Content uploaded by Marco Moser

Author content

All content in this area was uploaded by Marco Moser on May 13, 2022

Content may be subject to copyright.

2021-01-5108 Published 17 Dec 2021

Machine-Learned Emission Model for Diesel

Exhaust On-Board Diagnostics and Data Flow

Processor as Enabler

Marco Moser and Steve Kipping IAV GmbH

Kazuhiro Higuchi DENSO CORPORATION

Hiroyuki Hirayama NSITEXE Inc.

Citation: Moser, M., Kipping, S., Higuchi, K., and Hirayama, H., “Machine-Learned Emission Model for Diesel Exhaust On-Board

Diagnostics and Data Flow Processor as Enabler,” SAE Technical Paper 2021-01-5108, 2021, doi:10.4271/2021-01-5108.

Abstract

Conventional methods of physicochemical models

require various experts and a high measurement

demand to achieve the required model accuracy. With

an additional request for faster development time for diag-

nostic algorithms, this method has reached the limits of

economic feasibility.

Machine learning algorithms are getting more popular

in order to achieve a high model accuracy with an appropriate

economical eort and allow to describe complex problems

using statistical methods. An important point is the indepen-

dence from other modelled variables and the exclusive use of

sensor data and actuator settings.

e concept has already been successfully proven in the

eld of modelling for exhaust gas aertreatment sensors. An

engine-out nitrogen oxide (NOX) emission sensor model based

on polynomial regression was developed, trained, and

transferred onto a conventional automotive electronic control

unit (ECU) and also proves real-time capability. Within this

study several approaches are demonstrated to show the impact

on model accuracy with varying numbers of inputs and poly-

nomial structures. However, with increasing accuracy

demands, these models require additional resources of storage

and computational performance.

One solution is the use of a data ow processor (DFP,

developed by DENSO), which can process parallelizable algo-

rithms very eciently and only requires a fraction of the

computation time compared to a conventional ECU. It has

been shown that the computation time of the proven poly-

nomial regression model on the DFP can be reduced by a

factor of 20.

This paper provides an overview of applications of

machine learning algorithms and oers an outlook of the

advantages of a DFP.

Introduction

Motivation

The rising number of powertrain components leads to

an increasing demand for on-board diagnostics (OBD)

and, as a consequence, to the higher complexity of the

algorithms in automotive electronic control units (ECUs),

which need to beperformed in real time. On top of the more

stringent OBD limits and the upcoming on-board monitoring

(OBM) in future emission, legislations all over the world

demand an even further increased accuracy and therefore add

to this challenge.

Sensors are part of these powertrain components and are

needed to control processes or to detect failures. Sensor

models are required to diagnose these sensors or to provide

an alternative value if no sensor is available. Former sensor

models for the automotive exhaust line were very complex.

Exhaust lines and control strategies get even more complex

(i.e., due to dierent heating modes), emission standards get

tighter; hence, more sensors and models are needed, and

sensor models need to get more accurate while

maintaining robustness.

In addition, there is a high demand for a faster develop-

ment process, which is even more serious for the OBD, since

it is calibrated at the end of the development cycle.

Machine learning algorithms are getting more into the

focus and are increasingly used in ECUs in order to achieve

a high model accuracy with an appropriate economical eort.

An engine-out nitrogen oxide (NOX) emission sensor

model based on polynomial regression was successfully devel-

oped, and a soot model is currently under development. is

paper demonstrates the impact on model accuracy of dierent

© 2021 Moser, Kipping, Higuchi, Hirayama; Published by SAE International. This Open Access article is published under the terms of the Creative Commons Attribution

License (http://creativecommons.org/licenses/by/4.0/), which permits distribution, and reproduction in any medium, provided that the original author(s) and the source

are credited.

Downloaded from SAE International, Friday, May 13, 2022

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS 2

approaches with dierent numbers of inputs, polynomial

degrees, a nd polynomia l combinations of two and three inputs.

It has been proven that it is possible to use machine

learning models in different domains within the ECU.

However, with increasing accuracy demands these models

require additional inputs and enhanced polynomial

complexity and therefore need additional resources of storage

and computational performance of the already limited ECU.

For such purpose DENSO developed a data ow processor

(DFP), which can process parallelizable algorithms very e-

ciently and only requires a fraction of the computation time

compared to a conventional ECU. An evaluation of the poly-

nomial regression model on the DFP shows that the computa-

tion time can bereduced by a factor of 20 and allows to easily

implement multiple models to beperformed in parallel.

With a high number of machine-learned models which

are mostly developed in a highly parallelizable structure, opti-

mized ECUs could bevery helpful to enhance the overall

computation performance.

