U.S. GOVERNMENT RIGHTS
[0001] The invention was made with U.S. Government support
under contract NAS4-02038 awarded by NASA. The U.S. Government
has certain rights in the invention.
BACKGROUND OF THE INVENTION
(1) Field of the Invention
[0002] The present invention relates to a method, and an
apparatus for performing such method, for sequentially
building a hybrid model.
(2) Description of Related Art
[0003] A practical consideration for implementing a hybrid
engine model that incorporates both physics-based and
empirical components, involves the application of some form
sequential model building for the construction and
specification of the empirical elements. This arises for the
reason that sufficient engine data required to model the
entire flight regime for a given engine/aircraft application
is never available from one flight alone and may takes days or
weeks to assemble.
[0004] Such a consideration is of particular import when
constructing a hybrid gas turbine engine model consisting of
both physics-based and empirically derived constituents. A
typical architecture for such a hybrid model commonly used for
the purpose of engine performance monitoring is depicted in
Figures la and lb.
[0005] With reference to Fig. la, there is illustrated a
typical configuration wherein an empirical modeling process
captures the difference, or deltas, between the physics-based
engine model and the actual engine being monitored. The
empirical element can take many forms including, but not
limited to, Regression models, Auto-Regressive Moving Average
(ARMA) models, Artificial Neural Network (ANN) models, and the
like. The inclusion of an engine performance estimation
process in this architecture is not essential to the present
invention, but is included to depict a typical application for
which this hybrid structure is particularly helpful.
[0006] When the empirical model is complete, the hybrid
structure takes the general form shown in Figure lb. The
combination of the empirical element and the physics based
engine model provides a more faithful representation for the
particular engine being monitored. This provides more
meaningful residual information from which an engine
performance change assessment can be performed since potential
(physics based) model inaccuracies and shortcomings have been
effectively removed by virtue of the empirical element.
[0007] The scenarios illustrated in Figs, la-lb are typically
be performed on-board in real-time during actual engine
operation and flight. Referring to Figure la, such performance
necessitates the storage and retention of engine and flight
input data over a series of flights until such a time that
sufficient flight and engine regime data is collected to
complete the empirical model. This imposes an unrealistic
requirement in terms of storage capacity for an on-board
system.
[0008] What is therefore needed is a method for modeling the
performance of device such as an engine, preferably a gas
turbine engine, that does not require the storage and
retention of a large volume of data, such as engine and flight
input data over a series of flights.
SUMMARY OF THE INVENTION
[0009] Accordingly, it is an object of the present invention to
provide a method, and an apparatus for performing such method,
for sequentially building a hybrid mode.
[00010] In accordance with the present invention, a method for
modeling engine operation comprises the steps of : inputting a
first plurality of sensory data from a gas turbine engine into a
computer; partitioning a flight envelope into a plurality of
sub-regions according to the steps of: selecting a first sensory
parameter and a second sensory parameter, plotting each of said
first plurality of sensory data by using said first sensory
parameter as a first axis and said second sensory parameter as a
second axis, and dividing said first axis and said second axis
into a plurality of subdivisions to create a grid comprising a
plurality of sub-regions; assigning said first plurality of
sensory data into said plurality of sub-regions; generating an
empirical model of at least one of said plurality of sub-
regions; generating a statistical summary model for at least one
of said plurality of sub-regions; inputting an additional
plurality of sensory data from said gas turbine engine into said
computer; partitioning said second plurality of sensory data
into said plurality of sub-regions; generating a plurality of
pseudo-data using said empirical model; concatenating said
plurality of pseudo-data and said additional plurality of
sensory data; and outputting using said computer an updated
empirical model and an updated statistical summary model for at
least one of said plurality of sub-regions.
