Deep Learning Based Modeling of A Gas Turbine PDF

Indian Institute of Technology Bombay
Dynamic Modeling
Of Mini SR-30 Gas
Turbine Engine
Presented By:
Prof. P. S. V. Nataraj 1
Outline
▪ Quick glance at deep learning

▪ Introduction to gas turbine engine
▪ First principle based modeling
▪ Deep learning based modeling
▪ Results and validation
▪ Conclusion
2
Quick Glance at Deep Learning
A brief Introduction:
• 1943 - Walter Pitts and Warren McCulloch, gave us that piece of the puzzle
when they created the first mathematical model of a neural network.
• 1946 – John Mauchly & J. Presper Eckert develop world’s first digital
computer ‘ENIAC’ .
• 1952 — Arthur Samuel writes the first computer program capable of
learning.
• 1958 — Frank Rosenblatt designs the Perceptron, the first artificial
neural network.
What has fueled the development of deep learning?

1. Explosion of data.
2. Cheap computing cost – CPUs and GPUs.
3. Improvement of machine learning models.
3
Evolution of Deep Learning
4
Evolution of Deep Learning
Break through
Algorithm
Big data
CPU/GPU
5
Deep Learning revolutionized Machine learning:
• Deep learning don’t need to provide features ahead of time, it learns features
at different levels by itself.
• Same deep learning architecture can be trained to accomplish different
tasks.
6
Deep Learning revolutionized Machine learning:
• Deep learning don’t need to provide features ahead of time, it learns features
at different levels by itself.
• Same deep learning architecture can be trained to accomplish different
tasks.
7
Major area of research
8
Applications in various sector
9
SR-30 Gas Turbine Engine
Fig : Cross-sectional view

of laboratory SR-30
engine
Fig : schematic flow

diagram of laboratory
engine
10
Why Engine Model is Necessary??
To assess real world Ease of dynamic
phenomena simulation
Investigating strong Cost saving strategy

adverse dynamic for performance
conditions optimization
Fault diagnosis &

Sensor validation
detection
Model applications
Design and Condition monitoring
optimization of Or engine health
control system monitoring
11
Dynamic Model of SR-30 Engine
Objective:
• Develop a non-linear dynamic engine model.
• Simulate the steady state and transient performance.
• Integrate the developed gas turbine model with multi-
disciplinary systems.
Challenges:
• Experimental data
• Characteristics map of engine components.
• Tuning of characteristics maps.
• Simulate the model over full operating range.
12
Approaches for Modeling Dynamic Systems
13
Problem Statement
Fig : Illustration of input and output variables of the model 14

First Principle Modeling Approach
State variable method:
15
First Principle Based Engine Simulator
Fig : Component-wise Simulink Model of SR-30 Gas Turbine Engine 16

First Principle Based Engine Simulator
Fig : Closed loop Model of SR-30 Gas Turbine Engine along with dashboard tool 17
Motivation for Data Driven Techniques
1. White-box or First principles modeling approaches rely on
thermodynamic and energy balance equations. Hence, assumptions and
linearization methods are required to simplify and solve complex dynamics.
2. Models and control systems designed using simplified linearized
equations are not accurate enough to capture system dynamics precisely.
3. The unavailability of component maps is also one of the key reason to
shift on data driven modeling techniques.
4. Thus, Deep learning is a fair alternative to white box model as it is
independent of the system dynamics with an objective of maximize system
robustness, output power and efficiency.
18
Neural Network Architecture
The model can be mathematically represented
as:
y(t) = f (y(t - 1), y(t - 2),……, y(t – n y),
u(t -1), u(t - 2),……, u(t- nu))
where y (.) is Output, u (.) is Input and n
represents the Delay unit.
Fig : Network Architecture 19

Neural Network Architecture
How to build Deep neural network?
The model can be mathematically represented
as:
y(t) = f (y(t - 1), y(t - 2),……, y(t – n y),
u(t -1), u(t - 2),……, u(t- nu))
where y (.) is Output, u (.) is Input and n
represents the Delay unit.
Fig : Network Architecture 20

LSTM Network Architecture
• Forget gate:
• Input gate
• Cell memory state
• Output gate
Where: W is weight, b is bias, x is input Fig : Detailed schematic of LSTM block

data, y is target data.
21
Network Configuration
22
Model Validation Against Experimental Data
The validation results of First principle model as well as Deep learning based
model against experimental data is represented for shaft speed (full range RPM)
Fig: Shaft speed validation using First Principle model

23
and Deep Learning model
First Principle Model Validation Against
Experimental Data
24
Deep Learning Model Validation Against
Experimental Data
25
Deep Learning Model Validation Against
Experimental Data
26
Relative Error of predicted data against
experimental data
Parameter Deep Learning First Principle

Approach Approach
T2 0.002 0.0441
P2 0.0012 0.0112
T3 0.0031 0.1297
P3 0.0011 0.0255
T4 0.009 0.1428
P4 0.0036 0.0721
N 0.001 0.0014
27
Conclusions
• First Principle based method promises good dynamic behavior
when compared with the real time engine, provided that enough
information is available.
• The deep learning model is trained with a set of experimental data
which makes the model to learn a wide variety of engine behavior.
• The Deep Learning approach when compared with the First
Principle model against experimental data is found to be more
efficient in predicting behavior of system.
• LSTM performs good with MIMO system, however LSTM has its
own disadvantage: it is slower than other normal activation
functions which leads to the trial & testing process to be more slow.
28
Why Matlab & Simulink?
• Easy user-interface
• Less programming required while working in Simulink.
• Toolboxes are designed to integrate with parallel computing
environments, GPUs, and automatic C code generation.
• Documentation is written for engineers and scientists, not
computer scientists.
• Inbuilt functions are available that are required in day-to-
day computation.
• MATLAB App let you start working right away and then
automatically generate a MATLAB program to reproduce or
automate your work.
29
Future Steps in Modeling and Control of Engine
30
Acknowledgements
Experimental setup :
• Supervisor:
– P. S. V. Nataraj
• Co - worker
– Bhagyashri Somani (deep learning)
• Data collection from experimental setup:
– Swathi Surendran
– Sanjeet Kulkarni
31
32

Deep Learning Based Modeling of A Gas Turbine PDF

Uploaded by

Copyright:

Available Formats

Deep Learning Based Modeling of A Gas Turbine PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deep Learning Based Modeling of A Gas Turbine PDF

Uploaded by

Copyright:

Available Formats

Indian Institute of Technology Bombay

▪ Quick glance at deep learning

What has fueled the development of deep learning?

Fig : Cross-sectional view

Fig : schematic flow

Investigating strong Cost saving strategy

Fault diagnosis &

Fig : Illustration of input and output variables of the model 14

Fig : Component-wise Simulink Model of SR-30 Gas Turbine Engine 16

Fig : Network Architecture 19

Fig : Network Architecture 20

• Cell memory state

Where: W is weight, b is bias, x is input Fig : Detailed schematic of LSTM block

Fig: Shaft speed validation using First Principle model

Parameter Deep Learning First Principle

You might also like