Professional Documents
Culture Documents
Engineered Systems
DISSERTATION
Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy
in the Graduate School of The Ohio State University
By
2011
Dissertation Committee:
Kuldeep Agarwal
2011
ABSTRACT
manufacturing are examples of complex engineered systems because they have many
interconnected components and there is little knowledge about the interactions between
these different components. These processes consist of a raw material in the form of
manufacturing process and through a series of steps, converts it into a product. These
manufacturing processes are designed and operated to achieve a specific goal: the final
The current design cycle for these different manufacturing processes treats the different
system components (sub-processes) in isolation. Each engineer and personnel looks at the
problem in his own perspective and optimizes the solution pertaining to the requirements
To overcome these limitations, a new approach and methodology for the design of these
attribute into the various factors which affect it and determine its value. These factors are
ii
further decomposed into the physical phenomena which cause these factors to affect the
quality attribute. The physical phenomena are finally decomposed into the manufacturing
processes and material uncertainties which cause them to influence the quality
Novel approach of combining process models with the data obtained by testing sensors is
developed through the use of Bayesian Hierarchical modeling. Case studies involving
sheet, rod and aeroengine manufacturing are demonstrated. The developed Bayesian
iii
Dedication
iv
ACKNOWLEDGEMENTS
First of all I would like to give my deepest regards to my advisor, mentor and Guru Dr.
Shivpuri for not only providing me the opportunity to work on such interesting topic, but
also for guiding me throughout the 4 years of my graduate studies. He always provided a
logical way to overcoming the obstacles and this work would have been nowhere without
I would also like to thank my committee members Dr. Goel, Dr. Parthasarathy and Dr.
Brevick for providing their comments to improve this work. The work on Bayesian
modeling would not have been anywhere close to complete without the direction from
Dr. Goel.
I would also like to take this opportunity to thank Dr. John Merrill, Director, First Year
Engineering Program. He not only supported me as a GTA for 3 years but also gave me
I would like to thank many of my friends and colleagues who provided me the mental
Finally I would like to thank the two people who have contributed as much as I have to
this work – my wife and son. They always gave me support and smiled at all the
adversities so that I could see this day. I cannot thank them enough.
v
VITA
State University
PUBLICATIONS
1. Agarwal, K., Shivpuri, R., Zhu, Y., “Robust design of No-Twist-Mill parameters for
reduced geometric variation in the hot rolling of steel rods and coils”, Steel Research
International, 2011.
2. Agarwal, K., Shivpuri, R., Zhu, Y., “Role of Process Dynamics and Material
Instability in the Generation of Surface Defects during High Speed Hot Extrusion of
vi
3. Shivpuri, R., Singh, S., Agarwal, K., Liu, C., “Energy Release Rate based Approach
for the Wear of Punches inPrecision Blanking of High Strength Steel”, CIRP Annals,
2011.
4. Agarwal, K., Shivpuri, R., Zhu, Y., Chang, T.S., Huang, H., “Process knowledge
defects in Hot Rolling”, Expert Systems With Applications, Volume 38, Issue 6, June
5. Shivpuri, R., Cheng, X., Agarwal, K., Babu, S., “Evaluation of 3D printing for dies
6. Agarwal, K., Shivpuri, R. and Lembo, J., “Investigating Rapid Prototype Techniques
for Application to Dies for Short Run Forgings”, AFDM2002, The Second Int. Conf.
on Advanced Forming and Die Manufacturing, 17-19 June 2002, Pusan, Korea.
technique for application to dies for short run forging”, Solid Freeform Fabrication
8. Agarwal, K., Mathur, D., Shivpuri, R., Lembo, J., Stys, T. and Harris, W., “A Rapid
Die Manufacturing Technique for Short Run Forgings”, 24th Forging Industry
9. Mathur, D., Agarwal, K. and Shivpuri, R., “Microstructure Study during Hot
vii
Deformation of Aluminum Alloys, Ed. Z. Jin, A. Beaudoin, T.A. Bieler and B.
Radhakrishnan, TMS Annual Meeting, March 3-6, San Diego, CA, 2003.
10. Shivpuri, R., Agarwal, K., Mathur, D., Lembo, J. and Harris, W., “Forging of
Aluminum Helicopter Parts using Rapidly Prototyped Dies,” Aero Mat 2003, Dayton,
11. Shivpuri, R., Agarwal, K., Mathur, D., Lembo, J. and Harris, W., “Reduced Lead
Times for Forged Helicopter Parts”, AHS Forum 59, Proceedings of the American
12. Agarwal, K., Shivpuri, R., “The Role of Manufacturing Process in the Design for
13. Agarwal, K., Shivpuri, R., “Risk Based Process Design of Titanium Rotor Forgings
14. Agarwal, K., Shivpuri, R., Ai, X., Pauskar, P., “Bayesian Hierarchical Network based
15. Whitfield, C., Schlosser, P., Agarwal, K., Riter, E., “Advanced Energy Vehicle
ASEE Annual Conference & Exposition, Vancouver, BC, June 26-29, 2011.
viii
16. Agarwal, K., Shivpuri, R., Singh, A.P., Rath, S., Kumar, S., Mukherjee, D., Mathur,
A.S., “An On-Line Hybrid Mathematical Model for Quality Prediction during Hot
17. Agarwal, K. and Shivpuri, R., “Rapid Tooling for Forging Dies”, ASM Handbook,
Vol. 14A, Metalworking : Bulk Forming, Ed. S. L. Semiatin, 2005, pp. 645- 650.
FIELDS OF STUDY
ix
TABLE OF CONTENTS
ABSTRACT........................................................................................................................ ii
DEDICATION ................................................................................................................... iv
ACKNOWLEDGEMENTS ................................................................................................ v
VITA .................................................................................................................................. vi
LIST OF TABLES…………………………………………………………………….....xv
2.2 Modeling Microstructure Evolution in Carbon Steel during hot strip rolling ..... 31
x
2.3 Statistical Modeling of Sheet Manufacturing ...................................................... 41
3.2 Background.......................................................................................................... 67
xi
3.8 Bayesian Networks for surface defects ............................................................. 126
5.2.1 Manufacturing of aeroengine disk and introduction of hard alpha ........... 166
xii
5.4.2 Effect of processing parameters on movement and rotation of hard alpha 193
5.5 Effect of forging velocity and billet temperature on the movement and rotation of
CODE FOR BAYESIAN ANALYSIS AND FEM DATA FILES ................................ 250
xiii
LIST OF TABLES
xiv
Table 18: Confusion matrix(CM) of classification from MSVM (value on left are from
Table 20: Material Composition (%) used in the study .................................................. 107
Table 21: Material constants for flow stress equation .................................................... 111
Table 22: Summary of various models for prediction of mechanical properties ............ 130
Table 23: Sample file from the sensor data fusion ......................................................... 137
Table 24: Chemistry of the three materials in this study (percentage by volume) ......... 138
Table 25: Summary statistics for aloying elements in Material A.................................. 139
Table 26: Summary statistics for mechanical properties of Material A ......................... 140
Table 27: Material properties for three different materials considered in this research . 144
Table 28: Values of material constants for Ti-6Al-4V in FORGE 2008 ........................ 186
xv
LIST OF FIGURES
product ................................................................................................................................ 4
Figure 7: (a) Open Die foging (b) Closed die forging [65]............................................... 51
Figure 10: Coincident meshes (left) and non-coincident meshes (right) [74] .................. 60
Figure 13: Images of common defects found on hot rolled bars: (a) inclusion, (b) roll
marks, (c) seam, and (d) spall or pit. Source: Courtesy OG Technologies....................... 64
Figure 14: Common surface and internal defects in a typical continuous cast billet or slab.
........................................................................................................................................... 65
xvi
Figure 16: Schematic illustrating the formation of (a) underfill and (b) overfill in the
roughing passes that result in rolled seams in the finisher (Ji & Shivpuri [90]). .............. 70
Figure 18: Scale Behavior during a) first rolling pass b) fifth rolling pass [92] ............... 72
Figure 20: Modeling of scale during the rolling process [91] .......................................... 74
Figure 21: Typical location of cracks along the circumference of a bar [94] ................... 75
Figure 33: Different seam defects for the training dataset (a) Seam1 (b) Seam2 (c) Seam3
........................................................................................................................................... 96
Figure 34: Different longitudinal crack defects for the training dataset (a) Crack1 (b)
xvii
Figure 35: Different scale defects for the training dataset (a) Scale1 (b) Scale2 (c) Scale3
........................................................................................................................................... 97
Figure 36: Transverse cracks - corner (TCC) defects for the training dataset (a) TCC1 (b)
Figure 37: Transverse cracks – mid surface (TCMS) defects for the training dataset (a)
Figure 40: Flow stress curves for three grades at strain rate of 1/s and different
Figure 41: FEM simulation of the roughing mill using ROLPAS .................................. 109
Figure 43: Output profile of rod from FEM simulation and the calculation of ovality .. 112
Figure 45: Stresses during the rolling process illustrated by FEM model ...................... 113
Figure 46: Results from MCMC Simulation for parameters β12 and β23 for the seams
Figure 47: Results of model validation and error percentages for the models ............... 125
Figure 48: Simplified Graph for Seam Defect for Material A ........................................ 126
Figure 49: Multi-stage processing including melting, casting and rolling ..................... 128
Figure 51: Sensors in a hot strip mill and sensor data fusion ......................................... 136
xviii
Figure 52: Histograms and density curves of mechanical properties for Material A ..... 140
Figure 53: Histograms and density curves for process parameters for Material A ........ 141
Figure 54: Schematic of the deformation zone in hot rolling of flat products ................ 141
Figure 55: Variation in roll loads and torque due to the variation in chemistry of material
A...................................................................................................................................... 145
Figure 56: Median temperatures for material A during the finish rolling process ......... 148
Figure 57: MCMC Simulation results for β31 and β103 for YS model of Material A ...... 155
Figure 58: Actual vs Predicted YS and UTS for Material B .......................................... 156
Figure 62: Various fatigue failure in safety critical aircraft components [22]- (a) Tail
gearbox housing (b) Connecting rod of piston engine (c) Wheel Hub ........................... 163
Figure 63: Cutaway and Cross-section of a high bypass jet engine [23] ........................ 164
Figure 64: Four types of disks: low and high pressure compressor, and low and high
pressure turbine. The former are made from titanium and the latter of superalloys. Note
low pressure turbine drives low pressure compressor disk and high pressure turbine
xix
Figure 67: Discrete melt related defects in titanium alloys, hard alpha, pores and ceramic
particles (top); the probability of their occurrence as a function of defect size (bottom)170
Figure 68: Flow stress of low-frequency hard-alpha as a function of nitrogen content . 171
Figure 71: Overall Theme of the FAA Study [121] ........................................................ 173
Figure 72: Damage tolerance approach to determining the life of aeroengine disk [122]
......................................................................................................................................... 174
Figure 73: Crack Growth rates for Pratt and Whitney specimens in FAA study [121] .. 176
Figure 75: Experimentation on hard alpha during FAA study [120] .............................. 180
Figure 76: Physics Based Hierarchical Decomposition of Aeroengine Disk ................. 182
Figure 79: Half Section view of Diffusion zone (left) and hard alpha (right) ................ 189
Figure 80: Dogbone forging at 0% (left) and at 100% deformation (right).................... 190
Figure 81: Shapes of DZ (left) and HA (right) after 100% deformation ........................ 191
Figure 83: Seed micromechanics during pancake forging: location in the billet (left top),
its displacement and rotation (top right), and the coordinates of the seed. ..................... 194
Figure 84: Influence of processing parameters and Nitrogen content on movement of hard
xx
Figure 85: Displacement and rotation of the inclusion during disk forging ................... 197
Figure 86: Typical strains in the Titanium matrix and Hard alpha during the forging ... 197
Figure 87: Formation of cavities during upsetting of titanium [40] ............................... 198
Figure 89: Multi Body model for formation of cavities ................................................. 200
Figure 90: Effect of processing parameters on the width of cavity around hard alpha .. 201
Figure 92: Results of fatigue test used for data model [32] ............................................ 206
Figure 93: Histogram of various values of theta obtained from simulations .................. 210
Figure 94: Diagnostic and Posterior plots of MCMC for β and βTheta ............................ 214
Figure 96: Location of hard alphas for design study (Note that each billet contains only
Figure 97: Life estimates for different design variables ................................................. 220
Figure 98: Hip Implant inside the body (left) and the critical section which fails (right)
xxi
CHAPTER 1
INTRODUCTION
structure and is difficult to understand because of its intricacy. A complex system can be
The source of this complexity can be from too many interconnected components, too
much information or constraints, and many parameters. It can also stem from little
knowledge about the interaction between the components. In this thesis however, we will
use the term complex system to mean complex engineered system only.
1
Manufacturing processes like aeroengine manufacturing, steel sheet manufacturing or
steel rod manufacturing are examples of complex engineered systems because they have
many interconnected components and there is little knowledge about the interactions
between these different components [3-6]. These processes are characterized by the
following features:
- Each process consists of a raw material in the form of powder or a preform (pre-
- Each of these processes usually consist of several separate sub-processes which might
all of the sub-processes are used to produce variety of similar products (with some
- The output or finished product from this manufacturing process might be a final
product which is used directly, or might serve as a raw material for some other
manufacturing processes.
- These manufacturing processes are designed and operated to achieve a specific goal:
the final quality of the product being output from that process.
Table 1 summarizes these features for three different manufacturing processes and Figure
2
Manufacturing Input Sub-processes Configurations Output Quality Attributes
process
Aeroengine disk Ingot from Forging, Heat Different materials, Aeroengine disk Yield Strength, Fatigue life,
Finishing
Bar or rod Iron, alloying Melting, Continuous Different materials, Bar or rod Diameter, Ovality, Surface
Manufacturing elements like casting, rough rolling, different diameters defects, Yield Strength,
titanium etc.
3
Figure 2: Sequence of manufacturing processes that convert raw material to a finished
product
To achieve and maintain the quality of the products being produced from this sequence of
processes is a daunting and challenging task. This involves a series of steps and multiple
specification in mind. They also give a general list of specifications which should be
4
o The forging press manufacturer for aero-engine disk designs and builds his
builds his equipment based on the size requirements of the rolling mill.
modeling, FEM). This design gives the mean settings of different machines which
involve different forms of modeling and the set up specifications are meant to
optimize only that particular sub-process for which the modeling the done [7-8].
(CFD) based modeling is done for the continuous casting process, but
FEM based modeling is done for the rolling process. Each of these models
optimizes the sub-process which they are modeling and the output serves
- As the product goes from one machine to another, it usually changes in shape, size
and properties. However, the material state of the product remains constant and only
which result in final quality of the product, experimentation and modeling of internal
5
phenomenological models and modeling of material evolution during the process [9-
10].
evolution of different phases of steel from the furnace to the coiling. They
define the grain size, percentage volume etc. for austenite, ferrite etc. as
different phases, defects and grains during the forging and heat treatment
process.
life at which 90% of the components fail (L10 life) or the stress to life
- Since the mechanical and metallurgical designs do not capture the variability
manufacturing plants specify the control limits for the various parameters on the
machines which would result in acceptable part quality. The limits are maintained by
the automatic controls via the sensors and data acquisition system on individual
machines.
6
o The upper and lower control limits (UCL, LCL) are set on the temperature
of the billet after each rolling step, the rolling load at each step, the speeds
o The limits are set on the forging tonnage, the heat treatment temperature
- After the product is manufactured, extensive inspections and testing of coupons cut
from the product is done. This testing is to verify the desired properties and to satisfy
determine the fatigue life of the product. Also, extensive product testing is
An example of the flow of the various design and control activities for a specific case of
7
Figure 3: Design cycle for a complex engineered system
The current design cycle for these different manufacturing processes treats the different
system components (sub-processes) in isolation. Each engineer and personnel looks at the
problem in his own perspective and optimizes the solution pertaining to the requirements
8
or specifications of his particular department or sub-process only. An analogy which can
be given for this aspect is the example of “blind men touching different parts of the
elephant to figure out the overall shape of the elephant” (Figure 4). This approach leads
o In sheet manufacturing, some materials have a variation of ±20% from their mean
yield strength. Also, in some cases, more than 50% of the sheets have a yield
profitability.
o In aeroengine disks, there is a large scatter in the fatigue life of tested coupons.
Sometimes, this scatter is results in variation of 2-3 times the anticipated fatigue
life. This results in a conservative design of the aircraft and extensive inspections
9
- Extensive testing is done at the end of manufacturing to compensate for the lack of
understanding of the variation occurring during the process. The predictive models
built for the quality attributes are able to give a good estimate of the mean of the
populations set, but fail to account for the scatter in the properties.
o The most common predictive models for sheet manufacturing can predict
o Predictive models for bar manufacturing look at only one or two different
kinds of defects and can only predict whether the defect will occur or not.
o Almost all predictive models for aeroengine disk failure can give an
values [12-13].
- Due to the lack of understanding of the interactions between the different sub-
processes, the control of machines is done is at a local level (ie. Optimizing only that
particular machine or sub-process) in a hope that it will reduce the variation in the
quality attributes on the product. This leads to over-constraining the overall process
- Every time there are changes in processing like equipment change, material changes
etc. the whole design cycle has to be repeated again without using any prior
knowledge available through the previous processes. This wastes a lot of resources
and costs a lot of money. Machine degradation over time also changes the control
10
parameters on individual machines and leads to increased variations in product
quality.
Due to the various limitations of the existing design methodologies and practices for the
manufacturing processes discussed above, a new approach and methodology for the
design of these systems is presented in this dissertation. This methodology has the
- The quality attribute is decomposed into the various factors which affect it and
- These factors are further decomposed into the physical phenomena which cause these
- The physical phenomena are decomposed into the manufacturing processes and
- Interdependencies are determined between the different sub-processes and only those
simulations etc and the data obtained from the plant are used to find these
interdependencies.
- Design of the manufacturing process is carried out so that the desired quality
attributes can be achieved within the specification limits considering all the
11
An example of this decomposition for aeroengine disks and its life is shown in Figure 5.
In this approach, the life of an aeroengine disk (and its variability) is decomposed into
being due to three different states: the defect state (material impurities), the material state
and the loading. These states are then decomposed into the process physics which cause
the states. For example, the Grain size, microstructure and its texture etc. define the
material state; the applied stress on the disk defines its loading etc. These quantities are
further decomposed into the design parameters affecting them. For example, the forging
velocity affects the strain and strain rate in the defect which affects the defect state and
hence the life. Finally, the design parameters are decomposed into the actual
12
manufacturing processes which cause them. Thus, the forging velocity is controlled by
the forging process etc. In this manner, the life is linked to the actual manufacturing
process which affects it and the design of these manufacturing processes can be done to
enhance the life and understand its variability.Similar decomposition is done for sheet
Let [X] represent a probability distribution for a random variable X; [X|Y] represents the
for updating our prior distribution [X] based on data, say, Y. That is, the posterior
where, for a continuous variable X, [Y] = ∫[Y|X][X]dX. The basic concept of hierarchical
distributions:
In any process, we are interested in finding the output given the input and observables
(prediction) and also the inverse (design). In both the cases we need the joint probability
distribution of all the variables. Let us denote the data by D and the process by P.
13
We assume that there are statistical parameters Θ (typically error variance, but possibly
parameters accounting for factors other than variability) and physical parameters η. We
apply Bayes‟ theorem to model the joint uncertainty of all the quantities involved,
making some assumptions along the way, which are related to the physical process.
Thus,
The first term is the data model, the second term is a prior physical process model and
the last term is a prior model for the statistical and physical parameters, called the prior
geophysical systems [14-15] and has later been adapted by researchers in fields like
climate modeling and personal exposures [16-17], but has not been used in the
manufacturing domain. This has the advantage that all the unknowns are treated as
random and more complicated models can be easily incorporated in the same setup. The
methodology and the model formulation does not change as more and more data becomes
available and new knowledge about the process can also be incorporated in the physical
14
A new methodology for the prediction of quality attributes and the design in complex
which resulted in the product in study into account while designing the system. The
knowledge about various processes and their interaction as well as the data obtained
during the process is combined in a manner so that it can be intelligently used for the
purpose of design.
Previous research in the field of complex engineered systems has always focused on
individual elements of a system and optimization based on these elements. The design of
these systems has been traditionally based on skill sets of engineers and the overall
physics of the process is seldom taken into account. The role of manufacturing processes
has been recognized as being important to the design but has been rarely taken in the
design process.
incorporated into the design process through process physics. The complex system is
broken down into various elements based on the physics of the manufacturing processes
and these are linked in a Bayesian formulation to include the data obtained through the
15
(2) Modeling process physics into Bayesian description
convert it in a format which can incorporate both the data and the physical knowledge.
physics based format. In both these formats, continuous updating of the model is difficult
In this work, Bayesian Hierarchical Model (BHM) is developed to couple the process
physics and the data in a single framework. This framework is used to design a complex
system by integrating the experimental and plant data with the process models.
different complex systems: bar manufacturing, sheet manufacturing and aeroengine disk
manufacturing. In all these cases the basic framework of the decomposition remains the
Chapter 1 describes the problem statement, the current approaches in solving the problem
of design in complex systems and their limitations. This chapter outlines the approach
taken in this work and how it overcomes the limitations of previous work.
16
Chapter 2 describes the various methods of modeling of metal forming processes.
Physical modeling of bar rolling, sheet rolling and the forging process are illustrated.
Finite Element Method (FEM) and how it is used for modeling is also explained in this
chapter.
Chapter 3 illustrates the decomposition of bar manufacturing process for reduced surface
defects. Case study involving a plant is used to illustrate the concept of hierarchical
decomposition. Plant data is used to calibrate the model and model predictions are found
to be satisfactory. Bayesian networks are built to study the effect of rolling on the
Chapter 4 illustrates the same concept for sheet manufacturing process and the resulting
mechanical properties. Bayesian Hierarchical Model (BHM) is used to create the network
which shows the interactions between the rolling mill and the run out table on the yield
based process models and the experimental data is combined in BHM to predict the life
under different conditions. Based on the results, a design of the forging process is
Chapter 6 concludes the main work in this dissertation and points out the main
contributions.
Chapter 7 gives some ideas as to how this work can be extended in the present domains
17
CHAPTER 2
Tremendous amount of research has been done in the field of modeling of metal forming
processes like rolling, forging, sheet metal forming etc. Although it is not possible to
cover the entire field in this dissertation, this chapter will attempt to define some of the
most important methodologies and characteristics which are used to model these
processes.
As written by Wagoner et al. [18], “the rolling process can be defined as a continuous
process of plastic deformation for long parts of constant cross section, in which a
reduction of the cross sectional area is achieved by compression between two rotating
rolls (or more)”. The two major classification of rolling process are: Cold rolling and hot
rolling. Another important distinction is made according to geometry of the final product.
Flat rolling is performed with cylinders: this is also the case for sheet rolling or strip
rolling (in which the thickness is small), or slab rolling (in which the slab thickness is of
the order of 100 mm). Shape rolling allows the production of more complex workpieces
by using appropriate roll geometries: the cross section of the part can be a round, an oval,
18
Figure 6: Schematic of different rolling processes [19]
The quality of hot rolled sheet not only depends upon physical dimensions like thickness,
profile, flatness, but also depends on its intrinsic properties like the mechanical properties
(tensile strength, yield strength, elongation to fracture, etc.). The most important
objective of hot strip mill rolling process is to create steels having desired mechanical
properties of high strength, good ductility, good weldability and formability. In case of
controlled rolling, process parameters are chosen to suit a particular steel grade in a
particular mill in such a way that the desired mechanical properties can be achieved. The
19
parameters available are the temperature, the strain and the strain rate per pass, the
interstand and the pre-coiler cooling rates. The mechanical properties are determined by
microstructural characteristics, such as grain size and the volume fraction of each phase,
and material properties prediction models can be used to predict microstructural changes
and the resulting mechanical properties under various rolling and process conditions. The
strength of iron and steel increases with finer grain size, and each phase has a different
hardness. A system for predicting microstructural changes and the resulting mechanical
conditions, such as the processing and cooling conditions, and then uses a materials
prediction model to determine how to produce a coil or a plate with the desired
by predicting the microstructure after rolling, and then using these results as input to the
control of the actual rolling line so that the rolling conditions can be modified to obtain
the desired result. The evolution of the microstructure model can be well described:
beginning in the reheating furnace, where the material becomes austenitic and potential
precipitations dissolve; over the roughing and finishing mill with its recrystalization,
grain growth, precipitation, and strain hardening; to the cooling section where the hot
Microstructure modeling in hot strip mills began in the late 1970s pioneered by Sellars
and Whiteman[20-21] and empirical equations have been developed by many researchers
for various processes thereafter. Most of the work leading to development of models for
direct industrial application started in early 1990s. In 1991, Laasraoui and Jonas [22]
20
reported microstructure evolution model for hot rolling of steel with empirical
austenite grain size. They validated the model results using pilot mill data. Simulations
based on the above model showed that softening by dynamic recrystallization does not
occur in plain carbon steels under plate rolling conditions. However, when
recrystallization is incomplete, effective strains above 1.0 are easily reached. This
indicated that when the interpass times are short (as in hot strip rolling), and when both
such as Nb steel.
developed by Sellars (at University of Sheffield), Esaka and Yada (NSC), Kwon
(POSCO), Choque (IRSID), Roberts (Inst. for Metall.) and Hodgson (BHP) for kinetics
and control for factory automation. Beynon and Sellars [24] described details of a
Rolling) which had been validated not only with the data of C-Mn grade of steel, but also
21
microstructure-evolution model for hot rolling and its application to forecasting the
Nb microalloyed steels. Nanba et al. [26] of Kobe Steel developed mathematical models
to predict grain size distribution in the through-thickness direction in carbon steels. They
developed mathematical models for the critical strain for dynamic recrystallization, the
grain size of dynamic and static recrystallization, the fraction of dynamic and static
austenite. They found that grain size distribution in the through-thickness direction was
smaller when the initial austenite grain size was smaller or the rolling reduction was
larger. They also observed that the calculated values of the grain size distribution by
computer simulation corresponded well with the experimental values with rolling tests at
rolling reduction of 30 and 40%. The calculated values were larger than the experimental
ones at a large rolling reduction (about 50%). Kern et al [27] developed computer models
for the simulation of grain size development during thermo-mechanical rolling and the
resulting strength properties for the prediction of the material properties of microalloyed
HSLA steel plates. The main input variables for calculations were steel composition,
deformation-time temperature schedule during rolling and cooling rate after rolling. This
model was validated with industrial data. It was found that the standard deviation
between predicted and measured properties for YS and UTS was about 300 N/mm2.