[00011] In accordance with the present invention, a method for
modeling engine operation comprises the steps of :inputting a
3
first plurality of sensory data from a gas turbine engine into a
computer; partitioning a flight envelope into a plurality of
sub-regions according to the steps of: selecting a first sensory
parameter and a second sensory parameter, plotting each of said
first plurality of sensory data by using said first sensory
parameter as a first axis and said second sensory parameter as a
second axis, and dividing said first axis and said second axis
into a plurality of subdivisions to create a grid comprising a
plurality of sub-regions; assigning said first plurality of
sensory data into said plurality of sub-regions; generating an
empirical model of a portion of said plurality of sensory data;
generating a statistical summary model for said portion of said
plurality of sensory data; inputting an additional plurality of
sensory data from said gas turbine engine into said computer;
generating a plurality of pseudo-data using said empirical
model; concatenating said plurality of pseudo-data and said
additional plurality of sensory data; and outputting using said
computer an updated empirical model and an updated statistical
summary model for at least a portion of said sensory data.
[00012] In accordance with the present invention, an apparatus
for modeling engine operation comprises: means for inputting a
first plurality of sensory data from a gas turbine engine into a
computer; means for partitioning said first plurality of sensory
data into a plurality of sub-regions, said means for
partitioning comprising the following: means for selecting a
first sensory parameter and a second sensory parameter, means
for plotting each of said first plurality of sensory data by
using said first sensory parameter as a first axis and said
second sensory parameter as a second axis, and means for
dividing said first axis and said second axis into a plurality
of subdivisions to create a grid comprising a plurality of sub-
regions; means for generating an empirical model of at least one
of said plurality of sub-regions; means for generating a
statistical summary model for at least one of said plurality of
sub-regions; means for inputting an additional plurality of
sensory data from said gas turbine engine into said computer;
means for partitioning said second plurality of sensory data
into said plurality of sub regions; means for generating a
plurality of pseudo-data using said empirical model; means for
concatenating said plurality of pseudo-data and said additional
plurality of sensory data; and outputting using said computer an
updated empirical model and an updated statistical summary model
for at least one of said plurality of sub-regions.
[00013] In accordance with the present invention, a method of
constructing an empirical model comprises the steps of inputting
a first plurality of sensory data from a gas turbine engine into
a computer; partitioning an operating envelope into a plurality
of sub-regions according to the steps of: selecting a first
sensory parameter and a second sensory parameter, plotting each
of said first plurality of sensory data by using said first
sensory parameter as a first axis and said second sensory
parameter as a second axis, and dividing said first axis and
said second axis into a plurality of subdivisions to create a
grid comprising a plurality of sub-regions; assigning said first
plurality of sensory data into said plurality of sub-regions;
generating an empirical model of at least one of said plurality
of sub-regions; generating a statistical summary model for at
least one of said plurality of sub-regions; inputting an
additional plurality of sensory data from said gas turbine
engine into said computer; partitioning said second plurality of
sensory data into said plurality of sub-regions; generating a
plurality of pseudo-data using said empirical model;
concatenating said plurality of pseudo-data and said additional
plurality of sensory data; and outputting using a computer an
updated empirical model and an updated statistical summary model
for at least one of said plurality of sub-regions.
AMENDED!
^ACCOMPANYING
BRIEF DESCRIPTION OF THE*DRAWINGS
[00014] FIG. la A diagram of the architecture for
constructing an empirical model element.
[00015] FIG. lb A diagram of the architecture for hybrid
engine model after construction is complete.
[00016] FIG. 2 A diagram of an exemplary flight regime
partition of the present invention.
[00017] FIG. 3 A diagram of one embodiment of
architecture for implementing empirical model construction
using one possible method of performing the present invention.
[00018] FIG. 4 Comparative illustration of residuals
derived from an original multi-level perception (MLP) and the
Bootstrap MLP of the present invention.
[00019] FIG. 5 Illustration of the difference between an
original MLP and the Bootstrap MLP of the present invention.
[00020] FIG. 6 Diagram of one possible method of
performing the present invention.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT(S)
[00021] One possible embodiment of the present invention
teaches a methodology for constructing the empirical model
portion of a hybrid model, such as for an engine, in a
sequential manner without the requirement for storing all of
the original engine data previously collected and stored. The
method involves sequentially developing and storing a compact
statistical and parametric representation of the data, as it
is collected, and generating representative pseudo-data
samples from these models to be used in a piecewise model
building process. As used herein, wpseudo-dataw refers to a
generated data set having the same statistical and inter-
parameter dependencies as the original data set it is intended
to mimic.