Literature Review

A function development team of the Hubergroup developed

a neural network for the NOX sensor upstream of the Selective

Cataly tic Reduction (SCR) catalyst since a physical NOX model

did not show sucient accuracy to replace the NOX sensor

and therefore save hardware costs. e trained model is fed

with real data, which were not used for training. e accuracy

of the model increases with an increase of inputs. Failure

tolerances were not considered yet as well as the transfer to

dierent vehicle platforms [1, 2].

In 2016, the Innovative Transportation Research Institute

worked on dierent machine learning approaches to estimate

idling emissions in a gasoline vehicle during a mini durability

run of 1600km with dierent driving patterns. Data from a

portable emission measurement system was used to train CO2,

CO, HC, and NOX emissions. Among the ve dierent models

investigated, the boosted bagged decision tree was identied

to bethe best-t approach [3].

Bosch developed a method to integrate machine learning

know-how into the design of experiments steps to identif y the

most relevant combinations of inputs and achieved an accurate

machine learning model. e Gaussian process regression has

proven to bethe most favorable approach due to their natural

smoothness assumptions and the capability to consider the

temporal history of inputs. Bosch developed a new generation

of control units (Bosch MDG1, from device 3 upwards,

including an Adva nced Modelling Unit) to beable to calculate

the Gaussian regression model in real time [4].

In 2019, PSA Group and IFPEN developed in cooperation

a neural network and an empirical correlation to estimate

NOX emission during real driving conditions to substitute the

upstream NOX sensor [5]. e empirical model was used to

describe the combustible composition in the combustion

chamber and therefore to reduce the number of inputs into

the neural network. e validation was based on a wide

measurement campaign. e design of experiments was used

to train the model with 12 inputs. A network of one layer with

ve neurons was derived to prevent overtting and achieve

sucient accuracy [6].

e dissertation of Lumpp in 2011 describes the develop-

ment of an empirical NOX model with high accuracy. e basis

of the modelling here is a consideration of the processes in

the engine and the formation of relevant variables in the

combustion chamber by modelling with existing measured

variables. ese models are then used as features for the

empirical model. Global expert knowledge is therefore

required to create this NOx model. Furthermore, the inves-

tigations are based on test bench measurements of an engine

test bench. Its signicance for a eet is, therefore, limited [7].

In 2017, the University of Salerno investigated the possi-

bility of using a recurrent neural network (RNN) to model

NOX formation in the engine. e model was designed to adapt

to an existing NOX sensor over runtime (aging). An oxygen

model was created as an important input. ey also dealt with

a compensation of the temporal shis between the sensors.

e test environment consisted of a vehicle on a chassis dyna-

mometer and mainly standard cycles—casting doubt on the

robustness over all environmental variations [8].

Sequenz and Isermann discussed 2011in their paper the

use of “global-local” structures in ECUs for modelling emis-

sions. e aim was to replace the state-of-the-art two-dimen-

sional (2D) map-based approaches. Splitting the feature

dimensions into areas with good predictability is the idea

behind this. e Result is a model of models with dierent

parameter sets. For this evaluation it leads to a too complex

soware st ructure and is therefore not investigated f urther [9].

Outline of this Paper

e objective of this paper was to study the possibilities of

simple polynomial models for exhaust gas sensors that were

parameterized and optimized by machine learning algo-

rithms. Independence from other modelled variables and the

exclusive use of sensor data and actuator settings was a key

point. e goal was to t this model into the most commonly

used automotive control units.

Furthermore, the eciency impact of a new DFP with

such a model was studied.

is paper shows the possibilities of the interplay between

optimizations in the model structure and the machine

learning process and the development of suitable hardware

in order to make optimal use of the new functionalities. A

holistic approach to the interaction of hardware, soware,

and optimization processes is the key to improvements in

automotive control units.

PythonasBasis

Conventional Approach

e development process for new ECU functionalities was

performed in Matlab/Simulink®. Exist ing models can beeasily

adapted or new ideas can besketched and investigated. ECU

measurements were used to validate the simulation environ-

ment. Small errors were tolerated due to the fact that the ECU

functionalities were implemented in a fixed-point

Downloaded from SAE International, Friday, May 13, 2022

3

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS

environment on the ECU, which is a dierent hardware

compared to a discrete calculation on a conventional

personal computer.

However, these models are used to create and analyze

new concepts and also to derive an initial calibration.

Dependent on the data basis and the state of the calibration,

it is also possible to use the models to show the robustness of

a calibration. e bigger the data basis the more reliable is the

information regarding robustness. It helps to have data from

dierent vehicles, driver and test cycles, as well as test trips

from extreme conditions such as altitude and heat.

Especially for OBD reasons, the simulation environment

can also beused to perform tolerance analysis on relevant

sensor signals to prove diagnosis robustness against false

detection. In general, a Monte Carlo simulation is performed

with a huge number of runs to catch the most relevant combi-

nations to sensor tolerances. Dependent on the data basis and

the number of runs, this can bea time-consuming task, which

can take days or even weeks.