Yoshie et al [28] of Nippon Steel developed a mathematical model for predicting grain
22
multi-pass rolling was represented by the grain boundary area per unit volume and the
average dislocation density. They found that their model showed good result for
prediction of mechanical properties such as YS, UTS and Fracture Appearance Transition
Watanabe et al. [29] of Nippon Steel developed an integrated model for the prediction of
TMCP by linking the metallurgical model and the process models. Accuracy of the model
was examined by comparison between the observed and the calculated values about
microstructural parameters and mechanical properties concerning the plates which were
Corporation. Saito and Shiga [30] of Kawasaki Steel Corporation developed a computer
formulated into mathematical equations. Using this simulator they found that in high C-
high Mn steel, the transformed structure consists of ferrite and pearlite phases at lower
cooling rates and larger effective austenite interfacial area per unit volume, Sv, values.
The volume fraction of second phase increases with the increase of cooling rate and the
23
A review paper was published by Senuma et al. [31] about Mathematical Models for
models and their equations have been described in this paper. Suehiri et al. [32] have
described a microstructural evolution model for high carbon (0.5% C) steels. The
coefficients of the empirical equations for high carbon steels have been determined. The
mathematical model predicts the transformation kinetics during cooling with special
attention to pearlite transformation aiming at the application to high carbon steels. This
model is combined with the models for the calculation of austenite grain size after
reheating and the microstructural evolution of austenite during hot deformation. In order
to calculate an accurate cooling curve, the integrated model is coupled with a FEM model
taking into account the latent heat evolution of transformation. Good agreement between
calculated and measured temperatures on the run-out table was obtained. For 0.5%C
steel, the formation of bainite which deteriorates its quality can be avoided by the
condition that water cooling is stopped just before the transformation start and restarted at
precipitation kinetics during hot rolling HSLA steels. They prescribed two different
equations for precipitation kinetics: one for uncrystallized austenite and other for
cooling rate. Increasing the cooling rate lowers the incoherent fraction and promotes
coherent precipitation and consequently the final precipitation hardening. The coherent
precipitation kinetic was found to be very fast in ferrite: less than 10sec at 800 oC to
24
complete the coherent precipitation and 100 sec at the same temperature to have the full
incoherent precipitation in anNb steel. A very high cooling rate (at least 20 oC/s, but the
Hodsgon & Gibbs [34] developed mathematical models for each of the microstructural
events that occur during the hot rolling of a range of commercial steels. These models
have been incorporated into process models for the various mills to allow the prediction
of the evolving microstructure and the final mechanical properties. Mill data have been
used to verify, where possible, the accuracy of these models. The results to date indicate
that these models can accurately predict the final properties for C-Mn steels and
microalloyed steels containing Ti, V and Nb with a mean accuracy of ±10%. Medina and
Lopez [35] developed empirical model for static recrystallization in austenite and its
the hot strip mill. They found some important observations as described. The austenite in
the C-Mn steel cannot be heavily hardened, not even at rolling temperatures close to the
critical transformation temperature, Ar3. On the other hand, the austenite in the
microalloyed steel can be heavily hardened provided the steel is rolled at temperatures
below the critical recrystallization temperature (915OC), so the final ferrite grain size is
smaller than in the C-Mn steel. When the final temperature of the rolling cycle is higher
than the critical recrystallization temperature, the microalloyed steel also has a smaller
25
ferrite grain size than the C-Mn steel as its activation energy is still greater. The initial
austenite grain size has virtually no effect on the final ferrite grain size as its influence
temperature range where the activation energy is of the order of or higher than 108 J/mol,
then the resulting residual strain of austenite is roughly equal to the sum of the applied
strains. At the temperatures where controlled rolling of the austenite phase is usually
done, the two types of steel recrystallize fully at the roughing mill and the initial
microstructure at the finishing mill consists of completely equiaxial grains of smaller size
laboratory torsion tests for three low carbon steels. For the grade containing 0.17%C and
0.036 % Nb, it was determined that the Tnr =920 oC. For the material containing 0.09% C
and 0.037% Nb , this temperature was slightly lower, with Tnr=910 oC. For the steel
containing 0.12% C and no Nb, the Tnr value was much lower: Tnr = 832 oC. The Tnr
values for these steels were also established from mill log data and found to be virtually
identical to those evaluated using the torsion tests. The log calculations were carried out
by organizing the Sims roll force equations into desktop computer spreadsheet software.
It was also found that the overall levels of mean flow stress calculated from the mill logs
were close to those determined from torsion testing. The results of both the torsion testing
and mill log results compared reasonably well with previously developed correlations
between Tnr, and the chemical composition, and between Ar3 and the composition.
26
In 1995, Xu et al. [37] found the effect of Carbon content (4 types: 0.05%C, 0.15%C,
0.41%C, 0.84%C) on recrystallization behavior of steels. They found that the softening
curve associated with a DRV matrix has a sigmoidal shape and consists of two stages,
followed by complete softening. In contrast, the softening curve for a full DRX structure
consists of three distinct stages with three plateaus of softening, followed by incomplete
softening. For DRX matrices evolved at a strain corresponding to the first stress
minimum after peak, the softening rates in stage I and II increase and, in contrast, those
in stage III decrease with increase in carbon content and also with decrease in initial grain
size. Pietrzyk et al. [38] developed an integrated model consisted of thermal, mechanical
and microstructural sub models. This model, combined with accurate constitutive,
accurately predicted when using a heat transfer coefficient between the stock and the rolls
of 20kW/m2K. The model predicted the inhomogeneity of strain through the roll gap and
its effect on the rolling loads. Despite its limited validity due to the use of a power law in
the hot strength model, the rolling loads were accurately predicted when full
recrystallization was taking place. A recrystallization model was necessary to predict the
increase in rolling loads as partial recrystallization took place at lower temperature. The
loads and the austenite grain size than the Sellars model. The final ferrite grain size
through the thickness was accurately predicted using the Gibbs et al. grain size model.
27
Medina and Mancilla [39] developed a static recrystallization model for hot deformed
steels containing several alloying elements. They found that the use of steels with
suitable compositions has made it possible to identify the quantitative influence of each
the grain size. Exponent "n" in Avrami's equation on fraction of recrystallized volume
diminishes slightly with the temperature. Strain rate has less influence than strain and
grain size, but its influence is not negligible. Strain is not an independent variable; its
influence on recrystallization kinetics is slightly related with the grain size. The high
correlation index (r>0.96) between experimental t0.5 and calculated t0.5 demonstrated the
good prediction of the model. They [40] have also investigated static recrystallization
different strain and steel composition. They have observed that static recrystallization
Majta et al. [41] described modeling the Evolution of the Microstructure of a Nb Steel.
The model incorporating relations that describe the evolution of the microstructure of
element method, were presented in the paper. The model accounts for the restoration
mechanisms, the hardening events and the accumulation of strains. The reasonable
28
predictive capabilities of the technique have been demonstrated by comparing the
niobium steel to that of measurements. The capabilities of the model were further
the retained strains and the austenite grain distribution, was computed. Further, the effect
of the niobium content on its ability to retard grain growth was displayed in the paper. Li
et al [42] studied effect of initial grain size on the static recrystallization kinetics of Nb
microalloyed steels. Initial grain sizes were varied in experiments in the range of 10-71
micron. They determined the Avrami exponent (n) for the static recrystallization
of Mn.
Siciliano et al. [43] developed a mathematical model which correlates Mean Flow Stress
(MFS) with chemical composition, strain, strain rate and temperature. They compared
predicted MFS with the MFS value obtained from measured Roll force. They have found
that all the prediction falls in the range of ± 15%. They further developed an expression
for the redundant strain in flat rolling. For strip rolling, this falls in the range 0.07 to 0.16
times the nominal strain associated with each pass. With the aid of the above redundant
strain expression, together with factors for work roll flattening and forward slip, the data
from hot strip mill logs was analyzed in terms of MFS vs. l/T. This approach was
employed to determine the microstructural events that occur during hot rolling. A model
29
was developed to predict the MFS behavior of plain C-Mnsteels during strip rolling from
the strains, strain rates, temperatures, and interpass times. The model indicated that
few passes during the strip rolling of plain C-Mn grades. This is because the strains and
temperatures are relatively high and the strain rates are quite low; thus the DRX critical
strain can be readily exceeded in these passes. It should be noted, however, that the type
of DRX observed here is not associated with strain accumulation. This is in sharp
contrast to the case of Nb containing steels, where strain accumulation plays an important
role in dynamic recrystallization. Miami, Siciliano et al. [44] then developed a similar
model for Nb microalloyed steel. This model helped in good understanding of the
microstructural events that occur during the hot strip rolling of Nb grades. Data from mill
logs are analyzed in terms of MFS vs. 1/T. For this purpose, the approach of Sims was
used to convert the rolling load into MFS with the redundant strains and forward slips are
taken into account. The peak strain associated with the occurrence of dynamic
reflects the Nb level. The predictions obtained from the model described were in good
agreement with mill observations (±15%). In another paper, Kirihata, Siciliano et al. [45]
described a similar model developed for Multi-alloyed Medium Carbon Steel. On the
basis of the torsion tests performed and the mill log analyses carried out, they observed
that the activation energies for deformation, derived from the peak strain and steady state
stress behaviors, display different values. The activation energy associated with the
steady state stress is higher than that for the peak strain. Compared with plain C-Mn
30
steels, the kinetics of recrystallization in the multiply-alloyed C-Mn-Cr-Mo-V-Nb-Ni
steel are 100 times slower when dynamic recrystallization is taking place and 10 times
slower during static recrystallization. MFS predictions based on the new version of
were in good agreement (±15%) with the MFS values calculated from mill log data by
the Sims method. According to the model derived here, dynamic recrystallization takes
place in each of the first three stands because of the relatively large strains and the low
level of Nb addition. Dynamic recrystallization also occurs at the fifth stand, in this case
combining cooling control model with SPPC model. The earlier system was used to make
the coiling temperature constant along the length of the coil. But, as the latent heat
generation depends upon the chemical composition with phase transformation, this
2.2 Modeling Microstructure Evolution in Carbon Steel during hot strip rolling
During conventional reheating of a slab at the temperature approximately 1250 oC, only
large grains (>100 micron) are present in the slab, and it usually takes only a fraction of a
second for the material to recrystallize completely. Recrystallization time, however, may
increase significantly at lower reheating temperature and also during hot rolling. When
stage of rolling process with small applied reduction. However, the accumulation of
31
strain during subsequent passes leads to full recrystallization. Thus, the effect of initial
Manganese steels, when the grain size are below 100 micron, recrystallization is very
rapid above the temperature of 1000 0C and slows down below 950 0C.
Dynamic recrystallization
When recrystallization takes place under loading condition in hot rolling, it is called
dynamic recrystallization (DRX). When steel is deformed in the austenitic state at high
temperatures, the flow stress rises to a maximum and falls to a steady state. The initial
increase in stress is due to work hardening of austenite. Three types of strain are
(a) Peak Strain : The peak strain is the strain at which maximum flow stress is
produced. The maximum flow stress is also called peak stress ) . The general equation
for peak strain that has been used in the model is:
………... Equation 4
where
The values of the coefficients a,b,c,d and Q have been determined experimentally by
(b) Critical Strain : The critical strain corresponds to the start of dynamic
recrystallization. At this strain, the stress is equal to the critical stress ). Critical strain
(c) Steady-state Strain : This train corresponds to the end of dynamic recrystallization
(d) Dynamic recrystallized grain size (dd): When there is a dynamic recrystallization, the
grain size corresponding to steady state condition of flow stress is called the dynamic
Metadynamic recrystallization
Microstructure that are developed by dynamic restoration are not stable and at elevated
33
Two important parameters that define the MRX process are metadynamic
(a) Metadynamic recrystallization time (t0.5mx): This is the time required for completion
recrystallization, the general equation for metadynamic recrystallized grain size that
…………………Equation 5
where
The values of the coefficients a,b,c,d and Q have been determined experimentally by
many researchers. Some of these values have been given in the Table 3.
Static recrystallization
34
Microstructures that are developed by dynamic restoration and metadynamic processes
are further modified by static restoration processes. The static restoration processes are
static recovery, static recrystallization (SRX) and static grain growth. In hot rolling SRX
temperature. Some of the important features of SRX are: (i) A minimum amount of
deformation called critical strain is necessary to initiate SRX. (ii) The lower the degree of
deformation, the higher the temperature required to initiate SRX. This temperature is
greater number of deformed grains are recrystallized. The lowest temperature at which
two equations:
……………..Equation 6
……………..Equation 7
where
The values of the coefficients a,b,c,d and Q have been determined experimentally by
many researchers.
35
Grain growth after complete recrystallization
Grain growth after deformation is affected by both the amount of reduction and holding
time. Depending upon holding time and reduction, grain growth can be divided into 3
regions. In region-1, called mixed grain growth region, the grain growth starts from a
mixed structure that consists of either recovered grains and giant grains or recrystallized
grains. In region-2, called normal grain growth region, grain growth follows a pattern. In
region-3, called abnormal grain growth region, when large grains suddenly develop
among small grains. The size of the grains coalesced from the small grain structure
during this secondary recrystallization process is much larger than the size of the
coalesced from a large grain structure.The general equation for predicting grain growth is
given by:
………….Equation 8
where
The values of the coefficients a,b,n and Q have been determined experimentally by many
36
Researcher a b n Q/R (K-1)
Austenite Decomposition
Austenite transformed into various solid solutions and aggregates after cooling. They are
affected by the transformation temperature and cooling rate. Ferrite is the prime product
necessary to create conditions for a high ferrite nucleation rate during the transformation
of austenite to ferrite. The parameters mostly affecting the ferrite grain size are steel
temperature, cooling rate and coiling temperature. The general equation for predicting
ferrite grain size (in micron) is given by the following general equation:
……Equation 9
37
where
(micron)
The values of the coefficients a,b,c,d,e,f and g have been determined experimentally by
many researchers. Some of these values have been given in the Table 5.
Researcher a b c d e f g
(Ce>0.35)
(Ce<=0.35)
38
The principal metallurgical factors that affect resistance to deformation is the stress that
is required to create and move dislocations. A basic dislocation is the imperfections in the
grain structure of a pure metal without any alloying elements, and having a uniform
occurs in steels by the movement of numerous dislocations over large distances through
the crystal lattice. This process requires that there be an applied stress. The term
result of changes in the metal structure that impede the motion of dislocations. The ability
of dislocations to move through the alloy can be impeded by: (a) dissolved interstitial
atoms, (b) dissolved substitutional atoms, (c) interactions of dislocations with grain
boundaries, (d) interactions of dislocations with second-phase particles and (e) work
hardening. A general equation for yield strength of material can be expressed as follows
...............Equation 10
where
= precipitation strengthening
= dislocation strengthening
39
= work hardening strengthening
Similar equation exists for ultimate tensile strength of the material. Based on the above
Yield Stress
Yield stress (YS) of material is affected by many factors including chemical composition
of steel, ferrite grain size, material thickness, transformed volume fraction of ferrite.
…. Equation 11
Pickering
(Low Carbon)
where,
40
= ferrite grain size (micron)
….Equation 14
where,
With development of soft computing techniques like artificial neural network (ANN),
genetic algorithm (GA) etc, efforts have been made by many researchers to predict
41
mechanical properties of hot rolled steel. Andorfer et al [47] from Voest Alpine (VAI)
described an on-line microstructure model based quality control system which was
implemented in Hot Strip Mill, Voest Alpine Stahl Linz (VASL). Designated as VAIQ-
Strip, this model predicts mechanical properties along the length of the coil very
accurately. They have used regression with artificial intelligent methods for model
adaptation.
Bhadeshia [48] described the application of neural network in different fields of material
quantitative model as a function of large number of variables like percentage of Ni, Cr,
Co, Mo, W, Nb, Al, Ti, Fe, Mn, Si, C, B, and Zr concentrations, and temperature.
properties of steels by using neural networks and multivariate data analysis based on the
data obtained from Hot Strip Mill of Cockerill-Sambre Carlam Steel Plant. They
developed an ANN model with target variables as Yield strength [Re], Ultimate Strength
[R], Elongation [El] with input variables as thickness [Th], width [W], steel chemical
analysis ([C], [Mn], [P], [S], [Si], [Al], [V], [Nb], [Cu], [Cr],[Ni],[N],[Tl],[Sn],[B]
contents) and rolling process temperatures (reheating [TF], roughing [TS], finishing [TR]
and coiling [TC]. They found that Multivariate Data Analysis remains an essential step
before, during and after NN modelling. For a given set of variables selected for the
modelling, neural networks properly applied can suggest a limit value under which it is
not possible to go down and which is largely explained by the dispersion measures of
42
mechanical properties. Datta et al. [50] described an Petri Neural Network model (a
multilayered feed forward network model) used for predicting mechanical properties of
steel. They correlated the measured mechanical properties with chemical composition,
temperatures and % deformation in this model (total 14 input parameters). They found
that there is good agreement between measured and predicted properties (±10%).
Kong and Hodgson [51] developed both constitutive and artificial neural network models
to predict hot strength of austenitic steels with carbon content varying from 0.0037 to
0.79%. Due to the complexity of the stress strain behavior, both models were not able to
accurately predict the difference of the behavior between low and high carbon steels. The
accuracy of the ANN prediction was particularly poor for both training and test data sets,
which cannot be used for industrial application. In addition to higher prediction accuracy,
experimental data. Integrating constitutive and ANN models developed two new models.
As constants in constitutive model vary in a complex way, linear fitting of these constants
with statistical technique presents a high error and the features of the stress strain curves
caused by the variation in carbon content are not accurately predicted. Using a multilayer
ANN model, the constants in the constitutive model were predicted which is much more
accurate than the linear fitting. Consequently, the accuracy of the prediction on the stress
strain behavior with this integrated model was improved (±8%). They developed an
into the artificial neural network model and introducing more inputs from constitutive
43
model. The stress strain behavior varying with the carbon content was accurately
predicted. In comparison with the ANN model with using only chemical composition and
Femminella [52] described the importance of data pre-processing and model initialization
feature extraction for models constructed by the NF approach was obtained using input
which proved to be the most successful one in overall terms. A modeling cycle was
Wang et al [53] developed an artificial neural network model to describe the effect of the
carbon concentration and cooling rates on CCT diagrams. Using this model, it was found
that an increase of carbon concentration in steels gives rise to the decrease of ferrite start
temperature. The decrease rate is further dependent on the carbon concentration. The
calculated carbon dependence for continuous cooling conditions equals that calculated for
44
equilibrium conditions. The ferrite start temperature is also affected by cooling rate,
especially for high carbon steels. When the cooling rate is below 0.1oC/s, the influence is
not significant while above, the ferrite start temperature will decrease significantly with
increasing cooling rates. A decrease in carbon concentration will prolong the incubation
period of ferrite reaction. In contrast to the ferrite start temperature, the pearlite end
mechanical properties of steel. The grain size predicted by the empirical model was used
as an input to ANN model along with other parameters like chemical composition and
strip thickness. The model was validated with measured property data. It was found that
the model predicts mechanical properties within error band of ± 30 MPa and root mean
microstructure evolution model with deformation model to predict roll force accurately.
According to him, the microstructural evolution takes place entirely in the austenitic
condition. The initial deformation temperature and the rate of cooling during rolling
never bring the steel into the transformation temperature range until deformation is
complete. The final austenite grain size and the ferrite grain size, to which it transforms,
45
Lofler et al [56] described a 'Microstrcuture Monitor' developed by Siemens AG to
calculate steel properties such as yield and tensile strength for low-carbon manganese and
Agreement between measured data from several hot strip mills and calculated values
from the microstructure monitor is fairly good (±10%). With respect to yield and tensile
strength, the combined model comes close to the precision of the physical measurements.
The implementation of neural networks for steel property determination is quick and
easy. During the process of data collation, consideration can be given to complex
and mechanical properties of hot rolled steel strip. One important observation from this
paper is that the mechanical properties have been found out by linear regression with
In part-II [58], they described the verification and application of the model in industrial
condition. They found that the standard deviation between the calculated and observed
yield strength is 13 MPa. Similarly, the standard deviation between the calculated and
observed UTS is 9 MPa. The standard deviation between the calculated and observed %
elongation is 2.0%. They also conducted experiments to verify their model for prediction
46
of properties along length of coil and observed a variation of about 10 MPa at the strip
end.
process models. The individual model concepts for grain growth, recrystallization, and
precipitation and phase transformations were briefly discussed. The development from
have increased predictive capabilities for these models over a wider range of steel grades
and operational conditions. The challenges in the development of the next generation of
with novel processing routes and advanced high strength steels were evaluated. Further,
the majority of the currently employed models are on the macro-scale but future
microstructures rather than a number of average parameters (e.g. grain size, fraction
methodology that combines both empirical and soft-computing approaches as the most
probable way to develop modelling tools with predictive capabilities for industrial
to specific mill operating conditions. Datta et al [60] described a genetic algorithm based
multi-objective optimization technique used to train the ANN, known as predator prey
algorithm and a Pareto front was developed minimizing the training error along with the
network size. Genetic algorithm was successfully used for alloy design. Using the model
47
they have found that YS depends mostly on the solid solution hardening and the
precipitation hardening, but all these strengthening mechanisms have a negative effect on
the ductility of the steel. These factors in general and the role of individual variables in
particular, could easily be used to control the strength–ductility ratio and even the yield-
tensile ratio of the steel. They concluded that the pruning and predator prey algorithms
applied in this study have been demonstrated to be able to extract more knowledge from
the data, which is difficult to reveal by conventional ANN analysis. In this work the
relative relevance of the composition and TMCP parameters have been detected in a way
that it could be used for designing steel with tailored property balance.
Similarly, a model for the prediction of microstructure and mechanical properties during
the rolling of sheets was developed by AIST called Hot Strip Mill Model (HSMM) [61].
mill process cover each stage of rolling from the slab dropping out of the reheat furnace
until the finished product is coiled in the up/down coiler or delivered to the cooling bed.
The microstructure evolution calculations of the rolled material start from the first
horizontal rolling stand and continue until the finished product reaches its final
precipitation etc. This model requires a lot of training for a particular mill. According to
authors, the prediction errors for the ferrite grain size ere sometimes ±50% because of
48
errors in measurement. The errors for the yield and tensile strength are also about ±15%
for C-Mn, Interstitial free and HSLA steels. These errors result in the model being used
only for offline purposes. Also, no knowledge of interactions between the various mill
Similar model was developed by CRM and Thyssen Stahl called the STRIPCAM (Steel
et. al [62]. This model ran on personal computer and can consider C-Mn steels. Different
and phase transformation. By adding volume fraction of bainite formed during rolling,
the prediction errors of tensile and yield strength improved from 20% to 15%. The
equations for the transformation temperatures were very complicated involving 6-7
alloying elements and the volume fractions of austenite etc. Santos and Barbosa [63]
C-Mn and microalloyed steels. They carried out dilatometric tests on Gleeble 3500
machines to simulate the hot rolling process. Statistical regression was used to derive
alloying elements and also their second order values and their interactions. Using this
temperature, they derived the start of perlite formation and confirmed the findings of
other researchers about the recrystallization kinetics during the hot rolling. Their
49
Notable among the statistical models is the model by Danieli Automation called
transformation and precipitation. This model is proprietary software which is not portable
drawback of this model is that it needs to be calibrated for a long time for different
materials. The model accuracy is shown to be within ±6%. Even with the best predictions
among all the other work shown in literature, no insight into the rolling mill can be
obtained by this model. The interactions between the rolling mill, continuous casting and
Forging is a secondary manufacturing process which converts the products from some
compressed under high pressures to give shape to the products. Due to this deformation,
the forged parts have good mechanical properties because of alignment of material‟s
structure along the direction of deformation which eliminates the cast dendritic structure
Based on how metal flow is done, forging can mainly be classified as open die forging
and closed die forging (Figure 7). Open die forging is performed between two flat or
50
almost flat dies, with no walls on the side of the dies allowing metal to flow freely in
lateral direction. In closed die forging, two or more die blocks with negative shapes are
brought together to form a cavity, in which the metal being deformed undergoes plastic
deformation. As the metal flow is confined by die cavity, closed die forging can result in
more complicated shapes and much closer tolerances than open die forging.
