MAP-Controlled Neural Network for ICE Raw Emission Modelling

Abstract

IAV developed an combination of a MAP-based raw emission model, which is input to a Neural Network. Focus is on raw emissions from ICE vehicle (NOx, CO, HC, PM.....)
The MAP is calibrated with stationary measurements from a engine test bench with high robustness through DoE methods and a wide range of input variation (tolerances and calibration range).
The Neural Network is for dynamics and environmental corrections (temperature, pressure, humidity).
The outcome is a robust and accurate emission model which fits into current standard ECUs.

Full text

Standard
ECU
Map-Controlled
Neural Network
Humidity
Fuel Blend
Engine Test
Bench
Entire engine operating area
Accurate steady state
measurements
Variation of main inputs
(tolerances, possible
calibration area)
Altitude Climatic
Roller Test Bench
Vehicle environment
Pressure
Temperature
Humidity
Fuel blends
Map-Based
Model
Machine learning
Polynomial 
Ploy2Map
Robust inter- and
extrapolation
Well-known classic
ECU structure
IAV - 2023-09-13 - Marco Moser, Steffen Schaum
ETB
DoE
Design of
Experiment
Optimizing test
effort for the needs
of model learning
Reduction of
measurement effort
AC-RTB
MAP NN Neural
Network
Dynamics
modelling
Robustness to
environmental
influences
Limited by MAP-
based tolerances
Feature
Engineering
Environmental
influences
Fuel blend detection
Gradients
ICE Raw Emission
Modelling
±
Test Bench
Methodology
Model