Once the models are successfully validated, they can

beimplemented, which means the models are compiled into

machine code and tested furthermore via various test cases.

ECU Code in Python

e current development process for OBD functions consists

of the development, testing, calibration, validation, and trans-

lation of functionalities for an ECU up to the point when it is

ready for implementation. Hence, the compiled les for several

ECU functionalities are already available. So the idea was to

use the machine learning environment for the NOX model,

which is in Python already, instead of getting the calibration

into Matlab and a separate model.

Cython, which combines Python and C and acts as an

interface, was used to implement the existing source code into

Python. e original source les do not need to bechanged

at all. What needs to bedone is the creation of a setup le,

which “cythonizes” the source les and also contains compiler

information. It also handles the data transfer, all the datat ypes,

and the calibration. So the source code needs to bewell known

regarding inputs and outputs and their dataty pes. Aer execu-

tion the output file can be imported in Python as a

regular module.

e advantage of using this approach is that Cython code

itself is approximately three to four times faster than pure

Python code [10]. Compared to the conventional approach, it

is also approximately three to four times faster. So the ultrafast

simulation time (1 h measurement is performed in 0.5 sec)

helps especially for a huge number of data and runs. A

comparison against an ECU measurement reveals that the

accuracy is even increased compared to the conventional

approach. is is because the source code is executed in the

same manner as on the ECU with the correct datatype

and arithmetic.

It also allows a high potential for automation, especially

for auto-generated source code. Nevertheless, the analysis

capabilities in Python are almost endless due to the great

open-source community. Not to mention are the nonexistent

license costs compared to, for example, Matlab®.

However, the drawback is the initial setup, which can

be quite time consuming due to poor documentation.

Furthermore, this approach lacks the ability to change the

functionalities aerward once the source code is dened and

imported. Regarding measurement points, youare limited to

what is dened in t he source code, and it is not possible to add

additional signals as needed.

TestingEnvironment

e relevant ECU functionalities considered for this project

are developed in Matlab/Simulink®. is has the advantage

that new concepts can befast evaluated and benchmarked

against conventional or alternative approaches. It also helps

for further investigations regarding robustness and tolerance

analysis. From there the models are processed with dierent

toolchains into C-Code for ECU integration.

Another benet of Matlab/Simulink® models is that they

can also betransferred easily to rapid prototyping systems

such as ETAS ES910 and therefore betested in a real hardware

environment (Figure 1).

For such purpose, a demonstrator vehicle (mild-hybrid

diesel passenger car with a four-cylinder engine of 1.5 l

displacement) from previous work was used [11].

e aertreatment system consists of a close-coupled (cc)

lean NOX trap (LNT), ccSCR/SDPF combination, and an under-

oor SCR/ASC system and fu llls the EU6d and possible future

legislations. e engine-out NOX sensor, which is modelled

within this investigation, is located upstream of the LNT.

Within this context, the polynomial structure of the

engine-out NOX model was implemented in the demonstrator

vehicle and trained with on-road data. Due to the simple

nature of the polynomial structure (Equation 1), the model

was created within the Matlab/Simulink® environment using

several for-loops in an embedded Matlab function to calculate

the polynomial output instead of writing out the entire poly-

nomial structure. Dependent on the approach and the

maximum number to be combined for each polynomial

product, several for-loops are active.

Various measurements of urban and rural driving were

used to train the model. e preparation and selection of data

for machine learning represents a signicant part of the work

and is not described in this paper. Necessary steps include,

for example, gas-flow-dependent time shifting of sensor

signals to a common point, exclusion of ammonia slip via

FIGURE 1 Demonstrator vehicle with rapid prototyping

hardware for model implementation.

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS 4

low-pressure EGR, sensor status assessment, and

other considerations.

ModelDevelopment

NOX Sensor

e automotive diesel air path is controlled and diagnosed by

many dierent sensors (e.g., temperature, pressure, mass ow,

and gaseous species). e modelling of mass ow, tempera-

tures, and pressures is fairly simple because of the direct

physical correlation to its surroundings.

For gaseous species the lambda/O2 sensor is most

common. As for the former sensor types, lambda models are

also relatively easy to create. ere are only a few but major

dependencies to values like mass ow, EGR, and fuel mass.

All those sensors can be modelled with sufficient

physical approaches.

e NOX sensor signal on the other hand is much more

complicated to model. e NOX formation is a complex four-

dimensional process with many temporary local eects [12].

Building a reliable model is a costly and time-consuming

eort. e results of the simpler models needed for ECUs are

not sucient for future needs regarding control strategies

and OBD.

Considering this the NOX sensor is chosen to prove the

possibilities of machine-learned models for diesel air and

exhaust path. e acquired insights are applicable to most

other air path sensors.

e NOX sensor used in the demonstrator vehicle is a

state-of-the-art electrochemical sensor [13].