Figure 7: (a) Open Die foging (b) Closed die forging [65]
Due to great developments achieved in alloy development and machine design, research
51
The theory of plasticity governs the relationships between the processing and the internal
state variables during the deformation of metals and alloys. A lot of literature is available
which describes this theory in details [66-67]. Here, we will present only the most
The fundamental quantities that are used to describe the mechanics of deformation when
a body deforms from one shape to another under an external load are stress, strain, and
strain-rate. There are two basic theories which determine the relationships between the
observables and these internal variables – the infinitesimal deformation theory and the
solid formulation theory. The FEM methods and the metal forming research rely mainly
on the infinitesimal deformation theory and this will be used throughout this work. Under
this theory,
ε= = ln( ……………………….Equation 18
The main governing equations which determine the mechanics of plastic deformation
The unknowns in this formulation are the six stresses and the three velocity components.
The boundary conditions are specified in terms of velocity and friction. The solution of
these equations is very difficult to find analytically and hence many approximate methods
are used most popular of which is the Newton Raphson method. It is this method which is
plasticity equations. It requires that among admissible velocities ui that satisfy the
conditions, the actual solution gives the following functional (function of functions) a
…………………Equation 23
where ζ is the effective stress, έ is the effective strain-rate, Fi represents surface tractions,
and E(έij) is the work function. The solution of the original boundary-value problem is
then obtained from the solution of the dual variational problem, where the first-order
…………………Equation 25
The constant is removed by the penalized form of compressibility to arrive at the final
equation:
……………Equation 26
To solve this equation during the metal forming process, the billet is discretized into
several nodes and elements called a finite element mesh (Figure 8).
54
Figure 8: Finite Element mesh in a metal forming process [70]
After the mesh is formed, a set of nodal point velocities are defined as
where T represents the transposition and N is the total no. of node x dof of each node.
The variational equation defined above is now represented in terms of this velocity v and
……………………Equation 28
During the forging process, these equations are non linear in nature and hence cannot be
solved algebraically. Hence they are solved by Newton-Raphson method (Figure 9).
Taylor‟s expansion is used for linearization and an initial guess of v=vo is used.
55
…………………Equation 29
……………………………..Equation 30
Where K is called the stiffness matrix and f is the residual of the force vector at each
node.
The FEM procedures are implemented in a computer program in the following way.
2. Evaluate the elemental stiffness matrix for the velocity correction term Δv
3. Impose velocity conditions to the elemental stiffness matrix, and repeat step 2 over
5. Obtain the velocity correction terms by solving the global stiffness equation.
56
6. Update the assumed solution velocity by adding the correctional term to the assumed
7. When the converged velocity solution is obtained, update the geometry of the
workpiece using the velocity of nodes during a time increment. Steps 2 through 7 are
The important thing to note here is that all the internal state variables can now be defined
in a matrix format with respect to the velocity vector and at each time step, the state
variables are dependent on the previous time steps through these variational equations.
For example, the strain rate is defined by the following matrix equation in the FEM:
………………………Equation 31
Where,
………………Equation 32
To model the discrete defects inside the billet requires the use of something more than a
simple FEM formulation. This is accomplished by using the multi-body FEM formulation
inside a FEM software FORGE3. The complicated contact behavior between hard alpha
57
inclusions and matrix material is modeled using the penalty based contact approach. It
primarily used two body models. Our models of hard alpha were also developed using
Three situations in metal forming exist in which contact between two or more deformable
bodies needs to be modeled: (1) self contact when the material being worked is folded,
(2) the workpiece being worked is composed of different materials, and (3) there is
deformable body. Let a and b represent two bodies get contact with each other,
subscript a and b denote the variable associated with a and b , respectively. Assume
represents the interface between those two bodies ( a b ), the relative velocity
vs v v
a b
……………Equation 33
If n denotes outward normal of each body, the bilateral sticking contact condition can be
written as
v s 0
a a on boundary ……………Equation 34
σ n σ n 0
b b
v a n a v b n b 0 and 0 n (v n v n ) 0
a a b b
n
and
on
q 1
…Equation 35
ab K v s
a b
vs
58
where is shear stress, ab is the viscoplastic friction coefficient between two bodies,
n n n (σ n ) n (σ n ) n ……Equation 36
a b a a a b b b
v V ca2 ,
v * V ca2 , s : ε d v * ds ( v * v *) d s
0
T v * ds
a a b
Equation 37
s c
….
38
….Equation 39
vh V N n ( ) and v h V N n ( ) ……Equation 40
a a a b b b
n n
n n
where vh is the discrete velocity field, Vh is nodal velocity field and Nn is the shape
function.
59
Figure 10: Coincident meshes (left) and non-coincident meshes (right) [74]
Interface has two parts, a and b , that belong to two bodies a and b, respectively. In
continuous situation, they represent the same surface; however, after discretization, they
are not always the same. If the meshes of two contact bodies coincide, the calculation can
be done using the same nodes and elements. On the boundary , the normal component
of velocity and virtual velocity must be continuous, while other components must be
at time t may not be coincident at time t+Δt provided on the condition that an updated
Arbitrary Lagrangian Eulerian formulation (ALE) where the mesh velocities are
continuous.
Free surfaces of a and b may get into contact; this can be considered as a
meshing and remeshing problem. The calculation is more complicated when surface
mesh of boundary a and b are not coincident; the normal stress and normal
nodal contact condition, the meshes are either coincident at the interface as shown in
60
Figure 10 left, or they are totally shifted as shown in Figure 10 right.
In tradition slave master methods, the contact conditions can be enforced either on a :
n , [ v n v ( ( x n ))] n n 0 ……..Equation 41
a a b b a a
or on b :
n , [ v n v ( ( x n ))] n n 0 …………..Equation 42
b b a a b b
This traditional slave/master method provides the right number of contact conditions. But
this formulation is not symmetric, even this method is improved by using higher order
and more consistent projection method. A careful determination of slave body and master
body is required for this method. This choice also determined the solution accuracy of the
61
CHAPTER 3
3.1 Introduction
Bar manufacturing is a multistage process in which steel is melted and poured into a
continuous caster and formed into a billet. This billet then goes through a series of
roughing, intermediate and finishing passes through a hot rolling mill before being
formed into a rod which is used in downstream processes like forging (Figure 11).
Ladle
Tundish
Strands
Surface defects
Inspection
1)
Figure 1: Multi-stage continuous casting and rolling process
62
Each of the rough, intermediate and finish rolling steps consist of several machines (or
passes) to convert the shape from square to round and to obtain the desired size (Figure
12).
Due to ever increasing quality requirements, a very tight inspection is often performed
and defective products are either sent for rework or are scrapped. The scrapped products
due to surface defects account for more than 50% of the total scraps in hot rolling. This
is equivalent to about 200,000 tons of steel, or $120 million in lost yield per year for the
steel industry.
The surface defects most commonly detected on the hot rolled products are slivers (or
scales), heat checks, seams, cracks, cobbles and spalls (Figure 13). These various types of
defects occur not only in the different bars (individual runs) but also within the bars
(length of the product and along the diameter). In addition, they vary in size (length,
width, depth) as well as frequency of occurrence. The sequence of defect in a bar follows
63
no pattern, a seam may be followed by a scale defect or a cobble or scale may be
Figure 13: Images of common defects found on hot rolled bars: (a) inclusion, (b) roll
marks, (c) seam, and (d) spall or pit. Source: Courtesy OG Technologies.
Based on previous studies for industry and researchers, the reasons for the formation of
Category I: Causes related to the mill input: Defects and variability related to the
Category II: Causes related to the design of the mill: Design and selection of mill
Category III: Causes related to the mill settings and controls: parameters including
Category IV: Causes that represent noise factors or equipment failures that cannot be
64
In the following paragraphs we will discuss the issues related to the above categories
There are a lot of surface and interior defects formed during the continuous casting
process which are carried forward in the rolling and cause defects (Figure 14). These
defects occur in the continuous cast billets and slabs due to the variability in the melting
practice and or in the solidification of the melt into the solid strand, with their severity
and frequency related to the selection and variability in the casting process parameters.
Some of these defects start as small holes in the casting and then increase to seams or
heat checks during rolling. Sometimes, defects with small inclusions increase in size and
become scales in the preheat furnace. The variability of the defects in this incoming billet
to the rolling results in variability of defects formed during the rolling process.
Figure 14: Common surface and internal defects in a typical continuous cast billet or slab.
65
The design of the rolling process is usually done by mechanical and materials engineers.
They do the process modeling and finite element simulation of the rolling mill based on
certain assumed parameters. These models are deterministic in nature and are used to
verify the output mechanical and metallurgical properties of the bar. Because these
models do not capture the variation in the input parameters (the dimension of the billet,
temperature of billet, rolling speeds), the process is never truly under control and results
in defects.
Every equipment on the mill has sophisticated controls installed on them by the
equipment manufacturers. These adaptive controls try to keep the settings of the
individual parameters within limits based on prespecified values. Due to the variation of
the parameters, these controls are not able to effectively keep parameters within limits.
The SQC charts used by quality engineers do not address the multivariate nature and
interdependence of the parameters and are unable to keep the defects in check.
The equipment control is done based on the settings on the individual control
components. But due to the inherent nature of the rolling process and the environment it
operates in, there is a lot of noise in the variables. This noise is due to humidity,
temperature of environment, different control levels of the equipment etc. Some of the
noise is due to shift to shift variations like the operators, supervision etc. This noise can
66
cause the parameters to go beyond the settings and thus cause defect problems. In other
cases, there might be equipment malfunction or failure which might cause the settings to
go beyond the limits. For example, a faulty heater on the furnace can increase or decrease
the temperature of the furnace beyond limits, or wear of a roll might change the size
Section 3.2 discusses the background in modeling of bar manufacturing process – both
physical and statistical. Section 3.3 introduces the need for hierarchical modeling and
how decomposition can be done for surface defects in bar manufacturing. Sections 3.4-
3.6 describe the data model and process model needed for the hierarchical modeling and
3.2 Background
Previous work in the field of modeling of bar manufacturing process and the formation of
defects can be divided into two major categories: physical modeling and statistical
modeling.
Over the years, researchers have tried to come up with approaches to model the rolling
process and the defects. Mori [77], Park [78], Kiuchi [79] and Altan [80] have applied
classical FEM approaches to solve the design of rolling processes. Lenard[81] studied the
microstructural evolution during the hot rolling process. Sheppard[82] modeled the
67
recystallization during the rolling process to look at the metallurgical properties. Beynon
et al [83] have tried to understand the formation of scales in roughing process and predict
their formation. Kuhn [84] developed a methodology of workability during hot rolling
which helps in the prediction of cracks. Brimacombe [85], Das [86] have tried to model
the continuous casting process and show the effects of defects in continuous casting on
(ROLPAS) which can predict the microstructural evolution during the rolling and also the
profile of the bar being rolled. The non-isothermal deformation analysis in ROLPAS is
integrated into ROLPAS to enable modeling of austenite evolution. MICON uses the
microstructure evolution models to model the evolution of austenite during hot rolling.
The evolving austenite was found to significantly affect the flow stress of the material
while on the other hand; the material flow affects recrystallization kinetics. This situation
calls for an iterative approach in modeling metal flow and austenite evolution. For the
first pass, an initial preheated grain size is input to the program. After deformation and
heat transfer computations for each pass, the microstructure evolution module in
conjunction with the heat transfer analysis module computes recrystallized fraction and
the austenite grain size at each node in the interstand region. In the event of complete
68
recrystallization, grain growth after recrystallization becomes important in determining
the recrystallization kinetics of the next pass. Partial recrystallization is handled using the
rule of mixtures. The last section in the system is a module for modeling phase
transformation of austenite to ferrite, pearlite, bainite and martensite. This model also
These methods use the deterministic design of the mill parameters. They assume that the
material properties, the control parameters and the machine parameters are fixed and do
not change during the process. This is useful for the initial design of the mill, but does not
69
Apart from this generalized work in the area of bar rolling and microstructure – property
relationships, specific work has been done by researchers in modeling the individual
defects.
3.2.1.1 Seams
Seams are said to be longitudinal discontinuities in a steel bar which have closed during
rolling but not welded together. During the roughing process, if an under-fill or overfill
occur in a specific pass of a sequence, and the cross-section of the billet is deformed in a
direction perpendicular to the previous pass, then the material folds over causing seam(Ji
and Shivpuri [90]) to form. A schematic of the formation is shown in Figure 16. Thus in
practice, longitudinal defects near the parting line are said to be “seam” kind of defects.
This classification includes similar type of defects like overfill, laps etc. Topno et al. [91]
found that the surface defects, including seams, in bar rolling are caused by exogenous
entrapment (rolled scale), bad surface condition of input billets, non-uniform thermal
profile during rolling, incorrect guide setting, improper reduction schedule and improper
roll-pass design.
Figure 16: Schematic illustrating the formation of (a) underfill and (b) overfill in the
roughing passes that result in rolled seams in the finisher (Ji & Shivpuri [90]).
70
3.2.1.2 Scale
temperatures. According to the Fe–O phase diagram the following three kinds of oxides
exist at temperatures higher than 1100°F: wüstite (FeO), magnetite (Fe3O4) and hematite
(Fe2O3). Wüstite is the innermost phase adjacent to the metal and occupies about 95% of
the scale layer. Magnetite is the intermediate phase of scale it occupies about 4% of the
scale layer. Hematite is the outer phase of the scale and occupies 1% of the scale layer.
The scale formed at on the bar can be classified as primary, secondary and tertiary scale.
Primary scale forms during continuous casting of the billet as well as during pre-heating
in the furnace prior to hot rolling.After the billet leaves the furnace, it is subjected to
water sprays in the descaler. Some of the descalers use flat rolls to remove this scale.
These water sprays try to remove the scale formed in the furnace. But due to varying
adherence of scale for different materials and surface finish, they are unable to
completely remove the primary scales and it reaches the roughing mill. Secondary scale
is formed in the hot rolling process. As the billet enters the first pass of the roughing mill,
71
the scale breaks free from the billet (details X and Y in Figure 18). Some of the scale
particles fly out of the roll bite (detail Z), but most of them enter the roll and are pressed
by the roll. This depends on the adhesion of the scale to the steel surface, the temperature
if the initial stock and on the flow stress of the scale. The scale is subjected to both the
longitudinal tensile stresses and bending moment. Tertiary scale forms on the run out
table, during the coiling process and during cooling to room temperature. This scale is of
importance to the flat product (slab/sheet/strip) but not to the bar product.
Figure 18: Scale Behavior during a) first rolling pass b) fifth rolling pass[92]
There is a lower limit of the oxide scale thickness, for a particular steel grade and rolling
parameters, beneath which the scale comes into the roll gap without through-thickness
cracks. The longitudinal tensile stress at entry into the roll gap can favour through-
thickness cracks in the scale when the initial rolling temperature is low and the oxide
scale thickness exceeds its lower limit. The breaking up of the scale at the moment of the
roll gripping contributes to the scale failure for this temperature range. During the second
and subsequent passes, this scale is more pressed into the bar and loses its plasticity and
compressibility. By the time the bar enters the fifth or higher pass, the fracturing of scale
72
is minimum and the roll slides across scale particles called “sandpapering” of scale on
Beynon [91] studied the formation of scale during strip rolling in great detail and showed
the mechanism of modeling scale during the rolling process. The most important factors
in the modeling of scale behavior are the billet temperature, billet material and the scale
determine the ductility and fracture criteria of the scale. Longitudinal tensile deformation
ahead of entry into the roll, combined with the bending occurring at the moment the roll
grips the bar, induces the cracks in the scale at the entry zone (Figure 20). However, this
temperature range, the scale can either crack or can enter the roll without any pre-formed
cracks (Figure 19). The other important factor in the scale formation is the reduction
ratio and rolling temperature (Figure 19). Depending on the combination of rolling
temperature and the reduction ratio, the scale can either crack or enter without cracking.
73
Figure 20: Modeling of scale during the rolling process [91]
Once, the scale enters the roll, and then the ductility and fracture of the scale determine
whether it will be removed during the rolling process or will be embedded in the bar. Not
much work has been done in this area. Corus Research [93] looked at the aspect of scale
formation during the roughing stages based on the ductility and composition of scale.
Their work was more based on the metallurgical properties of scale and how that affects
their behavior during rolling. They did not try to predict the formation of scale or how
Surface crack in rolling is a ductile crack on the surface of billet. There are many criteria
for crack formation which have been proposed. These can be grouped into: (1) void
74
nucleation and growth; (2) damage mechanics; (3) cohesive zone; and (4) accumulated
strain.
During the continuous casting process, longitudinal cracks are formed due to the tensile
strains generated by mold and upper spray zones and also due to non uniform cooling.
Carbon levels, Mn/S ratio of the material, mold conduction etc. play a vital role in
Longitudinal cracks are formed when the workability of the bar is low and applied
stresses are high. This results in the formation of cracks or wrinkle kind of defects(Kwon
et. al, [94]). These defects form very early in the rolling mill (mostly in roughing passes)
and elongate in subsequent passes as the bar is being rolled. They are usually located at a
certain angle to the parting line and are symmetric across the parting line (Figure 21).
Longitudinal cracks are also present in the continuous casting billets but most of the
cracks of this type are usually removed by the scarfing or grinding before the start of
rolling.
Figure 21: Typical location of cracks along the circumference of a bar [94]
75
3.2.1.4 Transverse Cracks
Several types of cracks are formed during the continuous casting of billets due to
[95]). In particular, the transverse mid surface and corner cracks are mainly said to be
formed due to large surface gradients in spray zone and straightening within an
unfavorable range of temperatures during the continuous casting. These cracks are
usually removed by grinding and scarfing before the start of rolling. But, some of these
cracks are not eliminated and result in the formation of transverse cracks on the rolled
bars. These are strongly influenced by steel composition, Al, V, Nb, Mn being the most
important elements.
No-Twist-Mill (NTM) follows the roughing and finishing stands. This mill consists of
alternate rolls with round and oval pass design. The role of the NTM is to correct for the
dimensional variations that are produced due to the variations in the upstream processing
(billet casting, re-heating furnace, rolling stands etc.). This mill has cantilevered rolls of
smaller diameter which were arranged 45 degrees above and below the horizontal. The
perpendicular deformation between alternate rolls eliminates the need for twisting the rod
during sizing.
The dimensional quality of the output from the NTM has a significant bearing on the
quality of final product formed by forging processes etc. During the rolling process, the
NTM determines the precision of diameter and ovality through a series of reduction steps.
76
Any variation in cross section resulting from this process step affects the quality of
products formed by cold forging etc. It is necessary to control this process to achieve the
Considerable research has been done in the past regarding the modelling and analysis of
NTM processes. Boyomi and Lee [96] developed analytical models which predict the
model was based on the assumption of homogeneous deformation and roll bite geometry
during the deformation process. Effective strains and stresses were used to predict the roll
loads and torques. The model proved to be in good agreement with the experimental
results. They also discussed the effect of interstand tension on the roll loads, but failed to
calculate its effect on the geometry of the bar coming out. Also, they did not have any
results which would prove their theory of stresses in the rolling direction being a player
Shi [97] used state space approach to do fault diagnosis in a multi-stage manufacturing
process. This process looks at the fault occurring at a stage and tries to identify its root
cause based on matrix perturbation and Principal component analysis. Shi [97] modeled
the rolling mill as multi stage manufacturing process and used mill data to cluster the
defect-parameter relationships. Zhou [98] used multi level regression modeling to find
the relationship between seam defects and the mill settings on a hot bar mill. Li [99] used
77
logistic regression to find the root cause of defect formation during hot rolling. These
methods use the mill data from the sensors and try to predict the defect formation. The
important defect identification system used by them on a rolling mill are the HotEye
images[100] and other eddy current based inspection devices. The main drawback of
these techniques is that they lose the physics behind the process. They can be used once
we have obtained a lot of information about the process. They are not robust to design
The use of data mining techniques has also been widely used in the control and prediction
of defects in bar manufacturing processes [101]. These processes try to use the output of
sensors and by means of clustering techniques, group the defects into different classes.
Roning [102] implemented Self Organizing Maps [SOMs] to find out the scale defects in
hot strip rolling. This resulted in identifying the important parameters resulting in the
formation of scale during the rolling process but failed to give any insight into the control
requirements. Artificial Neural Networks (ANN) and genetic algorithms (GA) have also
been used to provide control of the manufacturing processes [103]. These models serve as
black box which take in the input and control settings and give out the output settings.
They need data for training purposes and then are used for fault diagnosis. These
techniques have been tried to control various elements of the rolling process but suffer
from the problem that cannot be used to predict once the design has changed.
78
3.3 Hierarchical Decomposition of the Bar Manufacturing process
To overcome the main disadvantages of the physical and statistical models and to design
the bar manufacturing process for reduced defects, the hierarchical decomposition of the
process was done (Figure 22). This decomposition is based on the same principles as
The surface defects are decomposed into the different kinds: seams, scale, cracks and
ovality. These are then decomposed into the physical quantities affecting them (eg. stress,
strain etc.). The physical quantities are decomposed into the design parameters on the
SURFACE
Level 1: Quality
DEFECTS
Attribute
Design
Level 3: Process Strain Heat
Stress Strain Grain Size
Physics Rate Transfer
79
Figure 23: Number of defects of different kinds in Material A, B and C
80
Histogram of aa
Normal Q-Q Plot
0.10
430
0.08
425
0.06
Sample Quantiles
Density
420
0.04
415
0.02
410
0.00
Histogram of bb
0.030
530
0.020
520
Density
0.015
Sample Quantiles
510
500
0.010
490
0.005
480
0.000
470
Figure 23 shows the plots of different defects in materials collected from a plant. This
figure shows that different materials have different dominant (most common or
prevalent) defects. For example, in Material A, seams are the dominant defects, while the
other defects (longitudinal cracks, transverse cracks etc.) are found in much less quantity.
81
Similarly, in Material B, longitudinal cracks are the most prevalent defects. Because of
this reason, we can assume that these defects are mutually exclusive in one material and
hence model for each material and defect can be built separately. Figure 24 shows the
distribution of each of the dominant defects. The defects are approximately normal
distributed as seen from the QQ plots. Hence, it was decided to build a simplistic normal
Bayesian Hierarchical Model discussed earlier. Let us denote the different levels in the
decomposition as follows:
S: Surface Defects in general (SSeam, SScale, STC, SLC, SOvality to represent them
individually)
Our aim is to find this joint distribution based on the observed data D. However, we
would like to use the Bayesian framework to learn about the underlying variables and
82
parameters. Using the likelihood in Equation 43 and the prior distribution on process
Sections 3.4 and 3.5 discuss the data collection and integration necessary for creating a
data model. These sections describe how the data was collected for calibrating and
validating the model and methods to convert this raw data into a usable format through
data mining. Section 3.4 describes the data collected and processed on the rolling mill
involving the temperature, roll loads, roll speeds etc. This corresponds to the Level 4 in
the decomposition. Section 3.5 describes the data obtained from inspection of bars and
how SVM technique was used to classify the different defects. This corresponds to the
Section 3.6 describes the FEM methodology which was used for creating the process
model (Level 3 in decomposition). Section 3.7 outlines how the data from sections 3.4-
There were three different kinds of data available for this study- data on rolling mill
(process settings and observed parameters), data on the surface inspection of bars and
data on the material of the bars. Section 3.4.1 describes the data on rolling mill, 3.4.2 on
material of bars and 3.4.3 on the surface inspection. Section 3.4.4 illustrates the
83
methodology of combining all these data in a single format (A database was created in
MS Access).
Data was available only from the furnace to the final bar in this study during the rolling
sub-process and was unavailable from continuous casting. There are numerous sensors at
each stage of the hot rolling sub-process which measure different variables on the billet
and the machines as it is being rolled like the temperature, rolling speeds, water spray
pressures etc. All these sensors take measurement every fixed unit of time (generally in
the order of 100 – 200 milliseconds). If a billet is passing through the sensor at that time,
it records the measurement for that billet, otherwise it records a 0. Data on the machines
this database. All the data is recorded and stored on a server in a relational database. Data
is recorded in order of 1-2 GB per day and about 500 different variables are recorded.
