A Scalable and Adaptive Tool for Rapid Process Modeling
Project Duration: October 2003 – April 2005
Sponsor: Ohio Aerospace Institute
Collaborator: Parker Hannifin, B. F. Goodrich
Aerospace
Synopsis
The goal of the proposed
project is to develop a software tool that will allow engineers to accurately
model complex processes in a short time and with low costs. The need for process modeling exists in many
aspects of an industrial organization, including product design, manufacture,
and service. The ability to quickly and
accurately model a process will lead to improved quality and productivity,
reduced costs, and shortened time to market.
Quarter I (10/15/03
– 1/14/04) Activity Summary
1.
Project kickoff.
Because of the long collaborative relation with Parker Hannifin,
conversation on identifying a target application started in August 2003, before
this project officially started. After
several teleconferences with Michael Benjamin and Ira Rubel. It was determined that probe array based fuel
measurement would be the target application.
Subsequently, Rich Orbell at Parker Hannifin’s Electronic Systems
Division provided background information and initial data for model building. In mid-November 2003, initial results were
presented to Rich Orbell and it was determined that additional data needs to be
collected, marking the official start of the project.
2.
Preliminary model based on small samples. It was
determined that the first task is to investigate the ability of a conventional
neural network based model for fuel measurement. It was found that all probes
(including redundant probes) need to be used to achieve required measurement
accuracy on the training data. However,
the model cannot achieve required accuracy on the testing data, as the testing
data contains certain level of noise. Click here for a
report.
3.
Follow-up model based on large samples. To improve accuracy on noisy data, a larger
amount of training samples were generated.
We also investigated the use of different inputs (use pitch and roll in
addition to probe and use total wet length of all probes). The results showed that a conventional neural
network based model was not able to achieve the desired accuracy of 0.16% error
on average and 0.8% error maximum. Click here for a
report. The question is whether the
desired accuracy can be achieved using Parker Hannifin’s current table look-up
method.
Quarter II (1/15/04
– 4/14/04) Activity Summary
- Assessment of possible model accuracy. It is apparent that neural network model
accuracy is adversely affected by noise in the data. Therefore, it is necessary to establish
an appropriate expectation of model accuracy with noisy data. The table look-up method has a high
accuracy, so the noisy data should be tested with the table look-up method
to establish a benchmark. Parker
Hannifin’s assistance is needed to establish such a benchmark. The high accuracy of the table look-up
method is partially attributed to that fact that it uses the sum of wetted
probe length as an input, which tends to cancel out white noise. Although this cannot be done with a
conventional neural network, it inspires the investigation of alternative
modeling approaches.
- Investigation of alternative modeling
approaches. We have designed a “reverse modeling” approach. The basic idea is as follows. The probe readings (wetted length) are a
result of the fuel volume and pitch and roll of the tank. Therefore, a “forward model” can be
build with volume, pitch and roll and input and probe readings as
output. As pitch and roll of the
tank can be obtained separately (without resorting to triangulation of
wetted probes), fuel measurement can be viewed as a “reverse modeling”
problem by finding the volume that produces matching probe readings. This approach can also handle the
situation when some probes malfunction and produces erroneous or no
readings. Click here
for details of this modeling approach.
- Development of reserve modeling approach. We are currently implementing the
reverse modeling approach, which requires two algorithms. The first algorithm is to allow building
the forward model with “don’t care” outputs, namely, the probes have no
readings because they are not wetted.
The second algorithm is to search for the input volume that
produces “matching” probe readings.
There is considerable flexibility in defining “matching,” e.g., the
sum of wetted probe length so white noise can be cancelled out.
Quarter III (4/15/04
– 7/14/04) Activity Summary
- Implementation of reverse modeling algorithms. Three different reverse modeling
algorithms were developed. The
first one, LMR, is base on the Levenberg-Marquardt optimization method
with a regular quadratic objective function (sum of squared error). The second, GDR, is based on first order
gradient descend search with a quadratic objective function. The third, GDR1, is based on first order
gradient descend search with a linear objective function (sum of
error). The rationale for
developing the third algorithm is that white noises tend to cancel out
when summed. There three algorithms
have been implemented in Microsoft Visual C++. Click
here for a description of these three algorithms.