Model Structure

e aim is to install the NOX sensor model in a standard ECU

(e.g., Bosch, DENSO, Delphi, etc.). Knowing the restrictions

of an automotive ECU helps to choose a reasonable model

structure. Some major restrictions for implementation are

•e model structure has to bexed. It is not possible to

change it during runtime.

•e impact on computation time is very important.

Every module in the ECU soware has a determined

time task to nish a single computation of the

entire function.

•e memory is limited. Only a few as possible feedbacks

should beused.

Because the target is to provide an analog output, it is

obvious to choose a regression model for the NOX sensor.

Feedforward models are preferred because of low

memory usage.

With these considerations, polynomial model structures

are used for this investigation. Another benet not to ignore

is the easy disclosure to authorities. As an alternative, neural

networks-like multilayer perceptron (MLP) or recurrent

(RNN) are also possible to implement on a standard ECU, but

they are much more dicult to explain in their behavior and

so they are not the rst choice here. e polynomial structure

used for the regression consists of independent linear terms

whose response to a change in an input is easily understood.

Polynomial models are described by their inputs

(features), the order of the terms, and the input combinations.

A polynomial of at most three input combinations is described

through the equation

p

xy

xx

x

ei

i

e

ee ij

i

ee ij

e

eeeijk

1

1

12

12 2

134

xxx

i

ee ij

ee ijk

e13 34 4

Eq. (1)

where

x are model inputs

y is the oset

i, j, k are the feature combination

e1, e2, e3, e4 are the polynomial degree

The following settings were used to investigate the

polynomial performance:

•e overall number of inputs into the model varied

between 16 and 24.

•e number of input combinations used to calculate a

product of the polynomial structure was either two

or three.

•e maximum order of the products of the polynomial

structure was set to ve.

•Dierent algorithms were taken into account for

model training.

For the size of the model, this results in a wide range. To

visualize the dierent sizes of the model, Tabl e 1 lists the

number of coecients.

All the dierent combinations, which were possible to

beimplemented in the hardware (microcontroller 256 k B) and

able to betrained on the given hardware (PC with 256 GB

RAM), are discussed in the following section.

Model Inputs

e common way to select features for machine learning

models is to take all available signals and do a correlation

TABLE 1 Model setup and related number of coecients as

characteristic value for the model size.

Input combination 2 Input combination 3

Inputs

order 16 20 24 16 20 24

2152 230 324 152 230 324

3 408 630 900 968 1770 2924

4748 1220 1752 3024 5780 9848

5 1280 2000 2880 6880 13,400 23,120

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

5

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS

analysis for ltering in respect of the target signal [14]. For

polynomial structures, the additional orders of the signals

and the combination of the signals and their higher orders

are needed for this evaluation.

ere are many signals available in the ECU, which can

beused for features of a NOX model, e.g., raw sensor signals,

ltered sensor signals, modelled sensors, targets and actual

values of actors, modelled mass ows, etc. A sensor is not

available for every useful variable, so it may benecessary to

fall back on a target variable of an actuator or a model. Existing

models are used here as a form of feature engineering, which

bundles information through preprocessing [15].

Fundamental understanding of the physicochemical

system leads to a faster progress in model development, and

that knowledge is available through former modelling

researches [12]. For our investigation, wecount on the most

common ltered sensor signals and target values of actuators.

Signals from models have to be verified. If they are

reliable, they can beused to reduce the dimensions of the

polynomial structure if the model itself combines features of

the machine learning model to a new signal [16]. e use of

models was explicitly not desired for this study, as the models

would have to beevaluated for this purpose.

For the demonstrator car wehave the following veried

signals from the air path to use for modelling the NOX sensor:

•Pressures: environment (p_0), inlet (p_ 2), outlet (p_3),

dierence pressure SCR-DPF (dp_SCRF), rail (p_rail)

•Temperatures: environment (T_0), intake (T_1), inlet

(T_ 2), outlet (T_3), turbine downstream (T_4), engine

(T_e ng), fuel (T_ fu e l)

•Mass ow: fuel (dm_fuel), intake (dm_maf )

•Other: rotation (n_eng), torque (M_eng), injection angle

(phi_fuel), oxygen (lambda), NOX SCR-DPF downstream

(r_NOx2), speed (vs)

•Target values of valves because of not existent actual

feedback: throttle valve (r_e mtv), high- and low-pressure

EGR valve (r_hp, r_lp), exhaust gas ap (r_et)

Figure 2 shows the result of a simple correlation analysis

of all reliable signals to the NOX sensor signal.

e direct correlation of the second NOX sensor and its

combinations with other features is reasonable. e same

applies to mass airow, fuel mass ow, injection angle, rotation

speed, and others. ese features have similar frequencies.