The speed of billet increases as it is being deformed in a rolling mill and it gets longer
and longer. Thus, for a particular billet, more measurements are taken by a sensor which
is located at the beginning of a mill and the number of observations per billet decreases as
it passes through the rolling process. To obtain the desired variables, a query is sent to the
database for variable names and the time interval required. The data is then transferred to
CSV format. A screen shot of the data is shown in Figure 25. Some of the product
characteristics like the temperature are measured by different sensors as the process is
carried on. By combining the measurement of all these sensors measuring the same
84
characteristic, it is possible to find the history of that variable throughout the process. A
ID TF T2 I2 G18
Figure 25: Multi Sensor data in CSV format (Note the values have been changed for
proprietary reasons)
85
3.4.2 Data on materials of bars
Due to the multiple configurations involving different materials, product sizes, shapes
and quantities being processed in the bar manufacturing process, a separate database
scheduling database. Each billet as it is formed from the continuous casting is given an
ID. This ID is carried forward to the rolling mill where more information is added to it
One of the most important quality parameters on a bar is the surface quality. Surface
quality of rolled bar affects its performance when it is used in downstream processes like
forging, drawing etc. The various defects which affect the surface quality are seams,
The inspection device used to measure these defects in this study is the HotEye® image
camera from O.G.Technologies [100]. Hot Eye™ system delivers high definition images
modulated with a unique signal, and then used for the illumination of the hot object. A
modified CCD camera is then used to take images of the object. These images can then
86
The approach of implementing imaging technology in bar mills includes the use of very
high-speed cameras and the proprietary HotEye™ technology. In order to process the
vast amount of image data in real-time with confined system costs, an ad hoc parallel
computing system is developed. Special optical design is adopted such that the images
taken from the non-flat surfaces are in high quality [104]. The camera records the images
and the related data on a server in real time which can later be downloaded. Data is
recorded at a rate of 1 GB/sec and stored in a relational database. The installation of the
camera, a sample image report and sample data set is shown in Figure 26 - Figure 29. The
data set consists of the length, width, longitudinal and transverse location of the defect
along with the severity. Based on the location and the dimension of the defect, it is
possible to find the type of defect (scale, crack etc.) at that location. Thus it is possible to
find the number of each kind of defect occurring on a bar and the spatial distribution of
HotEye Camera
87
Figure 27: HotEye Report
88
Width (W)
A
Minimum Enclosing
Area
Height (L)
Rolling
Direction
Defect A
Location (P)
Top
θ
Data from NTM mill was collected for a period of 6 months at the bar rolling mill. The
nominal diameter of incoming rod was 25.4 mm and that of the outgoing rod was 13.5
The main problems associated with data of such a magnitude and different forms are the
inconsistency in software and file formats. The data on rolling mill could only be
downloaded from the relational database in CSV format with a maximum of 60,000
observations and 20 variables in one file. Hence, for the same time period, in order to
89
record all 453 variables, 23 files were created. In all for the 962,134 records and 453
variables, 391 files were obtained. The data from the HotEye images was in MS Acess
format and was obtained in 1 single file. The data on materials was in Excel and Word
format.
Before any transformation or preprocessing could begin, we decided to convert all the
data into a single format. MS Access was chosen because of the ease in which data could
be imported into it and because it could handle all the data in study. The 391 files in CSV
format were converted into 1 single Access file, HotEye data was left as it was in the
Access format and the production and scheduling data was also converted into a single
Access file.
The file contained processing history only as time points. There is no barID available in
this database. Each of the variables Var1, Var2 etc. are time based history for a particular
bar. An example of couple of variables for few observations is shown in Figure 30. As
can be seen from the plots, there is a jump in values after certain time period. This
signifies the change from one bar to another. This feature of the data can be exploited to
work backwards from the inspection time to trace a particular bar. If a bar is observed by
automated inspection at time “t”, then based on observations by sensor just before the
inspection, the bar can be assigned the same barID. This can go iteratively backwards to
assign the barID on each set of values. Within each bar, a look at the trace of values can
90
Time plot for Variable 1 Time plot for Variable 2
BarID 1
2
2
BarID 1
1
Normalized variable
Normalized variable
BarID 2
1
0
0
-1
BarID 2
-1
-2
0 100 200 300 400 0 100 200 300 400
Time plot for Variable 1 for barID 1 Time plot for Variable 2 for barID 2
2
2
1
Normalized variable
Normalized variable
1
0
0
-1
-2
-1
0 10 20 30 40 50 60 0 5 10 15 20 25 30 35
Very initial data exploratory analysis showed a lot of noisy and missing data. The sensor
data had lot of missing values. This was due to the fact that the plant environment is very
harsh (high humidity, temperatures, dust etc.), which leads to malfunctioning of sensors
91
in between. To clear this issue, we decided to find outliers in this dataset. A sample
histogram of the subset of data from two of the sensors is shown in Figure 32.
Histogram of T1
Histogram of R12
250
150
200
100
150
Frequency
Frequency
100
50
50
0
0
-1 0 1 2 3 0 10 20 30 40 50
Based on domain knowledge, rules were formed to find a reasonable range of values of
the variables. For example, it is not possible to have the temperature of the billet at any
stage of the process to be more than 22000F or less than 15000F. All the values which fall
outside these ranges were removed. On an average about 4% of the values were found to
be outliers and hence the rest of the values were assumed to be correct.
A file of desired type can be represented as in Table 8. This file can be interpreted as
follows: each row depicts a single bar formed through the bar manufacturing process.
This row contains the information about the bar, its batch, its material and its processing
history through the process. It also contains all the quality attributes measured on that bar
92
by the different inspection techniques. To separate the different surface defects based on
the image data, a classification scheme was formulated. This is discussed in Section 3.5
.. … …. … … …. …. …. ….
To separate the different bars from one another and combine the information in a format
similar to table 2, an algorithm was devised and all the different files were merged
together into a single file with the format as specified in table 2. This file contained
information on about 10,144 billets with each billet occupying a row and had 475
variables for each of the billets. Each billet was now uniquely identified by its barID. In
this algorithm, we decided to take a 20% trimmed median of all the variables for the
(b) Ensure that a true representation of the population average is taken. Median is a
better statistic for the data than mean if the data is skewed or non symmetric.
93
3.5 Classification of Surface defect data
To classify the defect images into their defect types to completely fill the table 8, the
characteristics of the defects – their shape, and longitudinal and transverse location are
used for this purpose. The new transformed variables obtained from the original variable
set are:
3. Transverse location on the bar as cosine of the angle with the parting line (called
TL in the study)
Another variable directly used from the camera is the severity of the defect. This number
represents the average intensity of the pixels in the defect image. The scale of this
measurement is from 0-100 where 0 means very low intensity and 100 means maximum
possible intensity.
To understand the procedure for calculating these transformed variables, let us look at a
bar geometry and how the original variables are measured by the camera (Figure 29).
=L/W
94
Therefore 0 ≤ LL ≤ 100.
3) TL = cos θ
= (Radius of bar - Left Box) / sqrt( (Radius - Left Box)2 + Top Box2)
Therefore -1 ≤ LL ≤ +1
The sign of TL (+ve or –ve) is associated with the camera number. Readings by Camera
1 and 3 are +ve while those by 2 and 4 are negative. The characteristics of different
corners
circumference
length
95
Some of the images from the camera for each kind of defect along with the variables as
observed by the camera and the transformed variables are shown in Figure 33 - Figure 37.
Rolling
Direction
Figure 33: Different seam defects for the training dataset (a) Seam1 (b) Seam2 (c) Seam3
Rolling
Direction
Figure 34: Different longitudinal crack defects for the training dataset (a) Crack1 (b)
96
Rolling
Direction
Figure 35: Different scale defects for the training dataset (a) Scale1 (b) Scale2 (c) Scale3
Rolling
Direction
Figure 36: Transverse cracks - corner (TCC) defects for the training dataset (a) TCC1 (b)
97
Rolling
Direction
Figure 37: Transverse cracks – mid surface (TCMS) defects for the training dataset (a)
Thus the methodology of calculating the defect classes can be summarised as in Figure
Figure 38. A flowchart of the complete classification process from the images is shown in
Figure 39.
A look at the Table 10 - Table 14 gives some distinct features for the different defects:
a) Seams and Longitudinal cracks have LWR >4 and all others have LWR <4
98
b) Seams and Transverse corner cracks have 0.9 < | TL | < 1.0
ID (%)
ID (%)
99
Defect Height Width Top Left Location LWR LL TL Severity
ID (%)
ID (%)
ID (%)
100
Figure 39: Flowchart for classification using PK-MSVM
To form a good training set for the purpose of classification, 600 rolled bars of 5 different
materials were randomly selected over a period of 5 months. The HotEye® images of
these 600 bars along with the manual inspection by plant QC personnel were collected
101
and a dataset was formed by combining both these information. There were a total of
20,140 defects which were identified on these bars. This data set was divided into 15%
for the training and 85% for testing. Five-fold cross validation was used in the test
dataset.
There are many approaches for checking the accuracy of predictions from classification.
We are going to look at the confusion matrix from the classification at each stage and
calculate the error rate of classification from there. A typical confusion matrix (CM) for
Table 15.
Class Surface
Mid Surface
Corner
other words all the defects are equally important, but this can be changed to reflect
= 1 - sigma(Aii) / sigma(Aij)
For the implementation of the various SVMs, statistical software package „R‟ was used.
Three different techniques of implementing SVM are described in Agarwal et. al [105]. It
is shown that Hastie‟s algorithm performs the best for classification of defect data.
Recently, Lee et. al [106] have devised a technique for classification using multi class
complete solution path of the MSVM as a function of its tuning parameter. Lee‟s
algorithm, on the other hand is based on extending the binary hinge loss to the multi
class problem
2. Lee‟s algorithm relies on solving the dual and primal formulation of the quadratic
programming equation for the SVM based on multiple hinge loss, while Hastie‟s
algorithm relies on only one time initialization of the quadratic programming solver if
103
3. In Lee‟s algorithm, they assume the binary SVM as a special case and retain the same
problem structure. Hastie‟s algorithm on the other hand utilizes the idea that the SVM
The data set was classified using Lee‟s algorithm and its error rates were compared with
the previously best performing algorithm, e.g Hastie‟s algorithm. Table 16 - Table 19
show the results obtained from classification using Lee‟s algorithm (LL) and Hastie‟s
algorithm (HA).
104
Actual Defect Seam Longitudinal Scale Transverse Transverse
Class
Crack
Crack – Mid
Surface
Crack – Corner
Table 18:Confusion matrix(CM) of classification from MSVM (value on left are from
105
Actual Defect Class Seam Longitudina Scale Transverse Transverse
Mid Surface
Corner
Conclusions:
variables
3. The confusion matrix of PK-MSVM shows that the misclassification is reduced for
all the categories with LL algorithm. The 2 categories of Transverse cracks which had
106
The next section describes how process models were created for the bar manufacturing
process.
To understand the physical phenomena causing the formation of different surface defects,
it is necessary to formulate physical models for the process. In this case, finite element
method (FEM) based models were used to simulate the rolling process. Two different
software were used: ROLPAS for the roughing mill and FORGE3 for the NTM. Three
different grades of steel were used. Their chemical composition and the flow stress for
Material C Mn P S Si
107
Figure 40: Flow stress curves for three grades at strain rate of 1/s and different
temperatures
The simulation of the roughing mill with eight passes is shown in Figure 41. The output
from this simulation includes the stresses, strains, strain rates and temperature on billet
after each pass. The roll loads are also calculated at each pass.
108
Figure 41: FEM simulation of the roughing mill using ROLPAS
The FEM model of the NTM mill was created in FORGE3. The set up of the model is
shown in Figure 42. The model consists of 8 rolls in a configuration of four 2-roll stands.
The rolls are oriented alternating in horizontal and vertical plane to the incoming rod.
Each stand also has alternating round and oval pass design for effectively controlling the
dimension. The details of the roll pass design are not provided here due to proprietary
concerns.
109
Figure 42: FEM model of the NTM mill
The flow stress of the materials was modeled as Hansel Spittel rheology law in the form:
…………………Equation 45
…………………Equation 45 are
shown in Table 21. The rolls were modeled as rigid. The friction between the rolls and
the billet was taken to be Coulomb friction with a value of µ = 0.15. The heat transfer
coefficient between the billet and the rolls was taken to be 2.17 x 10-2 W/m2. The billet
was meshed with quadrilateral elements with a mesh size of 2 mm. The complete billet
110
Parameter Value for Value for Value for
Material A Material B Material C
A 2.173 x 105 2.62 x 105 2.13 x 105
m1 -0.00269 -0.0029 -0.00259
m2 -0.12651 -0.14227 -0.05216
m3 0.14542 0.14102 0.1378
m4 -0.05957 -0.05764 -0.0311
m5-m9 0 0 0
To calibrate and validate the ROLPAS model, the temperature of the billet at the exit of
each pass and the roll loads at each of the passes is compared against the plant data. The
friction factor and heat transfer coefficient is adjusted till the values match the real plant
data. These calibrated models are then used for the model building.
The FORGE3 model was validated with the help of plant data. Simulations were done for
all the three materials and the out coming diameter and ovality was evaluated with the
plant measurements. The profile of the rod from the FEM simulation and the
111
Dmin = Minimum Dia.
Figure 43: Output profile of rod from FEM simulation and the calculation of ovality
The results of validation are shown in Figure 44. The results show that the numerical
model correlates well with the actual plant data and can be used for process design. Since
it is not possible to simulate the roll bite automatically in FEM at the start of the
simulation, a manipulator is created at the front end of the billet which gives an initial
velocity to the rod. The initial velocity is equal to the velocity of the roll at the roll end (V
0.25
Material A -
0.2 Measured
Ovality (mm)
Material B -
0.15 Measured
Material C -
0.1 Measured
Material A -
0.05
Simulated
0 Material B -
Simulated
1 2 3
Material C -
Case Number Simulated
112
3.6.2 Model Results
After the model calibration and validation, various simulations were run to represent the
different materials and process settings. The following data was extracted for these
simulations:
c) Strain
d) Strain rate
Note that all these variables are extracted at each of the 8 roughing passes. Hence for
σy
σx
Figure 45: Stresses during the rolling process illustrated by FEM model
113
3.7 Bayesian Hierarchical Model for surface defects
Based on the hierarchical decomposition shown in Figure 22, the posterior distribution
identified in Equation 44and the data available from both the plant and the process
models (Sections 3.4 – 3.6), the joint PDF for the surface defects model is written as:
1) S: SSeam, SScale, STC, SLC : The defects of each kind on a billet. Note that since one
2) I: Strain (ε), Strain Rate (έ), Stress in rolling direction (ζy), and Stress in transverse
3) F: Chemistry (C, Mn, P, Sul, Si, V), Temperature (T), Roll Load (L), Roll Speed
(RS), Gauge Position Variable (G), Gauge Tracking system (X) for each pass i, where
i=1, 2, ….., 8
a) DSeam, DScale, DTC, DLC : Number of defects of a particular kind. Since there is a
dominant defect in one material, model is built for each of these defects
separately.
b) Dεi, Dέi, Dζyi, Dζxi : Data on strain, strain rate and stresses for each pass, i=1, 2,
….., 8. As described in Section 3.6, the roughing mill had 8 passes. Hence, the
strain, strain rate and stresses can be obtained for each of these passes.
c) DTi, DLi, DRSi, DGi, DXi : Data on temperature, roll load, roll speed, gauge position
114
d) DC, DMn, DP, DSul, DSi, DV: Data on the elemental composition of the material
Since the NTM is a separate section of the bar rolling, the defects (ovality in this case)
originating from NTM are modeled separately. For the NTM mill, the following variables
are used:
1) S: SOvality
2) I: Incoming ovality (In), Strain (ε), Strain Rate (έ), Stress in rolling direction (ζy), and
Stress in transverse direction (ζx) for each pass j, where j=1, 2, 3, 4. As shown in
Section 3.6, the NTM has 4 alternating passes of horizontal and vertical
configuration.
3) F: Temperature (T), Roll Load (L), Roll Speed (RS), for each pass j, where j=1, 2, 3,
b) Dεi, Dέi, Dζyi, Dζxi : Data on strain, strain rate and stresses for each pass, j=1, .., 4
c) DTi, DLi, DRSi : Data on temperature, roll load, roll speed for each pass, j = 1, …, 4
Since each defect model can be created separately, we will demonstrate the data model
for seams here and the models for other defects can be built in a similar manner.
115
[DSeam, Dεi, Dέi, Dζyi, Dζxi, DTi, DLi, DRSi, DGi, DXi, DC, DMn, DP, DSul, DSi, DV | SSeam, ε, έ,
The basic idea for hierarchical modeling is that given an ordering of variables in a
model. Using this idea, we will try to simplify the conditional structure in the above
1) First assumption we make is that the data on the strain given the strain rate is
strain data to other process variables like strain rate, temperature comes from the
2) The data on the process parameters given the parameter itself is conditionally
independent of other process variables. Thus the data on Rolling speed given the
rolling speed process is conditionally independent of the data on gauge position and
so on.
Based on the abovementioned assumptions, the data model in Equation 46 is rewritten as:
[DSeam, Dεi, Dέi, Dζyi, Dζxi, DTi, DLi, DRSi, DGi, DXi, DC, DMn, DP, DSul, DSi, DV | SSeam, ε, έ,
[DSeam | SSeam ] [Dεi | ε] [Dέi | έ] [Dζyi | ζY] [Dζxi | ζX] [DTi | T] [DLi | L] [DRS | RS] [DGi | G]
[DXi | X] [DC | C] [DMn | Mn] [DP | P] [DSul | Sul] [DSi | Si] [DV | V]…………..Equation 47
116
From the exploratory data analysis, these different data variables look normal distributed
in nature. Hence, we assume a prior distribution which is normal for each of these
And so on.
Here i = 1, 2, …., 8 represents the different passes of the roughing mill and
The data model captures the uncertainty inherent in the observation of processes
imperfectly (measurement error, location error, sampling error etc.). The variance in each
of these conditionally independent subsets represents these uncertainties for each of the
by the HotEye camera, ζ2εirepresents the computational error due to FEM modeling etc.
We can create the process models in two different ways. In one method we can consider
all the defects together and then create a common process model containing all the
different defects. In another method, we can select each of the defects one by one and
then create process models for them individually. The second method is simpler and can
117
be understood in a slightly easy manner, hence in this section we will build the process
As described in section 3.2.1.1, seams are underfill or overfill of material during the
rolling and are caused due to strain and strain rates during the process. The strains and
strain rates are influenced by the temperature, roll load, roll speed, gauge tracking and
gauge postion. Hence the process model for seams is written as:
[S, I, F] = [S | I, F] [I | F] [F]
1) The physical mechanism which forms the seams is the temperature, strain and strain
rate during the rolling. This is relevant based on previous observations by researchers
in which it is seen that the seams are formed because of the metal flow during the
roughing passes.
2) The temperature, strain and strain rate are dependent on the rolling speed, gauge
location, gauge tracking and the roll load. This assumption comes directly from the
theory of plasticity.
Therefore,
118
[SSeams, I, F] = [SSeams | ε, έ, T] [ε, έ, T | RS, G, X, L, C, Mn, P, Sul, Si, V] [L, RS, G, X,
3) Given the chemistry, speed, gauge location, gauge tracking and the load; the strains,
strain rates and temperature are conditionally independent. The strain and strain rate
depends on the rolling schedule and the speed. Hence, they are conditionally
Therefore,
4) The roll load, gauge location and gauge tracking are conditionally independent of
each other given the roll speed. The roll load depends on the strain and strain rate
which are dependent on the speed and rolling schedule. Similarly, the gauge location
and tracking are dependent on the rolling speed. The chemistry is independent of
everything else as it is dependent on the continuous casting process and not the
rolling process.
Therefore,
P, Sul, Si, V] [L | RS] [G | RS] [X | RS] [RS] [C] [Mn] [Sul] [P] [Si] [V]………Equation
55
119
For each of these conditionally independent subsets we assume linear relationship which
relates the response to the exploratory variables via regression parameters. Thus:
[SSeam(k) | εk, έk, Tk] ~ N (µSSeam + (β11 εk1 + β12 εk2 + … + β18 εk8 + β21 έk1+ β22 έk2+ …… +
[εk | RSk, Gk, Xk, Lk] ~ N (µε+(β41 RSk1 + ….. + β48 RSk8 + β51 Gk1+ … + β61 Xk1+ …. +
[έk | RSk, Gk, Xk, Lk] ~ N (µέ+ (β81 RSk1 + …… + β91 Gk1+ ……. + β101 Xk1+ β111 Lk1) ,
ζ2Sέ) ………………Equation 58
[Tk | RSk, Gk, Xk, Lk Chemistryk] ~ N (µT+ (α31 Ck + ….. + α36 Vk + β121 RSk1 + ……. +
[Lk | RSk] ~ N (µL + β161 RSk1+ ….. + β168 RSk8, ζ2SL) ………Equation 60
[Gk | RSk] ~ N (µG + β171 RSk1+ ….. + β178 RSk8, ζ2SG) ……..Equation 61
[Xk | RSk] ~ N (µX + β181 RSk1 + ….. + β188 RSk8, ζ2SX) …...Equation 62
For each of these relationships, polynomial and other complicated relationships can be set
up, but for simplicity purposes and the accuracy of results obtained from these simple
Also note that the superscripts 1, 2…., 8 represent the eight passes of the roughing mill.
120
As described in section 3.2.1.4, LC is mainly formed due to the stresses in the
longitudinal direction during the rolling and the temperature of the billet (which affects
the flow stress). Hence the process model for longitudinal cracks is written as:
[S, I, F] = [S | I, F] [I | F] [F]
= [SLC | ζY, T, L, RS, G, X, C, Mn, P, Sul, Si, V] [ζY, T | L, RS, G, X, C, Mn, P, Sul, Si,
Again, the Bayesian idea is that given this ordering of variables, we can assume
1) The physical mechanism which forms the LC is the temperature and the longitudinal
researchers in which it is seen that the LC are formed because of the metal flow
2) The temperature and the stresses are dependent on the rolling speed, gauge location,
gauge tracking and the roll load. This assumption comes directly from the theory of
plasticity.
Therefore,
[SLC, I, F] = [SLC | ζY, T] [ζY, T | RS, G, X, L, C, Mn, P, Sul, Si, V] [L, RS, G, X, C, Mn,
121
3) Given the speed, gauge location, gauge tracking and the load; the stresses and
schedule, chemistry and the speed. Hence, they are conditionally independent given
these parameters.
Therefore,
[SLC, I, F] = [SLC | ζY, T] [ζY | RS, G, X, L, C, Mn, P, Sul, Si, V] [T | RS, G, X, L, C, Mn,
4) The roll load, gauge location and gauge tracking are conditionally independent of
each other given the roll speed. Similarly, the gauge location and tracking are
Therefore,
[SLC, I, F] = [SLC | ζY, T] [ζY | RS, G, X, L, C, Mn, P, Sul, Si, V] [T | RS, G, X, L, C, Mn,
P, Sul, Si, V] [L | RS] [G | RS] [X | RS] [RS] [C] [Mn] [P] [Si] [Sul] [V] …Equation 67
The linear relationships are now assumed to be similar in structure to those for seams
except that the strains and strain rates are replaced by the stresses.
The above specified model was created in WinBugs. Models were created for different
surface defects and different materials and the data from the ROLPAS simulations and
the plant data were used as inputs to the models. WinBugs follows the MCMC
122
simulations based on the general idea presented by Gelfand and Smith. The data on bars
was randomized and 80% of the data was used for model building and the remaining 20%
was left for finding the prediction errors. Different starting values for the chain were
simulated.
The results of the MCMC simulations for various parameters of the model described
above for the seams for Material A are shown in Figure 46: Results from MCMC
Simulation.
'alpha' 'mu[1,1]'
0.020
0.003
0.010
0.000
0.000
Figure 46: Results from MCMC Simulation for parameters β12 and β23for the seams
'mu[1,2]' 'mu[1,3]'
model for Material A
0.030
0.008
0.015
0.000
0.000
280 300 320 340 360 380 250 300 350 400 450 500
123
'mu[1,4]' 'mu[1,5]'
0.006
0.010
3.7.4 Model Validation
The model was developed in section 3.7.3 with 80% of the data and the remaining 20%
of the data was used for model validation. About 300 bars of each type were selected for
this analysis. The results of model validation for Materials A, B and C for the different
surface defects is shown in Figure 47. The values of the actual defects were scaled to be
124
Material A - Seams
1
Histogram of err * 100
0.8
14
Predicted
12
0.6
10
0.4
Frequency
8
0.2
6
4
0
2
0 0.2 0.4 0.6 0.8 1
Actual
0
-10 -5 0 5 10 15
err * 100
0.8
10
Predicted
0.6
8
0.4
Frequency
6
0.2
4
0
2
Actual
0
-6 -4 -2 0 2 4 6
err * 100
Material C - Ovality
1 Histogram of err * 100
10
0.8
Predicted
0.6
0.4
6
Frequency
0.2
4
0
2
Actual
-10 -5 0 5 10
err * 100
Figure 47: Results of model validation and error percentages for the models
125
3.8 Bayesian Networks for surface defects
directed acyclic graph is drawn for the different materials and surface defects. The arrows
represent the conditional independence structure of the model and the values represent
the median with the 95% credible intervals for the parameters linking the variables. The
regression parameters are all scaled to be between [0, 1] so that easy comparison can be
made.
For the sake of simplicity, only the variables for which the parameters do not contain a 0
in their credible intervals are shown in these graphs. A representative graph is shown in
Figure 48.
Seam
ε2 ε3 έ2 έ3 T2
G2 X2 L1 L2 RS1 RS2 C Mn Si
126
Based on this simplified graph, some of the observations which can be made are:
1) The numbers of seam defects in Material A are influenced mainly by the strain and
2) The other roughing passes do not play a major role in causing the formation of seam
defects in this material. Thus, the material flow in the first two passes determines
3) The temperature in roll stand 2 also influences the formation of seams. This
temperature probably affects the metal flow in the second and third roll stand and
4) The strain and strain rates in the first two stands are affected by the gauge location
and gauge tracking. Thus the location of the billet inside the roll stand is critical to the
formation of seams. If the billet is not centered properly inside the box or oval pass, it
5) The chemistry of the material (C, Mn and Si in this case) affects the temperature of
the rod during the rolling. This is due to the fact that during the melting and
continuous casting, the alloying elements vary with different heats. This variation
causes a difference in flow stress of the rods during the rolling which affects the
formation of seams.
