Neuro Emission Controller for Spark Ignition Engines

1
Neuro Emission Controller for Spark Ignition
Engines

• Jagannathan Sarangapani and Jim Drallmeier

Department of Electrical and Computer
Engineering Department of Mechanical and
Aerospace Engineering The University of Missouri
at Rolla 1870 Miners Circle, Rolla, MO,
65409 Tel [phone] Email sarangap_at_umr.edu
Santa Fe, New Mexico June 29th-July 1st
This research is supported in part by NSF award
ECS0296191 and ECS0327877
2
Outline

• Motivation
• Lean Combustion Emission control.
• Exhaust Gas Recirculation (EGR) Emission control.
• Lean Controller implementation.
• Conclusions.
• Future work.

3
I. Introduction

• Motivation.
• Cyclic dispersion in heat release during lean
  engine operation and EGR
• Neural network (NN) universal approximation
  property.
• Uniformly ultimate boundedness (UUB).

4
I. Motivation

• Lean operation on a spark ignition (SI) Engine
  can reduce emissions (HC, CO NOx) by as much as
  30 and also it improves fuel efficiency by as
  much as 5 10.

• Engines operating with high EGR (exhaust gas
  circulation levels) can further reduce emissions
  by as much as 50 to 60.
• Major problem in operating the engine in either
  regimes (extremely lean or with high EGR) is
  cyclic dispersion of heat release.
• Other major problem most controller designs are
  performed in continuous time domain whereas
  discrete-time controller development is necessary
  in order to implement it using the embedded
  hardware.

Fig. 1. Emission profile (Heywood, 1998).
5
I. Cyclic Dispersion in Heat Release
Similarity between Lean and EGR
Fig. 3. High EGR levels
Fig. 2. Extreme lean operation
6
I. Multilayer Neural Networks

• Functional Approximation
• Learning and Adaptation
• Parallel Processing
• Noise Filtering

7
I. Neural Network (NN) Approximation
(1)
where w and v are target weights, denotes
the activation functions, and is the
NN functional reconstruction error.
(2)
where is the estimated weights.
B. Igelnik and Y. H. Pao, Stochastic choice of
basis functions in adaptive function
approximation and the functional-link net, IEEE
Trans. Neural Networks, vol. 6, 1995.
8
I. Uniformly Ultimately Boundedness
Given the Nonlinear System
(3)
for any
and
, there exists a
, such that
for all
9
II. SI Engine Dynamics without EGR
(4) (5) (6)
total mass of air in the cylinder
total mass of fuel in the cylinder
small fresh fuel changes, the control
combustion efficiency
Fig. 4. Illustrations of engine dynamics.
equivalence ratio
C. S. Daw, C. E. A. Finney, M. B. Kennel and F.
T. Connolly, "Observing and modeling nonlinear
dynamics in an internal combustion engine",
Phys. Rev. E, vol. 57, no 3, pp.2811 2819, 1993.
10
II. Model Output Vs. Experimental Data
Fig. 5. Experimental data.
Fig. 6. Simulation result.
11
II. NN State Feedback Lean Emission Control

• Control objective.
• Stabilize the engine operation at lean condition
  by precisely controlling the equivalence ratio
  using NNs.
• Assumption.
• The measurement of the total mass of air and fuel
  in the cylinder at every combustion cycle is
  available.
• Relaxed later

12
II. Controller Development

• Compact Representation of the Engine Model

(7)
(8)
(9)
(10)
(11)
13
II. Control Objective

• Reducing the cyclic dispersion by minimizing the
  variations in equivalence ratio through forcing
  both the states to be bounded close to their
  respective targets.

(12)
14
II. Tracking Error-Based NN Control

• Definitions of the system errors.
• Virtual control design.
• Control input design.

(13) (14) (15) (16)
15
II. NN State Feedback Lean Emission Control

• Theorem.
• Given the system (1) and (2), let the
  disturbance and NN approximation errors be
  bounded. Let the controller be provided as in
  Fig. 4. Take the first NN weight tuning be
• with the second NN weight tuning be
  provided by
• Then the system errors and the NN weights
  estimation are UUB provided the design parameters
  are selected as

(17) (18) (19) (20)
P. He and S. Jagannathan, Neuro emission
controller for minimizing cyclic dispersion in
spark ignition engines, in Proc. Int. Joint
Conf. Neural Networks, Portland, OR, 2003, pp.
15351540.
16
II. NN Lean Emission Controller Structure
Fig. 7. NN state feedback lean emission
controller.
17
II. NN Lean Emission Controller Structure
Neural Network Controller
Engine
Control Inputs
Measurements
18
II . Testing and Verification

• Simulation parameters.
• a) Desired equivalence ration
  .
• b) Residual gas fraction .
• c) Fresh air and fuel and
  .
• d) Desired target values for total air and
  fuel and .
• e) Controller gains .
• f) NN adaptation gains
  .
• g) Hidden layer nodes 15.
• h) Initial hidden layer weights selection
  uniformly distributed in 0,1.
• i) Initial output layer weights selection
  zeros.
• j) The cycles 1000.