Other signals (e.g., engine temperature) are known to have a

high physical inuence on NOX formation [12] but, because

of their different frequency, are not detectable as high

performers with correlation analysis.

e next step can bean embedded feature selection to

nd not relevant features through direct model learning.

Learning algorithms like Lasso Lars use regularization to

penalize not important features to zero. Aer one model

learning, the most penalized features can be excluded.

erefore the number of features can bereduced even further.

ResultsandDiscussion

e following section considers the variations of the model

structure in terms of number of features, feature combination,

polynomial degree, and choice of optimization algorithm. It

is necessary to solve the conict between the highest possible

accuracy of the model and a resource-saving model structure.

e results show that high accuracy does not necessarily

require the most complex model structure.

Feature Selection

Every evaluated signal enhances the model to more accuracy

but also leads to a bigger structure too—so the number of

features is a trade-o between accuracy and size of the model.

As long as the accuracy of the model is high enough, fewer

inputs are preferable. Correlation analysis, model training

with dierent subsets of features, and expert system analysis

can help to take out the most unimportant features to reduce

the size of the model.

For the evaluation, wechose feature sets of 24, 20, and 16

signals. A set of 24 was the maximum number over all of the

available measurements and 16 was the set of features which

was the minimum set to describe the modelled system.

Figure 3 shows the inuence of the reduction of features

on model size and accuracy. Both models were trained with

FIGURE 2 Correlation to NOX sensor signal (features 24,

pol. order 3, pol. combinations 2).

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 3 Correlation to NOX sensor signal with varying

features 16 and 24 (pol. degree: 5, pol. combinations: 2).

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS 6

the same dataset and Ridge Regression for the learning algo-

rithm. e model structure was with a combination of two

features and a maximum polynomial degree of 5 for each

polynomial term. e reduction from 24 model inputs to 16

inputs reduces the model size from 2880 to 1280 terms. But,

at the same time, the standard deviation in the NOX model

oset increases from 15.4 ppm to 22.1 ppm.

e nal choice of the model structure depends on the

restrictions of the ECU and the needs for the use case of

the model.

Feature Combination

e correlation analysis shows that the polynomial combina-

tion of features leads to a gain of information for the

NOX model.

e rst rows of the list in Table 2 indicate that a combi-

nation of two or three features to one polynomial term has

high informative content. erefore, it is sensible to use

combinations of at least two features.

For the model integration into the ECU, it is important

not to exceed a certain size of the model. For each combination

order, the number of polynomial terms and thus the structure

doubles (see Equation 1). With the number of features going

up to 24, combinations of 3 features are the acceptable

maximum for a module in a standard ECU to run in the 100

ms time task.

Figure 4 shows an example of a polynomial input combi-

nation of 2 and 3 for 20 features and a polynomial degree of

4. e standard deviation is improved from the combination

of 2 to the combination of 3 features from 21.4 ppm to

15.2 ppm.

e additional polynomial combinations lead to a model

enlargement. For this example the number of polynomial

terms rises from 1220 to 5780. is has to be taken into

account for assessing the best specic solution for the target

ECU. e model with a feature combination of 3 is more than

four times as big as the model with combinations of 2 features,

and for standard ECUs the total number of 5780 terms and

parameters for one module would be a hindrance

for implementation.

Polynomial Degree

e degree of a polynomial is the maximum sum of the expo-

nents of a polynomial term. Table 1 shows a degree of 5in the

third row. Higher degrees are not evaluated because of the

high impact on the size of the polynomial; see Equation 1.

Figure 5 shows an example of a polynomial degree of 2

and 5 for 20 features and a feature combination of 2. e

standard deviation is reduced for the change from degree of

2 to degree of 5 from 31.5 ppm to 19.9 ppm.

e scatter plot of all test points and the histogram of

Figure 5 shows an increase of accuracy—but the higher poly-

nomial degree leads to an increased number of polynomial

terms (degree 2, 230 coecients vs. degree 5, 2000 coe-

cients). On the side the results indicate that, despite the very

small model with a polynomial degree of 2, a usable model

could bebuilt.

Algorithms

is evaluation is a test of dierent useable standard regres-

sion algorithms (e.g., Linear Regression, Ridge Regression,

Lasso Lars, SGD Regressor, and MLP) from Scikit-Learn [17].

e target was not to nd the optimal algorithm, but to prove

that there are different useful algorithms for different

TABLE 2 Sorted correlation to NOX sensor signal (features:

24, pol. order: 5, pol. combinations: 3).

