Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson...

39
Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor, Materials Science & Aerospace Engineering, Iowa State University

Transcript of Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson...

Page 1: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

Exploring the Implications of Bayesian Approach to Materials State Awareness

R. Bruce ThompsonDirector, Center for Nondestructive Evaluation

Professor, Materials Science & Aerospace Engineering,

Iowa State University

Page 2: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 2

Outline

Interpretation of Current Status of and Future Needs for Prognosis

Microstructural Characterization Sensors Integration within Bayesian Framework A Conceptual Illustration Conclusions

Page 3: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 3

L. Christodoulou and J. M. Larsen, “Using Materials Prognosis to Maximize the Utilization Potential of Complex Mechanical Systems,” Materials Damage Prognosis, J. M. Larsen, L. Christodoulou, J. R. Calcaterra, M. L. Dent, M. M. Derriso, J. W. Jones, ad S. M. Russ, Eds. (TMS, 2005).

Page 4: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 4

L. Christodoulou and J. M. Larsen, “Materials Damage Prognosis: A Revolution in Asset Management,” Materials Damage Prognosis, J. M. Larsen, L. Christodoulou, J. R. Calcaterra, M. L. Dent, M. M. Derriso, J. W. Jones, ad S. M. Russ, Eds. (TMS, 2005). (adapted from Cruse)

Long term Advanced MaterialState Sensing

MesomechanicalDamage Models

CharacterizeMaterial

Microstructures

Full-Authority DigitalEngine Control (FADEC)

Math Model Mission Simulation

Ready

Short term

Application

Decision Capability forLegacy Engines

Lifing Algorithms

Analytical Stress Model

Installed AutonomousSensorsLong term

Logic for Integrated, Automated Prognosis System

Page 5: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 5

New Ingredients

“In many ways, materials damage prognosis is analogous to other damage tolerance approaches, with the addition of in-situ local damage and global state awareness capability and much improved damage predictive models”

L. Christodoulou and J. M. Larsen, “Materials Damage Prognosis: A Revolution in Asset Management,” Materials Damage Prognosis, J. M. Larsen, L. Christodoulou, J. R. Calcaterra, M. L. Dent, M. M. Derriso, J. W. Jones, ad S. M. Russ, Eds. (TMS, 2005).

Page 6: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 6

In principle, we simply need to execute the following strategy

This would be a “done deal” if the input data were correct/complete and models were of sufficient accuracy and computationally efficient.

Utopian View

DamageProgression

Model

DamagedState

InitialState

OperationalEnvironment

FailureModel

ExpectedLifetime

FailureCriteria

Page 7: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 7

Barriers to Reaching Nirvana

Missing information Do not currently determine the initial state of individual

components/structures/systems with high precision Have not traditionally monitored the operating environment

of individual components Damage progression models have traditionally been

empirical (e.g., Paris Law) Difficult to incorporate the missing information if it were

available Uncertainty

There will always be uncertainty in the input data Variability

Even if we eliminate uncertainty, we would have to take variability into account

Page 8: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 8

Examples of Research Underway and Gaps Operational environment

Temperature, strain and chemical sensors under development State sensing data

Global Structures: strain, displacement, acceleration Propulsion: vibration analysis

Local Guided waves to sense structural changes Moisture Ultrasonic, eddy current, … to sense microstructure

Damage models Under refinement in many programs

Page 9: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 9

Long Term Microstructural Sensor Needs Improved sensor and data interpretation

procedures to monitor evolution of microstructure during damage A key will be a well-developed, quantitative

understanding of relationship of sensor response to microstructural changes

Physics-based models of the sensing process Must work subject to practical constraints

Access Survivability Simplicity of implementation

Page 10: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 10

Systems perspective to integrate all of the NDE state data with damage model predictions Depot, field, on board sensors Global, local sensors Measurements of initial state, damage state

Must recognize fundamental difference in data structure for traditional (depot and field) and on board NDE measurements

Long Term Integration Needs

Space

Traditional NDE providesinformation as a functionof position at discrete timesT

ime

On board sensors provide information as a function of time at discrete locations

Space

Traditional NDE providesinformation as a functionof position at discrete timesT

ime

On board sensors provide information as a function of time at discrete locations

Traditional NDE providesinformation as a functionof position at discrete timesT

ime

On board sensors provide information as a function of time at discrete locations

Page 11: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 11

Outline

Interpretation of Current Status of and Future Needs for Prognosis

Microstructural Characterization Sensors Integration within Bayesian Framework A Conceptual Illustration Conclusions

