Sie sind auf Seite 1von 13

Uncertainty Quantification in Prognostic Health

Management Systems
H. Heath Dewey Derek R. DeVries Scott R. Hyde
Northrop Grumman Northrop Grumman Northrop Grumman
Innovation Systems Innovation Systems Innovation Systems
9160 N Hwy 83 9160 N Hwy 83 9160 N Hwy 83
Corinne, UT 84307 Corinne, UT 84307 Corinne, UT 84307
435-863-2375 435-863-6693 435-863-6307
Heath.Dewey@ngc.com Derek.DeVries@ngc.com Scott.Hyde@ngc.com

Abstract – Uncertainty plays a role in nearly all aspects of Systems engineering plays a vital role in this technical
prognostic health management (PHM) systems. Aleatory process. The systems engineering processes “are used to
uncertainty from inherently-variable inputs such as material define the requirements for a system, to transform the
properties, epistemic uncertainty from a lack of knowledge requirements into an effective product, to permit consistent
about the system and its inputs, and ontological uncertainty due
reproduction of the product where necessary, to use the
to completely unknown factors must all be accounted for in
order to provide the most accurate assessment of the health of product to provide the required services, to sustain the
the monitored system. Northrop Grumman Innovation Systems provision of those services, and to dispose of the product
(NGIS) develops, produces, and provides sustainment of solid when it is retired from service.” [1]
rocket motor systems for the aerospace and defense industry
and has extensive experience in applying uncertainty These processes define the activities to optimize the designs
quantification (UQ) principles to complicated numerical and reduce the risks associated with technical decisions
simulations and analyses. In this paper, lessons learned by during the development, manufacturing, fielding, operation,
NGIS on UQ simulations and analyses are presented, and their and removal from service of the products. These processes
applicability to PHM systems is explored. Methods for also define, verify, validate, and control the “the timeliness
measuring and tracking uncertainty through the PHM
and availability, the cost effectiveness, and the functionality,
predictive train are presented, as is a Monte-Carlo-based
method for performing prognostic numerical calculations, reliability, maintainability, producibility, usability and other
which accounts for and quantifies epistemic, ontological, and qualities” [1] of the system. An extremely important part of
aleatory uncertainty. this process is to accurately ascertain the ability of the
product to meet the required performance throughout its life,
including the capture of information about the product
TABLE OF CONTENTS through testing, component and system model analysis, and
collection of information from the system’s performance.
1. INTRODUCTION ....................................................... 1 The captured system information contains key performance
2. ARCHITECTURE OF A PHYSICS-BASED PHM information or parameters as defined by the system’s
SYSTEM ....................................................................... 2 components and material performance models.
3. UNCERTAINTY QUANTIFICATION .......................... 2
4. SOURCES OF UNCERTAINTY.................................. 3 If information is collected incorrectly or inaccurately or the
method of collection does not provide the information to
5. CATEGORIES OF UNCERTAINTY .......................... 4 accurately model the behavior, then the performance of the
6. UNCERTAINTY IN PHM SYSTEMS ........................ 5 model will have increased uncertainty (may not be
7. CALCULATION WITH UNCERTAINTY ................... 8 understood or cannot be quantified) or may not represent the
8. CONCLUSIONS ..................................................... 10 expected behavior at all.
REFERENCES............................................................. 11
In today's environment of customers demanding more
BIOGRAPHY .............................................................. 12 efficient and affordable designs while maintaining or
increasing reliability, with lower mass and lower overall cost
of ownership, it is imperative that we improve and
1. INTRODUCTION standardize our technical assessment approach. This paper
Northrop Grumman Innovation Systems (NGIS) has long discussed the approaches that are being used to obtain the
understood that collecting accurate health state information necessary information and quantify the uncertainty as part of
about a product is crucial to the success of our products. a Prognostic Health Management (PHM) system used to
ascertain the health state information of the system and its
components.
978-1-5386-6854-2/19/$31.00 ©2019 IEEE
1
In Box 2, carefully designed experiments are conducted,
2. ARCHITECTURE OF A PHYSICS-BASED PHM sometimes on idealized as well as real systems, to measure
SYSTEM the parameters in the evolution model built from the causal
information of Box 1. These experiments must often span
The architecture of a general physics-based PHM system is several years, since the evolution of many parameters is very
shown in Figure 1. It is referred to as the predictive train; a slow.
sequence of models and data sources that start with the causal
state variables on the left and end with predicted asset In Box 3, the evolution model is exercised to predict the
performance on the right. The physics-based models can be evolution in time of the causal state variables given an
categorized into three groups: (1) the evolution models that applied set of boundary condition histories (the history
describe how the state variables evolve under the influence operator). If some assets have seen a harsher boundary
of the history operator, (2) the conversion models that convert condition history, they age out sooner than other assets from
from state variables to the properties required by the the same set that have not seen the harsher boundary
performance assessment models, and (3) the performance conditions.
assessment models that predict how the asset will operate
with that given set of properties. The right-most activity is the In Box 4, the state variables are converted into the properties
uncertainty roll-up that outputs the desired results of the PHM needed by the performance assessment models.
system: when will the asset fail, what component will fail
first, and what are the uncertainties and confidence limits in In Box 5, the properties from any point in time are fed into
these predictions? the performance assessment models that predict asset
performance.
The five numbered boxes constitute the predictive train. The
uncertainty roll-up on the right is the critical activity that Lastly, all uncertainties from all data sources and all models
provides the desired output – probability of failure with along with the ascertained uncertainty for each critical
confidence limits. parameter are rolled into a final analysis called the
uncertainty roll-up, which provides probability of failure (or
conversely, reliability) versus time or usage. [2] [3] [4]