Rank Polynomial term Correlation to NOX

1 r_NOx2 0.85

2 r_et^2*r_NOx2^1 0.84

3 r_et^4*r_NOx2^1 0.84

4 r_et^3*r_NOx2^1 0.84

5 r_et^1*r_1p^1*r_NOx2^1 0.83

6 r_1p^2*r_NOx2^1 0.83

7 r_et^2*r_1p^2*r_NOx2^1 0.83

8 r_et^3*r_1p^1*r_NOx2^1 0.83

9 r_et^1*r_1p^3*r_NOx2^1 0.83

10 r_et^1*r_1p^2*r_NOx2^1 0.83

11 r_1p^4*r_NOx2^1 0.83

12 r_1p^3*r_NOx2^1 0.83

13 r_et^2*r_1p^1*r_NOx2^1 0.83

14 r_et^1*r_NOx2^1 0.83

15 r_1p^1*r_NOx2^1 0.83

16 dm_maf^1*r_et^1*r_1p^1 0.76

17 dm_maf^1*r_et^3*r_1p^1 0.76

18 dm_maf^1*r_1p^2 0.76

19 dm_maf^1*r_et^2*r_1p^2 0.75

20 dm_maf^1*r_et^2*r_1p^1 0.75

21 dm_maf^1*r_et^1*r_1p^2 0.75

22 dm_maf^1*r_et^1*r_1p^3 0.75

23 dm_maf^1*r_1p^3 0.75

24 dm_maf^1*r_1p^4 0.75

25 dm_maf^1*r_et^2 0.74

26 r_lp^1*r_NOx2^2 0.74

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 4 Correlation to NOX sensor signal with varying

polynomial combinations 2 and 3 (features: 20, pol. degree: 4).

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

7

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS

problems. e dierent algorithms need specic constructions

of the ECU soware related to the specic optimization of the

weighting factors.

To understand the selection of used algorithms here is a

short description of the tested algorithms:

•Linear Regression: e weighting factors (coecients) of

the polynomial are optimized to minimize the residual

sum of the error between the target sensor and model

approximation. ere is no limit to the value area of the

weighting factors, so it is not optimal for the planning of

a xed-point-based structure for automotive ECUs.

Besides this, this algorithm is sensitive to random errors

in the target.

•Ridge Regression: is algorithm extends the Linear

Regression with an L2-norm regularization. e size of

the coecients is penalized. is results in a set of small

weighting factors, and most of the factors are none zero.

e weak point here is the small quantization dierences

between the coecients which can lead to noise eects

in the model prediction.

•Lasso Lars Regression: is algorithm extends the Linear

Regression with an L1-norm regularization. Weighting

factors near zero are penalized and set to zero.

Correlating features are treated equally with

equiangularity between the weighting factors. is

together tends to give a set of good-natured coecients

concerning the value area. e algorithm can

besensitive to noise eects.

•SGD Regressor: Stochastic Gradient Descent estimates

the gradient of the model error at every new sample. is

makes it t for large datasets of more than 100 k samples.

e regularization is tunable between L1- und L2-norm

so that the set of resulting coecients is adjustable to the

needs of the soware structure inside the ECU.

•MLP: is algorithm is not a polynomial approach like

the others. An MLP is, in general, a neural network. It

serves here as a benchmark.

In the investigations carried out here, the machine

learning algorithms were not optimized with respect to

their parameters so that the results correspond to a basic

series of experiments. The tuning of the so-called hyper-

parameters for the different estimators can be done by

manually varying the parameters or by using approaches

for parameter optimization from Scikit-Learn (e.g.,

GridSearch or HalvingRandomSearch).

For an example this evaluation was done with a polyno-

mial order of 5 and 20 inputs with Ridge Regression and Lasso

Lars as learning algorithms. e 2D histogram in Figure 6

shows a better result for the Ridge Regression algorithm. e

drawback of this algorithm is the granularity of the weighting

factors. Ridge Regression (L2 regu larization) uses much more

weighting factors and tries to use smaller va lues. is can bea

problem for the quantization of computation steps inside

the ECU.

e Lasso Lars algorithm with its L1 regularization tends

to have more weighting factors deactivated with 0—very

important features are favored with higher weighting factors.

e advantage of the algorithm can beseen in Figure 7. Lasso

Lars sets most of the weighting factors (about 10,000) to 0,

which can be seen in the histogram plot (right side, green

plot). The other algorithms show a more or less

equal distribution.

In addition, Lasso Lars a lso provides the smallest average

quantization error (le side, green plot). is means, for the

FIGURE 6 Correlation to NOX sensor signal with dierent

learning algorithms: Ridge and Lasso Lars (features: 20, pol.

degree: 5, pol. combinations: 2).

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 5 Correlation to NOX sensor signal with varying

polynomial degrees 2 and 5 (features: 20, pol.

combinations: 2).

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 7 Quantization error at ECU implementation.

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS 8

implementation into an ECU, more accuracy of the model

output because of more reliable accuracy of the

model coecients.

Results of Evaluation

Summarizing the different investigations there are the

following results.

Figure 8 shows the accuracy of the dierent learning a lgo-

rithms. Linear Regression, Ridge Regression, Lasso LARS,

and SGD Regressor make use of a polynomial which is easy

to implement in common ECUs.

