Content uploaded by Patrick Stracke
Author content
All content in this area was uploaded by Patrick Stracke on Nov 22, 2022
Content may be subject to copyright.
Raw-Emission Modelling in the Context
of EU7 OBM/OBD
Patrick Stracke, Marco Moser, Philipp Brinkmann, Dr. Maximilian Brauer
Content
8th International MinNOx Conference for Sustainable Mobility, October 20222
1. EU7 Requirements with Focus on OBM / OBD
2. OBM challenges and ICE system complexity
3. Emission Modelling Approaches
4. Raw Emission Modelling Diesel
5. Conclusion & Outlook
EURO 7: More Stringent Air Pollutant Standards for ICE
8th International MinNOx Conference for Sustainable Mobility, October 20223
Stricter limits (e.g. NOx: -50% / PM: -56% / PN: -83%)
New pollutants (NH3/ NMOG / CH4/ N2O)
Almost all driving conditions (focus on RDE and PEMS)
Wide range of ambient conditions (-10..+45°C, 2200m)
Extended durability (e.g. 240tkm or 15 years)
On-Board Monitoring (emission compliance over lifetime)
On-Board Diagnostic (Fault detection / Repair in workshop)
European Green Deal as driver for new emission standards
EURO 7: More Stringent Air Pollutant Standards for ICE
8th International MinNOx Conference for Sustainable Mobility, October 20224
Stricter limits (e.g. NOx: -50% / PM: -56% / PN: -83%)
New pollutants (NH3/ NMOG / CH4/ N2O)
Almost all driving conditions (focus on RDE and PEMS)
Wide range of ambient conditions (-10..+45°C, 2200m)
Extended durability (240tkm or 15 years)
On-Board Monitoring (emission compliance over lifetime)
On-Board Diagnostic (Fault detection / Repair in workshop)
Potential Requirements
Monitoring tailpipe emission over lifetime
Regulated emission →e.g. NOx, PM, NH3
OBM limits →not defined yet
MIL activation and enforcement to repair
Identification of multiple faults for repair
Anti-Tampering
European Green Deal as driver for new emission standards
Raw Emission Modeling
•Complex processes
•Cylinder pressure sensor
usually not available
•Real-time capability vs.
accuracy
EAT Modeling
•Complex system setups
•Emissions might be affected by
more than one component
•Real-time capability vs.
accuracy
Sensor / Model
•Availability of sensors limited
•New sensor technologies not
expected
•Model-based monitoring requires
raw emission and EAT models
OBM Challenges and Complexity of Potential ICE Systems
8th International MinNOx Conference for Sustainable Mobility, October 20225
IAV/AECC/IPA Gasoline Zero-Impact Demonstrator
Engine
Engine
HCA
SCR
ASC
SCR SCRF
EHC
PNA
TWC GPF TWC
ASC
IAV Diesel Zero-Impact Demonstrator
12
31
32
1
2
3
→Emission modelling from engine to tailpipe required if no sensors available
EHC
Good prediction accuracy and fast computation speed (3.)
Engine: Raw emission models already in serial application (4.)
(specific emission predicted for control or OBD purpose)
EAT: Initial investigations for catalyst and sensors done (4.)
(End2End models and component specific model structure evaluated)
Modelling Approaches
8th International MinNOx Conference for Sustainable Mobility, October 20226
Computation Speed
Degree of Complexity / Accuracy
Mathematical space ECU capability
Physico-chemical space
2. Reaction
Kinetics
(Low-
Dimensional)
4. Science
informed
Machine
Learning
3. Machine
Learning
Map-
Based
1. Reaction
Kinetics
(Micro-, Macro-
kinetic)
Physico-chemical approach
→IAV OBM concept focus on both approaches for emission modelling
→Machine learning for raw emission and physico-chemical approach for EAT components
Engine and EAT models standard for CAE investigations (1.)
Engine: Model reduction is emission-specific limited (2.)
(0D/1D model for NOx possible; 3D model for HC, CO, PM required)
EAT: Model reduction for catalyst and sensor feasible (2.)
(Simplification of models done at IAV →0D/1D model)
Good performance in unknown operation conditions (2.)
Machine learning approach
Raw Emission Modelling –Machine Learning Algorithms
8th International MinNOx Conference for Sustainable Mobility, October 20227
Polynomial Regression
Feed Forward Neural Networks
Gradient Boosting Trees
Recurrent Neural Networks (RNN)
+ Structure
+ Computation speed
- Time intensive training
- Memory (polynomic terms)
+ Higher accuracy
+ Less memory requirements
- Data pre-processing very important
+ Higher accuracy
+ Learn temporal dependencies
- ECU limitations (complex math operat.)