,

.
19
II. Lean Emission Controller Performance
Fig. 9. Heat release with NN controller.
Fig. 8. Heat release without control.
20
II. Equivalence Ratio Error and Control Input
Fig. 10. Equivalence ratio error.
Fig. 11. Control Input.
21
II. Conventional Controller
Fig. 12. Controller performance.
Fig. 13. Control input.
22
II. Reinforcement Learning-Based NN Control

• Performance index.
• Definition of the utility function .
• Definition of the strategic utility function
  .
• Critic signal is to approximate the
  strategic utility function.

(21) (22) (23)
23
II. Reinforcement Learning-Based NN Control

• Controller development.
• Definitions of the system errors.
• Virtual control design.
• Control input design.

(24) (25) (26) (27)
24
II. Reinforcement Learning-Based NN Control

• Theorem.
• Given the system (2) and (3), let the
  disturbance and NN approximation errors be
  bounded. Let the controller be provided as in
  Fig. 12. Take the first and second action NN
  weight tuning be
• with the critic NN weight tuning be
  provided by
• Then the system errors and the NN weights
  estimation are UUB provided the design parameters
  are selected as

(28) (29) (30) (31)
P. He and S. Jagannathan, Reinforcement-learning
neural network-based control of nonlinear
discrete-time systems in non-strict form,
submitted to Proc. IEEE Conf. Decis. Contr.,
Bahamas, 2004.
25
II. Reinforcement Learning NN Controller Structure

• Contribution.
• Optimization of certain long-term system
  performance index is undertaken
• Demonstration of the UUB of the overall system is
  shown even in the presence of NN approximation
  errors and bounded unknown disturbances.
• The NN weights are tuned online instead of
  offline training

Fig. 14. Reinforcement learning based NN
controller.
26
II. Reinforcement Learning-Based NN Control

• Simulation parameters for the engine dynamics.
• a) Desired equivalence ration
  .
• b) Residual gas fraction .
• c) Fresh air and fuel and
  .
• d) Desired target values for total air and
  fuel and .
• e) Controller gains .
• f) NN adaptation gains
  .
• g) Hidden layer nodes 15.
• h) Initial hidden layer weights selection
  uniformly distributed in 0,1.
• i) Initial output layer weights selection
  zeros.
• j) The cycles 1000.

,

.
27
II. Reinforcement Learning-Based NN Control
Fig. 15. The heat release with NN controller.
Fig. 16. The equivalence ratio error.
28
II. NN Output Feedback Lean Emission Control

• Control objective.
• Stabilize the engine operation at lean condition
  by precisely controlling the equivalence ratio
  using NN heat release observer and NN.

29
II. NN Output Feedback Lean Emission Control

• Simulation parameters.
• a) Desired equivalence ration
  .
• b) Residual gas fraction .
• c) Fresh air and fuel and
  .
• d) Desired target values for total air and
  fuel and .
• e) Controller gains .
• f) NN adaptation gains
  .
• g) Hidden layer nodes 15.
• h) Initial hidden layer weights selection
  uniformly distributed in 0,1.
• i) Initial output layer weights selection
  zeros.
• j) The cycles 10000.

,

.
30
II. NN Output Feedback Lean Emission Controller
Performance
Fig. 17. Output Feedback Controller Performance.
31
III. EGR Emission Control

• SI engine dynamics with EGR.
• NN EGR emission controller Design

32
III. SI Engine Dynamics with EGR
33
III. NN Emission Control with EGR

• Control objective.
• Stabilize the engine operation with high levels
  of EGR by precisely controlling the equivalence
  ratio.
• Assumptions.
• All the states are available for measurement and
  the amount of EGR is precisely controlled.

34
III. EGR Emission Controller Development

• Controller development is quite similar to the
  lean operation except the EGR is taken as an
  additional input.
• EGR system will have a separate controller to
  accurately allow the EGR into the intake

35
III. EGR Emission Controller Structure
Fig. 18. EGR controller structure with EGR.
36
III. NN Emission Control with EGR

• Simulation parameters.
• a) Desired equivalence ratio One
  .
• b) Residual gas fraction F 0.15
  .
• c) Molecular weight of fuel, air and EGR
  114, 28.84 and 30.4 .
• d) Total gas mole 0.5 ratio of
  hydrogen/carbon 1.87.
• e) Controller gains .
• f) NN adaptation gains
  .
• g) Hidden layer nodes 15.
• h) Initial hidden layer weights selection
  uniformly distributed in 0,1.
• i) Initial output layer weights selection
  zeros.
• j) The cycles 10000.
• k) EGR 21.2 with a variation of 0.004.

,

.
37
III. EGR NN Emission Controller Performance
Fig. 19. EGR Controller Performance.
38
IV. Embedded Emission Controller Hardware
Fig. 21. SI engine at UMR.
Fig. 20. PC 770 SBC.
39
V. Conclusions.

• Emission Controller so far indicates that
  significant reductions in emissions can be
  possible by operating the engine at extreme lean
  operation along with high EGR levels
• Neural network controllers can successfully limit
  the amount of cyclic dispersion in heat
  release
• Hardware Implementation of the controller is
  currently being addressed on an SI Engine

40
VI. Future Work.

• Time Delays due to the Feedback has to be
  compensated.
• Develop and Implement the proposed output
  feedback NN controller on the hardware.