[00022] One consideration that must be addressed in the
practical implementation of the hybrid model system described
above is that measurement residuals are likely to vary with
flight condition (e.g. mach and altitude) for the same engine
power condition. As a result, the present invention teaches
the partitioning of the flight envelope to allow individual
empirical representations to be derived in lieu of using one
empirical model for the entire flight regime.
[00023] Thus, one possible method of performing the present
invention supports an incremental approach to empirical
modeling such that it does not expect that an engine will
experience the entire flight regime in a single flight. In a
preferred embodiment, the present invention partitions the
flight envelope into sub-regions as a function of pertinent
independent flight parameters. With reference to Fig. 2,
there is illustrated an exemplary partition scheme wherein
ambient pressure (Pajnb) and Reynold's Index ( Rel ) are chosen
as the defining parameters 21. In such a scenario, it is
possible to effectively capture inlet temperature and pressure
variations, altitude and mach number effects.
[00024] Individual points 23 represent where measurement data
is available and residuals, representing the difference
between the physics based model and the actual sensor
measurements, are computed. Groupings of points obtained from
measurements from a particular flight or a portion of a flight
experiencing a well defined flight regime tend to form
discrete clusters but can overlap with data recorded from
other flight regimes. Over time, the grid 25 become more
complete and the individual (regional) models can be built
each corresponding to a discrete region 27. Each region 27 is
represented by an individual empirical element that takes the
form of, but is not limited to, a Multi-Level Perceptron
Artificial Neural Network (MLP ANN) for each residual
measurement under consideration. The evaluation of a partition
model requires continuous interpolation between models of
adjacent regions 27 so that the empirical estimates can be
continuously generated as an engine traverses several flight
regions 27 in real time.
[00025] The completed empirical model is formed by the
concatenation of the individual sub-region models with an
appropriate regime recognition logic controlling the sub-model
evaluation and interpolation where required. An empirical
model is considered complete when all previously or presently
observed data reside in a sub-region that has been modeled.
[00026] The partitioning of the flight envelope contributes to
the concept of sequential modeling in that it allows the
construction of a predefined series of sub-models to represent
the model space. Since the grid 25 is pre-defined (in order to
limit the number of such sub-models), it is conceivable, and
in fact likely, that insufficient data within a given grid
element, or region 27, will be collected during a single
flight to properly model the subspace. It should be clear
that, no matter what particular modeling methodology is
utilized, the entire set of data populating the grid 25 must
be maintained for the proper modeling of a given sub-region
27. As noted, prior art methodologies for modeling and entire
flight envelope would require the storage of the entire
partitioned flight envelope resulting in the impractical
storage of a large volume of data. While illustrated in
exemplary fashion as formed of sixteen sub-regions 27, in
practice, the grid 25 is not so limited.
[00027] The method of the present invention avoids the problem
of storing prohibitive volumes of flight regime data by
compressing the flight data in the form of statistical and
correlative information at the conclusion of each (MLP ANN)
training session. Then, after the next flight when new data is
introduced (within a given sub-region) a set of pseudo-data is
generated (with proper sample size) having the same
statistical and inter-parameter dependencies as the original
data. This pseudo-data is combined with the newly acquired
data to form a new set upon which the next sequential model is
obtained, after which, the concatenated data set is compressed
as before awaiting the next iteration in this process.
[00028] One possible implementation to capture the statistical
and parametric properties of the data collected during a given
flight is a radial basis function (RBF) ANN, although other
modeling functions could be used which is sufficient to
provide a statistical and correlative model for each dependent
measurement residual that captures the correlation of the
parameter with the set of independent input commands driving
the engine and engine models. The RBF ANN can be used to (re-
generate) a statistically and parametrically consistent sample
of pseudo-data.
[00029] The general process proceeds as follows as illustrated
with reference to Fig. 6:
First, at step 1, data is collected for an initial flight
forming a sample of N"™" data points. At step 2, the
collected data is partitioned into pre-defined sub-regions,
{R,} . Then, at step 3, for each sub-region {R,} for which there
is data (sample of N?™") , an empirical model (e.g. MLP ANN)
MLP, is generated. At step 4, for each sub-region R, for
which there is data (sample of Nf™), a statistical summary
model (e.g. RBF ANN) RBF, is generated. Note that
Y^N?™" =Neurn" . Then , at step 5, {MLP,} and {RBF,} are stored
along with attendant sample sizes Nf" = Nf"", for each region
{R,} . At this point, at step 6, data is collected for a
subsequent flight yielding a sample of j^eumnt data points.