MLP uses a dierent structure with hidden layers, and

this gives the advantage of much more weighting factors (here

a factor of 3 through 3 hidden nodes). us the algorithm here

primarily serves as a benchmark and is otherwise outside of

the evaluation.

The best results without optimization of the hyper-

parameters produce the Linear Regression algorithm. e

larger the polynomial in terms of degree, feature combination,

and number of inputs, the more accurate the model becomes.

Linear Regression has not had the high number of hyper-

parameters like the other algorithms. us the results of the

learning experiments already represent the optimum of this

algorithm. e disadvantages of the wide value area of the

coecients together with its strong correlation to huge poly-

nomials make it not a favorite for ECU implementation.

Ridge Regression performs even better with the rising

number of inputs, feature order, and input combination.

Depending on the capacity of the ECU, this algorithm has a

good performance for the NOX model, but is not as good as

the simple Linear Regression. e set of hyper-parameters

leaves room for optimization.

Lasso Lars seems to take no advantage of more complex

polynomial structures—the models’ accuracy remains rela-

tively constant. But through its behavior to set many weighting

factors to zero, it is possible to take unimportant inputs out

of the model or reduce computation load by ignoring zero set

computations. ese are the performance gains of Lasso Lars:

The polynomial can remain very simple for the same

model accuracy.

e SGD Regressor has the worst behavior in this test

environment. Because of the large number of hyper-param-

eters and the algorithms’ ability to produce an optimum

between the L1-regularization of Lasso Lars and the

L2-regularization of the Ridge Regression, SGD has the

greatest potential here to provide the optimal

learning procedure.

e MLP shows stable results over the dierent combina-

tions of polynomial order, number of inputs, and input combi-

nations. With enough capacity in the ECU, the algorithm is

a stable candidate for a reliable model.

Figure 9 shows the comparison between the best evalu-

ation performer (polynomial with 24 inputs, degree of 4,

feature combination of 3 and Linear Regression) and worst

performer (polynomial with 16 inputs, degree of 2, feature

combination of 2 and SGD). e big dierence between the

two test sets can beseen in the scatter plot as well as in the

standard deviation (73 ppm vs. 8 ppm). The scatter plot

together with Figure 8 shows that an even larger polynomial

can no longer bring any signicant improvements since the

model already represents the sensor very accurately.

e nal choice of model structure and learning algo-

rithm depends on many factors, and there is more than one

possibility to nd a good model. For this evaluation with no

direct restrictions, the Linear Regression with 24 inputs, input

combination of 3, and polynomial order of 5 leads to the

best result.

For series ECU implementation, a polynomial with 20

inputs, a polynomial degree of 4, and a feature combination

of 2 was favored because of the smaller model structure.

Because of its ability to shrink the parameters to a small

number, Lasso Lars was chosen to calibrate the model.

In Figure 10 a time trace of a measurement of a real street

drive is shown. e high accuracy of the machine-learned

model is clearly visible. e integrated values (“NOX cum”)

show only minor dierences. With such a model, OBD and

system controls are feasible on a high level. is approach is

not limited to engine-out emissions but can also beused for

any other sensor in the powertrain such as an air mass

ow sensor.

FIGURE 8 Accuracy of dierent machine

learning algorithms.

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 9 Comparison between the best and worst

test performers.

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

9

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS

Humidity and Air Quality: The

Missing Inputs

It is well known that humidity has an impact on NOX forma-

tion [17, 18]. Especially for high temperatures (>30°C), the

impact can bevery high because the absolute humidity (mass

of water in the intake air) has a wide range. Most of the current

passenger cars do not have a humidity sensor, so this impor-

tant input is missing. e air ow meter itself is sensitive to

absolute humidity, so newer versions of air ow meters have

a humidity sensor implemented [4]. It is very sensible to insta ll

such a sensor in future cars—not only for NOX modelling but

for emission controls too.

Another factor is the intake air quality. Combustion

processes in direct proximity (e.g., trac jam in close areas

of cities or long tunnels) or decomposition processes (e.g.,

swamp areas) can lead to lower oxygen levels which lead to

lower NOX formulation (like an external/outside exhaust gas

recirculation). Current passenger cars are not equipped with

an intake oxygen sensor.

Both disturbances of the NOX formulation result in devia-

tions of the oxygen level in the exhaust gas. It is possible to

detect this deviation and adjust the computation of the NOX

model. Because of insucient data of such environments, it

is not possible to show the eect here.

For ozone, it is possible to have an eect too [19]. e high

reactive gas species could lead to dierent combustion eects.

Because of a missing sensor, this could not be evaluated in

this research.

Adoption of Other Sensors

e model structure is usable for other air path sensors too.

Air ow meter or pressure or temperature sensors are candi-

dates for this kind of modelling.