Figure 1. Architecture of a physics-based PHM system

In Box 1, the causes of property evolution are found. If an 3. UNCERTAINTY QUANTIFICATION


asset component is evolving due to chemical
reaction/diffusion processes, chemists will define the causes Uncertainty quantification (UQ) is the combination of
of evolution. If the evolution is due to cumulative exposure verification (assessment of mathematical accuracy) and
to electromagnetic fields or to mechanical fatigue, other validation (assessment of applicability) of mathematical
scientists are needed to define the causes of evolution. For models of real-world phenomena. Understanding this
most materials, there is a rich literature describing causal requires quantifying the uncertainty of the numerical
processes. The literature is a good starting point, but must solutions of those models with assessments of the variation
usually be adapted and augmented for the specific assets of in those models, the solutions, the model inputs (boundary
interest. conditions, physical parameters, etc.), and in the phenomena
in order to quantify the variations in the numerical model
outputs.
2
When one applies uncertainty quantification to an Model input uncertainty is the uncertainty in any input to the
engineering analysis, they hope to assess the uncertainty in mathematical models used to perform the analysis. In
the real-world physical phenomena and the uncertainty in the engineering analyses, input quantities such as material
mathematical models of those phenomena and their properties, initial conditions, and boundary conditions fall
numerical implementations in order to better understand the under the umbrella of model input uncertainty. Model input
variability in the real-world physical responses that are being uncertainties are estimated by a variety of methods typically
analyzed. involving a form of validation testing of the input parameter.
As an example, materials may be tested repeatedly in a
Uncertainty quantification is now required during product laboratory in order to estimate the probabilistic distribution
analysis in many industries for several reasons. First, the of the material properties under differing conditions.
globalization of all industries, as well as increased
competition in specific industries including aerospace and Model discretization uncertainty is the uncertainty in the
defense, is forcing companies to reduce product development implemented mathematical model due to the finite resources
costs in order to stay competitive. This reduction in costs available on computer systems. Specifically, sources of
manifests itself in many ways including fewer tests of real- discretization uncertainty in engineering analyses typically
world physical assets and more virtual tests via analysis. In include mesh (spatial) discretization uncertainty, time step
order for the results of these virtual tests to be applicable to (temporal) discretization uncertainty, and uncertainties due to
the physical products being tested, the uncertainty in the finite floating-point number precision. Estimation of model
analyses must be quantified. discretization uncertainty sources typically involves a
verification process where the results of an analysis are
Secondly, as many higher-performance products are compared to an analytical solution (e.g., method of
developed, normalization of margins-of-safety or safety manufacturer solutions) if possible or another numerical
factors across the asset are required in order to keep mass to analysis performed with parameters that utilize a different
a minimum while still maintaining the required reliability of level of discretization (e.g., grid refinement studies).
the asset. The uncertainty in the performance of the product
must be quantified in order to enable this normalization. Model form uncertainty is the uncertainty in the degree to
which the implemented mathematical model represents the
A significant amount of research has been performed on real-world behavior of the physical objects that are
uncertainty quantification in recent years. This paper extends represented in the model. Simplifications in theory and
the work of Ewing, Liechty, and Black [5] from general implementation in order to make the numerical solutions
engineering analysis to the field of PHM and provides details tractable (assumptions) are an example of a source of model
and examples of uncertainty sources that are often form uncertainty. This form of uncertainty is typically
encountered with PHM systems. estimated by validation, comparison of the output of the
mathematical model to the real-world response that the model
Specific to PHM, uncertainty quantification has always been is attempting to represent. This validation is often challenging
a requirement since the purpose of a PHM system is to as measurements of the real-world response are subject to
ascertain the reliability of an asset via probabilistic methods. additional measurement uncertainties that must be accounted
Those working in the field of PHM have traditionally for and subtracted from the calculated model form
quantified sources of uncertainty from the aleatory risk of a uncertainty.
component in their analyses while completely ignoring other
sources of uncertainty that may be occur from epistemic One should note that the chosen numerical model has a great
risks. Previous PHM work at Northrop Grumman Innovation effect on the level of uncertainty from each of the
Systems has shown that some sources of uncertainty, aforementioned sources. As an example, machine learning
typically on very sensitive parameters such as activation algorithms have very low model form uncertainty (at least
energy material properties, have previously been ignored but after sufficient learning has been done) since they operate by
may dominate all other uncertainty in the system, including building functions that match measured data. However, the
asset-to-asset variation. In order to accurately predict the same machine learning algorithms may have undefined or
performance of an asset, PHM systems must quantify all hard to calculate model discretization uncertainty because the
sources of uncertainty and utilize those uncertainties in all underlying functions that have been generated may be
aspects of the prognostic calculations. arbitrarily discretized by the algorithm. As such, care must be
taken when choosing an algorithm to utilize one that has
measureable or estimable uncertainties from each source.
4. SOURCES OF UNCERTAINTY
Specific to PHM systems, additional care must be taken to
In engineering model analyses, there are many sources of choose models that have uncertainties that can be quantified
uncertainty that factor into the overall uncertainty in the far into the future. For example, a PHM system may be tasked
result. All of those sources of uncertainty can be with monitoring a fleet of assets over an expected lifetime of
characterized as being a subset of one of three uncertainty several decades. In that instance, models must be utilized that
sources: model input uncertainty, model discretization have inputs, discretization, and forms that can be measured
uncertainty, and model form uncertainty [5].
3
or estimated for a significant span of time. Further detail on of the sample variance distribution. This results in
this problem and possible solutions to it are given below.
𝜎̃ 2 → 𝜎𝑠2 (𝑛 − 1)𝐼𝑛𝑣𝜒𝑛−1
2 (6)
2
5. CATEGORIES OF UNCERTAINTY where 𝐼𝑛𝑣𝜒𝑛−1 is the Inverse Chi-Squared Distribution with
degrees of freedom n-1.
Uncertainty in engineering and PHM analyses can further be
categorized into three classes based on the underlying cause Similarly, from the Central Limit Theorem, the distribution
of uncertainty: aleatory uncertainty, epistemic uncertainty, of the sample mean as a function of the number of samples is
and ontological uncertainty [5]. given by