Page 12: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 12

Detailed Understanding of Microstructure must be a Key Ingredient in Development of State Awareness Strategies

An idealized scenario

Generally, each link has it challenges Non-uniqueness Inadequate sensitivity to key parameters Limitations of the theory base

Force a stochastic approach

Page 13: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 13

Need for Microstructural Characterization Tools as Well as Flaw Detection Tools Need to be able to assess the progression of

damage before cracks form Quantification of initial state Check of evolution of damage when possible

Validation of prognostic calls

Page 14: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 14

Incidentsoundpulse

100 m

Single crystal

(“grain”)

Grain boundary echoes

Characterization of Grain Morphology

The reflection of sound at grain boundaries results in “noise” seen in UT inspections

Page 15: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 15

Time Domain Waveforms

Page 16: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 16

Characterization of Grain Structure

Grain noise inhomogeneity provides information about microstructure

Page 17: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 17

Characterization of Grain Structure

Ultrasonic backscattering controlled by grain size

Theoretical base exists to quantify relationship (single scattering assumption)

Page 18: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 18

Characterization of Grain Structure

Determining grain size and shape from single sided backscattering measurements

Page 19: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 19

Characterization of Grain Structure Results obtained on rolled and extruded

aluminum

Page 20: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 20

Characterization of Fatigue Damage Normalized Harmonic Ratio -vs- Percent Low Cycle Fatigue Life*

Ni-based Aero Engine Alloy

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

0 20 40 60 80 100 120Fatigue Life (Percent)

N3 - 51 ksi - =180kN4 - 47 ksi - =302kN7 - 47 ksi - =290k

* 100 % is last data point prior to first detection of surface crack

N

fN

fN

f

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

0 20 40 60 80 100 120

Nor

mal

ized

Har

mon

ic R

atio

(A2/

A12 )/

(A2/

A12 ) u

nfa

tigu

ed

Page 21: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 21

The Way Forward

Significant benefits can be obtained from further developing nondestructive microstructural characterization tools Best developed if seek relationship to

microstructure rather than properties Need physics-based, rather than empirical

understanding Needs collaboration of measurement and materials

experts

Page 22: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 22

Some Open Questions

Role of precipitates and grain boundary decorations in ultrasonic and backscattering measurements

Role of dislocations in attenuation measurements

Relative roles of dislocations and microcracks in harmonic generation

Page 23: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 23

Outline

Interpretation of Current Status of and Future Needs for Prognosis

Microstructural Characterization Sensors Integration within Bayesian Framework A Conceptual Illustration Conclusions

Page 24: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 24

The Bayesian Approach

The essence of the Bayesian approach is to provide a mathematical rule explaining how you should combine new data with existing knowledge or expertise From an intuitive perspective, we can consider the “utopian view” that we

discussed previously as existing knowledge The new data are the results of NDE measurements about initial state,

operational environment, or the state of damage evolution This approach addresses the non-uniqueness problem that plagues

the interpretation of many NDE measurements A framework for data inversion

Enabling technologies are Physics-based models of the NDE measurement process High speed computational capability that makes implementation practical

(not the case a decade ago)

Page 25: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 25

Traditional Data Inversion Consider a model relating input parameters (state of

material or flaw) x, to experimental observations, y, where y and x are vectors

In principle, y might be a global or local variable One way to “invert” data is to adjust x to maximize the

pdf, p(y/x) One seeks parameter values that maximize the probability of

the observed data We do this all at the time in making least square fits to data Need more observations than unknown parameters in order

for this to work

y = m xobservation

material state parameters (e.g., flaw size)

Page 26: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 26

Likelihood: Direct Use in Inversion

In the language of the likelihood approach, is proportional to the likelihood function

Sometimes written We seek to choose the values of x such that the likelihood

is maximized These values are considered best estimates of x

In special cases, this approach is equivalent to the more familiar least squares fitting procedures y normally distributed about mean values No systematic errors in models (model predicts mean

values) No truncated or censored data

p y x x ;yL or L x

Page 27: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 27

Limitations of this Approach to Inversion This approach (including least squares fitting) breaks down if

Data is not sufficient to determine parameters without auxiliary information or assumption (i.e., solutions of inverse problem would not be unique)

One wishes to incorporate knowledge from past experience in a systematic way

One wishes to estimate probability of parameter values (not just most likely values)

Bayes Theorem provides a path forward Allows direct incorporation of physical understanding of

processes (e.g., as incorporated in physics-based simulation tools)