- Brainstorming Additional Modeling Methods. Preliminary results with reverse
modeling algorithms do not indicate superior performance. This is because reverse modeling is
highly dependent on the accuracy of the forward model. A neural network forward model is
difficult to achieve optimal accuracy.
Therefore we brainstormed additional modeling methods. One is to apply reverse modeling to the
table lookup model itself, as it is the most accurate model. This would be achieved using stochastic
optimization techniques such as particle swarm or genetic algorithms. Another is to develop different neural
network models for different flight situations, since it was observed that
probe readings behave different in severe flight situation. The other is to use historical
information (progressive decrease of fuel volume during flight) to remove
noise from probe readings. Click here
for a summary of additional modeling methods.
Quarter IV (7/15/04
– 10/14/04) Activity Summary
- Reverse modeling results. The three reverse modeling algorithms
were applied. For noise free test
data, reverse modeling actually produces slightly inferior results
compared with the original neural network model. For noisy test data, the GDR algorithm
improves the original neural network model results slightly. The advantage of reverse modeling is
more evident when redundant probes are removed. However, we should note that none of the
reverse modeling algorithms produces results that meet the accuracy
requirements, while the original neural network model meets the accuracy
requirements when applied to noise free test data. This may be caused by the lack of access
to an accurate forward model (the forward model used for reverse modeling
is developed using neural networks and has inherent inaccuracy). Click here
for a summary of reverse modeling results.
- Study of model robustness. Since reverse modeling shows promise
when redundant probes are removed, we believe it would functional well in
the presence of faulty probes.
Dealing with faulty probes is a strong motivation for starting this
project since the table lookup method is very accurate and handles noisy
data reasonably well. Its
disadvantage is that it would not produce good results when there are
faulty probes. There was an
expectation that a neural network model can deal with faulty probes
gracefully. However, this was found
not to be true. The performance of
the neural network model deteriorates quickly even with only two faulty
probes. On the other hand, particle
swarm optimization (PSO) proves to be very robust. Gradient search based reverse modeling
can be used to fine tune the result of PSO to achieve slight performance
improvement. Click here for
a summary of model robustness study.
Quarter V (10/15/04
– 1/14/05) Activity Summary
- Comparison of different modeling techniques. We have explored different modeling
techniques and it is time to consolidate the efforts. Reverse modeling requires the use of
pitch and roll, whereas forward modeling may or may not use pitch and roll
as input. Therefore, we
consolidated all previous efforts to develop three models; namely, (1)
forward model without pitch and roll as input (FMO), (2) forward model
with pitch and roll as input (FMI), and (3) reverse model (REM). We evaluate the performance of these
three models with noise-free data and noisy data and under the assumption
of the presence of faulty probes. Click here
for a summary of result comparison.
- Methods to improve model accuracy and robustness. Since none of the models meet accuracy
requirements when data contains noise, we believe data preprocessing
methods must be developed to reduce noise.
One such method is to sample multiple readings from a probe and
take the average as model input. We
also believe that Parker’s table look-up method can be modified to deal
with faulty probes gracefully. This
requires knowing pitch and roll and constructing a look-up table for each
probe. We will discuss the
feasibility of such an approach with Parker engineers.
Quarter VI (1/15/05
– 4/15/05) Activity Summary
- Fuel measurement project wrap-up. We discussed project findings and
recommendations with Parker engineers.
Our findings showed that simply applying a neural network modeling
approach does not lead to an algorithm that tolerates noise and probe
failure. This supports Parker
engineers' skepticism on some competitors' unsubstantiated claims. Our recommendations are also consistent
with Parker engineers' knowledge.
The fuel measurement project was successfully concluded. Click here for
a summary of the project.
- Modeling software tool development. In parallel to this project, we worked
with VerTech LLC to apply rapid process modeling in machinery diagnosis
and prognosis under a SBIR (Phase I and II) project supported by the
National Institute of Standards and Technology. A software tool, DP, was developed. An installation package of the software
tool is available upon request. Click here for the instruction manual.