Aleatory uncertainty is uncertainty due to real-world physical 𝜎2 (7)


variation. Uncertainties that are categorized as aleatory 𝜇̃𝑠 → 𝑁 (𝜇, )
𝑛
cannot be reduced or decreased with additional knowledge
about the physical system being analyzed. An example of This equation can similarly be inverted to produce the
aleatory uncertainty is inherent variation in material probability distribution function for the mean of the
properties that are caused by inhomogeneity or random underlying data as a function of a sample of the sample mean
distribution of molecules in the structure of the material. distribution as
Aleatory uncertainties can be statistically estimated through
testing of the real-world physical systems bounded by 𝜎2 (8)
statistical distributions of the varying parameters. Note that 𝜇̃ → 𝑁 (𝜇𝑠 , )
𝑛
these statistical distributions will contain uncertainty due to
sampling of the parameters, and this uncertainty must also be Epistemic uncertainty is uncertainty due to a lack of
captured. For example, if the distribution of the underlying knowledge, which can be reduced or eliminated completely
physical parameter can be represented as a normal by incorporating additional knowledge into the model. An
distribution, example of an epistemic uncertainty is the initial state of the
system being analyzed. The system has a definite initial state,
𝑉̃ → 𝑁(𝜇, 𝜎 2 ) (1)
but that state may not be adequately measured or completely
and a finite-number of measurements of that parameter are known when the analysis model is exercised on the numerical
taken, the form of the normal distribution can be estimated model of the system. If additional measurement were
using further statistical distributions. In this case, the sample performed to identify the correct initial state prior to
distribution of the underlying parameter can be estimated execution of the analysis, this epistemic uncertainty would be
from the individual, repeated measurements and the number eliminated. Epistemic uncertainty should be estimated to
of measurements as produce a statistical distribution. This becomes somewhat
difficult as one is asked to estimate something that they, by
𝑉̃𝑠 → 𝑁(𝜇𝑠 , 𝜎𝑠2 ) (2) definition, have no knowledge of. As such, often uniform
distributions are utilized so the uncertain parameters are
where the sample mean and unbiased sample variance are assumed to fall somewhere within the specified bounding
box.
1 (3)
𝜇𝑠 = ∑ 𝑉𝑖 Ontological uncertainty is uncertainty due to completely
𝑛
𝑛 unknown factors; Donald Rumsfeld’s “unknown unknowns”
[7] applied to engineering analyses. Such uncertainty may be
and due to phenomena that are unexpected and thus not accounted
for in the engineering model but are nonetheless present in
1 (4)
𝜎𝑠2 = ∑(𝑉𝑖 − 𝜇𝑠 )2 the physical system. This is especially problematic for PHM
𝑛−1 systems where the phenomena that are well-understood may
𝑛
only be dominant in the near term, and unknown or poorly
From Cochran’s Theorem [6], the distribution of the unbiased categorized phenomena dominate the response of the system
sample variance due to the finite number of samples is in the far term. Such uncertainties are notoriously difficult to
quantify, but some possible methods for roughly estimating
𝜎2 (5) them is discussed below.
𝜎̃𝑠2 → 𝜒2
(𝑛 − 1) 𝑛−1
One of the difficulties in categorizing an uncertainty source
2
where 𝜒𝑛−1 is the Chi-Squared Distribution with degrees of is that there is an apparent overlap in the definitions of
freedom n-1. The actual variance, 𝜎 2 , is not known, only a aleatory and epistemic uncertainty, which is best illuminated
sample of the sample variance distribution, 𝜎𝑠2 , has been via example. Material properties, as is discussed above, are
obtained. However, the above equation can be inverted to classified as aleatory due to the inherent inhomogeneity in the
calculate the probability density function for the actual material structure within the bounds of the measured
variance of the underlying distribution based on this sample information. However, hypothetically one could measure the
4
complete microstructure of the material and incorporate this it uses. That number of bits and the digital sensor value can
new knowledge into the engineering model, which would then be used to create a uniform distribution of the value of
instead define the uncertainty as epistemic because only the the actual underlying physical parameter.
lack of knowledge about the actual microstructural variation
across the entire model is causing uncertainty. To avoid this Similar to sensor quantization uncertainty, floating-point
overlap of terms, one must define the knowledge that is numerical truncation and round-off errors introduce
missing when dealing with epistemic uncertainty must be uncertainty into all numerical calculations. This model
relative to the engineering model [5]. This means that if the discretization epistemic uncertainty source is due to the finite
model, as defined and implemented, is capable of precision with which floating-point numbers are represented
incorporating the additional knowledge about the physical in computer systems. As an example, one may be surprised
system without modification, the uncertainty is defined as to find that the sum of 1,000,000 and 1 in single-precision
epistemic; if the model must be modified in order to floating point arithmetic does not yield 1,000,001 but instead
incorporate the additional knowledge about the physical yields 1,000,000. This is because a single-precision floating
system, the uncertainty is defined as aleatory; and if the point number on the order of a million does not have
model cannot be modified because the phenomena or sufficient precision to differentiate between the numbers
response are unknown, the uncertainty is defined as 1,000,000 and 1,000,001; so the summation has no effect
ontological. So, if the engineering model associated with the whatsoever.
analysis used in the example above assumed homogeneous
materials, as is commonplace in engineering analyses, the This uncertainty source is present in all numerical
material property variation due to inhomogeneity is aleatory calculations, but typically does not have a large effect unless
uncertainty. one deals with both large and small numbers simultaneously.
Quantification of this uncertainty source can be done via
inspection. Inspection of the numerical models must be made
6. UNCERTAINTY IN PHM SYSTEMS by one familiar with floating-point arithmetic and who can
evaluate where in the model such uncertainty may arise. They
Prognostic health management systems have many of the can then estimate the distribution of the error based on the
same sources of uncertainty as other engineering disciplines, numerical precision.
some of which are associated with the use of the same design
models for application into prognostic systems; however, Using our example above, the summation of a number on the
there are other uncertainty sources that are unique and of order of a million with a number on the order of one using
particular interest (or concern) with PHM systems. Several of single-precision floating point arithmetic will introduce
these uncertainty sources are discussed in this section. uncertainty in the result that can be represented by a uniform
distribution with a width of approximately four. Thankfully,
While typical engineering analyses deal with models of when using double-precision floating point arithmetic, as is
systems with varying inputs based on engineering knowledge common and recommended for engineering analyses, these
and assumptions, PHM systems deal with inputs to the uncertainties are many orders of magnitude smaller;
models that come from physical sensing devices. These however, they still exist and should be accounted for.
sensors introduce several types of uncertainty that are not
present in other engineering analyses. Sensors also introduce uncertainty due to natural variation of
the readings they provide. For example, a relative humidity
One such sensor-based source of uncertainty is quantization sensor may only be accurate to within 3%, so if a humidity
error. Sensors are typically analog devices that detect and sensor gives a reading of 30% there is no way to know
measure some varying parameter in the physical world. These whether the actual relative humidity was 27%, 33%, or
analog data are captured by sensors that are measuring an anywhere in between. Similar to sensor quantization
electrical representation of a physical parameter and the uncertainty, this model input uncertainty is epistemic in
capture analog information if conditioned and filtered. This nature. Quantification of this uncertainty source is simple, as
conditioned data is then converted into digital data via nearly all sensor manufacturers state the accuracy (or
analog-to-digital converters (ADC). These ADCs have a variation) of their sensors in their product documentation and
fixed precision, typically in the range of 8 to 16 bits, and thus data sheets.
a significant amount of precision in the analog value will be
lost by quantizing the conditioned value. One method to reduce the uncertainty caused by sensor
variations is to utilize multiple redundant sensors at each
