleet)
Ha
Confirmed
1 16.55 $15,719.57 1 16.55 15719.57 344.0 5,694.0 $5,407,531.22 Yes
2 11.81 $5,966.33 4 47.23 23865.31 340.0 16,057.2 $8,114,206.52 Yes
3 11.24 $5,605.15 5 56.21 28025.73 1476.0 82,959.8 $41,365,979.90 Yes
4 20.63 $60,787.06 2 41.26 121574.12 107.0 4,415.1 $13,008,430.84 Yes
5 14.91 $17,648.65 2 29.82 35297.30 199.0 5,934.0 $7,024,162.70 Yes
6 16.82 $16,768.53 1 16.82 16768.53 3369.0 56,651.0 $56,493,164.65 Yes
7 25.04 $48,178.56 2 50.08 96357.12 2941.0 147,292.0 $283,386,301.34 Yes
8 10.15 $4,849.24 1 10.15 4849.24 42.0 426.5 $203,668.27 Yes
9 15.80 $16,260.25 2 31.59 32520.51 643.0 20,312.7 $20,910,687.60 Yes
10 26.69 $82,081.53 2 53.37 164163.06 2018.0 107,710.1 $331,281,048.34 Yes
11 14.86 $14,981.51 2 29.72 29963.02 3878.0 115,235.6 $116,196,596.12 Yes
12 11.14 $5,625.79 5 55.72 28128.93 156.0 8,692.1 $4,388,113.46 Yes
13 10.00 $2,230.67 2 20.00 4461.34 552.0 11,040.0 $2,462,658.23 Yes
582,420.0 $890,242,549.20
Total
Savings for
20 year
timeframe
(in 2019
Dollars)
92
5.5 Conclusions
An integrated methodology of modeling (for coating and structure) presented in this
work demonstrated a significant cost savings (material and man-days) when optimizing
tank inspection periodicities, given the use of high performance UHS/HS coatings,
balancing risk and incorporating other stated factors.
5.6 Body of Knowledge
This work adds to the body of knowledge by providing proven analysis results to
optimize USN surface ship tank inspection periodicities for the USN maintenance
community. It also provides a novel methodology in doing so.
5.7 Future Study
To further this work, it is recommended that other naval communities like aircraft
carriers and submarines have tank inspection periodicities analyzed with this
methodology as well. It is also recommended to further study optimized results to
minimize the likelihood of Type 1 errors. Lastly, it is recommended that other
applications of OLR, where appropriate, be explored for potential cost savings within the
wider USN surface ship maintenance community, where inspection results are ordinal in
nature.
93
References
American Bureau of Shipping (ABS). “The Application and Inspection of Marine
Coating Systems”, 2017a, pp. 1-122.
American Bureau of Shipping (ABS). “Maintenance and Repair of Protective Coatings”,
2017b, pp. 1-50.
Balkey, Kenneth; Ayyub, Bilal; Gore, Bryan. “American Society of Mechanical
Engineers (ASME) Research: Risk-Based Inspection Guidelines”, Mechanical
Engineering. Vol (112), ProQuest pp. 68.
Carrier Planning Activity (CPA). “Request Technical Concurrence for Changes to
Inspection Survey Periodicities Listed in Technical Manual for Corrosion Control
Assessment and Maintenance Manual (CCAMM), Rev 3”, Official Letter
Correspondence, December 6, 2016, pp. 1-8.
Csapo, Paul. “USN Surface Ship Tank Inspection Rate”, email communication, February
14, 2019.
Eliasson, Johnny. “The Future of Ballast Tank Coatings”, www.paintsquare.com,
Accessed July 15, 2019.
Gardiner, C.P., Melchers, R.E., “Corrosion analysis of bulk carriers, Part I: operational
parameters influencing corrosion rates”, Marine Structures, 2003, Vol (16), pp.
547-566.
Geis, Alan; Haase, Jake; Cone, Caron. “Colonial employs risk management to distribute
load of tank-inspection program”, Oil & Gas Journal, June 10, 2002, Vol (100),
ProQuest pg. 66.
94
Gudze, M.T., Melchers, R.E.. “Operational based corrosion analysis in naval ships”,
Corrosion Science, 2008, Vol (50) , pp. 3296-3307.
Guedes Soares, C.; Garbatov, Y.. “Reliability of maintained, corrosion protected plates
subject to non-linear corrosion and compressive loads”, Marine Structures, 1999,
Vol (12), pp. 425-445.
Guedes Soares, C.; Garbatov, Y.. “Structural maintenance planning based on historical
data of corroded deck plates of tankers”, Reliability Engineering and System
Safety, 2009, Vol (94), pp. 1806-1817. DOI 10.1016/j.ress.2009.05.013.
Ifezue, D.; Tobins, F.H.. “Risk-Based Inspection of a Crude Oil Import/Export Line: The
Corrosion Engineer’s Role”, Journal of Failure Analysis and Prevention, 2014,
Vol (14), pp. 395–404. DOI 10.1007/s11668-014-9830-6. 2014.
Ingle, Mark. “High Solids Coatings Performance and Service History”, NAVSEA
Technical Paper, 2011a, pp. 1-8.
Ingle, Mark. “Are New Coatings living up to expectations?”, NAVSEA brief, 2011b,
MegaRust Conference, pp. 1-26.
Ivosevic, S.; Mestrovic, R.; Kovac, N.. “An approach to the probabilistic corrosion rate
estimation model for inner bottom plates of bulk carriers”, Brodogradnja
Shipbuilding, 2017, Vol (68), pp. 57-70.