- Time intensive training
+ Fast training time
+ Computational speed
- Structure / Parameter complex
-Generalization
→Focus in first step on Feed Forward Neural Networks for raw emission modelling
Inputs
ploynomic terms
ŷ = θ0 + θ1x1 + θ2x2 + θ3x1x2 + θ4x21 + θ5x2
Raw Emission Modelling –Science-Informed Model
8th International MinNOx Conference for Sustainable Mobility, October 20228
Features Layer 1 Layer 2 Outputs
Feature Selection
✓System knowledge for
physico-chemical
relevance
Data Preparation
✓Feature and emission
synchronization
✓Emission and signal
plausibility
✓Time-based gradients
HC
CO
NOx
Model Training
✓Fuel specific compatibility check of
emission ratios (COMPA equation
based on Brettschneider-λ)
Post Processing
✓Compensation of Model
Inaccuracies (e.g. dynamic
amplification, low pass)
→Knowledge of the physico-chemical process is essential for machine learning models
Raw Emission Modelling –Data Collection and Preparation
8th International MinNOx Conference for Sustainable Mobility, October 20229
Emission formation
•High transient conditions
•Long term effects from EGR
→Target of model prediction
Measurement location
•Position of sensor or analyzer
•Low response time and gas runtime
•Turbulent and diffusive mixing
→Plausibility of emission ratios (COMPA equa.)
→Synchronize features and emission
1
2
Engine
HCA
SCR
ASC
SCR SCRF
EHC
PNA
HP-EGR LP-EGR
1 2
t1
Loss of
information
t2
→Synchronization of features and emission essential for model training
IAV Diesel Zero-Impact Demonstrator
NOx meas. shift
Raw Emission Modelling –Humidity Impact on Emission
8th International MinNOx Conference for Sustainable Mobility, October 202210
Humidity vs. Temperature
How does humidity affect
the environment oxygen?
→Environmental conditions for raw
emission modelling very important
•Variation of environmental conditions
(temperature: -10°..+30°C, abs. humidity: 2g/m³..35g/m³)
•Humidity increase cause decrease of
environment oxygen →Impact on calorific
properties of cylinder gases
•Emission formation affected by humidity and
environment oxygen
1
2
1
2
3
IAV Diesel Zero-Impact Demonstrator
3
Raw Emission Modelling –Neural Network for Diesel Demonstrator
8th International MinNOx Conference for Sustainable Mobility, October 202211
→Multi-target Neuronal Network show
good potential for emission prediction
Multi-Target Neural Network
HC
CO
NOx
Humidity vs. Temperature
•Variation of environmental conditions
(temperature: -10°..+30°C, abs. humidity: 2g/m³..35g/m³)
•Cumulative emission of NOx, HC and CO in
WLTC cycle
•Good baseline of emission prediction
(NOx: -6% / HC: -2% / CO: -8%)
•Deviations mainly caused at high transient
conditions
1
2
2
1
NOx
HC
CO
IAV Diesel Zero-Impact Demonstrator
Raw Emission Modelling –Neural Network for Diesel Demonstrator
8th International MinNOx Conference for Sustainable Mobility, October 202212
•Good prediction of NOx emission even
partially at high NOx peaks
•HC emission very good predicted except very
transient conditions (high HC peaks)
•High dynamics and concentration levels of CO
emission →Further improvement possible
1
2
→Already good emission prediction for
initial multi-target model
→High potential of further improvement
(e.g. data base, recurrent neural network, feature
engineering, hyperparameter optimization etc.)
→Machine learning approach applicable
to all emissions (e.g. PM emission)
1
23
3
HC
IAV Diesel Zero-Impact Demonstrator
NOx
CO
Raw Emission Modelling –Feature Engineering for Modularization
8th International MinNOx Conference for Sustainable Mobility, October 202213
Emission
Model
Emission
Model Time-Based
Gradients
Physico-chemical
Submodels Humidity
Fuel
Feature
Replacem.
Sensor vs.
Model
→System-specific feature engineering required
→ML models for humidity and fuel derivative →Proof of Concept
Fuel Blend Density Detection
Model of Absolute Humidity
▪NOx/ NH3sensing
▪Limitation due to cross-sensitivity
and dew point release
▪PM sensor not measuring
continuously
Conclusion & Outlook
8th International MinNOx Conference for Sustainable Mobility, October 202214
EU7
OBM/OBD
Development
On-Board-
Monitoring
EAT
Modelling
System
Monitoring
Sensor
Based
Monitoring
Raw
Emission
Modelling
▪Emission compliance over lifetime with focus on OBM
▪Potential OBM emission (e.g. NOx, PM, NH3)
▪MIL activation and enforcement to repair
▪Science-Informed ML algorithm
▪Good emission prediction of
Neural Network
▪Applicable to diesel and
gasoline emissions
▪Availability of sensors / new technologies limited
▪Emission modelling from engine to tailpipe
▪OBM defines requirements for OBD
▪Detection of multiple faults and partial
degradation
▪Enhancement of existing monitors
▪Development of new monitoring
strategies required
▪Physico-chemical EAT models
▪Good emission prediction and
real-time capability
▪Available for diesel and
gasoline applications
Contact
Patrick Stracke
IAV GmbH
Carnotstrasse 1, 10587 Berlin (GERMANY)
patrick.stracke@iav.de
www.iav.com
Many Thanks to
Marco Moser, marco.moser@iav.de
Philipp Brinkmann, philipp.brinkmann@iav.de
Dr. Maximilian Brauer, maximilian.Brauer@iav.de
Andreas Kuhrt, andreas.kuhrt@iav.de
MinNOxexhibition: Light-duty gasoline and diesel demonstrators
IAV/AECC/IPA Gasoline Zero-Impact Demonstrator IAV Diesel Zero-Impact Demonstrator