Because the system is capable of reading the actual analog location. Because the sensors are in the same location, they
sensor value, but knowledge about that value is lost, this should be measuring approximately the same underlying
model input uncertainty source is categorized as an epistemic physical value. The different sensor readings can then be
uncertainty. Quantification of this uncertainty source can be interpreted as statistical samples of a single value. With a
straightforward by tracking uncertainty sources of each of the sensor with a fixed accuracy, such as the humidity sensor
elements of the measurement system. For example, the ADC example above, the values from the redundant sensors will
that is used in the sensor itself (if digital) or the external ADC fall within a uniform distribution centered on the actual
will be documented with the number of bits of precision that physical value. Again, as per the Central Limit Theorem, by
5
averaging these multiple sensor readings, a better estimate of algorithms take into account both mechanical and chemical
this physical value can be obtained. aging of the materials in the system, both of which are very
sensitive to the early environmental conditions that the
Uncertainty caused by sensor variation has an additional materials experience. If the environment around these
complexity that is especially prevalent in PHM systems: the materials is not monitored by the PHM system, as is likely
sensors age and their variations will change with time. Over since the materials begin to age as they are being
the potentially long service lives of PHM systems, these manufactured and continue aging during storage, shipment,
sensor variations can increase by orders of magnitude. and assembly with the monitored asset, this can present a
Unfortunately, few sensor manufacturers measure the significant amount of uncertainty in the subsequent PHM
degradation of their sensors over such long periods of time analyses.
(potentially decades), and so quantification of this
uncertainty source is problematic and can introduce Unless the environment that the materials see from their
ontological uncertainty risk into the measure information. beginning of life can be monitored by the PHM system,
Manufacturers may state an expected lifetime or mean time which will likely require a significant expense as well as
between failures of their sensors or components, but this does coordination with material manufacturers and logistics
not give sufficient information to determine the sensor companies, the best way to quantify the uncertainty in the
variation with time. initial conditions of the materials is to utilize material batch
testing. This involves measuring the different batches of
It is suggested that builders and operators of PHM systems materials that will be utilized in the system as they are
work with the sensor manufacturers to quantify the variation delivered and utilized in manufacturing the system. Once the
of the sensors over long periods of time. This can be time- initial state of and the uncertainty in the material properties
consuming though, as decades of testing may be required in are known, the PHM system can utilize these values as the
order to quantify the variation of a sensor before the sensor starting point for the prognostic analyses. In addition to
can be utilized within a PHM system. Alternately, a PHM making the question of the environmental conditions prior to
system builder should consider incorporating into their monitoring moot, this technique will also quantify the
systems multiple redundant sensors of each type and, in uncertainty in the homogeneity and batch-to-batch
addition to averaging their readings to reduce the sensor consistency of the materials, both of which introduce further
variation uncertainty as described above, measure the model input epistemic uncertainty.
variation between each redundant sensor as time passes.
Just like when utilizing multiple sensors to reduce the sensor
While this will not necessarily allow an accurate assessment variation uncertainty, when testing material properties one
of the sensor variation uncertainty to be projected into the must take multiple measurements of each property in order to
future because such variation is typically nonlinear, it will quantify the uncertainty in both the material and in the
allow for the current sensor variation to be measured and a measurement process. The process of calculating the
rough approximation of that uncertainty to be projected (with underlying distributions of those parameters is identical, and
high additional uncertainty) into the future. This will amount for the case that the underlying distribution is normal results
to the statistical distributions of sensor variation uncertainty in a normal distribution of the mean and a Chi-Squared
(e.g., the uniform distribution between 27% and 33% distribution of the variance, again as per Cochran’s Theorem
discussed above) to be projected out into the future, yielding [6] (Equations 1-8). This means that after measuring the
distributions that expand with time. material properties multiple times, one does not know the
exact underlying distribution of the physical property, but
Because sensor variation uncertainties can become large and they do know the statistical distribution of that underlying
dominate the uncertainty in the PHM analysis, it is physical distribution, which can be utilized within the
recommended that sensors also be a part of the PHM analysis subsequent PHM analyses.
and tracking, allowing for condition-based maintenance or
replacement of the sensors when variation uncertainty Similar to initial material property variability, the initial
becomes large as measured by the redundant sensors. A geometry of the monitored system is likely to vary for a
discussion on the logistics of this is not made here, but the variety of reasons including manufacturing tolerances and
implications on the statistical measures of the uncertainty are processes. For many systems, this relatively small variability
important to recognize and track. Replacing sensors once in geometry will not have a large impact on the uncertainty
sensor variation uncertainty has the effect of limiting the in the result of the PHM analysis, but in other cases it may,
spread of the statistical distributions of the sensor variation so it is recommended that an estimate of this uncertainty be
uncertainty. This will keep the sensor variation uncertainty incorporated into the PHM analysis. Whether this model
from dominating the PHM analysis, and thus it is highly input uncertainty source is aleatory or epistemic depends
recommended that such sensor replacement strategies be upon the model being utilized in the analysis. For most
incorporated into any PHM system that must be operational models, the geometry of the system is assumed to be fixed,
for a significant amount of time. and thus the uncertainty is aleatory; however, other models
may treat the geometry as variant, and thus the uncertainty is
Another source of epistemic uncertainty that is important in epistemic. In either case, the best method of quantifying this
PHM systems is that of initial conditions. Many PHM aging
6
uncertainty is to measure critical dimensions on each asset of a propeller may be dependent on many factors that are
and account for the measurement uncertainty in the incorporated into the PHM model, but if that propeller is
uncertainty rollup. allowed to operate in a condition that was not considered, like
a sandstorm, there will be a great deal of uncertainty between
Many PHM systems monitor all assets in the fleet of assets; the predicted health of the propeller and the observed health.
however, some systems monitor only a subset of the fleet
assets in order to determine the overall health of the fleet. Like any other ontological uncertainty, estimating this source
This can be done statistically, but it does introduce further of uncertainty is very difficult. One method of coming up
epistemic model input uncertainty to the analyses. Just like with a very rough estimation is to take all of the physics that
the sampling of material properties discussed above, the state are built into the PHM model and quantify the change in the
of the individual assets in the fleet can be estimated from the response of the model when combinations of those physics
state of the assets that were sampled. If the distribution of the are disabled. Extending the propeller example above, if the
states of the assets is assumed to be normal, then equations 1 propeller model included phenomena such as rain impact and
through 8 can again be utilized to calculate distributions of icing, then differences in the model when these phenomena
the underlying distributions. In practice, the distribution of are included and excluded will give a rough order-of-
the states of the assets is highly unlikely to be a normal magnitude estimation of the uncertainty due to other similar
distribution though, so additional statistical methods not phenomena, such as sand impacts, not being included in the
discussed here should be employed. model.