Ncum»t will vary from flight to flight. Then, at step 7, the
flight data is partitioned into pre-defined sub-regions, {R.} .
Then, at step 8, for each sub-region R{, for which there is
data (sample of Nf"™") , MLP, is used to generate pseudo-data
of sample of size Nj"" . Next, at step 9, for each sub-region
R, used in step 8, the current data and pseudo-data is
concatenated to form a data set of size N,= Nf"™'+N!"1'and this
data set is used to generate both a new empirical model (e.g.
MLP ANN) MLP,, and a new statistical summary model (e.f. RBF
ANN) RBF, . At step 10, the generated {MLP,} and {RBF,} are
stored, along with attendant sample sizes NfMUt=N,, for each
region {R,} . Lastly, a determination is made as to whether all
sub-regions {R,} have been adequately modeled. If all sub-
regions {R,} have been adequately modeled, the process is
terminated. If not, steps 7-10 are repeated until all sub-
regions {#,} have been adequately modeled.
[00030] A general architecture 33 supporting the above
procedure is depicted in Fig. 3 and can be used to refine the
architecture in Figure la for developing the empirical model
in a sequential manner using the bootstrap methodology. In a
preferred embodiment, architecture 33 is formed of a general
purpose computing device (not shown) adapted, through the
implementation of hardware and software, to carry out the
storage and retrieval of inputs, outputs, and intermediate
computational results, as well as to perform computations upon
data. For example, a computer, such as a personal computer or
other such electronic computing device formed of a central
processing unit and a data storage and retrieval device, may
be used to provide a means for partitioning the sensory data
into sub-regions, a means for generating an empirical model of
at least one the sub-regions; a means for generating a
statistical summary model for at least one of the sub-regions,
a means for collecting additional sensory data, a means for
partitioning the additional sensory data into the sub-regions,
a means for generating pseudo-data using the empirical model,
and a means for concatenating the pseudo-data and the
additional sensory data to generate an updated empirical model
and an updated statistical summary model for at least one of
the sub-regions. In addition, any sensor, such as a
thermometer or other sensory device adapted to sense an
environment parameter may be utilized as a means for
collecting the sensory data.
[00031] The process outlined above provides a foundation for
an on-board implementation of the architecture presented in
Fig. 3 for developing a hybrid engine model. To illustrate the
efficacy of one possible method of performing the present
invention, an empirical model using engine residual data was
created and then re-created using bootstrap pseudo-data as
outlined above. The salient features of this experiment are
illustrated with reference to Figure 4 below. The chart in the
upper left hand corner contains the Nl residuals 41 between
the engine and the physics-based engine model, as well as
several of the input parameters driving the engine and model
(e.g. low pressure compressor speed (Nl) excursion from Idle
to take-off to Idle (43), stator vane angle (SVA) (45), and
various bleed commands, etc). The chart below it represents
the original residuals and the MLP model of the residuals
(47) . The chart in the upper right represents bootstrap data
(following the above procedure) for this same excursion. The
scrambled appearance arises from the fact that there is no
memory of time sequence for the data in the RBF representation
that is used to manufacture the pseudo-data. It is as if we
took the original data (left-hand chart) and permuted it. The
chart in the lower right reflects the MLP modeling (49)
accomplished using just the scrambled bootstrap data alone,
superimposed on the original residual sequence (41). Comparing
the two lower charts demonstrates the efficacy of the
procedure. Figure 5 depicts the original residual sequence
(41), the original model MLP (47), and the bootstrap modeled
MLP (49). The difference between the two MLPs 47, 49 is quite
small.