127
CHAPTER 4
4.1 Introduction
The hot strip mill (HSM) involves a series of processing steps to convert the material and
mechanical state of the cast slab into that of the hot rolled steel coil. In the hot rolling of
steel, steel is melted with different alloying elements like Si, V, Cr and Mo, and poured
into a continuous caster with multiple strands. The slabs produced by the caster then go
hot rolling mill before being coiled. The rolled strips are sent to the customer for further
processing into sheets for autobodies, sheets and piling for railways etc. (Figure 49)
Run Out
Table
of microstructures and mechanical properties like Ultimate Tensile Strength (UTS), Yield
Strength (YS) and Elongation that are possible when it is subjected to controlled
processing and heat treatments. The properties of rolled steel products used in the various
applications depend upon the composition of the material, deformation history in the
rolling mill and the transformations after rolling. It is these properties which determine
the use of the sheets. Due to high pressures of cost reductions, it is necessary to predict
these properties as the coils are being rolled so that testing can be minimized and controls
can be improved.
Section 4.2 describes the different modeling techniques used in the past for prediction of
hierarchical decomposition for mechanical properties. Section 4.4 discusses the data
model and how data from the plant was collected and integrated into a single format.
Section 4.5 describes the physical model used for the hierarchical model and section 4.6
4.2 Background
Lots of research has been done in the last 60 years in the development of methodology,
relationships and prediction of mechanical properties in hot strip rolling. Most of it has
Table 22.
129
Model Type Basic Features Prediction
Errors
Both these methods have serious shortcomings. The mathematical models try to use
laboratory experiments and the models developed there in the actual mills. This results in
poor predictions (with errors up to ±20% in some cases). To reduce these errors,
researchers are trying to introduce more complexity into the models by adding more
this results in slight improvement of predictions (the errors are still very high), but the
models now become very complicated. Another shortcoming of this modeling approach
is that one parameter is changed at a time during the experimentation and its effect is
looked at the final properties (for example, effect of alloying element like silicon, effect
of cooling rate in ROT etc.). The interactions between the different processing steps are
130
ignored in the experimentation and this causes the errors in the predictions. Due to lack of
understanding of these interactions, no decision about the controls and design can be
made. It is important to note that the empirical relationships derived from experiments are
just statistical fits to the data for the particular material and conditions for which the
experiment was done. It does not reflect the true behavior of the material under the actual
Empirical models, although, have a better prediction than the mathematical models, but is
really black-box in nature. The models based mainly on neural network type non linear
regression give no idea about the interactions between the different processes and
mechanical properties. These models require huge amounts of data for training and
predictions. Moreover, every time there is a slight change in the mill setup (new
machines, machines degrading over time), the models need to be reformulated for the
predictions.
To overcome the main disadvantages of the physical and statistical models and to design
the bar manufacturing process for reduced defects, the hierarchical decomposition of the
process was done (Figure 50). This decomposition is based on the same principles as
The mechanical properties are decomposed into the different kinds: yield strength (YS),
ultimate tensile strength (UTS) and elongation. These are then decomposed into the
131
physical quantities affecting them (eg. stress, strain etc.). The physical quantities are
decomposed into the design parameters on the rolling mill which are connected to the
individual sub-processes.
Design
Level 3: Process Strain Heat
Stress Strain Grain Size
Physics Rate Transfer
Hierarchical Model discussed earlier. Let us denote the different levels in the
decomposition as follows:
132
[M, I, F] = [M | I, F] [I | F] [F] ..………………………………………Equation 68
Our aim is to find this joint distribution based on the observed data D. However, we
would like to use the Bayesian framework to learn about the underlying variables and
parameters. Using the likelihood in Equation 68 and the prior distribution on process
Section 4.4 discusses the data collection and integration necessary for creating a data
model. This section describes how the data was collected for calibrating and validating
the model and methods to convert this raw data into a usable format through data mining.
Section 4.4.2 describes the data collected and processed on the rolling mill involving the
temperature, roll loads, roll speeds etc. This corresponds to the Level 4 in the
decomposition. Section 4.4.3 describes the data obtained from inspection of sheets. This
Section 4.5 describes the process modeling methodology which was used for creating the
process model (Level 3 in decomposition). Section 4.6 outlines how the data from
sections 4.4 - 4.5 was used for creating the Bayesian Hierarchical model.
133
4.4 Data Collection and Integration
The framework described in this study was developed at a hot strip mill. This is a 2000
mm (78.74”) wide continuous mill that has three 100 ton/hr continuous pusher type
reheat furnaces, five 4-high roughing stands, seven 4-high finishing stands and three
pneumatic coilers. The Roughing Mill (RM) has one vertical stand for slab control and
sizing, one 2-high stand and four 4-high universal stands with edgers and roller tables
with hydraulically operated side guards. The Finishing Mill (FM) has seven stands in
tandem with hydraulic AGC in the last 4 stands. It has online profile and flatness
measurement system.
The Run-Out-Table (ROT) has 200 m length with 554 rollers set 360 mm apart. Its
maximum rolling speed is 19 m/s. The laminar cooling system (LCS) has 74 m cooling
zone with 13+13 banks with 4+16 headers per bank, 76+38 nozzles per header, 450oC
minimum coiling temperature and ± 15oC accuracy. The quality and dimensional
parameters are as follows: thickness tolerance (±) 0.025 to 0.05 mm, width tolerance -0.0
to +10 mm, strip crown 20 to 50μm, finishing temperature ±15oC (95% strip length),
coiling temperature ±18oC (95% strip length), flatness 5 mm/m for > 4.0 mm strip, coil
telescopicity within 25 mm, and coiler idle time 5 s for 2 coiler operation.
A modern hot strip mill has multiple sensors and data acquisition systems installed for
monitoring and control. These are the chemical sensors measuring the chemical
different stages, load cells measuring the roll loads and torques and optical gages
134
measuring the thicknesses and width along the rolling process (Figure 51). The sensor
- Massive Size: Each sensor records observations every few seconds. The typical size
of the database with all the readings is usually 10-20 GB per day.
- Spatiotemporal in nature: The data collected by these sensors is both time based and
location specific.
- Noisy: These sensors are located in an environment which has high temperatures and
dust. All these result in a lot of noise in the data which needs to be cleaned and
This recorded data is then transferred to a central computer which stores it in a database
format. The format of this data varies from mill to mill based on the automation system,
but the basic features remain the same. Different parts of the mill generally use different
databases to store this information. For example, in the steel melting shop the chemical
composition is usually stored in a heat by heat manner. On the other hand, in the rolling
mill the information is stored on a coil by coil basis. It is therefore necessary to combine
all the information from these various sources into a common format which can be used
for the development of the framework. In the previous work done by researchers using
plant data, they have pre-selected some of these variables in their prediction models. This
practice introduces bias in the modeling and we can miss some of the important
collected all the available data from the mill before starting any model building activity.
135
Figure 51: Sensors in a hot strip mill and sensor data fusion
All the data available from the sensor fusion was transferred in a text file. Each row of
the text file represented one coil and the different columns represented the different
process settings and variables. In total there were 150 variables in the file including:
1. Chemistry (Percentage of C, Mn, Si, S, P, Nb, Ti, V, Al, Cr, Mo, Ni, Cu)
7. Rolling speeds, load and torque at each stand of the finishing mill
9. Temperatures at the front and tail of the coil at each finishing stand
136
10. Temperature at the entry and exit of run out table
A sample file is shown in Table 23. Note that the temperatures, loads etc. were averaged
by the data acquisition system before storing in the database. Hence, within coil
Data over a period of 2 years for many different materials was collected. Since the data
set obtained was large enough for analysis, it was decided to remove all the coils which
either had missing data value or had a data value which was an outlier. All these two
cleaning steps were completed; three materials having the most information were selected
137
for further analysis and study. Two of these materials were C-Mn steels and one was
C Mn Si P Al Nb Ti V Cr Cu
Material 0.05- 0.3- 0.1- 0.01- 0.01- Trace Trace Trace 0.005- 0.01-
Material 0.05- 0.1- 0.05- 0.01- 0.02- Trace Trace Trace 0.005- 0.01-
Material 0.1- 0.5- 0.1- 0.01- 0.01- 0.01- 0.01- 0.01- 0.005- 0.01-
C 0.25 1.5 0.2 0.04 0.05 0.2 0.1 0.1 0.015 0.02
Table 24: Chemistry of the three materials in this study (percentage by volume)
Before any model is built, it is necessary to look at the data to see the kind of distribution
it follows. The first analysis which was done was to look at the distribution of different
alloying elements in the different heats of Material A. The summary statistics are shown
in Table 25.
138
Alloying Min. 1st Quantile Median Mean 3rd Quantile Max.
Element
Next, the mechanical properties for the different sheets were explored and their variation
was looked into. The summary statistics for the mechanical properties for Material A are
shown in Table 26. The histograms and the density curves for UTS and YS are shown in
Figure 52. This figure shows that the properties follow close to a Gaussian distribution.
139
Property Min. 1st Quantile Median Mean 3rd Quantile Max.
Histogram and Density of UTS for Material A Histogram and Density of YS for Material A
0.020
0.020
0.015
0.015
Density
Density
0.010
0.010
0.005
0.005
0.000
0.000
Figure 52: Histograms and density curves of mechanical properties for Material A
Some of the different process parameters were also explored to look at their variation
(Figure 53).
140
Histogram and Density of Torque at Stand 8
Histogram and Density of Temperature at Stand 10
0.008
0.07
0.06
0.006
0.05
0.04
Density
0.004
Density
0.03
0.002
0.02
0.01
0.000
0.00
930 940 950 960 970 350 400 450 500 550 600 650 700
Figure 53: Histograms and density curves for process parameters for Material A
Figure 54: Schematic of the deformation zone in hot rolling of flat products
141
Before the hierarchical model can be built, it is necessary to convert the process
accomplish this, empirical models of the sheet rolling process were used.
To study the sheet manufacturing process empirically, the following assumptions are
made:
Roll flattening does not occur in the arc of contact (rolls can be assumed to be rigid).
Material does not undergo work-hardening during its passage between the rolls.
Compression rate from the point along the arc of contact does not have any effect on
To accurately predict the grain size, phase transformation and mechanical properties
during the sheet manufacturing process, the two most important parameters to be
calculated during the rolling are the strains and strain rate at each pass. The calculation of
142
Strain rate (έ) = , given by Wusatowski […] …………Equation 71
r = (h1-h2) / h1
Q = activation energy for plastic deformation which has been determined by experiments
T = temperature in K
C, α, n = Material constants
There is no direct way to measure the parameter Z. Hence an indirect method is used to
calculate this. The roll load is measured with the sensors and it can be defined per unit
width as
P = ζLQp ………………….Equation 73
Therefore,
143
P2 = ζMQp ……………………………….Equation 75
Using the Equation 77, legacy data is used to find the constants C, α and n for different
materials using least square regression and Z is then calculated as each coil is rolled. The
Table 27. The values in parenthesis for the constants represent their standard errors. The
variation in the material constants is due to the variation in the chemical composition
Material Constants
C α n
Table 27: Material properties for three different materials considered in this research
144
The plots depicting the variation in loads and torque in the finishing mill due to the
variation in chemistry and the resulting flow stress for material A are shown in Figure 55.
Two different sheet thicknesses of 5 mm and 16 mm are shown to illustrate the difference
Figure 55: Variation in roll loads and torque due to the variation in chemistry of material
Heat transfer plays a significant role in the hot rolling process as the temperature affects
both the sheet and the rolls. The two areas during sheet manufacturing where heat
transfer plays a role in determining the mechanical properties are the rolling (roughing
and finishing) and the run out table. The different mechanisms by which heat transfer
To understand and decompose these mechanisms, the temperature history of the sheet
was explored. The median temperatures at the head and tail of the sheet during the
finishing passes (both at entry to a pass and its exit) are shown in Figure 56. This plot
shows that there is little temperature drop during the rolling process itself due to an offset
between the deformation heating and the roll chilling, while most of the heat loss occurs
between the head and tail of the sheet. Similar trends were observed during the roughing
passes. Hence the primary mechanism of heat transfer during the roughing and finishing
This heat transfer in practice depends on the temperature of water being sprayed, the
pressure at which it is being sprayed and the number of nozzles. However, to model these
phenomena is a complex procedure. Moreover, the settings on the mill usually remain
constant and hence if an estimate of the heat transfer coefficient can be found, that can be
used for the model building. The most common form of calculating this heat transfer
146
where,
v = strip speed
t = strip thickness
The header flow rate and is assumed to be constant and hence can be combined with the
coefficient K. Also, the strip speed remains the same for a particular thickness. Hence it
Therefore,
h = K tb (T1-T2)c …………………………..Equation 79
The Equation 79 was used for the different materials and the constants K, b and c were
determined. The assumption made here is that this heat transfer coefficient is same for all
the passes.
147
Figure 56: Median temperatures for material A during the finish rolling process
Based on the hierarchical decomposition shown in Figure 50, the posterior distribution
identified in Equation 68 and the data available from both the plant and the process
models (sections 4.4 – 4.5), the joint PDF for the mechanical properties model is written
as:
2) I: Strain (ε), strain rate (έ) and Heat Transfer (h) for each pass i = 1, 2, …. 8, Furnace
Roll Speed (RS), Roll Torque (RT), Temperature at each stand (T1, T2, …., T8),
f) Dεi, Dέi, Dhi : Data on strain, strain rate and heat transfer for each pass, i=1, 2, …..,
h) DTF, DLi, DRSi, DRTi, DT1i : Data on temperature, roll load, roll speed, roll torque,
ROT Temperature and Coiling Temperature (Either for each pass, i = 1, 2, ….., 8
or one value)
i) DC, DMn, DP, DSul, DSi, DV: Data on the elemental composition of the material
The data model for each of the mechanical properties can be built separately. The data
[DYS, Dεi, Dέi, Dhi, Df, DR, DLi, DRSi, DRTi, DT1i, DC, DMn, DP, DSul, DSi, DV | MYS, ε, έ, h, f,
Using the same idea as discussed in Chapter 3, we will try to simplify the conditional
149
1. First assumption we make is that the data on the strain given the strain rate is
data to other process variables like strain rate, temperature comes from the process
2. The data on the process parameters given the parameter itself is conditionally
independent of other process variables. Thus the data on rolling speed given the
rolling speed process is conditionally independent of the data on rolling torque and so
on.
Based on the abovementioned assumptions, the data model in [DSeam, Dεi, Dέi, Dζyi,
Dζxi, DTi, DLi, DRSi, DGi, DXi, DC, DMn, DP, DSul, DSi, DV | SSeam, ε, έ, ζY, ζX,
[DYS, Dεi, Dέi, Dhi, Df, DR, DLi, DRSi, DRTi, DT1i, DC, DMn, DP, DSul, DSi, DV | MYS, ε, έ, h, f,
[DYS | MYS ] [Dεi | ε] [Dέi | έ] [Dhi | h] [DR | R] [DLi | L] [DRS | RS] [DRTi | RT] [DT1i | T1]
[DC | C] [DMn | Mn] [DP | P] [DSul | Sul] [DSi | Si] [DV | V]………………..Equation 81
From the exploratory data analysis, these different data variables look normal distributed
in nature. Hence, we assume a prior distribution which is normal for each of these
150
[Dέi(k)| έi(k)] ~ N (έi(k), ζ2έi)………………………………….Equation 84
And so on.
Here i = 1, 2, …., 8 because there are eight passes in the rolling mill and
The data model captures the uncertainty inherent in the observation of processes
imperfectly (measurement error, location error, sampling error etc.). The variance in each
of these distributions represents these uncertainties for each of the variables. For
example, ζ2YSis the error in the testing of the coupons to obtain the yield strength and
We can create the process models in two different ways. In one method we can consider
all the properties together and then create a common process model containing all the
different properties. In another method, we can select each of the properties one by one
and then create process models for them individually. The second method is simpler and
can be understood in a slightly easy manner, hence in this section we will build the
Using the hierarchical decomposition, the process model for YS can be written as:
151
[M, I, F] = [M | I, F] [I | F] [F]
= [MYS | ε, έ, h, f, R, TF, L, RS, RT, T1, C, Mn, P, Sul, Si, V] [ε, έ, h, f, R| TF, L, RS,
RT, T1, C, Mn, P, Sul, Si, V] [TF, L, RS, RT, T1, C, Mn, P, Sul, Si, V]….Equation 86
Again, the Bayesian idea is that given this ordering of variables, we can assume
1) The physical mechanism which determines the YS is the heat transfer, strain,
strain rate, furnace and the ROT during the rolling. This is relevant based on
2) The heat transfer, furnace, ROT, strain and strain rate are dependent on the rolling
speed, rolling torque, temperature and the roll load. This assumption comes
Therefore,
[MYS, I, F] = [MYS | ε, έ, h, f, R] [ε, έ, h, f, R | TF, L, RS, RT, T1, …, T8, C, Mn, P, Sul,
Si, V] [TF, L, RS, RT, T1, …, T8, C, Mn, P, Sul, Si, V]……..Equation 87
3) Given the chemistry, speed, load, torque and the temperatures; the strains, strain
rates, heat transfer, furnace and ROT are conditionally independent. The strain
and strain rate depends on the rolling schedule and the speed and not on chemistry
Therefore,
152
[MYS, I, F] = [MYS | ε, έ, h, f, R] [ε | L, RS, RT] [έ | L, RS, RT] [h | TF, L, RS, RT, T1,
…, T8, C, Mn, P, Sul, Si, V] [f | TF] [TF, L, RS, RT, T1, …, T8, C, Mn, P, Sul, Si, V]
………Equation 88
4) The roll load, torque and temperature are conditionally independent of each other
given the roll speed. The roll load depends on the strain and strain rate which are
dependent on the speed and rolling schedule. The temperature of the furnace is
everything else as it is dependent on the continuous casting process and not the
rolling process.
Therefore,
[MYS, I, F] = [MYS | ε, έ, h, f, R] [ε | L, RS, RT] [έ | L, RS, RT] [h | TF, L, RS, RT, T1,
…, T8, C, Mn, P, Sul, Si, V] [f | TF] [L | RS] [RT | RS] [TF] [RS] [C] [Mn] [Sul] [P] [Si]
[V]………Equation 89
We now assume linear relationships between the different response variables and the
[MYS(k) | εk, έk, hk, fk, Rk] ~ N (µYS + (β11 εk1 + β12 εk2 + … + β18 εk8 + β21 έk1+ β22 έk2+ ……
+ β28 έk8 +β31 hk1 + ….. + β38 hk8 + β4 fk + β5 Rk) , ζ2MYS) ………….Equation 90
[εk | RSk, RTk, Lk] ~ N (µε+ (β61 RSk1 + ….. + β68 RSk8 + β71 Lk1+ … + β81 RTk1+ …. +
[έk | RSk, RTk, Lk,] ~ N (µέ+ (β91 RSk1 + …… + β101 RTk1+ ……. + β201 Lk1) , ζ2Mέ)
………………Equation 92
153
[hk | TFk , RSk, Tk, RTk, Lk Chemistryk] ~ N (µh + (α31 Ck + ….. + α36 Vk + β121 RSk1 +
……. + β131 Tk1+ ……. + β141 RTk1+ β151 Lk1) , ζ2Mh) …………..Equation 93
[Lk | RSk] ~ N (µL + β161 RSk1+ ….. + β168 RSk8, ζ2ML) ………Equation 94
[RTk | RSk] ~ N (µRT + β171 RSk1+ ….. + β178 RSk8, ζ2MRT ) ……..Equation 95
Note that the superscripts 1,2,….,8 represent the eight rolling passes for each coil. k=1,
The above specified model was created in WinBugs. Models were created for different
mechanical properties and different materials and the data from the empirical
relationships and the plant were used as inputs to the models. WinBugs follows the
MCMC simulations based on the general idea presented by Gelfand and Smith. The data
was randomized and 80% of the data was used for model building and the remaining 20%
The results of the MCMC simulations for various parameters of the model described
154
Figure 57: MCMC Simulation results for β31 and β103 for YS model of Material A
The model was developed in section 4.6.3 with 80% data and the remaining 20% of the
data was used for model validation. The data set for each material had around 5000 coils.
The result of model validation for Material B for the different mechanical properties is
155
Figure 58: Actual vs Predicted YS and UTS for Material B
directed acyclic graph is drawn for the different materials and mechanical properties. The
arrows represent the conditional independence structure of the model and the values
represent the median with the 95% credible intervals for the parameters linking the
variables. The regression parameters are normalized between (0,1) for ease of
comparison.
For the sake of simplicity, only the variables for which the parameters do not contain a 0
in their credible intervals are shown in these graphs. A representative graph is shown in
Figure 59.
156
YS
ε1 ε2 έ2 έ5 h f R
1) The YS for Material A is influenced by all the major variables – strain, strain rate,
2) The influence of heat transfer on the YS seems to be the most as it has the
3) The strain in the first and second stand and the strain rate in the second and fifth
Similar graph was created for Material B which is shown in Figure 60. Yield Strength for
material B is affected by only the strain, strain rate and the furnace. Heat Transfer during
the finish rolling, ROT, coiling and the chemistry do not affect the YS directly.
157
YS
ε1 ε2 ε3 έ1 έ2 έ3 f
158
CHAPTER 5
5.1 Introduction
With increased societal reliance on complex engineered systems, their performance and
failure involves significant human and environmental risk, and life cycle costs. The
higher the complexity of the product and the larger the number of associated components,
the higher is their probability of failure or malfunction. The failure of the component
leads to the failure of the entire system. Often the causes of these failures are trivial but
their impact huge. Examples of high profile malfunctions and failures include nuclear
meltdown such as Chernobyl or near meltdown at Three Mile Island, shuttle disasters
blow out preventer in oil rig: Deepwater Horizon-BP [108] and several disasters in
commercial aviation including those due to engine failure (Sioux City, Iowa) or cabin
failure (fuselage, door failures etc.) due to fatigue crack growth. While nuclear and
aerospace malfunctions may have very high visibility in the eye of the public their
probability of occurrence is often very low. On the other hand, automotive malfunctions
may have lower visibility but much higher probability of occurrence leading to higher
fatalities and fiscal impact. Examples of these include tire blowout accidents associates
159
with Firestone/Ford and gas pedal related accidents with Toyota. Finally, oil rig
explosions/tanker failures have large environmental impact that lasts for a long time
(Figure 61).
An aircraft is a complex engineered system which consists of numerous sub systems and
components. At the first level, an aircraft can decomposed into assemblies: propulsion
system, the structure (wing, cockpit etc.) and the landing gear system. At the second
level, the propulsion system can be decomposed into components: aeroengine and the
propulsion controls. The aeroengine itself can be decomposed into parts: blades, disks,
160
gears etc. Thus, the whole aircraft can be broken down into assemblies, subassemblies,
Airworthiness is a term used to describe whether an aircraft has been certified as suitable
for safe flight. This certification is obtained by performing required maintenance actions
every 100 flight hours, with more intensive maintenance inspections at longer intervals.
“The application of airworthiness defines the condition of an aircraft and supplies the
basis for judgment of the suitability for flight of that aircraft, in that it has been designed
approved standards and limitations, by competent and approved individuals, who are
The pilot in command of a civil aircraft is responsible for determining whether that
aircraft is in condition for safe flight. The pilot in command shall discontinue the flight
There are two kinds of repairs which are done to restore a aircraft to airworthiness: major
161
• Might appreciably affect airworthiness by (1) changing weight, balance,
improperly done; or
operations.
The other aspect which goes into keeping an aircraft airworthy is scheduled
maintenance. This means required inspections (annual / every 100 hrs.), replacement of
The various categories in which aircrafts have failed in the past can be divided into the
following:
a) Bird Hazards
b) Fuel Exhaustion
c) Landing/Takeoff
d) Structural failures
f) Wind shear
All these categories fall into the life cycle of aircraft as design/manufacturing issues,
162
Based on the various aircraft accidents and the analysis of different failed components in
an aircraft, FAA has identified safety critical parts on an airplane. According to FAA,
Flight Safety Critical Aircraft Part (FSCAP) is “Any part, assembly, or installation
These parts include landing gear; aeroengine and its components like turbine blades,
gears, disks; windscreen etc. A nice analysis of fatigue failures in different safety critical
parts has been provided by Bhaumik et. al [112]. They identify the potential causes of
Figure 62: Various fatigue failure in safety critical aircraft components [22]- (a) Tail
gearbox housing (b) Connecting rod of piston engine (c) Wheel Hub
5.2 Background
163
One of the most important components in the list of safety critical items is the aircraft
aeroengine (Figure 63 and Figure 64). The aeroengine consists of many sub-assemblies
which can cause catastrophic failure of the aircraft in flight. These include the turbines,
Figure 63: Cutaway and Cross-section of a high bypass jet engine [23]
Analysis of years of aircraft accident data shows that the leading cause of engine related
accidents for turbofan engines is the uncontained failure of safety critical parts [113]. The
failure of safety critical parts can present a significant hazard to an aircraft by releasing
fragments that can penetrate the cabin or fuel tanks, damage control surfaces, or sever
inspections at each piece-part opportunity during their service lives, using methods that
detect flaws that could lead to failure. In daily operation, many engine parts are exposed
to high thermal and mechanical loads. As a result of these loads, cracks can form. If these
164
cracks are not detected, they can grow and lead to part failure. Cracks can also form for
(7) Corrosion
Figure 64: Four types of disks: low and high pressure compressor, and low and high
pressure turbine. The former are made from titanium and the latter of superalloys. Note
low pressure turbine drives low pressure compressor disk and high pressure turbine
crash in Sioux City in 1989. On July 16, 1989, United Airlines Flight 232 suffered a
catastrophic failure in one of its engines. This engine failure destroyed the plane‟s
hydraulic systems and the plane lost control. It broke up on the runway during the
emergency landing at Sioux City, Iowa, leading to the death of 111 people. This crash
was a result of fatigue failure initiated by the presence of hard-alpha inclusion in titanium
Examining the processing history of the turbine disc (Figure 66), it seen that it is
manufactured in a sequence of steps starting with vacuum melted (VIM/VAR) ingot that
is converted into a billet using hot cogging, hot forged, heat treated, ultrasonically
inspected for defects, machined to final shape, finish machined (drilling or holes and
166
creep feed grinding of fir tree for attachment), surface finished and finally inspected
before shipping.