Kent, Chris. “Critique of SURFMEPP Time Directed Maintenance Strategy”, Official
email communication, May 2019, pp. 1.
Lucas, K.E.. “Comprehensive Monitoring and Evaluation of Ballast Tank Coatings
Integrity for Life Prediction and Condition Based Maintenance”, NRL Technical
Paper NRL/MR/6126--00-8492, 2000, pp. 1-15.
95
McCullagh, Peter. “Regression Models for Ordinal Data”, Journal of Royal Statistical
Society, Vol 43, Issue 2, 1980, pp. 109-142.
Melchers, R.E.. “Probabilistic model for marine corrosion of steel for structural reliability
assessment”, Journal of Structural Engineering, 2003, Vol (129), pp. 1484-1493.
Melcher, R.E.; Jiang, X.. “Estimation of Models for durability of epoxy coatings in water
ballast tanks”, Ships and Offshore Structures, 2006, Vol (1), pp. 61-70,
DOI:10.1533/saos.2004.0006.
Muhammet, Gul; Erkan, Celik; Emre, Akyuz. “A hybrid risk-based approach for
maritime applications: The case of ballast tank maintenance”, Human and
Ecological Risk Assessment: An International Journal, 2017, Vol (23), pp. 1389-
1403, DOI: 10.1080/10807039.2017.1317204.
Naval Sea Systems Command (NAVSEA), “Corrosion Control Assessment and
Maintenance Manual (CCAMM)”, USN Technical Pub T9630-AB-MMD-
0102015, 2015, pp. 1-178.
NAVSEA. “Engineering for Reduced Maintenance / Corrosion Control Initiatives”,
NAVSEA Business Case Study, 1997, pp 1-55.
NAVSEA. “Impact of Corrosion Brief”, NAVSEA presentation, 2016, pp. 1-16.
NAVSEA. “Naval Ship Technical Manual (NSTM) Chapter 100 Hull Structures”, USN
Technical Pub S9086-DA-STM-010, 2017a, pp. 1-140.
NAVSEA. “Naval Ship Technical Manual Chapter 631 Preservation of Ships in Service –
Surface Ship / Submarine Applications”, USN Technical Publication S9086-VD-
STM-030/CH-631V3R1, 1996, pp. 1-70.
96
NAVSEA. “Ultra High Solids (UHS) Tanks and Coating Metrics”, NAVSEA
Presentation, 2017b, pp. 1-15.
Noor, Norhazilan; Smith, George; Yahaya, Nordin. “The Weibull Time-Dependent
Growth Model of Marine Corrosion in Seawater Ballast Tank”, Malaysian
Journal of Civil Engineering, 2007, Vol (19), pp. 52-65.
Noor, Norhazilan; Smith, George; Yahaya, Nordin; Noor, Shadih. “A Probabilistic Time-
Variant Corrosion Wastage Model for Seawater Ballast Tank”, Arabic Journal of
Science and Engineering, 2013, Vol (38), pp. 1333–1346, DOI 10.1007/s13369-
013-0608-z.
Paik, Jeom Kee; Jae Myung Lee; Joon Sung Hwang; Park, Young II. “A Time-Dependent
Corrosion Wastage Model for Structures of Single- and Double-Hull Tankers and
FSOs and FPSOs”, Marine Technology, 2003, Vol (40), pp. 201-217.
Parson, Evan; LaMontagne, Kevin; McGaulley, Wayne. “The Annual Cost of Corrosion
for Navy Ships”, NAVSEA Cost Report, 2011, pp. 1-81.
Parsons, Dan. “A Fresh Coat of Paint Can Save Navy Billions”,
www.nationaldefensemagazine.org, accessed January 15, 2019.
Qualified Products Database, Defense Logistics Agency,
https://qpldocs.dla.mil/default.aspx., accessed July 15, 2019.
Rizzo, CM; Paik; Brennan, Carlsen; Daley, Garbatov; L. Ivanov ; B. C. Simonsen , N.
Yamamoto; H. Z. Zhuang. “Current practices and recent advances in condition
assessment of aged ships”, Ships and Offshore Structures, Vol (2), pp. 261-271,
DOI: 10.1080/17445300701423486. 2007.
97
Slebodnick, P.F; Santagelo, K.; Thomas, K.; Cuzzocrea, J.; Zuskin, D.; Brinkerhoff, B..
“Eleven Years of Ultra High Solids (UHS) Coating Usage in the U.S. Navy and
Related Advances”, NRL Technical Paper NRL/MR/6130--08-9153, 2008, pp. 1-
446.
Stein, G. “The Flow must Go On”, www.paintsquare.com, date accessed July 15, 2019.
Surface Maintenance Engineering Planning Program (SURFMEPP), “Surface Ship Tank
and Void Time-Directed Maintenance Strategy”, Official Letter Correspondence,
2015, pp. 1-27.
UCLA. Institute of Digital Research and Education [web course], retrieved October 1,
2019, https://stats.idre.ucla.edu/stata/dae/ordered-logistic-regression/
Wang, G.; Spencer, J.; Haihong, S.. “Assessment of Corrosion Risk to Aging Ships
using an Experience Database”, Proceedings of 22nd International Conference on
Offshore Mechanics and Arctic Engineering, 2003, pp. 149-159.
Wheat, David; Thill, Tony. “Preventing Tank Corrosion”, Chemical Engineering, 2011,
Vol (118), pg. 47.