[00032] This strategy of employing pseudo-data to
incrementally build the hybrid portion (MLP) within each
flight envelope partition works because the model does not
explicitly use time as a modeling parameter. If one were to
take the original residual (time) sequence and scramble it in
any order, one would obtain the same" empirical model MLP
(assuming one uses the same (typically random) initial
weights). The small difference between the original MLP 47 and
the Bootstrap MLP 49 is caused by the statistical variability
in the pseudo-data generation using the radial basis functions
(RBFs) from the RBF model. The bootstrap data is statistically
consistent with the original (time-sequenced) data, but of
course, not identical. Repeating this process for the
remaining gas path parameters, provides similar results.
[00033] The effect (of using bootstrap data) on estimating
module performance deltas as depicted in Figure lb is
negligible. One is aided in practicing the present invention
by the fact that one is modeling parameter residuals. The gas
path parameters of the real engine, of course, have a time
dependency, since this does represent a real physical process.
Fortunately, the physics-based engine model also must share
the same time dependency. The difference between the two, in
effect, cancels the time dependency in the residuals.
[00034] It is apparent that there has been provided in
accordance with the present invention a method for
sequentially building a hybrid engine model which fully
satisfies the objects, means, and advantages set forth
previously herein. While the present invention has been
described in the context of specific embodiments thereof,
other alternatives, modifications, and variations will become
apparent to those skilled in the art having read the foregoing
description. Accordingly, it is intended to embrace those
alternatives, modifications, and variations as fall within the
broad scope of the appended claims.
We Claim :
1. A method for modeling engine operation comprising the
steps of:
inputting a first plurality of sensory data from a
gas turbine engine into a computer;
partitioning a flight envelope into a plurality of
sub-regions according to the steps of:
selecting a first sensory parameter and a second
sensory parameter,
plotting each of said first plurality of sensory
data by using said first sensory parameter as a first
axis and said second sensory parameter as a second
axis, and
dividing said first axis and said second axis into
a plurality of subdivisions to create a grid comprising
a plurality of sub-regions;
assigning said first plurality of sensory data
into said plurality of sub-regions;
generating an empirical model of at least one of
said plurality of sub-regions;
generating a statistical summary model for at
least one of said plurality of sub-regions;
inputting an additional plurality of sensory data
from said gas turbine engine into said computer;
partitioning said second plurality of sensory data
into said plurality of sub-regions;
generating a plurality of pseudo-data using said
empirical model;

concatenating said plurality of pseudo-data and
said additional plurality of sensory data; and
outputting using said computer an updated
empirical model and an updated statistical summary
model for at least one of said plurality of sub-
regions .
2. The method as claimed in claim 1, comprising the
additional step of repeating steps 8 through 11 until an
updated empirical model and an updated statistical summary
model is generated for each of said plurality of sub-
regions .
3. The method as claimed in claim 1, wherein said
selecting said first sensory parameter and said second
sensory parameter comprises selecting ambient pressure and
Reynold's Index.
4. The method as claimed in claim 1, wherein said
generating said empirical model comprises generating a
multi-level perceptron artificial neural network (MLP ANN).
5. The method as claimed in claim 1, wherein said
generating said empirical model comprises concatenating a
plurality of said empirical models each corresponding to
one of said plurality of sub-regions.
6. The method as claimed in claim 1, wherein said
generating said statistical summary model comprises
generating a radial basis function (RBF) ANN.
7. The method as claimed in claim 1, wherein inputting
said plurality of sensory data comprises collecting a
plurality of residuals each formed from the difference
between an engine measurement and an output of a physical
model of said engine.
8. A method for modeling engine operation comprising the
steps of:
inputting a first plurality of sensory data from a gas
turbine engine into a computer;
partitioning a flight envelope into a plurality of
sub-regions according to the steps of:
selecting a first sensory parameter and a second
sensory parameter,
plotting each of said first plurality of sensory
data by using said first sensory parameter as a
first axis and said second sensory parameter as a
second axis, and
dividing said first axis and said second axis into
a plurality of subdivisions to create a grid
comprising a plurality of sub-regions;
assigning said first plurality of sensory data
into said plurality of sub-regions;
generating an empirical model of a portion of said
plurality of sensory data;
generating a statistical summary model for said
portion of said plurality of sensory data;

inputting an additional plurality of sensory data
from said gas turbine engine into said computer;
generating a plurality of pseudo-data using said
empirical model;
concatenating said plurality of pseudo-data and
said additional plurality of sensory data; and
outputting using said computer an updated
empirical model and an updated statistical summary
model for at least a portion of said sensory data.