Due to the reactive nature of titanium to oxygen and nitrogen in the atmosphere, several
defects or anomalies are generated during the melt processing of titanium alloys into
167
ingots. These include macrosegregation of oxygen, beta stabilizing elements, and
addition to these melt related defects; microstructural, textural and damage-type (cracks,
pores etc) defects may form during ingot to billet conversion process in alpha/beta
titanium alloys [115]. The two melt related inclusion systems, hard-alpha inclusions and
high-density inclusions (HDI), come principally from the raw materials used. High
interstitial defects result from high nitrogen or oxygen concentrations in sponge, master
alloy, or revert [116-117]. The occurrence of HDI in the final product is rather rare with
improved melting practices such as hearth or scull melting. The so-called low-density
ways, but the principal source with high nitrogen contents is in the original manufacture
of titanium sponge. This alpha stabilized phase is very brittle and after mechanical
working normally contains cracks and may retain some porosity from the original sponge
In large-diameter billet used for applications such as disk forgings, the specification
detection limit is approximately 1,100 µm, with a 50 percent probability of detection for
a defect of 800 µm. This value is uncomfortably large in respect to the fracture design of
many engine parts. The defect has a very low occurrence rate in presently-melted alloys,
with detection at the final forging shape showing approximately one defect per 105 parts
examined. However, because the defect has a long history of creating in-service failures,
168
being continuously implemented. Even the strictest control in double or triple melting
cannot guarantee total elimination of hard alpha inclusions. Currently, the detected
this type of defect still leads to great concern due to high risk associated with it. Hard
substantially higher hardness and low ductility than the material from the surrounding
region. Figure 67 includes examples of typical hard alpha inclusions with a dimension of
200 µm and defect exceedance curve in rotor disks. Figure 68 contains the flow stress of
hard alpha inclusions for different nitrogen content [118] and Figure 69 contains the flow
Usually, hard alpha inclusions include voids or cracks, which make them easily detected
also been proved to be difficult due to the background noise generated by structure in
thick sections. Other NDT techniques such as thermoelectric detection have been studied
to deal with uncracked inclusions. Hard alpha inclusions generally have a TiN core
some cases, nitrogen stabilized ϒ phase may be there instead of TiN. The brittle
inclusions are generally the initiation site of a crack. A study in component failure has
shown the early fatigue initiation caused by an inclusion with a diameter of only 1 mm.
The hard alpha inclusions can also lead to voids during forging, thus increasing the size
169
of potential initiation site. This has also been confirmed in [30] by observing that
Figure 67: Discrete melt related defects in titanium alloys, hard alpha, pores and ceramic
particles (top); the probability of their occurrence as a function of defect size (bottom)
170
Figure 68: Flow stress of low-frequency hard-alpha as a function of nitrogen content
171
Figure 70: Defects in rotor grade titanium material
A 10 year study was conducted by the Federal Aviation Agency (FAA) to study the effect
of forging parameters on the distribution of hard alpha and its effect on fatigue life. The
final report of the study was published in 2008 and can be read for the detailed study
[120-121]. The main conclusions of the study are summarized in the following (Figure
71).
172
Figure 71: Overall Theme of the FAA Study [121]
There has been a lot of research into predicting the fatigue life of aeroengine disk and
also to model the failure of the disk. These approaches can be broken into three major
categories:
173
Figure 72: Damage tolerance approach to determining the life of aeroengine disk [122]
cooling plates and turbine disks, are subjected to cyclic stresses caused by engine
maneuvers. These cyclic stresses can exceed the yield strength of the material at
stress-concentration sites, such as bolt holes, bores, and can thus lead to low cycle
fatigue cracking. A safe life method is conventionally used to determine the in-
service life of rotating gas turbine components subject to low cycle fatigue. In this
procedure a safe service life for the component is obtained from component and
specimen tests of the appropriate material under sample loading conditions. The
component life is calculated by fitting a statistical model to the fatigue results, and
the life is determined based upon the probability of crack initiation after a period of
normally restricted to the number of flight hours which are calculated to produce
174
a 0.8mm surface crack in one component out of an assumed population of 1000
identical parts under the same usage This crack size is considered detectable with
high reliability by the most commonly used inspection technique - liquid penetrant
The cyclic stress intensity factor (ΔK) ahead of the crack tip is given by
………………..Equation 96
where λ is a factor that depends on the crack shape, structure and gradient of stresses and
………………Equation 97
where C and n are determined experimentally. This is an average model because C and n
The fatigue life model for analysis is commonly referred to as the Walker strain
………………Equation 98
where E is the material modulus, Δε denotes the strain range, and m is an empirical factor
to account for the effects of mean stress [34]. The effect of temperature is included
175
Figure 73: Crack Growth rates for Pratt and Whitney specimens in FAA study [121]
One of the main probabilistic fatigue life tools developed is DARWIN by South West
accurate tool for engineers to compute the probability-of-fracture of a rotor disk with and
without inspection. The hard alpha defects may be randomly distributed within a disk,
and, therefore, a zone-based risk integration approach is used to account for the
uncertainty of location. A zone is a grouping of material such that all sub-regions in the
zone have a generally uniform stress state, and the same fatigue crack growth properties,
words, the risk computed for any sub-region of material of the zone will be the same (or
nearly so); and, therefore, the subregions are grouped into a zone. Thus, the disk cross
section is divided into a manageable number of zones of approximately equal risk using
the finite element mesh and stress results as the framework for the zone discretization.
176
This “zone-based” system reliability methodology accounts for:
• possibility that a hard alpha anomaly developed during the titanium melt process
177
The methodology assumes, at most, one significant anomaly exists in the disk (the
probability of two or more anomalies in a disk is negligible). Also, it assumes that the
…Equation 99
1. In rolling contact fatigue (RCF) testing the stress is kept uniform while in low
cycle fatigue (LCF) testing the strains are kept constant. In DARWIN only the
3. It is assumed that the risk will be same for a subregion which might not be the
heterogeneity in physics based models. The main research work in this area are:
a) Representative volume element (RVE) by Vornoi cell tessellation (Ghosh et. al.
178
element level in FEM. This heterogeneity is then randomized over the entire
[125]
c) Converting the heterogeneity into statistical form through spectral methods has
used to model the randomness in material and the variability is captured through
d) Stochastic Grid Collocation approach has been used by Zabaras et. al. [127] to
capture the variability at the material level. The texture and grain orientations are
captured in a stochastic manner and are then randomized over the entire material
5.2.4 Experimentation
FAA conducted a series of experiments by artificially seeding Titanium billets with hard
alpha and subjecting them to different forging parameters to study the effect of forging on
the movement of hard alpha. A schematic of this experimentation is shown in Figure 75.
Based on these experiments, a numerical microcode was developed with the purpose of
179
Figure 75: Experimentation on hard alpha during FAA study [120]
Developed and tested microcode to predict shape and orientation of hard alpha
defects
Generated constitutive property data for hard alpha and diffusion zone
180
Developed life prediction tool DARWIN
To overcome the main disadvantages of the physical and statistical models and to design
the aeroengine disk manufacturing process for reduced failure, the hierarchical
decomposition of the process was done (Figure 76). This decomposition is based on the
In this decomposition, the fatigue life of one aeroengine disk is decomposed into being
due to three different states: the defect state (material impurities), the material state and
the loading. These states are then decomposed into the process physics which cause the
states. For example, the Grain size, microstructure and its texture etc. define the material
state; the applied stress on the disk defines its loading etc. These quantities are further
decomposed into the design parameters affecting them. For example, the forging velocity
affects the strain and strain rate in the defect which affects the defect state and hence the
life. Finally, the design parameters are decomposed into the actual manufacturing
processes which cause them. Thus, the forging velocity is controlled by the forging
process etc.
181
Figure 76: Physics Based Hierarchical Decomposition of Aeroengine Disk
Due to the complicated nature of the interactions between the different variables, the
1) The loading on the disk is assumed to be constant. This loading is based on the testing
2) The material state of the disk is influenced mainly by the heat treatment process.
After the forging process, the microstructure and the residual stresses are dependent
heavily on the heat treatment time and temperature [128]. To simplify the analysis,
3) Only the effect of defect state is studied on the life of the aeroengine engine.
Furthermore, the defect state is characterized only by single hard alpha particle in the
182
titanium matrix for the purpose of illustration. This assumption can be easily relaxed
Let us assume that the fatigue life of the component is measured and is denoted as L.
183
M: Defect state for the hard alpha (like X-movement of hard alpha, rotation of hard alpha
etc.)
I: Internal State variables for both the matrix and the inclusion (strain, strain rate etc.)
which affect M
To understand the mechanism which affects the variables defined above, we need to
model the joint probability distribution of all these parameters, which can be denoted by
[L, M, I, F]. Through conditional distributions, this joint pdf can be decomposed as:
Our aim is to find this joint distribution based on the observed data D. However, we
would like to use the Bayesian framework to learn about the underlying variables. Using
the likelihood in Equation 106 and the prior distribution on process parameters F, the
Here:
[D | L, M, I, F] = Data model
184
To capture the process model (level 3 in decomposition), numerical or virtual modeling
of the forging process is done using the FEM. The aim is to capture the internal state and
the process which affects the defect state during the process and hence affects the life.
Moreover, we have assumed a single defect or hard alpha anomaly which influences the
life. To model this defect inside the titanium matrix, we use multi-body approach of FEM
in which the titanium matrix and the hard alpha anomaly are treated as two separate
bodies with interactions. This approach helps us in understanding the behavior of hard
A comprehensive study of behavior of hard alpha during forging is included in the result
of Turbine Rotor Material Design (TRMD) project. In that project, the constitutive
properties of hard alpha anomaly were studied by compression test, indirect tension test,
indentation test, and plane strain compression test. A constitutive relation was developed
with consideration of pressure, nitrogen content, and fracture. The constitutive properties
of hard alpha titanium have been studied in TRMD project and published in [120]. The
engineering stress strain curves of hard alpha titanium under different test conditions for
compression tests are shown in Figure 75. Materials with different nitrogen content show
that under normal forging condition material with less than 4% nitrogen can be plastically
185
deformed. The flow data of Ti-6Al-4V from material data base of FORGE is represented
m4
A1 e e
……………………Equation 102
m 1T m2 m3
f
where σf is the flow stress, ε is the strain, is strain rate, T is the temperature, A1, m1,
m2, m3, m4 are material constants with values shown in Table 28. Note that these values
A1 m1 m2 m3 m4
The comparison of flow stress of Ti-6Al-4V with hard alpha Ti is made at the same
temperature and strain rate. The stress and strain in Hansel-Spittel law from f
A1 e
m 1T
m 2
m 3
e
m
4
……………………Equation 102 are true stress and strain; to compare the results from
compression test of hard alpha as shown in Figure 75, the stress and strain have to be
converted to engineering stress and engineering strain using the following equations:
where ε is true strain, εe is engineering strain, σ is true stress and σe is engineering stress.
By comparison, it is found that the flow stress of hard alpha is about 2 to 4 times higher
186
than the flow stress of Ti alloys.
To accurately predict the behavior and location of such anomalies during forging, multi
In addition in the FAA reports results of experiments are reported in which hard-alpha
seeds are artificially integrated in titanium-64 billets. Seeds with different nitrogen
content and orientation were placed in designed locations for various billets and
isothermally forged into dogbone shaped forgings which are further hot-die forged into
disk forgings. Ultrasonic inspections were made before and after forging, and selected
seeds were cutup to measure the dimension. One of the forging results reported are for
the forging labeled SB-1 (seeded billet 1). Six different seeds were placed in six billets in
different orientations. The details are shown in Figure 78. Note that only one seed was
placed in each billet because the probability of having two seeds in one billet is
negligible.
187
Figure 78: Description of different seeds in SB-1
Model Validation
Seed 3 was chosen for the calibration of model because it was located near the boundary
and is most likely to flow the most during the forging. This seed had 12%N in the core
188
and 2%N as a diffusion zone and was located axially in the billet. The core (12%N) was a
cylinder of height 0.100” and diameter of 0.100”. The diffusion zone (2%N) was a
cylinder of height 0.200” and diameter 0.200”. They are shown in Figure 79. The billet
Figure 79: Half Section view of Diffusion zone (left) and hard alpha (right)
The forging was set up using the same conditions as the experiments (isothermal, with
billet at 9500C and dies at 6000C, forging speed of 5 mm/s and with a friction factor of
m=0.3.). The model was set up as 1/8th of the total model to optimize the computation
(b) Axial, radial and hoop dimensions of the diffusion zone (DZ) after deformation
(c) Axial, radial and hoop dimensions of the hard alpha (HA) after deformation
189
Results
The dogbone forging at 0% deformation and 100% deformation is shown in Figure 80.
The final location of the centroid of HA from ultrasonic inspections of forgings and
Table 29.
190
The final shapes of the DZ and HA are shown in Figure 81.
(a)
0.14”
0.07”
0.18”
0.09”
0.22” 0.10”
The final dimensions of the deformed HA and DZ in all the directions are compared with
191
(a)
FORGE3 Model
Predictions FORGE3 Mo
Predictions
GE3 Model
dictions FORGE3 Model
Predictions
5.4.1 Conclusions
192
Based on the results of the FORGE3 model the following conclusions can be made:
1. The model predicts the movement (by measurement of final location) of hard
2. The model predicts the deformation of diffusion zone in all 3 dimensions very
closely to the measurements. The axial measurement is off by -16%, the radial
3. The model predicts the deformation of hard alpha in all 3 dimensions very closely
to the measurements. The axial measurement is off by -9%, the radial and hoop
Thus, the developed FORGE3 model can be used to study the effect of forging
5.4.2 Effect of processing parameters on the movement and rotation of hard alpha
The numerical model was further used to study the effect of processing conditions such
as the nitrogen content of hard alpha (affects the flow stress), preheating temperature of
the workpiece and velocity of the die, results are included in Figure 83. In pancake
forging, the cylindrical billet (diameter 162.6 mm) is reduced in height from 289.6 mm to
99.4 mm. Since this step is done in an isothermal forging setup, the starting temperature
of both the billet and the environment is the same with the temperature of the flat dies at
6000C. For numerical experiments, two billet preheating temperatures were taken 8500C
and 9500C, and three die speeds 5 mm/s, 70 mm/s and 700 mm/s to represent different
193
forging equipment. The dimensions of the alphas seed were height 5.08 mm and diameter
5.08 mm.
Figure 83: Seed micromechanics during pancake forging: location in the billet (left top),
its displacement and rotation (top right), and the coordinates of the seed.
194
Figure 84: Influence of processing parameters and Nitrogen content on movement of hard
alpha
195
It is seen that forging temperature and die velocity do not affect the displacement of the
inclusion but considerably affect the rotation and the deformation of the inclusion. This is
primarily due to increase in flow stress (and rotation) with lower temperatures and higher
die velocities. Increase in nitrogen content of the hard-alpha inclusion increases its flow
stress with respect to the parent titanium alloy. This causes both the rotation and
The second step in forge processing is forging the pancake into a compressor disk. Figure
this step when the original pancake was forged at 8500C and 5 mm/s. It is seen that die
speed has a major influence on the displacement and deformation of the inclusion at a
fixed preheat temperature, and a small effect on inclusion rotation. On the other hand,
preheat temperature mainly affects the inclusion deformation. Looking at the parent
material (matrix), we observe that both the strain and stress in the matrix increase with
die velocity at 9500C and 2% nitrogen, the stress jumps 50%. It is to be noted that the
strain rate hardening response of the parent titanium alloy matrix is much higher than that
196
Figure 85: Displacement and rotation of the inclusion during disk forging
Figure 86: Typical strains in the Titanium matrix and Hard alpha during the forging
197
5.4.3 Cavitations of hard alpha
Another major cause of limiting the fatigue life of the aeroengine disk is the formation of
cavities around the hard alpha which act as locations where the fatigue crack initiates and
propagates. These cavities are formed due to the difference in the flow stress of titanium
matrix and the hard alpha anomaly (Figure 87). The hard alpha is usually surrounded by
multiple grains of titanium which consist of different alpha and beta phases. These
different phases have different flow stress behavior and hence each of the grain around
The forging of titanium has the following features which are captured in the model:
198
• Prior hot working with recrystallization annealing in single β phase field,
microstructures.
Using the self consistent model and the different flow stress behaviors of hard alpha and
the surrounding grains, multi-body simulations were done in the region around the hard
alpha. This model determines the velocity and temperature conditions under which the
cavities form and also finds the size of the cavities with respect to the hard alpha
movement. Furthermore, the effect of grain boundary sliding on the formation of cavities
can be studied using this model. This grain boundary sliding is modeled using the
interface friction at the different bodies modeled as grains in this model (Figure 89).
199
Grains representing α + β phase
of Ti-6Al-4V
Volume fraction of alpha – 80%
Volume fraction of alpha – 65%
Discrete Anomaly
Anisotropy :
• Difference in flow stress based on volume fraction of alpha
phase
• Grain Boundary sliding modeled by friction criteria
Grain boundary
sliding with m=0.5 No grain boundary sliding
Cavitation
200
0% Nitrogen
3
1.5
850 C
1
950 C
0.5
0
0 0.2 0.4 0.6 0.8 1
2% Nitrogen
3
Width of Cavity (mm)
2.5
1.5
850 C
1
950 C
0.5
0
0 0.2 0.4 0.6 0.8 1
4% Nitrogen
3
Width of Cavity (mm)
2.5
2
1.5
1 850 C
950 C
0.5
0
0 0.2 0.4 0.6 0.8 1
Figure 90: Effect of processing parameters on the width of cavity around hard alpha
201
The points in Figure 90 and physical process information imply that following patterns
between the processing and width of cavity around the hard alpha can be seen:
2) The width decreases as the temperature increases given a strain rate. This is due to a
decrease in the differential between the flow stress of hard alpha and the titanium, as
temperature increases.
3) The width increases as the Nitrogen content in the hard alpha increases. This is
because the differential flow stress between the hard alpha and the titanium matrix
5.5 Effect of forging velocity and billet temperature on the movement and rotation
202
Let us assume that the defect state is completely characterized by the location and
rotation of hard alpha. Also, let us assume that these parameters are only affected by the
forging velocity and billet temperature. Thus, our posterior model can be written as:
Where:
M has two components representing the X-movement (X) and rotation (Theta) for
I has two components IM and IH representing Strain (ε) and Strain rate (έ) for the
matrix and hard alpha respectively. The values for matrix will have subscript m.
DX and DTheta represent the X-movement and rotation data for the hard alpha
respectively. Although the hard alpha moves in both X and Y direction, we are
observations relating the fatigue life to position of hard alpha as reported in FAA
report [120]
Dε and Dέ represent the strain the strain rate data for hard alpha
Dmε and Dmέ represent the strain the strain rate data for titanium matrix
203
The different components of the Hierarchical Bayesian Model are:
Data model: [DL, DX, DTheta, Dε, Dέ, Dmε, Dmέ, DV, DT | L, X, Theta, ε, έ, εm, έm, V, T]
Section 5.4 described the process model and the section 5.5.1 describes how the data
model is created.
To simplify the conditional structure in the above equation based on process knowledge
we do the following:
[DL, DX, DTheta, Dε, Dέ, Dmε, Dmέ, DV, DT | L, X, Theta, ε, έ, εm, έm, V, T]
= [DL | L] [DX | X] [DTheta | Theta] [Dε | ε] [Dέ | έ] [Dmε | εm] [Dmέ | έm] [DV | V] [DT | T]
……….Equation 105
1) The data for the X location and the theta are independent of each other given their
process models. Hence, the dependence between these data comes from the process
2) The strain and strain rate are independent of each other given their process models. If
we know the forging process model, we can determine the strain and strain rate.
3) The forging velocity and the billet temperature are set individually by the engineers.
204
Furthermore, we assume each of these independent pdfs to be normally distributed.
Although more complicated models can be built, for sake of simplicity normal models are
The life data used for DL was used from the TRMD report based on their experimentation
and is given in Figure 92. These data points also have the process conditions which were
used to generate the other data. The different variances in the data model above represent
the errors in the measurement due to sensors, experimentation and computational issues.
For example, the strain and strain rate are derived from FEM simulations which might
205
have errors due to convergence of the numerical solution.ζ2Xand ζ2Theta represent the
Figure 92: Results of fatigue test used for data model [32]
206
We assume that the fatigue life is completely defined by the location and rotation of hard
alpha based on the FAA experiments. Also, conditional on the strain in hard alpha, strain
rate of hard alpha, strain of titanium, strain rate of titanium and the process parameters, X
a) X and Theta are can be completely calculated if we know strains and strain rates in
the hard alpha and the titanium matrix surrounding it during the forging process.
b) Strains do not depend on the strain rate during the forging process but only on the
Therefore,
Finally, the forging velocity and billet temperature are process settings and can be set
independently of each other. The forging engineer can set the velocity on the press and
temperature of the billet in the furnace without affecting each other. The final process
The aim now is to define each of these individual conditional probability distributions.
a) [L | X, Theta]
This conditional distribution defines the relationship of the Life with the location and
rotation of hard alpha. We assume linear relationship between the life and the location
[L | X, Theta, µL, βX, βTheta, ζ2LL] = [Li | Xi, Thetai, µL, βX, βTheta, ζ2LL]....Equation
119
[Li | Xi, Theta, µL, βX, βTheta, ζ2LL] ~ N (µL + (βXi + βTheta XTheta) , ζ2LL)…Equation 120
Although this seems to be a very simplistic model and might not represent the reality, but
due to the small amount of data available for model validation, this model was chosen. In
future, more complicated models which might include higher order effects of
b) [X | ε, έ, εm, έm]
208
The X movement of hard alpha during the forging process is dependent on the internal
state variables (strain and strain rate) both of the matrix and the hard alpha itself. There is
another thing to notice here. This movement is not only defined by the final strains and
strain rates, but on the path and the transition strains and strain rates. This can be defined
where t=1, 2,….R represents the time steps during the forging process (t is usually in
Or,
[XR | εm, έm, ε, έ] = [X1] [Xt+1 | Xt, εmt, έmt, εt, έt] ……………….Equation 122
Where, [X1] is the initial location of the hard alpha prior to the forging. This is usually
defined by the melting process which determines distribution of a defect inside a billet.
These distributions are obtained by extensive ultrasonic testing on the billets after
melting.
209
This formulation is consistent with the theory of plasticity and FEM formulation as
defined in Chapter 2. From the FEM data we can find the constants F, G, GM, H and HM.
For example, in the case when V=70 mm/s, T=900oC these constants are:
Histogram of cc
65
0.15
Density
Sample Quantiles
60
0.10
0.05
55
0.00
55 60 65 -2 -1 0 1 2
Based on the various FEM simulations with different velocities, temperature and %N of
hard alpha, the values of the rotation (Theta) were plotted (Figure 93). This histogram
shows that the theta for the entire range of processing conditions lies between 45 and 76
degrees. Moreover the values follow a distribution close to Gaussian. The rotation is not
only defined by the final strains and strain rates, but on the path and the transition strains
210
and strain rates. This can be defined as a state transition probability through the Markov
Chains.
where t=1, 2,….R represents the time steps during the forging process (t is usually in
Or,
[ThetaR | εm, έm, ε, έ] = [Theta1] [Thetat+1 | Thetat, εmt, έmt, εt, έt] …..Equation
125
Again, [Theta1], the initial rotation of hard alpha before the forging process is determined
From the FEM data we can find the constants FA, GA, GMA, HMA and HA. For this
purpose, after the simulations were done, the output file was stored in CSV format with
the required strain, strain rate results at certain time steps. These were then used as inputs
in the WinBugs file. The code and the output file as given in Appendix.For example, in
211
d) [ε | V, T, φ] and [εm | V, T, φ]
The strains inside the billet during the forging process can be divided into three different
types: the elastic strain (εe), [the plastic strain (εp) and the thermal strains (εt). During the
forging process, the billet undergoes plastic deformation, hence the elastic strains are
negligible. Moreover, the disk forging of Titanium is an isothermal forging process and
therefore the thermal strains are also negligible. Therefore, the only strains we will
consider here are the plastic strains (εp) which are dependent only on the preform initial
The distribution [ε | V, T, φ] therefore reduces to [ε | φ]. Hence the strains are constant
for a given preform design when only the velocity and temperature are changed.
e) [έ | V, T, φ] and [έm | V, T, φ]
Analogous to the strains, we assume that the component of strain rate (units of 1/sec)
which affects the movement and rotation of hard alpha is the plastic strain rate (έp). This
212
f) [V | φ] and [T | φ]
Both these functions determine the prior distributions of the design variables or the
process conditions. These prior distributions are based on the engineering settings of the
plant and represent the variation occurring during the forging process. To simplify the
analysis, we assume these to be normally distributed with a certain mean and variance.
The above specified model was created in WinBugs. FEM simulations of the movement
and rotation of hard alpha were done in FORGE and the data from these simulations was
used as the dataset for model building. The effect of two different velocities (5 mm/s and
70 mm/s) and two different temperatures (8500C and 9500C) was studied.