For large assets being monitored by PHM systems such as One additional model form uncertainty source should be
aircraft, ships, and rocket motors, the variation of discussed; that of human errors in the development of the
environmental conditions across the asset can be a significant models and their numerical implementations. The estimation
source of epistemic uncertainty. Quantification of this of this uncertainty source is nearly impossible, so it is best to
uncertainty is a two-fold process. First, the effects of varying strive to eliminate this source instead of quantifying it. While
boundary conditions on the model must be ascertained by total elimination of unknown human errors is not a feasible
running several analyses with differing variations of goal, utilizing several different processes while developing
conditions across the asset. Then the actual variation of the models and implementations can reduce the likelihood of
conditions across the asset must be measured using multiple these sorts of errors significantly.
sensors positioned on and around the asset. The latter
uncertainty can then be quantified and used as an input into It is recommended that formal peer reviews be held at all
the former uncertainty to calculate an overall quantification levels of development, from theory and model development
of this uncertainty source. to algorithm development to software development. Many
errors can be discovered simply by having one or more peers
In a PHM system, large amounts of data including sensor data step through and review the development process. Specific to
values and calculated asset health data must be transferred software development, the use of test-driven development is
across networks and stored in a data warehouse. There are another method to reduce errors. This development
many different methods of transferring and storing this data, methodology uses requirements to build test cases, which
and several of those methods involve compressing the data in must be peer reviewed, and then development is done in order
order to reduce network bandwidth and storage space [8] [9] to make the numerical models successfully run the test cases,
[10]. and thus successfully implement the numerical methods and
fulfil the requirements of the program. This is backwards
This data compression may have the effect of introducing a from the usual software development methodology where
form of epistemic discretization uncertainty into the data. test cases are written after the fact to test the software that
Determining the uncertainty that is introduced due to this was built to the requirements. Formalizing the requirements
compression can be problematic because internal details of of the numerical implementation and adding the additional
the compression algorithm and how it operates on specific layer of peer review of these requirements ensures that the
data values must be known. Because storage space is software is developed correctly and in alignment with the
relatively cheap in comparison with the costs associated with developed models.
introducing significant uncertainty into an analysis and the
resulting potential over-design of a product, lossless One final area of uncertainty needs to be accounted for, that
compression algorithms can be utilized in PHM system data of ontological uncertainty associated with the unknown
warehouses, which can assist in eliminating this source of unknowns. One method of accounting for this is by
uncertainty. understanding margins in the design and using these to
account for unknown effects. This method can compensate
Similar to the above, one must consider uncertainty that for the ontological risk in evolutionary systems, but to
enters into the model due to unknown boundary conditions account for these risks is difficult for new revolutionary
and physical mechanisms. These model input and model form systems due to the lack of baseline data.
ontological uncertainties are typically due to the asset being
in an environment or encountering new physics that were not
considered when building the model. For example, the health
7
7. CALCULATION WITH UNCERTAINTY
Once all of the uncertainty sources in an engineering or PHM
analysis have been identified and quantified, those
uncertainties must be utilized within the engineering
calculations and models in order to roll up the uncertainty and
determine the variability and repeatability of the analysis due
to those uncertainties.