9. The method as claimed in claim 10, wherein said
inputting said sensory data comprises collecting sensory
data from a gas turbine engine.
10. The method as claimed in claim 10, wherein said
generating said empirical model comprises generating a
multi-level perceptron artificial neural network (MLP ANN).
11. The method as claimed in claim 10, wherein said
generating said statistical summary model comprises
generating a radial basis function (RBF) ANN.
12. The method as claimed in claim 1, wherein inputting
said plurality of sensory data comprises collecting a
plurality of residuals each formed from the difference
between an engine measurement and an output of a physical
model of said engine.
13. An apparatus for modeling engine operation comprising:

means for inputting a first plurality of sensory
data from a gas turbine engine into a computer;
means for partitioning said first plurality of
sensory data into a plurality of sub-regions, said
means for partitioning comprising the following:
means for selecting a first sensory parameter and
a second sensory parameter,
means for plotting each of said first plurality of
sensory data by using said first sensory parameter as a
first axis and said second sensory parameter as a
second axis, and
means for dividing said first axis and said second
axis into a plurality of subdivisions to create a grid
comprising a plurality of sub-regions;
means for generating an empirical model of at
least one of said plurality of sub-regions;
means for generating a statistical summary model
for at least one of said plurality of sub-regions;
means for inputting an additional plurality of
sensory data from said gas turbine engine into said
computer;
means for partitioning said second plurality of
sensory data into said plurality of sub regions;
means for generating a plurality of pseudo-data
using said empirical model;
means for concatenating said plurality of pseudo-
data and said additional plurality of sensory data; and
outputting using said computer an updated empirical
model and an updated statistical summary model for at
least one of said plurality of sub-regions.
A method of constructing an empirical model, comprising
steps of:
inputting a first plurality of sensory data from a
gas turbine engine into a computer;
partitioning an operating envelope into a-
plurality of sub-regions according to the steps of:
selecting a first sensory parameter and a second
sensory parameter,
plotting each of said first plurality of sensory
data by using said first sensory parameter as a
first axis and said second sensory parameter as a
second axis, and
dividing said first axis and said second axis into
a plurality of subdivisions to create a grid
comprising a plurality of sub-regions;
assigning said first plurality of sensory data
into said plurality of sub-regions;
generating an empirical model of at least one of
said plurality of sub-regions;
generating a statistical summary model for at
least one of said plurality of sub-regions;
inputting an additional plurality of sensory data
from said gas turbine engine into said computer;
partitioning said second plurality of sensory data
into said plurality of sub-regions;
generating a plurality of pseudo-data using said
empirical model;

concatenating said plurality of pseudo-data and
said additional plurality of sensory data; and
outputting using a computer an updated empirical
model and an updated statistical summary model for at
least one of said plurality of sub-regions.

The present invention discloses a method for
modeling engine operation comprising the steps of: inputting
a first plurality of sensory data from a gas turbine engine
into a computer; partitioning a flight envelope into a
plurality of sub-regions according to the steps of:
selecting a first sensory parameter and a second sensory
parameter, plotting each of said first plurality of sensory
data by using said first sensory parameter as a first axis
and said second sensory parameter as a second axis, and
dividing said first axis and said second axis into a
plurality of subdivisions to create a grid comprising a
plurality of sub-regions; assigning said first plurality of
sensory data into said plurality of sub-regions; generating
an empirical model of at least one of said plurality of sub-
regions; generating a statistical summary model for at least
one of said plurality of sub-regions; inputting an
additional plurality of sensory data from said gas turbine
engine into said computer; partitioning said second
plurality of sensory data into said plurality of sub-
regions; generating a plurality of pseudo-data using said
empirical model; concatenating said plurality of pseudo-data
and said additional plurality of sensory data; and
outputting using said computer an updated empirical model
and an updated statistical summary model for at least one of
said plurality of sub-regions. An apparatus for modeling
engine operation is also disclosed.