Significant computations may be required “Computational plenty” is reducing this objection

Page 28: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 28

Bayes Theorem for Continuous Variables

Note: Physical understanding of the measurement, ideally as captured by a physics-based model, enters through the likelihood p(y/x). “How likely was the observed state data for possible states in the prior distribution”

( / ) ( )( / )

( / ) ( )

f y x f xf x y

f y s f s ds

Likelihood of x p(y/x) Prior distribution of x

Posterior pdf Normalization

Page 29: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 29

Summary of Bayesian Approach

Advantages Framework to utilize “prior” knowledge

Update beliefs about probability of state in light of new evidence, the measurement results y

Provides “posterior” (probability distribution of state), not just most likely state

Depends in a simple way on the “likelihood”, something that can be computed from forward models

Issues Significant computations Dependence on the prior

Posterior may not be highly sensitive to this Sensitivity studies needed

Page 30: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 30

An Intuitive Description

The prior contains our knowledge about the materials state that is expected to be present In one way or the other, we often make such assumptions

in a less formalized way “If the defect were a crack, it would have the following size”

We use the measurement results to determine which of those possible states are most consistent with the data In essence, ruling out the portions of the prior distribution

that are inconsistent with the observations The posterior is the sharpened distribution of states

that emerges

Page 31: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 31

Generalization to Failure Prediction

Probabilistic model for P(x,y,c) x: state of defect y: measured data c: 1 if piece survives under specified conditions

0 if piece fails under specified conditions From this model, want to infer the probability of failure (c) given the NDE data

failure model NDE data inversion

Note: P(x/y) will depend on the accept/reject criterion

( / ) ( / ) ( / )P c y P c x P x y dx

Richardson

Page 32: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 32

Effects of Randomness and Completeness

One measurement Failure uncertainty Measurement uncertainty

One measurement Failure perfect Measurement perfect

Complete measurement Failure uncertainty Measurement perfect

false accepts false accepts false accepts

fals

e re

ject

s

fals

e re

ject

s

fals

e re

ject

s

Page 33: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 33

Outline

Interpretation of Current Status of and Future Needs for Prognosis

Microstructural Characterization Sensors Integration within Bayesian Framework A Conceptual Illustration Conclusions

Page 34: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 34

Waspalloy Disk

“The scatter in material behavior is attributed to the inhomogeneous microstructure elements with metals.”

L. Nasser and R. Tryon, “Prognostic System for Microstuctural-Based Reliability”, DARPA Prognostics web site(with reference to work at Cowles, P&W)

Page 35: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 35

Microstructural Fatigue Model

Page 36: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 36

Potential Sensor Assistance at Various StagesStage of Fatigue

Potential Measurement Status of Scientific Foundation

Implementation Issues

Crack nucleation

Grain size determination by UT backscatter after manufacturing

Well established for single phase materials

Effects of precipitates and grain boundary decorations under study

No major “show stoppers”

Short crack growth

Ultrasonic harmonic generation

Mechanisms for engineering materials under study (dislocations vs. microcracks as sources)

Very challenging measurement on wing

Long crack growth

Deploying tradition NDE in-situ

Broad foundations in place

Effects of morphology e.g., closure, subject of ongoing study

Challenging measurement of wing

Page 37: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 37

At the End of the Day(In this or other applications) When we balance

Our improving but incomplete understanding of failure processes

The ideal characterization procedures based on understanding of the measurement physics

The measurement possibilities as constrained by practical constraints

We will be making prognoses based on incomplete information

Exact data inversion will not be possible Suggest use of Bayesian statistics to eliminate

possible outcomes inconsistent with sensor data

Page 38: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 38

Outline

Interpretation of Current Status of and Future Needs for Prognosis

Microstructural Characterization Sensors Integration within Bayesian Framework A Conceptual Illustration Conclusions

Page 39: Exploring the Implications of Bayesian Approach to Materials State Awareness R. Bruce Thompson Director, Center for Nondestructive Evaluation Professor,

AFOSR Prognosis Workshop_February 2008 39

Conclusions

Realizing a full Materials State Awareness capability will require a wide range of inputs Mesoscopic damage models Sensing of operational parameters of individual components Advanced material state sensing

Needs physics-based understanding of relationship to microstructure

Constrain by access, survivability, need for simplicity Bayesian statistics provides an attractive framework

for integrating these disparate inputs Enabled by physics-based models of the measurement

process A conceptual example based on aircraft engine disks

was provided