The traditional method for utilizing such uncertainties in an


engineering analysis is to sample all of the variables based on
random variation for each statistical distribution within a
single Monte Carlo loop, as seen in Figure 2. The underlying
models are then run repeatedly with the different sets of
sampled inputs, and the outputs from the models can be
analyzed in order to determine the statistical variation of
those outputs. While this method does indeed provide the
needed quantification of the variation of the outputs, it does
not separate the epistemic uncertainty sources from the
aleatory uncertainty sources. This means that the variation of
the outputs due to natural variation of the inputs (aleatory
uncertainty), which are real variations that will be observed
in the physical system, is not separated from the variation of
the outputs due to lack of knowledge (epistemic uncertainty),
which will not manifest itself in the physical system and will
only affect the repeatability of the analysis.

Figure 3. Two-tier Monte Carlo algorithm that gives


information about the variability and repeatability of
the uncertainty quantification analysis

In a PHM system, the underlying model typically provides a


pass/fail of the asset. For example, a load may be calculated
based on boundary conditions and compared to a material
allowable to calculate whether the material fails. This simple
calculation of pass/fail is typically run forward in time to
create a Boolean pass/fail curve. The inner Monte Carlo loop
Figure 2. Typical single-tier Monte Carlo algorithm used then samples the aleatory uncertainty sources, runs this
in uncertainty quantification analyses simple calculation many times, and statically analyzes the
resulting Boolean curves to determine a single probability of
In an attempt to address these shortcomings of a single Monte
failure that varies with time. The outer Monte Carlo loop is
Carlo loop, it has been suggested that a two-tier Monte Carlo
then used to sample epistemic uncertainty sources and
loop be utilized in order to separate the epistemic and aleatory
generate many different probability of failure curves that
uncertainties [5]. In such an analysis, an inner Monte Carlo
represent the variability in the analysis due to a lack of
loop is used to sample aleatory uncertainty sources in order
knowledge about the physical system.
to capture the variation that is expected to be observed in the
physical system. A second outer Monte Carlo loop is then Sample results from such an analysis can be seen in Figure 4.
utilized to sample the epistemic uncertainty sources in order The first step in analyzing these results is to generate a P-box
to capture the lack of knowledge about the physical system [11] around the probability of failure curves with a specific
and quantify the repeatability of the analysis. Figure 3 gives confidence. This is done by determining the area within the
a representation of such a two-tier Monte Carlo analysis. plot that contains the specific percentage of the curves. The
upper and lower limits of this P-box are shown in Figure 4 as
orange curves.

8
Figure 4. Sample probability of failure curves from a
Figure 5. Identifying the range of probabilities of failure
two-tier Monte Carlo analysis of an asset monitored by a
at four years from credible P-box with 90% confidence
PHM system. The gray curves are individual samples
from the outer Monte Carlo loop; the orange curves
define the 90% confidence P-box.

One can now intersect horizontal and vertical lines in the plot
with these credible P-box curves and draw statistical
conclusions from those intersections. For example, drawing
a vertical line on our sample plot at four years and
intersecting it with the 90% credible P-box shows that the
range of probabilities of failure that we are 90% confident
that the physical system will have at four years is 0% to 25%
(Figure 5). If instead we are interested in replacing or
refurbishing our asset when it reached a probability of failure
of 50%, we could draw a horizontal line on the plot at 50%
probability of failure and intersect that with the credible P-
box to show that the range of the lifetime of the asset prior to
it needing refurbishment or replacement with 90%
confidence is 4.2 to 5.7 years (Figure 6).

Figure 6. Identifying the range of refurbishment ages


from credible P-box with 90% confidence

Until now we have only been varying aleatory and epistemic


uncertainty sources, and ontological sources have not been
considered. One could utilize a three-tier Monte Carlo
algorithm to also sample these ontological uncertainties and
roll them into the analysis, but the difficulty in quantifying
such uncertainty sources renders the benefits of performing
such an analysis negligible. It is recommended that the two-
tier Monte Carlo loops be utilized in all cases.

