PresentationPDF Available

Machine Learning for Exhaust AfterTreatment

Authors:

Abstract

Machine Learning for Exhaust AfterTreatment
Machine Learning for EAT
Marco Moser (IAV), 2019
Content
IAV 01/2019 Marco Moser -Machine Learning for EAT
2
Machine Learning for Real Time NOxEstimation
Motivation
IAV Data Processing Tool
Machine Learning (Regression)
Results
Machine Learning for OBD Pinpointing
Motivation
IAV OBD Concept
Machine Learning (Classification)
Results
Conclusion and Next Steps
Machine Learning for Real Time NOxEstimation
NOXModel
Challenges for Exhaust After Treatment Models
IAV 01/2019 Marco Moser -Machine Learning for EAT
4
Complex physical problem with correlation to given data leads to Machine Learning
NOX-Model
Ambiance
physical
modelling
complex characteristic
map structures
machine
learning
NOXModel
IAV Approach
IAV 01/2019 Marco Moser -Machine Learning for EAT
5
Challenge at hand:
Real-time modelling of engine out NOxemissions for OBD or EAT control
Low accuracy and high calibration effort of standard (map-based) models in
ECU’s
IAV Approach:
-Machine learning using real driving data
-Emission estimation by means of neural networks or polynomial models
-Real-time capability and fast implementation on engine controller
Benefits:
-Real-time estimation of real-driving emissions
-Accurate NOxraw emission models for NOx sensor monitoring
-Elimination of NOxsensor Cost reduction
NOXModel
IAV Data Processing Tool Input Definition
IAV 01/2019 Marco Moser -Machine Learning for EAT
6
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
Selection of NOxmodel inputs through dendrogram with engineering knowledge
Filtered Inputs for Machine Learning Model
Ambiance
Data
Analysis Expert
Knowledge
Correlations
of Inputs to
target value
Measurements
NOXModel
Data Processing Tool for NOxModel
IAV 01/2019 Marco Moser -Machine Learning for EAT
7
Measurement data
with defined Inputs
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
*.dat
ppm
ppm
Polynomial Regression with
IAV Data Processing Tool
Dispersion plot for
showing the data
coverage
Scatter-Plot, absolute
Distribution (Sensor
vs. Model)
Histogram with
Gauss-Fit:
Distribution Ratio
(Sensor vs. Model)
DCM
DCM ready-to-use with INCA etc.
without additional (manual) steps
NOXModel
Why Polynomial for the First Steps?
IAV 01/2019 Marco Moser -Machine Learning for EAT
8
Scalable in required computing power and memory
workload in relation to precision
Data and software structures are independent
That’s not the case regarding Random Forests
(dynamic software structures required)
Mathematic problem is convex when alternative
methods are used
That’s not the case regarding many alternative training
methods (neuronal network)
Polynomial preferred for current projects (ECU boundaries)
Deep Learning , Goodgellow, Bengio, page 81
global
minimum
Good local
minimum
bad local
minimum
Not Convex
global
minimum
Convex
Convex means: The optimization of the cost
function has only one minimum which is the only
result at the same time.
NOXModel
Challenges for Machine Learning
IAV 01/2019 Marco Moser -Machine Learning for EAT
9
“normal” Regression insufficient generally non-
linear results with correlating inputs
Enhancers
Big Data needed for all possible operation modes
DOE with experts
Model Function has to be complex enough to prevent
underfitting
Data Analysis
Model Function has to be simple enough to prevent
overfitting
Cost Functions
Complex Data Processing for reliable model needed
NOXModel
IAV Data Processing Tool - Development
IAV 01/2019 Marco Moser -Machine Learning for EAT
10
Python Scripts
Result relies on defined software environment
Freely changeable, expert knowledge assumed
Runnable *.exe
No installation needed
Software environment defined within program
independent of used PC
Configuration possible (.ini)
User-friendly and less prone to errors
Server Based Evaluation (SaaS, „Software as a
Service“)
Currently in development
Useable for series application
exe
NOXModel
Machine Learning - Data Analysis
IAV 01/2019 Marco Moser -Machine Learning for EAT
11
Tools available for detailed analysis of dependencies
NOXlevel
NOXmodel
inputs
NOXModel
Machine Learning - Data Covering
IAV 01/2019 Marco Moser -Machine Learning for EAT
12
Measurements have to cover complete operation range
Missing data can be compensated with Altitude-Climate-RTB
Ambient Pressure [hPa]
600
650
700
750
800
850
900
950
1000
1050
Am bient Temperature [°C]
-20
-10
0
10
20
30
40
area with low quantity of data
NOXModel
Results Learning Process
IAV 01/2019 Marco Moser -Machine Learning for EAT
13
Very good results able to stand every comparison
Diagnose Factor []
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
Number of Diagnoses []
200
400
600
800
1000
1200
1400
1600
1800
Frequency []
10
20
30
40
50
60
70
80
90
100
Diagnose Factor []
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
NOXModel
Results - Sensor Diagnosis
IAV 01/2019 Marco Moser -Machine Learning for EAT
14
Reliable NOXModel for whole operating range
All Operation Modes / Climate Conditions
Cross influences of unavailable data
(e.g. cylinder pressure, humidity)
NOXModel
Results - Reliability of Machine Learning
IAV 01/2019 Marco Moser -Machine Learning for EAT
15
Reliable Machine Learning process
Model data set
Ratio model / sensor (log) Ratio diff.