The results of the MCMC simulations with the different variables are shown in Figure 94.
213
MCMC simulation plot for theta[2] MCMC simulation plot for theta[4]
400
500
350
400
300
300
Theta[4]
Theta[2]
250
200
200
100
150
0
100
Iteration Iteration
Quality Index in Batch 1 for Material B Quality Index in Batch 3 for Material B
0.020
0.03
0.015
Density
Density
0.02
0.010
0.01
0.005
0.000
0.00
240 260 280 300 320 340 360 380 240 260 280 300 320
N = 11001 Bandwidth = 2.404 N = 11001 Bandwidth = 1.476
Figure 94: Diagnostic and Posterior plots of MCMC for β and βTheta
graph is drawn for the life. The arrows represent the conditional independence structure
of the model and the values represent the median with the 95% credible intervals for the
parameters linking the variables. The values of the parameters are calculated by taking
(parameter)]. This is done so that the different parameters can be compared easily.
214
For the sake of simplicity, only the parametersthat do not contain a 0 in their credible
intervals are shown in these graphs. A representative graph is shown in Figure 95.
The network shows that the X-location plays a more important role than the rotation in
215
5.8 Design of Forging Process
Let us consider a practical situation which occurs very commonly in the forging of
aeroengine disks. The forging company receives a billet after melting and casting. During
the ultrasonic inspection it observes a defect located in the billet and the x-ray analysis
reveals that it is hard alpha. The company knows the exact location of the defect in the
billet and also its chemical composition. At this stage it has 2 options:
b) To go ahead with the forging of the billet in a manner such that the life of the final
The first option results in a loss of several million dollars because of the cost of the
material and initial processing. Hence, the company must decide on option 2. In this case,
the design of the preform and the forging process becomes very critical. The various
This case study gives an example of such a situation by considering various scenarios and
216
In this case study, 4 different locations of the hard alpha in the initial billet and two
different chemical compositions (2%N, 4%N) were chosen and the design was run for
these locations. It is assumed that a melted billet will have only one of these defects and
will have only one of the chemical compositions. The design however can be extended
The hard alpha is assumed to be of the same size and shape in all the locations. This is
based on inspection results and studies by previous researchers. The hard alpha is
assumed to be a cylinder with diameter of 5.08 mm and height of 5.08 mm. These
locations are identified as Loc1, Loc2, Loc3 and Loc4 in Figure 96. All the hard alphas
detection = 0.5.
Loc1
Loc2
Billet
Loc3
Loc4
Figure 96: Location of hard alphas for design study (Note that each billet contains only
217
The Bayesian network estimated was used for the design of the forging process in all the
four cases. The aim of the analysis was to maximize the life. Predictive modeling is easy
in Bayesian case because we can specify the inputs as stochastic nodes without any data
value and the output will the predictive distribution for that particular node. A series of
values for the design parameters (height, diameter, velocity and temperature) were
selected for each location of hard alpha and the predictive distribution of the life was
obtained in each case. The life was normalized to 1.0 for easy comparison. The values
Diameter: 240, 200, 180, 170, 160 mm corresponding to each height (When height =
Some of the results of the effect of different process parameters on the life are shown in
Table 30.
218
Location %N Height Diameter Velocity Temperature Life
hours)
The expected life is then plotted with respect to the different design variables and the
219
Location 1 - 2%N
1
0.9
Height (mm)
Location 1 - 4%N
1
0.9
Normalized Life
0.8
0.7
0.6
0.5
0.4 Different Velocities
0.3 Different Temperature
0.2
0.1
0
150 200 250 300
Height (mm)
Location 4 - 2%N
1
0.9
Normalized Life
0.8
0.7
0.6
0.5
0.4 Different Velocities
0.3 Different Temperature
0.2
0.1
0
150 200 250 300
Height (mm)
1) Based on the location and chemical composition of hard alpha, the design of the
2) Some of the preform designs exhibit higher variability with different velocities
For example, if it is found that the melted ingot contains hard alpha with a composition of
2%N at location 1, the maximum life which can be obtained for the disk is 0.72. To
obtain this life, the preform has the height of 210 mm and diameter of 180 mm.
Furthermore, the velocity and temperature of the forging do not play a major role in the
life for this preform design. On the other hand, if a velocity of 50 mm/sec is desired for
some other aspect of the forging, than the preform height can be set at 180 mm. At this
preform height, the life of the disk can be maximized for the above said velocity. Hence,
there is a strong interplay between the preform design and the forging parameters and
both these need to be carefully selected based on the incoming billet for optimal design of
the disk.
221
CHAPTER 6
CONCLUSIONS
structure and is difficult to understand because of its intricacy. The source of this
complexity can be from too many interconnected components, too much information or
constraints, and many parameters. It can also stem from little knowledge about the
steel rod manufacturings are examples of complex engineered systems because they have
many interconnected components and there is little knowledge about the interactions
The current design cycle for these different manufacturing processes treats the different
system components (sub-processes) in isolation. Each engineer and personnel looks at the
problem in his own perspective and optimizes the solution pertaining to the requirements
In this work, a novel approach of designing complex engineered systems by including the
manufacturing processes is formulated. The physical models are combined with the data
obtained to create Bayesian Hierarchical Models which are used for prediction and
design. Case studies involving surface defects on bars, mechanical properties in sheet
222
manufacturing and defects in aeroengine manufacturing are described. The main
incorporated into the design process through process physics. The complex system is
broken down into various elements based on the physics of the manufacturing processes
and these are linked in a Bayesian formulation to include the data obtained through the
In all the three case studies explained in this thesis, the decomposition follows the same
architecture. On the top level is the class of the quality attributes which is important for
the product (defects, mechanical properties or life). On the second level are the different
features which make up the quality attributes (seams, scales etc as defects, YS, UTS as
mechanical properties). On the third level are the quantities based on process physics
which directly affect theses attributes. These quantities can be observable quantities like
the inclusions in aeroengine disk, or can be hidden engineering quantities like the strain,
strain rate in bar and sheet manufacturing. The forth level are the manufacturing design
parameters which affect the process physics and the final level connects the
In case on bar manufacturing, the surface defects are the class of quality attributes which
can be decomposed into individual defects like seams, scales etc. Each of the these
defects is then related to the strain, strain rate, stress and heat transfer occurring at
223
different roll stands. These physical quantities are then related to the roll loads,
temperatures; roll speeds etc. which are the settings on the individual roll stands. These
design parameters are then decomposed into the manufacturing process affecting them.
For example, the seams were found to be influenced mainly by the second and third
roughing roll stand. The empirical relationships and FEM models are used along with the
In the case of aeroengine disk, the decomposition is done based on the defect state (hard
alpha), the material state and the applied load. The defect state is dependent on the
location and rotation of hard alpha. This state depends on the strains, stresses etc. formed
during the forging process which affect the fracture, cavitation etc. FEM models of the
forging process are built which show that the defect state is very tightly controlled by the
initial design of the forging process and the process parameters like the velocity and
temperature.
In this work, Bayesian Hierarchical Models (BHM) are developed to couple the process
physics and the data in a single framework. This framework is used to design a complex
system by integrating the experimental data with the Finite Element based process
models of forging process. Data is extracted from these FEM models and used as inputs
The two main aspects of these BHM are the data model and the process model. The data
models are built on the data collected at the individual manufacturing processes through
224
different sensors. This step requires the use of various data mining tools because the
- Massive Size: Each sensor records observations every few seconds. The typical size
of the database with all the readings is usually 10-20 GB per day.
- Spatiotemporal in nature: The data collected by these sensors is both time based and
location specific.
- Noisy: These sensors are located in an environment which has high temperatures and
dust. All these result in a lot of noise in the data which needs to be cleaned and
built over a period of time by various researchers which try to explain the quality in
terms of physical phenomena. These models are used as process models at various
b) FEM models: These are computer generated models of the entire manufacturing
process under study. These models are based on the process physics and can be used
to understand the process by looking at the engineering quantities. In this work, FEM
models are built for the bar manufacturing and the aeroengine disk manufacturing
process. These models help in giving valuable insight into the process. The FEM
models of the bar manufacturing help in understanding the formation of seam, cracks
etc. while the models for aeroengine disk help in understanding the movement of hard
alpha defects.
225
(3) Validating the hierarchical decomposition on different processes
In this dissertation, the methodology developed is illustrated on the design and prediction
1) Data is collected on the plant. This data contains both the process information and the
2) The data is converted into a form so that it can be used in this work. In the case of bar
and sheet manufacturing, single text files are created for the data. These files contain
3) Process models are built for the three cases and these models are calibrated and
4) Bayesian Hierarchical Models (BHM) are built using both the process models and the
plant data.
5) The prediction results of these BHM are compared with the data.
226
In all the cases, the BHM performs well and the errors are low. But, in the case of
aeroengine disk manufacturing, the amount of data available is very limited. Hence, the
model for aeroengine disk might have a high generalization error. As more and more data
227
CHAPTER 7
FUTURE WORK
The future work based on the current work can be subdivided into two main categories:
one which is a direct extension of this work and one where the present work can be
In the model building of aeroengine disk, the following work needs to be done:
(a) Include the effect of heat treatment process variables on the defect state
The present work assumes that the material state is constant as heat treatment process
is assumed to be same. This assumption should be relaxed and the effect of heat
treatment time and temperature should be studied on the material state in both the
(b) Include multiple hard alpha defects in the design of the forging process
The present work assumes that there is only one hard alpha defect in the billet. If
more than one defect is present, it is necessary to understand the interaction between
these different defects and how they would affect the life.
228
(c) Include the effect of small anomalies on the design and prediction of life
The present work considers only large hard alpha defects in titanium matrix. There
are other defects present in titanium billets which are much smaller in size and are
numerous in numbers. Modeling of such defects should be done to see their effect on
In the model building of bar manufacturing, the following work needs to be done:
The FEM models need to be built which can simulate the formation of defects like
The FEM models should include the material information in a non homogenous
manner. This can be accomplished by building material models which can represent
decomposition can be applied to the design of other complex systems. One example
229
Design of process for hip implant
According to this year‟s National Joint Registry (NJR) Annual Report, UK, performance
Revision rates (how likely it is that a patient will need an operation to remove and usually
replace a prosthesis) for the devices are as high as 13.61% for MoM. Rates for the non-
MoM hip prosthesis types ranged from 3.31% to 4.94%.Figure 98 illustrates a typical hip
implant and its location inside the body. The figure on the right illustrates the critical
Figure 98: Hip Implant inside the body (left) and the critical section which fails (right)
[…]
To improve these revision rates, the forging process which manufactures these implants
must be designed so that each implant has a property compatible with the patient
receiving it. A typical sequence in the forging of hip implants is shown in Figure 99.
230
Figure 99: Preforming stage of hip implants [...]
In a manner similar to the aeroengine disk, the main failure modes of the implant can be
determined and set as the quality attributes. These attributes can be decomposed into the
material properties and the mechanical properties causing them. These material and
mechanical properties can then be related to the manufacturing process through the use of
process physics and plant data and related to the requirements of the patient. A sample
Design
[2] Adnan Akay and Judy Raper, “Complex Engineered and Natural Systems”,
[3] Lenard, J.G., Pietrzyk, M., Cser, L., “Mathematical and physical simulation of the
[5] Ginzburg, V., “Metallurgical design of flat rolled steels”, CRC Press, 2004
[7] Phadke, S., Pauskar, P., Shivpuri, R., Proc. 9th ISPE Int. Conf. on Concurrent
329-337.
[8] Shivpuri, R., Pauskar, P., Annals of CIRP, Vol. 48, p. 191-194
Metallurgical Defects, ASME J. Eng. Gas Turbines Power, 2007 (129) 827–835.
[10] Wu, Y.T., Enright, M.P., Millwater, H.R., 2002. Probabilistic methods for design
232
[11] Harrison GF, Tranter PH, Shepherd DP, Ward T, Application of multi-scale
[12] Witek, L., Failure analysis of turbine disc of an aeroengine, Eng. Fail. Anal.
2006(13) 9-17.
[13] Semiatin, S.L., Nicolaou, P. D., Thomas, J. P., Turner, T. J., Defect occurrence
[14] Berliner, L.M, R.F. Milliff, and C.K. Wikle, Bayesian hierarchical modeling of
doi:10.1029/2002JC001413
[15] Berliner, L. M., Hierarchical Bayesian time series models, in Maximum Entropy
and Bayesian Methods, edited by K. Hanson and R. Silver, pp. 15–22, Kluwer
[16] Cressie, N., Buxton, B. E., Calder, C. A., Craigmile, P. F., Dong, C., McMillan,
N. J., Morara, M., Santner, T. J., Wang, K., Young, G., and Zhang, J. (2007).
[17] Santner, T. J., Craigmile, P. F., Calder, C. A., and Paul, R. (2008). “Effect and
233
[18] R.H. Wagoner and J.L. Chenot, “Fundamentals of metal forming”, John Wiley
[19] R.H. Wagoner and J.L. Chenot, “Metal Forming Analysis”, Cambridge University
Press, 2001
[20] Sellars, C.M., Whiteman, J.A. 1979, "Recrystallization and grain growth in hot
[21] Sellars, C.M., "The physical metallurgy of hot working", Hot Working and
Forming Processes, Sellars, C.M. and Davies, G.J. (eds.) London: Inst. of Metals,
[22] A. Laasraoui and J.J. Jonas, "Prediction of Temperature Distribution, Flow Stress
and Microstructure during the Multipass Hot Rolling of Steel Plate and Strip",
[23] Ohjoon Kwon, "A Technology for the Prediction and Control of Microstructural
[24] John H. Beynon and C. Michael Sellars, "Modelling Microstructure and Its
Effects during Multipass Hot Rolling", ISIJ International, Vol. 32 (1992), No. 3,
pp. 359-367
Recrystallization
234
[26] Shigenobu Nanba, Mitsuru Kitamura, Masao Shimada, Masaaki Katsumata,
Ttsuyoshi Inoue, Hiroki Imamura, Yasushi Maeda and Shigeo Hattori, "Prediction
after Hot Rolling in Carbon Steels", ISIJ International, Vol. 32 (1992), No. 3, pp.
377-386
[27] Andreas Kern, Joachim Degenkolbe, Bruno Musgen and Udo Schriever,
[28] Atsuhiko Yoshie, Masaaki Fujioka, Yoshiyuki Watanabe, Kiyoshi Nishioka and
235
[31] Takehide Senuma, Masayoshi Suehiro and Hiroshi Yada, "Mathematical Models
[32] Masayoshi Suehiro, Takehide Senuma, Hiroshi Yada and Kazuaki Sato,
High Carbon Steels", ISIJ International, Vol. 32 (1992), No. 3, pp. 433-439
Microalloyed
Steel at the Hot Strip Mill", ISIJ International, Vol. 33 (1993), No. 5, pp. 605-614
236
[37] Zhou Xu, Guo-Rong Zhang and Taku Sakai, "Effect of Carbon Content on Static
1102
[42] G Li, T. M. Maccagno, D.Q. Bai and J. J. Jonas, "Effect of Initial Grain Size on
Modeling of the Mean Flow Stress, Fractional Softening and Grain Size during
the Hot Strip Rolling of C-Mn Steels", ISIJ International, Vol. 36 (1996). No. 12,
pp. 1500-1506
237
[44] K. Minami, F. Siciliano Jr., T. M. Maccagno and J. J. Jonas, "Mathematical
Modeling of Mean Flow Stress during the Hot Strip Rolling of Nb Steels" ISIJ
Modelling of Mean Flow Stress during the Hot Strip Rolling of Multiply-alloyed
Medium Carbon Steels", ISIJ International, Vol. 38 (1998), No. 2, pp. 187-195
[46] Ohjoon Kwon, Kyung Jong Lee, Ki Bong Kang, Jae Kon Lee, "Control of the
structure and properties of hot strip", Asia Steel, 1998, pp. 205-209
quality Control system for hot rolled strip", Steel Times International, Sept., 1998,
pp. 19-22
[50] S. Datta, J. Sil and M. K. Banerjee, "Petri Neural Network Model for the Effect of
of HSLA Steels", ISIJ International, Vol. 39 (1999), No. 10, pp. 986-990
Neural Network Models to Predict the Hot Strength of Steels", ISIJ International,
238
[52] O. P. Femminella, M. J. Starink, M. Brown, I. Sinclair, C. J. Harris and
Modelling
[53] Jiajun Wang, Pieter J. van der Wolk and Sybrand van der Zwaag, "Effects of
[54] Riidiger Doll, Giinter Sorgel, Matthias Daum and Custav Zouhar, "Control of
mechanical
[55] K. N. Melton, "The deformation models needed by the steel industry", Phil.
579–585
[57] Dongsheng Liu, Fateh Fazeli and Matthias Militzer, "Modeling of Microstructure
Evolution during Hot Strip Rolling of Dual Phase Steels", ISIJ International, Vol.
239
[58] B. Koczurkiewicz, "The model of prediction of the microstructure austenite C-Mn
Carbon Sheet Steels", ISIJ International, Vol. 47 (2007), No. 1, pp. 1–15
[60] R.A. Shulkosky, D.L. Rosburg, J.D. Chapman, K.R. Barnes, MS&T Conference,
(2003)
PUBLICATIONS/media/LONDON_96.pdf
[63] A.A. dos Santos and R. Barbosa, Steel Research Int. 81 (2010), 55
System for Controlling Mechanical Properties of Hot Rolled Coil", Proc. Int.
[65] Mao Y, Forging process design for risk reduction, Doctoral dissertation, The Ohio
[66] Hill, R. (1998). The Mathematical Theory of Plasticity. OUP (Oxford Classics
Series).
[67] Khan, A.S. and Huang, S. (1995). Continuum Theory of Plasticity. JohnWiley &
240
[68] Kobayashi, S., Oh, S., Altan, T., “Metal forming and the finite element method”,
[69] FORGE 2008 reference guide, 2008, Transvalor SA, Cedex, France.
[70] Fourment L., Chenot J.L., Mocellin, K., Numerical formulations and algorithms
for solving contact problems in metal forming simulation, Int. J. Num. Meth.
[72] H. Fujii, T. Ohashi, T. Hiromoto, “On the formation of internal cracks in the
continuously cast slabs”, Trans. ISIJ, 1978, vol. 18, pp. 510-518
[73] C. Ouchi and K. Matsumoto, “Hot ductility in Nb-bearing high strength low alloy
[75] J. Li, J. Shi, T.S Chang, (2007), “On Line Seam Detection in rolling processes
926-933.
[77] K. Mori, “General purpose FEM simulator for 3-D rolling”, Advanced technology
of plasticity, 4:1773-1778.
241
[78] J.J. Park, Oh, S.I., “Application of 3D FEM analysis to shape rolling process”,
[80] T. Altan, S. Kobayashi, N.S. Kim, “Three dimensional analysis and computer
simulation of shape rolling by the finite and slab element method”, Int. Journal of
[81] J.G. Lenard,J. Biglou, “A study of dynamic recrystallization during hot rolling of
with empirical models”, J. of Mat. Processing Tech., 130-131, 2002, Pg. 250-253
[84] H.A. Kuhn, “Workability theory and application in Bulk Forming processes”,
[85] Wei et al, “The study of dual response surface method in MDO”, Third
242
[87] P. Pauskar, R. Shivpuri, H. Cho and N. Kim, An FEM based model of metal flow
[88] Pauskar, P., Shivpuri, R., 1999, Microstructure and Mechanics Interaction in the
Modeling of Hot rolling of Rods, CIRP Annals, Vol. 48, Issue 1, p. 191-194
[89] Shivpuri, R., Shin, W., Lee, S.M., Altan, T., 1998. Finite slab element
[90] M. Ji and R. Shivpuri, Reduction of random seams in hot rolling through FEM
based sensitivity analysis, Material Science and Engineering A 425 (1–2) (2006),
pp. 156–166.
[92] Ginzburg, V., “Metallurgical design of flat rolled steels”, CRC press, 2005
[93] Ginzburg, V.B., Steel-rolling technology theory and practice, Marcel Dekker,
[94] H.C. Kwon, H.W. Lee, H.Y. Kim, Y.I. Im, H.D. Park and D.L. Lee, Surface
wrinkle defect of carbon steel in the hot bar rolling process,Journal of Materials
243
[95] Bayoumi, L.S., Lee, Y., Kim, H.S., 2002. Effect of roll gap change of oval pass
[96] Bayoumi, L.S., Lee, Y., 2004. Effect of interstand tension on roll load, torque and
[97] Shi, J., Li, J., 2007. Knowledge Discovery from Observational Data for Process
[98] N. Jin, S. Zhou and T.S. Chang, Identification of impacting factors of surface
[99] J. Li, J. Shi and T.S. Chang, On line seam detection in rolling processes using
[100] Huang, H., Gutches, D., & Chang, T.S. (2004). Imaging based in-line surface
defect inspection for bar rolling. In Proceedings of the AIST iron and steel
[101] Jia, H., Murphey, Y. L., Shi, J., & Chang, T. S. (2004). An intelligent real-time
vision system for surface defect detection. In Proceedings of the 17th international
244
[102] N. Jin, S. Zhou and T.S. Chang, Identification of impacting factors of surface
[103] J. Li, J. Shi and T.S. Chang, On line seam detection in rolling processes using
[104] Huang, H., Gutches, D., & Chang, T.S. (2004). Imaging based in-line surface
defect inspection for bar rolling. In Proceedings of the AIST iron and steel
[105] Karatzoglou, A., Smola, A., Hornik, K., & Zeileis, A. (2005). Kernlab – Kernel
stepwise procedure for building and training a neural network. In: J. Fogelman,
Verlag (1990).
[107] U. Kreßel, Pairwise classification and support vector machines. In: B. Scholkopf,
[108] http://accidents-ll.faa.gov
[111] http://en.wikipedia.org/wiki/Airworthiness
245
[112] S.K.Bhaumik, M. Sujata, M.A.Venkataswamy, “Fatigue failure of aircraft
[113] Francis Fagegaltier, “Engine Failure Training”, EATS 2009 – Prague, Powerpoint
presentation
Critical Innovations for Turbine Engines”, Key Engineering Materials Vol. 380
(2008) pp 113-134
[115] Boyer, R.R., 1998. Titanium and Titanium Alloys, in: Davis, J.R. (Eds.), Metals
[116] Donachie, M.J., 2000. Titanium: a technical guide, 2nd ed. ASM Int., Materials
Park, OH.
[117] Chan K. S., Perocchi, L., Leverant, G. R., Constitutive properties of Hard-Alpha
[118] Bazinger, T. J, Gigliotti Jr., M. F. X., Bewlay, B. P., Srivasta, S. K., Method for
[119] Semiatin, S.L., Nicolaou, P. D., Thomas, J. P., Turner, T. J., Defect occurrence and
[120] Semiatin, S.L., Montheillet, F., Shen, G., Jonas J.J., Self consistent modeling of
the flow behavior of Wrought alpha/beta titanium alloys under isothermal and non
246
[121] FAA/DOT, Turbine Rotor Material Design-Phase I: Final report. US Department
[122] FAA/DOT, Turbine Rotor Material Design-Phase II: Final report. US Department
[123] A.K. Koul, N.C. Bellinger and A. Fahr, “Damage-tolerance-based life prediction
[124] D.P. Shepherd, S.J. Williams, “New Lifing methodology for engine fracture
critical parts”, RTO AVT Symposium on “Ageing Mechanisms and Control: Part
B – Monitoring and Management of Gas Turbine Fleets for Extended Life and
[127] Kumar RS, Wang A-J, McDowell DL, “Effects of microstructure variability on
induced by grain size and orientation uncertainties, Acta Materialia 55, 2279-
2290.
[129] Tin S, Lee PD, Kermanpur A, Rist M, McLean M, Integrated Modeling for the
247
Manufacture of Ni-Based Superalloy Discs from Solidification to Final Heat
[130] Semiatin, S.L., Goetz, R.L., et al, Cavitation and Failure during hot forging of Ti-
[131] Torster, F., Baumeister, G., Alrecht, J., Lutjering, G., Helm, D., and Daeuler, M.A.
Influence of grain size and heat treatment on the microstructure and mechanical
[132] Zhou, Z., Roters, F., Raabe, D., Introduction of a Texture Component Crystal
[133] R Development Core Team (2008). R: A language and environment for statistical
[134] Gelman A, Carlin JB, Stern HS, and Rubin DB. Bayesian Data Analysis, London:
[135] Gelfand A.E., Smith A.F.M, “Sampling based approaches to calculating marginal
densities”, Journal of the American Statistical Association, Vol. 85, No. 410. pp.
398-409.
[136] Gelfand A.E., Smith A.F.M, “Gibbs sampling for marginal posterior
1747-1766.