In the typical PHM model, there are a very high number of


inputs (e.g., material properties, boundary conditions that
vary with time, etc.). This means that a large number of
Monte Carlo iterations must be utilized in both tiers of the
two-tier approach in order to attain proper convergence. This,
combined with the complexity of the underlying models,

9
means that run times for such analyses can easily become convergence. In order to reduce this number while not greatly
intractable. affecting the statistical results, one should seek to eliminate
variation in input parameters that do not have a large effect
There are several methods that can be utilized to reduce the on the results when varied within their statistical
run times of such analyses. One such method is to increase distributions. To this end, sensitivity studies must be
the processing power of the computer system that is running performed on the inputs. In such studies, the analysis models
the analyses. This is typically done by one of two methods: are run while varying individual parameters within their
utilizing parallel processing and utilizing general purpose statistical distributions (e.g., a normally-distributed variable
graphics processing units. may be varied +/- 3σ). The variation of the results due to these
input variations are then tabulated, and those parameters that
Parallel processing is very simple to implement for a Monte have a negligible effect on the results can be treated as
Carlo algorithm because the algorithm itself is constant in the Monte Carlo analysis.
“embarrassingly parallel”. That is to say that the iterations
within the algorithm are completely independent of one Different sampling methods can also be utilized with the
another, and thus they can be run simultaneously and get Monte Carlo algorithm to aid in convergence. The Latin
near-perfect speedup without much additional software Hypercube Sampling (LHS) method may provide better
development effort. The use of parallel processing can reduce sampling, and thus require fewer iterations to attain
run times of Monte Carlo algorithms by near a factor of the convergence, but the magnitude of the improvement is highly
number of cores that are utilized. If one has access to a very dependent on the model. If the model has a high number of
large high performance computer system, this can be up to input parameters and there are significant nonlinearities
seven orders of magnitude [12]. However, the costs between the input parameter terms (both of which are
associated with running analysis on such large systems can common in PHM models), the improvements are
be prohibitive, especially with a PHM system where the significantly reduced. Furthermore, because additional work
analyses need to be continuously rerun as new sensor data is must be done to generate the sampling points with LHS, it is
acquired. possible that run times with LHS may be slower than with
standard Monte Carlo random sampling. This is an area
Calculation with general purpose graphics processing units where additional research should be done in order to quantify
involves leveraging the high-performance graphics cards on any performance improvements when used with PHM
modern computer systems for computation instead of models.
graphics-related tasks. Table 1 compares the processing
power of a modern central processing unit (CPU) with a
modern graphics processing unit (GPU) by comparing the 8. CONCLUSIONS
number of floating point operations per second (FLOPS) that
each can theoretically achieve, showing that speedups of In the aerospace industry, systems are counted on to work
nearly two orders of magnitude can be achieved by utilizing reliability. Most missions involve critical capabilities either
a GPU instead of a CPU. Note that actual FLOPS figures are for human life, scientific studies, or defense applications.
likely to be significantly below these theoretical peak figures. Solid rocket motor systems in particular are counted on to
Also note that GPUs are more optimized for lower-precision work the first time every time. This means we must
calculations, whereas CPUs see no theoretical benefit in understand the product thoroughly and ensure that the system
utilization of half-precision floating-point operations because capability can meet or exceed the demands of the mission.
their hardware does not directly operate on them [13]. Use of PHM systems that accurately capture and quantify
uncertainty are key to gaining this understanding.
Table 1. Comparison of floating-point calculations per
second (FLOPS) for the Intel i7-8700K CPU and the Uncertainty quantification is a key aspect of all engineering
NVIDIA Quadro RTX 6000 GPU analyses, but it is especially important in PHM because the
uncertainty in the outputs of the engineering models are one
Peak FLOPS of the primary outputs that must be analyzed and acted upon.
Double Single Half In order for the uncertainty in a PHM analysis to be properly
Type Model
Precision Precision Precision accounted for, all sources of uncertainty must be identified,
Intel categorized, and quantified. A discussion of some common
CPU i7-8700K 5.64×1011 1.13×1012 1.13×1012 PHM-related uncertainty sources, the categorization of those
[14] sources, and methods to quantify them has been presented.
NVIDIA
Quadro It is recommended that the asset health calculations in a PHM
GPU 7.40×1012 1.63×1013 3.26×1013
RTX 6000 system with proper uncertainty quantification be performed
[15] with a nested two-tier Monte Carlo loop in order to quantify
the variability and repeatability of the analysis based on
Another method to reduce the run times is to reduce the epistemic and aleatory uncertainty sources. A variety of
number of inputs to be varied, and thus reduce the number of methods have been presented that have the potential to
Monte Carlo iterations that will be needed to attain dramatically reduce the run times of the Monte Carlo
10
implementation, but additional research needs to be done in REFERENCES
this area specific to PHM analyses in order to quantify the
potential gains. [1] International Council on Systems Engineering
(INCOSE) SE Handbook Working Group, "INCOSE
The output from the two-tier Monte Carlo analysis in a PHM Systems Engineering Handbook v. 3.2.2," 2011.
system is typically a set of reliability curves that vary with
time. The smooth change in reliability with time is due to [2] S. Hyde, D. Richardson, B. Allen and B. Goldberg,
aleatory uncertainty sources, whereas the variation between "Model Based Design Influence on Program Testing
reliability curves is due to epistemic uncertainty sources. Programs Part I," in AIAA Missile Sciences
Post-processing these curves allows for calculation of both an
Conference, AIAA Defense and Security Forum,
asset health assessment and the confidence in that
Laurel, MD, 2015.
assessment. It is suggested that such a method for quantifying
uncertainty be utilized within all PHM systems, particularly
those monitoring high-value assets, in order to produce the [3] R. Smith, "Chapter 1," in Uncertainty Quantification:
highest quality actionable quantifications of health for the Theory, Implementation, and Applications,
assets being monitored. Philadelphia, SIAM, 2014.

[4] D. DeVries, S. Hyde and I. L. Davis, "A Systems


Engineering Approach to Evolution of Physics-based
Prognostic Health Management of Aging Solid Rocket
Motor System Assets," AVT-268 RSM, Utrecht,
Netherlands, 2017.