Real example NOxModel:
100 data sets learned with
70% of the measurements
NOXModel for OBD:
OBD simulated with test trips
of 1 year
Cumulated results
Reliable
Exact
Easy to calibrate
1.1
0.9
NOXModel
Results – Outlook for Machine Learning
IAV 01/2019 Marco Moser -Machine Learning for EAT
16
NOXModel for SCR Control:
In pipeline
Higher time grid
Signal memory for dynamics
More dynamic NOXmeasurement
NOXsensor too slow
NOXmodel for SCR control in pipeline
Model
Sensor
time
NOx
IAV 01/2019 Marco Moser -Machine Learning for EAT
17
Polynomial regression (e.g. 2nd order):
ŷ = θ0 + θ1x1 + θ2x2 + θ3x1x2 + θ4x21 + θ5x2
+ Implementation
+ Computational requirements
-Pure steady state model, no internal dynamical states
Regular fully connected networks:
+ higher accuracy than polynomials
+ less memory requirements than
polynomials
-computational expensive due to
several matrix multiplications
-long execution durations
Convolutional neural networks (CNN):
+ learn temporal dependencies
+ faster to train than RNNs
-functional memory
-long execution duration
+ learn temporal dependencies
(internal state)
+ lowest RMSE and cumulative
error
-Slow to train
Recurrent neural networks (RNN):
NOXModel
Machine Learning - Approaches
best alternative for polynomial
NOXModel
Results – NOx-Model for Commercial Vehicles
IAV 01/2019 Marco Moser -Machine Learning for EAT
18
Summary:
an average cumulative error under 2%
a correlation with the expected NOxof 0,99
a 3σ of the relative error distribution at around +/-10%
For polynomial models and RNN real-time capability is given
NOxemission models by means AI-Methods are promising alternatives for real time estimation
NOXModel
Results - NOx-Model for Commercial Vehicles (Polynomial vs. RNN)
IAV 01/2019 Marco Moser -Machine Learning for EAT
19
Polynomial regression RNN (LSTM)
NOXModel
Results - NOx-Model for Commercial Vehicles (RNN)
IAV 01/2019 Marco Moser -Machine Learning for EAT
20
Low ambient temperatures
High ambient temperatures
High altitudes
NOXModel
Polynomial Regression vs. RNN
IAV 01/2019 Marco Moser -Machine Learning for EAT
21
Polynomial regression RNN (LSTM)
Polynomial Regression and Neural Network both reliable
Machine Learning for SCR OBD
Machine Learning for SCR OBD
Motivation – State of the Art Diagnosis
IAV 01/2019 Marco Moser -Machine Learning for EAT
23
State of the Art: Unattached Component Diagnosis
Fault set by efficiency diagnosis inhibits detection of a sensor failure (and further results)
Diagnosis A (e.g. efficiency diagnosis passive)
Diagnosis B (e.g. NOxsensor downstream)
Efficiency diagnosis active
t
catalyst classified “ok”
by intrusive measures
Diagnosis C (e.g. AdBlue)
Diagnosis D (e.g. NOxsensor upstream)
Some diagnosis require intrusive measures to
validate the passive result
time for intrusive measures has to be reserved
„Competition of diagnosis“ first who
sets a fault can lock/inhibit other results
Existing information are
not used (e.g. „sensor
failure expected“)
Efficiency diagnosis active
Machine Learning for SCR OBD
Motivation – State of the Art Diagnosis
IAV 01/2019 Marco Moser -Machine Learning for EAT
24
System diagnosis based on collected component diagnosis
All existing information are used to generate the decision
The coordinator gives a reasoned result to the error logic, not the fastest one
result on hold,
then discarded OR
CORRECTED
t
Diagnosis A (e.g. efficiency diagnosis passive)
Diagnosis B (e.g. NOxsensor downstream)
Diagnosis C (e.g. AdBlue)
Diagnosis D (e.g. NOxsensor upstream)
Depending results are released when all involved
diagnosis got a result
Diagnosis with higher quality: every diagnosis can
use/observe the whole drive-cycle
Passive approach saves time: No additional time
needed for intrusive measures
IAV 01/2019 Marco Moser -Machine Learning for EAT
25
Modular concept with evaluation and combination of all known information to pinpoint defect components (like an
engineer with all that data)
Machine Learning for SCR OBD
IAVs Holistic OBD Concept
Signal Processing
NOx model /
sensors
Efficiency
calculation
Release
coordination
Filters
Delays
Signal qualities
StatisticsHolistic Observer
Criteria evaluation
Collect & spread
information
Result decision
… AI
Error-/ OK
message from
collected data
Symptom Modules
Efficiency
Sensor diag.