248
APPENDIX
249
APPENDIX A
250
dat=cbind(vel,xdis,life)
library(BRugs)
N=5
#WinBUGS setup
modelCheck(fileName = "model.txt") # check model file
modelData(file.path("data.txt")) # read data file
modelCompile(numChains=1) # compile model with 1 chain
modelInits("inits.txt") # read init data file
modelGenInits()
#Run the chain for 5500 iteration and look at trace plots
ITER <- 10000
modelUpdate(ITER)
samplesHistory('*')
model
{
for(i in 1:N){
251
life[i] ~ dnorm(mu.l[i],var.l)
xdis[i] ~ dnorm(mu.x[i],var.x)
vel[i] ~ dnorm(mu.v[i],var.v)
mu.l[i] <- beta0 + beta1*xdis[i]
mu.x[i] <- alpha0 + alpha1*vel[i]
mu.v[i] ~ dnorm(0.0, 1.0E-10)
}
var.l ~ dgamma(0.001,0.001)
var.x ~ dgamma(0.001,0.001)
var.v ~ dgamma(0.001,0.001)
beta0 ~ dnorm(0.0,1.0E-10)
beta1 ~ dnorm(0.0,1.0E-10)
alpha0 ~ dnorm(0.0,1.0E-10)
alpha1 ~ dnorm(0.0,1.0E-10)
mu0=100
nu0=1000
alpha1=0.001
beta1=0.001
alpha2=0.001
beta2=0.001
alpha3=0.001
beta3=0.001
iter=20000
k=dim(cube)[3]
sigma.1=matrix(0,iter,3)
sigma.2=matrix(0,iter,5)
sigma.3=matrix(0,iter,8)
sigma.4=matrix(0,iter,8)
sigma.5=matrix(0,iter,53)
tau.1=rep(0,iter)
tau.2=rep(0,iter)
tau.3=rep(0,iter)
252
tau.4=rep(0,iter)
tau.5=rep(0,iter)
m=rep(0,iter)
alpha=rep(0,iter)
mu.1=rep(0,iter)
mu.2=rep(0,iter)
mu.3=rep(0,iter)
mu.4=rep(0,iter)
mu.5=rep(0,iter)
theta.1=matrix(0,iter,3)
theta.2=matrix(0,iter,5)
theta.3=matrix(0,iter,8)
theta.4=matrix(0,iter,8)
theta.5=matrix(0,iter,53)
len.m1=rep(NA,3)
len.m2=rep(NA,5)
len.m3=rep(NA,8)
len.m4=rep(NA,8)
len.m5=rep(NA,53)
for(i in 1:3)
len.m1[i]=length(na.omit(cube[,i,1]))
for(i in 1:5)
len.m2[i]=length(na.omit(cube[,i,2]))
for(i in 1:8)
{
len.m3[i]=length(na.omit(cube[,i,3]))
len.m4[i]=length(na.omit(cube[,i,4]))
}
for(i in 1:53)
len.m5[i]=length(na.omit(cube[,i,5]))
alpha[1]=100
m[1]=10000
tau.1[1]=1000
tau.2[1]=1000
tau.3[1]=1000
tau.4[1]=1000
tau.5[1]=1000
253
sigma.1[1,]=1000
sigma.2[1,]=1000
sigma.3[1,]=1000
sigma.4[1,]=1000
sigma.5[1,]=1000
mu.1[1]=100
mu.2[1]=100
mu.3[1]=100
mu.4[1]=100
mu.5[1]=100
theta.1[1,]=100
theta.2[1,]=100
theta.3[1,]=100
theta.4[1,]=100
theta.5[1,]=100
for(i in 2:iter){
for(j in 1:3){
tmp3=len.m1[j]/sigma.1[i-1,j] + 1/tau.1[i-1]
tmp2=(sum(cube[1:len.m1[j],j,1])/sigma.1[i-1,j] + mu.1[i-1]/tau.1[i-1])/tmp3
theta.1[i,j]=rnorm(1,tmp2,1/sqrt(tmp3))
}
for(j in 1:5){
tmp3=len.m2[j]/sigma.2[i-1,j] + 1/tau.2[i-1]
tmp2=(sum(cube[1:len.m2[j],j,2])/sigma.2[i-1,j] + mu.2[i-1]/tau.2[i-1])/tmp3
theta.2[i,j]=rnorm(1,tmp2,1/sqrt(tmp3))
}
for(j in 1:8){
tmp3=len.m3[j]/sigma.3[i-1,j] + 1/tau.3[i-1]
tmp2=(sum(cube[1:len.m3[j],j,3])/sigma.3[i-1,j] + mu.3[i-1]/tau.3[i-1])/tmp3
theta.3[i,j]=rnorm(1,tmp2,1/sqrt(tmp3))
}
for(j in 1:8){
tmp3=len.m4[j]/sigma.4[i-1,j] + 1/tau.4[i-1]
tmp2=(sum(cube[1:len.m4[j],j,4])/sigma.4[i-1,j] + mu.4[i-1]/tau.4[i-1])/tmp3
theta.4[i,j]=rnorm(1,tmp2,1/sqrt(tmp3))
}
for(j in 1:53){
tmp3=len.m5[j]/sigma.5[i-1,j] + 1/tau.5[i-1]
tmp2=(sum(cube[1:len.m5[j],j,5])/sigma.5[i-1,j] + mu.5[i-1]/tau.5[i-1])/tmp3
theta.5[i,j]=rnorm(1,tmp2,1/sqrt(tmp3))
254
}
s=rep(NA,5)
s[1]=sum((theta.1[i,]-mu.1[i-1])^2)
s[2]=sum((theta.2[i,]-mu.2[i-1])^2)
s[3]=sum((theta.3[i,]-mu.3[i-1])^2)
s[4]=sum((theta.4[i,]-mu.4[i-1])^2)
s[5]=sum((theta.5[i,]-mu.5[i-1])^2)
tau.1[i]=1/rgamma(1,alpha2+0.5*3,beta2+0.5*s[1])
tau.2[i]=1/rgamma(1,alpha2+0.5*5,beta2+0.5*s[2])
tau.3[i]=1/rgamma(1,alpha2+0.5*8,beta2+0.5*s[3])
tau.4[i]=1/rgamma(1,alpha2+0.5*8,beta2+0.5*s[4])
tau.5[i]=1/rgamma(1,alpha2+0.5*53,beta2+0.5*s[5])
tmp3=3/tau.1[i] + 1/m[i-1]
tmp2=(sum(theta.1[i,])/tau.1[i] + alpha[i-1]/m[i-1]) / tmp3
mu.1[i]=rnorm(1,tmp2,1/sqrt(tmp3))
tmp3=5/tau.2[i] + 1/m[i-1]
tmp2=(sum(theta.2[i,]) / tau.2[i] + alpha[i-1] / m[i-1]) / tmp3
mu.2[i]=rnorm(1,tmp2,1/sqrt(tmp3))
tmp3=8/tau.3[i] + 1/m[i-1]
tmp2=(sum(theta.3[i,]) / tau.3[i] + alpha[i-1] / m[i-1]) / tmp3
mu.3[i]=rnorm(1,tmp2,1/sqrt(tmp3))
tmp3=8/tau.4[i] + 1/m[i-1]
tmp2=(sum(theta.4[i,]) / tau.4[i] + alpha[i-1] / m[i-1]) / tmp3
mu.4[i]=rnorm(1,tmp2,1/sqrt(tmp3))
tmp3=53/tau.5[i] + 1/m[i-1]
tmp2=(sum(theta.5[i,]) / tau.5[i] + alpha[i-1] / m[i-1]) / tmp3
mu.5[i]=rnorm(1,tmp2,1/sqrt(tmp3))
tmp.a=alpha1+2.5
tmp.b=beta1+0.5*((mu.1[i]-alpha[i-1])^2+(mu.2[i]-alpha[i-1])^2+(mu.3[i]-alpha[i-
1])^2+(mu.4[i]-alpha[i-1])^2+(mu.5[i]-alpha[i-1])^2)
m[i]=1/rgamma(1,tmp.a,tmp.b)
tmp3=5/m[i] + 1/nu0
tmp2=((mu.1[i]+mu.2[i]+mu.3[i]+mu.4[i]+mu.5[i])/m[i] + mu0/nu0)/tmp3
alpha[i]=rnorm(1,tmp2,1/sqrt(tmp3))
for(j in 1:3)
255
sigma.1[i,j]=1/rgamma(1,alpha3+0.5*len.m1[j],beta3+0.5*sum((cube[1:len.m1[j],j,1]-
theta.1[i,j])^2))
for(j in 1:5)
sigma.2[i,j]=1/rgamma(1,alpha3+0.5*len.m2[j],beta3+0.5*sum((cube[1:len.m2[j],j,2]-
theta.2[i,j])^2))
for(j in 1:8)
sigma.3[i,j]=1/rgamma(1,alpha3+0.5*len.m3[j],beta3+0.5*sum((cube[1:len.m3[j],j,3]-
theta.3[i,j])^2))
for(j in 1:8)
sigma.4[i,j]=1/rgamma(1,alpha3+0.5*len.m4[j],beta3+0.5*sum((cube[1:len.m4[j],j,4]-
theta.4[i,j])^2))
for(j in 1:53)
sigma.5[i,j]=1/rgamma(1,alpha3+0.5*len.m5[j],beta3+0.5*sum((cube[1:len.m5[j],j,5]-
theta.5[i,j])^2))
burn=2000
iter=13000
plot(density(theta.2[burn:iter,4]),type="l")
plot(density(theta.2[burn:iter,5]),type="l")
plot(density(theta.3[burn:iter,1]),type="l")
plot(density(theta.3[burn:iter,2]),type="l")
plot(density(theta.3[burn:iter,3]),type="l")
plot(density(theta.3[burn:iter,4]),type="l")
plot(density(theta.3[burn:iter,5]),type="l")
256
plot(density(theta.4[burn:iter,1]),type="l")
plot(density(theta.4[burn:iter,2]),type="l")
plot(density(theta.4[burn:iter,4]),type="l")
plot(density(theta.4[burn:iter,5]),type="l")
boxplot(mu.1[burn:iter],mu.2[burn:iter],mu.3[burn:iter],mu.4[burn:iter],mu.5[burn:iter],o
utline=FALSE,main="Posterior distribution of Theta's",ylim=c(0,600),ylab="Quality
Index",names=c("A","B","C","D","E"),xlab="Material")
boxplot(theta.1[burn:iter,1],theta.1[burn:iter,2],theta.1[burn:iter,3])
summary(mu.2)
10.2127 105.161
25.7375 14.8595 100.878
36.5818 13.0907 94.5419
30.5041 6.39036 101.689
39.6123 0 96.9613
33.469 0 102.623
37.4097 13.5358 78.8759
46.1026 10.6148 84.4603
46.0629 3.92841 81.1142
53.273 11.8919 83.129
40.0372 0.887925 82.4881
40.1289 1.41275 82.4797
40.6594 1.22728 82.4716
40.4555 0.762992 82.0532
40.5064 0.235275 82.3861
40.0084 0.411922 82.3399
40.4412 0.481929 82.6298
40.879 1.20414 83.4173
41.1623 0.644899 83.2984
40.8447 0.845184 83.553
40.8312 0.906175 83.2796
40.7129 0.652144 82.4823
257
40.8418 0.260191 82.517
46.0687 19.5963 84.3468
39.1396 2.64206 83.7632
38.9702 2.81152 84.7428
38.9732 2.54031 83.9901
39.3506 2.58081 84.558
38.0398 2.75274 83.5967
38.1069 2.57656 83.9931
37.8115 2.54593 83.6555
38.1086 2.58414 83.5535
39.3071 2.24713 82.6007
39.6332 2.04035 82.658
39.2572 2.03622 82.649
39.5702 1.95657 82.4846
40.9599 1.90984 82.1831
40.4787 1.66456 82.6939
40.8884 1.40464 82.7621
38.8636 1.12608 82.0795
39.4331 1.27716 82.1941
40.0411 1.59944 80.9535
38.6252 2.3729 81.7028
37.832 1.46649 82.6255
37.4713 1.62698 82.4499
37.5964 1.79584 82.646
37.6605 1.87011 82.207
37.6278 1.22228 82.4519
37.2738 2.1341 82.5367
37.028 1.72188 82.4548
38.393 2.60918 83.8482
39.3233 2.26307 82.6719
38.7484 2.87431 83.7494
38.6829 2.5788 83.7138
38.7802 2.68548 84.0758
38.64 2.57842 84.0389
38.5989 2.49667 82.5021
38.838 2.02337 82.4395
39.3184 2.63173 82.4609
40.3646 0.746725 82.4846
39.1397 0.782387 82.6048
39.5806 0.653681 82.6075
39.1507 0.65685 82.2877
20.5368 5.57533 84.8566
26.4135 0 84.2682
14.5577 0 82.5437
258
28.244 0 76.0797
40.1528 1.59876 82.0837
39.5677 0.588994 81.8837
38.639 0 81.3967
40.2484 0.941102 82.6291
40.547 0.815202 82.6442
39.8377 0 81.5759
40.5856 0 81.8422
40.759 0.652385 81.6603
38.3472 0 82.2083
37.5237 0 82.6147
38.2052 0 82.6054
38.247 0.321969 82.3236
38.2854 1.08617 82.29
39.1161 0 82.0826
40.3636 1.70872 82.5045
40.2964 2.10108 82.6081
38.8847 0 82.6026
39.3794 0 82.6069
37.0704 2.46138 83.2328
36.2331 2.54126 83.2351
36.6298 2.28245 83.3501
37.2742 2.72177 83.9195
38.9747 2.53511 84.7066
39.2947 2.57354 85.0676
38.9761 2.53594 85.2288
41.0266 0 82.0228
41.1182 0.827315 82.3653
41.112 0.261802 82.5472
42.0391 0 81.9321
40.2246 0 82.2625
40.0159 2.03434 84.2304
40.1403 2.11629 84.1972
40.0163 2.03367 84.627
40.1895 1.86037 84.4357
40.2602 1.87224 82.6827
40.0185 1.70054 82.6506
40.0463 2.0808 82.6805
38.7218 1.43164 82.3921
38.1762 1.87529 82.4617
38.2148 1.69338 82.6292
38.6942 2.56894 84.5327
38.9741 2.53589 84.4335
40.1436 2.27198 82.3107
259
39.9336 2.04882 82.5837
37.7537 0 81.9715
40.1564 0.497887 82.6229
40.7897 1.156 82.6821
37.2588 2.35643 83.5135
37.4943 2.58372 83.4203
37.5342 2.45795 83.4132
37.2517 2.37058 83.0802
34.647 20.0035 146.357
42.8679 13.5326 147.597
41.9973 15.2319 138.389
40.6975 23.4967 139.62
36.5939 0.597633 82.3971
35.9998 0.95655 82.4024
36.309 1.37345 82.0856
35.6314 0.832439 82.1483
41.1852 0.814143 82.9241
40.9829 0.592438 82.676
40.8835 0.827581 82.8549
14.4787 0 72.5505
20.7377 11.9729 78.7569
10.9673 6.33196 81.1596
41.0092 0 83.2079
41.1486 0.129921 83.2132
41.3199 0 83.1064
41.1376 0.314018 82.9619
40.545 1.86822 82.8086
41.1946 1.50936 82.771
41.3409 2.25103 82.8936
37.7241 2.25255 82.5435
37.9582 2.17071 82.6584
40.1571 2.96943 82.6604
41.6227 3.59018 80.5979
42.5776 2.61804 83.4428
42.6345 1.23587 82.8936
41.609 0 82.7258
41.05 0 82.6803
45.1186 0 82.9672
38.9839 2.60433 82.6992
39.0173 2.2924 82.6636
41.6802 0.773448 81.1899
38.6624 0.868293 82.6061
39.0038 1.56345 82.6255
39.1301 1.79937 82.6362
260
39.3066 1.69783 82.5072
39.9401 2.30576 82.8376
39.7991 2.24351 82.6883
39.8678 1.01544 82.6235
39.6373 1.22915 82.4465
37.6478 2.51886 83.0208
37.8087 2.55784 83.2684
37.7984 2.60392 82.7102
37.6721 2.75407 82.8697
41.2143 0.300816 83.4313
39.6862 1.51045 82.631
41.2819 0 83.4141
36.9767 2.53266 82.6986
36.754 2.05505 83.1302
36.7254 2.09228 82.7208
43.4583 0 80.9494
42.819 0 82.7535
38.1767 2.31883 82.503
37.688 2.74481 82.4534
38.9741 2.55612 83.2722
38.686 2.59536 83.1822
38.723 2.58102 83.4273
38.5472 2.76522 83.345
41.0148 0.296239 82.6731
40.7587 0.502032 82.6513
35.6056 4.28907 82.2932
36.2594 2.67938 84.4268
39.2556 2.44598 83.7146
39.2567 2.44019 84.1173
41.7109 0.578845 83.1827
39.5236 2.34523 84.1836
39.7202 2.38107 84.5503
39.5239 2.34311 84.6297
39.786 2.18012 84.2417
38.6748 2.57165 85.4792
38.6683 2.57088 85.0874
38.7208 2.70294 85.4341
38.5934 1.88684 82.6372
38.2952 2.0275 82.6439
39.2922 2.4249 84.5687
39.7198 2.71621 84.2433
38.3918 2.62004 83.361
38.138 2.59635 83.207
38.7552 2.29573 82.6625
261
38.3925 2.66894 82.7027
21.3811 0 93.1342
39.5245 2.34914 83.8155
39.8769 2.28693 83.7804
36.6666 1.32768 82.5319
36.3837 1.18818 82.6729
36.7099 1.24618 82.6581
36.6365 1.50212 82.6729
40.0456 1.31608 82.6364
36.7079 2.72797 81.8683
37.5512 2.68337 79.6892
40.471 2.49925 83.529
40.6616 1.58616 83.2116
40.4358 1.63038 83.3933
40.3021 2.06356 83.1344
39.0333 2.68494 83.0664
41.069 0 80.9598
36.3295 1.0836 82.5396
36.7832 0.990342 82.6444
39.6854 0 82.2262
39.743 0.297644 82.6122
38.9763 2.56911 83.0112
36.3046 1.66386 82.7167
36.4011 2.04553 83.0537
36.5386 1.85297 83.02
36.5193 1.87679 82.7114
38.7682 2.56087 84.2649
38.6292 2.57896 85.8181
38.3976 2.60333 85.6014
39.8753 0 82.614
38.6852 2.61391 82.9338
39.5485 2.40676 82.6982
34.9901 0 82.1973
34.9771 2.3148 82.4559
35.5593 0 80.7421
32.5752 0 81.4404
38.9777 2.53895 85.7719
40.9615 0.587575 82.8613
40.6771 0 77.88
35.0817 0 77.1659
37.5671 0 80.0328
39.5384 2.38558 82.8808
39.5752 2.57163 83.1603
39.2328 2.48818 82.9191
262
36.4783 2.69038 82.5151
35.9069 2.15193 82.8749
40.6672 0 82.6393
38.6275 2.87962 82.7406
39.8857 2.15276 83.0092
39.6349 2.43115 83.5621
40.7631 0 82.3972
38.102 0.598513 82.6032
37.9671 1.03241 82.6132
40.2814 1.96857 83.6569
40.5747 1.72666 83.8266
40.2515 1.79991 84.0889
40.2213 1.83663 83.6755
40.9863 0.283284 83.0401
37.339 2.12462 82.6727
37.6175 2.36053 82.6849
40.3911 1.44205 82.6681
38.0962 2.63417 82.7041
39.7269 2.216 84.6644
39.5246 2.34313 84.8655
39.7816 2.18159 84.9553
41.0249 0 82.9597
36.4823 0 81.0659
35.8619 0 81.6113
39.276 2.44632 83.3709
40.9763 0.274379 83.3287
38.3902 1.40685 82.6215
40.4296 1.62446 83.8334
40.4276 1.62291 84.1294
37.4272 1.56398 82.6424
40.6007 0.245501 82.6371
40.8792 0.796029 83.057
40.0165 2.03585 83.8838
40.4468 1.64 83.107
40.2558 1.83003 83.0568
40.4666 1.65539 82.8301
40.0945 0.255304 82.6206
39.1602 2.54336 82.6984
39.753 2.20619 83.7951
40.9436 0.583504 83.2329
40.0202 2.04104 83.5422
37.9975 1.85491 82.6412
38.8355 2.55131 84.8791
38.6592 2.57167 84.7827
263
38.378 0.342526 82.6042
37.8344 0.395917 82.6103
39.9523 2.29619 83.2384
40.0247 2.04728 83.3006
39.5553 1.1403 82.6207
39.8768 2.13953 83.311
39.5288 2.36189 83.2081
40.7586 1.13986 83.0294
40.7423 1.13253 83.5486
39.5235 0 80.41
37.577 1.21869 82.6273
40.9376 0.581756 83.4948
38.9329 2.03642 82.6449
40.2147 1.85659 83.3148
38.1372 2.88459 82.8808
38.105 2.60837 82.9743
40.7651 0.703615 82.6597
39.2726 2.43175 85.0479
40.0369 2.06415 82.8696
30.9856 4.86521 79.3406
40.2699 0 82.625
37.7003 0.890233 82.6149
37.4335 0.747991 82.6166
41.7363 2.97032 92.0481
49.8873 3.33287 88.3553
46.325 0 93.4975
55.3413 0 94.374
48.6374 0 84.5032
56.6001 0 84.916
47.4269 16.4341 92.1545
41.323 23.8578 103.661
40.7397 1.13094 83.8318
35.4072 20.4424 94.2455
49.6062 20.9546 89.5803
76.6762 10.4362 107.968
86.6286 12.0947 105.426
84.9299 20.9712 106.817
85.5705 17.9529 111.066
39.2953 2.42461 85.3809
39.5259 2.34366 85.3005
33.695 8.51611 121.036
32.4942 0 124.789
30.6928 8.34977 129.324
24.5523 6.44869 120.552
264
49.0209 0 79.1898
39.5251 2.35191 83.5953
38.1016 2.81442 83.2095
36.2998 0.768913 82.6617
36.625 0.622891 82.6459
36.257 0.423438 82.6612
37.3604 1.88733 82.6567
38.397 2.60204 85.3664
38.9736 2.54847 83.5499
47.2777 27.2958 98.8739
51.3764 9.38095 93.9971
40.6231 1.38619 82.8544
40.3804 1.20233 82.6472
46.0314 5.70711 72.9451
38.4939 2.30218 82.6616
38.3917 2.6329 83.0725
40.7368 1.1292 84.1778
40.9908 0.982775 84.174
40.9043 1.3055 84.6457
40.7352 1.12805 84.6239
27.0427 0 97.8649
38.2105 2.36235 82.6668
37.465 2.46132 82.8854
72.8001 33.0423 102.157
59.8235 34.5391 102.335
66.4759 38.3799 110.004
65.6018 37.8752 93.7356
36.2149 1.80838 83.5993
39.748 2.36579 85.467
39.8422 2.14588 85.6296
39.9883 2.36434 85.9892
39.5264 2.3439 85.6104
84.8932 21.1774 118.03
75.539 11.1674 123.239
73.7285 12.23 115.346
86.6774 11.6995 117.383
49.3466 7.6771 119.25
62.328 7.9547 119.574
56.1623 8.47389 112.459
55.282 17.9324 113.4
66.2242 12.0136 107.565
65.0922 16.1525 117.415
73.7044 22.2354 108.83
66.2261 23.2841 113.314
265
9.08239 5.24373 110.855
16.9679 9.79641 117.535
7.91913 4.57212 120.248
12.1732 0 120.673
35.8118 0.515833 82.5214
35.9327 0 82.0105
30.4273 17.5672 84.1297
21.5555 12.4451 89.4576
54.5013 14.807 138.05
66.6497 20.0044 143.103
50.7884 22.5374 138.494
50.0592 21.3946 148.965
10.5397 6.08512 89.9826
71.8971 31.9361 116.848
78.605 38.3884 110.342
75.7745 43.7484 113.868
40.6197 1.33759 83.3406
57.8515 33.4006 113.817
64.9986 25.5208 105.857
57.5638 0 123.633
49.6211 0 112.554
38.5333 0 119.23
36.6268 0 131.311
72.8772 23.6173 121.425
83.512 26.0482 114.589
82.9252 27.5464 119.048
40.5577 1.46075 83.089
37.3679 9.66936 113.552
46.5451 17.2806 121.61
-1.43203e-005 1.4671e-021 115.772
40.5885 1.37712 83.6569
40.2022 23.2107 111.502
36.9122 21.3113 122.251
48.746 28.1435 117.41
24.8289 14.335 123.915
40.5841 1.37605 83.9951
46.4924 0 125.606
51.0797 29.4909 136.848
41.3674 23.8834 131.903
48.3849 0 103.173
13.5408 7.81777 130.488
29.2238 16.8724 135.036
37.2234 1.29111 82.6388
37.178 1.16283 82.3605
266
24.2283 7.50458 91.3914
9.83679 0 90.8675
36.969 0.404422 82.6313
36.9932 0.801885 82.635
37.0535 1.71468 82.6598
37.0524 1.99048 82.678
36.8633 2.09995 82.541
32.4751 0 92.61
40.7556 1.36916 84.0556
46.6306 26.9222 148.654
59.1977 26.9601 144.922
39.0887 3.29282 85.6227
40.4118 3.88216 84.0502
38.6006 3.51419 83.1623
38.9324 9.33703 85.7909
40.766 2.89778 84.0668
42.5354 3.93592 85.1354
41.1375 0.717524 83.7953
33.3669 2.64519 84.8769
40.4267 1.62189 84.4019
40.5783 1.69975 84.4843
11.4614 0 142.837
18.0249 5.27394 141.154
9.74618 5.62697 143.65
36.7749 0 82.6454
37.0971 0 82.4249
37.0477 0.296468 82.4031
-1.43203e-005 1.4671e-021 138.129
41.3993 1.41157 83.9456
35.7335 2.19887 83.9445
41.2159 0.193177 83.7219
41.8123 0 83.6108
41.3085 0 83.7291
37.4385 2.90038 83.126
37.5411 2.47274 83.1337
36.5536 1.88368 82.5407
35.9772 1.59379 82.5411
37.5129 0.801459 82.0343
36.5107 0.173551 82.2681
36.2461 0.203412 82.4855
267