[5] M. Ewing, B. Liechty and D. Black, "A General


Methodology for Uncertainty Quantification in
Engineering Analyses," Journal of Verification,
Validation and Uncertainty Quantification, vol. 3, no.
2, 2018.

[6] W. G. Cochran, "The distribution of quadratic forms in


a normal system, with applications to the analysis of
covariance," in Mathematical Proceedings of the
Cambridge Philosophical Society, Cambridge, 1934.

[7] D. H. Rumsfeld, 12 February 2002 News Briefing on


Weapons of Mass Destruction in Iraq, Washington
D.C., 2002.

[8] B. Furht, "A Survey of Multimedia Compression


Techniques and Standards Part I JPEG Standard,"
Real-Time Imaging, vol. 1, no. 1, pp. 49-67, 1995.

[9] N. Kimura and S. Latifi, "A Survey on Data


Compression in Wireless Sensor Networks," in
International Conference on Information Technology:
Coding and Computing, 2005.

[10] W. Su, O. B. Akan and E. Cayirci, "Compression


Techniques for Wireless Sensor Networks," in
Wireless Sensor Networks, Springer Science+Business
Media, 2006, pp. 207-232.

[11] S. Ferson and W. T. Tucker, "Sensitivity Analysis


Using Probability Bounding," Reliability Engineering
& Systems Safety, vol. 91, pp. 1435-1442, 2006.

11
[12] "TOP500 June 2018," TOP500.org, June 2018. BIOGRAPHY
[Online]. Available:
Mr. Heath Dewey is a Software
https://www.top500.org/lists/2018/06/. [Accessed 28
Engineer in the Propulsion Systems
September 2018]. Division of Northrop Grumman
Innovation Systems (formerly ATK
[13] P. Konsor, "Performance Benefits of Half Precision and Orbital ATK) at Promontory,
Floats," 15 August 2012. [Online]. Available: Utah. He received a B.S. from Iowa
https://software.intel.com/en-us/articles/performance- State University in 2004 and a M.S.
benefits-of-half-precision-floats. [Accessed 28 from the University of Illinois at
September 2018]. Urbana-Champaign in 2006, both in
Aerospace Engineering. Since that time, he has worked in
[14] Intel, "Intel Core i7-8700K Processor," [Online]. his current position developing modeling and analysis
Available: https://ark.intel.com/products/126684/. tools that are used in simulation of complex physics and
[Accessed 28 September 2018]. phenomena that occur inside solid rocket motors during
firing and flight. In recent years, he has also led several
[15] E. Kilgariff, H. Moreton, N. Stam and B. Bell, software engineering teams in developing data collection
"NVIDIA Turing Architecture In-Depth," 14 and prognostic health management systems for monitoring
September 2018. [Online]. Available: strategic solid rocket motor fleets. He and his wife Kendra
have been married for 12 years and have two children.
https://devblogs.nvidia.com/nvidia-turing-
architecture-in-depth. [Accessed 28 September 2018].
Mr. Derek R. DeVries, P.E. is a
Senior Fellow and Propulsion
Systems discipline owner of Avionics
and Control Disciplines for Northrop
Grumman Innovation Systems He is a
licensed Professional Engineer,
Senior Member of the Institute of
Electrical and Electronics Engineers,
and U.S. Air Force veteran. He was
awarded the NATO Science and Technology
Organization’s 2016 Scientific Achievement award for
Application of Integrated Munition Health Management.
He was honored as an invited Luminary speaker at the
international PHM conference in 2015. He has over 30
years of experience in the aerospace industry in operation,
integration, and development of space and strategic
launch systems. He received his B.Sc. degree in Electrical
Engineering from the University of Utah and his M.Sc.
degree in Electrical Engineering from Utah State
University. He served as a member of the Industrial
Advisory Board for the University of Utah Department of
Electrical Engineering and an American Institute of
Aeronautics and Astronautics Standards Committee
Member for development of the industry standard “S-122-
2006 Direct Current Power Systems for Earth-Orbiting
Satellites.” He is currently a member of the Society of
Automotive Engineers PHM Standards Committee. He has
14 U.S. and foreign patents and numerous published
papers. He was awarded the AIAA 2001 Joint Propulsion
Conference Arthur D. Rhea Award for Best Paper of the
Year presented in Ordnance Components and Systems.
Mr. DeVries is a leader in developing technologies in
aerospace systems including stewarding the development
of new significant aerospace projects in component and
system development and integration. He has and continues
to support active development, operational, and
sustainment activities of Air Force and Navy Long Range
Missile Systems, NASA Space Shuttle and Deep Space
12
missions and development of their next generation launch
vehicles. He has led or been an expert member of several
investigation teams for flight failures charged with
determining root cause and then recommending and/or
leading corrective actions for return to flight. He and his
beautiful wife Jodie have been married for over 35 years
and have five children and nine grandchildren.

Mr. Scott R. Hyde is a Senior Staff


Scientist at Northrop Grumman
Innovation System’s Flight Systems
Group. Throughout his career he has
held many different positions including
Design Engineer, Chief Engineer,
Department Manager, Technology
Program Manager, Business
Development Manager and Senior Scientist. He has over 30
technical publications. Scott received his B.S. in 1986 and his
M.S. in 1992 in Structural Engineering from Utah State
University in Logan, Utah. He has a passion for Integrated
Vehicle Health Management technologies and has
contributed to their development and implementation
throughout his aerospace career. He has spent 30 years
working on improving the management of solid rocket motor
systems. He led a multidiscipline team of engineers and
scientists who developed, verified and validated a
fundamental physics based methodology for predicting the
complex behavior of solid rocket motor propellant. He and
his wife Tawni have been married for more than 35 years and
have four children and are enjoying their first grandson.

13

Das könnte Ihnen auch gefallen