Adaptation
Stochastic f.
Machine Learning for SCR OBD
IAVs Holistic OBD Concept Symptom Modules
IAV 01/2019 Marco Moser -Machine Learning for EAT
26
Current Diagnosis as basis
Sensors gains / offsets
Actors observers
System learn values
special evaluation algorithms for error highlighting
Information gain through new sensors
NH3sensor
SCR load sensor
Holistic observing of all symptoms leads to understanding of system condition
Symptom Modules
Efficiency
Sensor diag.
Adaptation
Stochastic f.
Machine Learning for SCR OBD
IAVs Holistic OBD Concept Observer
IAV 01/2019 Marco Moser -Machine Learning for EAT
27
Evolutions:
Logic Operations:
„If diagnosis does not work as expected no release
condition“
Manual structure specification based on expert
knowledge
Decision Tree:
Generated automatically
Criteria for differentiation based on measurement
Linear decision making
Concepts with different complexity and quality feasible
Holistic Observer
Criteria evaluation
Collect & spread
information
Result decision
… AI
kein Fehler
Fehler A
Fehler B
Fehler C
Fehler n
kein Fehler
Fehler A
Fehler B
Fehler C
Fehler n
Machine Learning for SCR OBD
IAVs Holistic OBD Concept Observer - Classification
IAV 01/2019 Marco Moser -Machine Learning for EAT
28
Evolution to “Artificial
Intelligence” ..
Classification by machine
learning
Faults can be separated by
looking from “the right side”
(same color = same fault,
but with different intensity)
n-dimensions for
visibility reduced to 3
dimensions
(isomap algorithm)
Holistic Observer
Criteria evaluation
Collect & spread
information
Result decision
… AI
Machine Learning for SCR OBD
Results – Different Classification Algorithms
IAV 01/2019 Marco Moser -Machine Learning for EAT
29
“AI” confirms symptom logic
Neural Network
SVMKNN
No false pass, no false MIL
(Faults which were not in training data are
“no fault”)
Machine Learning for SCR OBD
Results Two Errors
IAV 01/2019 Marco Moser -Machine Learning for EAT
30
Good detection of two errors
Machine Learning for SCR OBD
Conclusion
IAV 01/2019 Marco Moser -Machine Learning for EAT
31
Modular Concept:
existing diagnosis concepts/modules can be used and
integrated
new modules can be added
Combines information from all relevant diagnosis
modules:
more robust diagnosis results, i.e. less risk of false MIL
allows for detailed pinpointing
AI is a promising candidate for the coordination subsystem
(standard) ECU fitting code available
Cloud-based AI possible
IAVs Modular SCR OBD Concept
IAV Holistic OBD
Implementation Master Plan
IAV 01/2019 Marco Moser -Machine Learning for EAT
32
Atomic: Car ECU
Simple AI approaches
Self observer
Simple history and statistics
Global: Outsourcing of Holistic Observer
Observing car as a part of the complete fleet
Service Cloud
restricted
completely unrestricted
Cloud
IAV 01/2019 Marco Moser -Machine Learning for EAT
33
-In-Service Compliance: OEM has to make sure
RDE conformity every car at every time
-Fleet data collected in Cloud permanently
screened
-Special use case: simple (cheap) sensor set
vs. exact (costly) sensor set
-Fewer cars with expensive set
-Statistic comparison of both sets
-Modelling: cheap vs. costly sensor set.
-goal: interval estimation for emission in real
drive
-Predictive Maintenance: car interval
estimation near threshold car gets call for
garage
-Garage data usable to individualize model of
single car (compensation of individual
tolerances) results in more precise
measurement for that car
Service Cloud
Contact
Marco Moser
IAV GmbH
www.iav.com
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.