As an example Figure 11 shows the modelling of a

pressure sensor upstream of the engine. For this model the

specic pressure sensor switched its input status with the NOX

sensor. e accuracy of this model is very high.

e same result is valid for other sensors or actors in the

air path, Figure 12 shows an example for the air ow sensor.

So the same model structure is usable for dierent sensors,

and the building of soware with many parallel models is

easily feasible.

In the end, there is a way to use many machine-learned

models with thousands of parallelizable computational steps.

The disadvantage here is the high load on the central

processing unit (CPU) that has to behandled.

DFPasanEnabler

About DFP

In the automotive eld in recent years, there has been a need

for processors that enable unprecedentedly sophisticated and

complex operations. e CPU, which is widely used today, is

connected to storage devices and I/O devices and is able to

handle various types of control instructions. However, due to

serial processing, there is a problem that real-time perfor-

mance is impaired as the amount of calculation increases. On

the other hand, a graphics processing unit (GPU), which is

FIGURE 11 Model accuracy for pressure sensor

(engine inlet).

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 10 Time trace of NOX model accuracy.

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 12 Model accuracy for the mass air ﬂow sensor.

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS 10

becoming more and more popular for image processing, is a

processor that can process large-scale routine operations in

parallel at high speed. However, GPU tends to belarge in size

and consumes a large amount of power. In this context, DFP

has been developed as a processor that realizes high-speed

and power-saving processing while minimizing the size by

maximizing the utilization eciency of the parallel processing

mechanism. DFP is designed to beimplemented in a micro-

computer for automotive applications, and it is expected to

beone of the technologies to achieve an unprecedented high-

performance ECU by sharing the role with the CPU.

Architecture and Mechanism

of Parallel Processing

Combination

As shown in Figure 13, a DFP consists of three types of

processor cores: four scalar processing units (SPUs), a vector

processing unit (VPU), and a control core unit (CCU).

e SPU is an operations core and is able to execute

dierent processes independently in parallel (a function

similar to that of a multi-CPU). e VPU is a 512-bit wide

vector operations core, and it can execute up to 16 operations

concurrently if the data length is 32-bit (a function similar to

that of a GPU). e CCU is responsible for dynamically allo-

cating task processing to free SPUs to improve computing

eciency. Fi gu re 14 shows the outline of the parallel processing

ow in the DFP.

Each SPU executes independently assigned task

processing in parallel, and if there is a vector operation

instruction in it, the VPU executes the process in place of the

SPU. Sharing the VPU with each SPU in this way makes it

possible to achieve high throughput. By combining the task

parallelism and the data parallelism by vector operations, the

DFP achieves highly ecient processing while maintaining

the compact size and lower power consumption required for

use in a microcomputer for automotive applications. For the

polynomial models used in this paper, the operation of each

term and the sum of all terms were parallelized and processed

by the VPU. Moreover the entire process was divided into

multiple tasks that were assigned to and executed by indi-

vidual SPUs. is improved the utilization eciency of the

VPU and increased execution speeds.

Execution Time Assessment

e DFP execution time was measured by using the evaluation

board equipped with DFP and by means of clock counter value

at the measurement points on the soware framework. In the

same manner, as a benchmark, the execution time was

measured with an ECU equipped with an existing microcom-

puter for automotive applications (CPU clock frequency of

400 MHz). Figure 15 shows the relationship between predic-

tion error (Root Mean Square Error, RMSE) of NOX model

with dierent polynomial inputs and order and the execution

times of DFP and CPU.

Prediction error decreases by increasing the number of

inputs and order of polynomial model, but as a trade-o, it

can beseen that the CPU execution time increases exponen-

tially. is conrms that the calculation order O of the poly-

nomial model is O((np)c) for the combination of the number

of inputs n, the order p, and the number of combinations c.

On the other hand, the DFP keeps the execution time low by

performing parallel processing even though it was for the

same calculation order. Focusing on the polynomial model

with RMSE = 25 ppm from the viewpoint of OBD purpose,

the CPU execution time was 250 μs, while the DFP execution

time was about 1/20 of this, at 12 μs. is result shows that

the DFP enables large-scale processing such as polynomial

FIGURE 13 Architecture of DFP.

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 14 Overview of parallelization process.

© Moser, Kipping, Higuchi, Hirayama.

FIGURE 15 Result of execution time (DFP vs. CPU).

© Moser, Kipping, Higuchi, Hirayama.

Downloaded from SAE International, Friday, May 13, 2022

11

MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS

models to beexecuted in the ECU w ithout impairing real-time

performance. Furthermore, the result shows that an even more

complex and larger number of models can also behandled.

SummaryConclusion(s)

e target of this work was to evaluate the possibilities of

bringing a machine-learned emission model with high

accuracy to a standard control unit of a passenger car. e

challenge here is to nd an optimum of the conict between

the accuracy of the model and the correlating complexity of

its structure.

is paper shows the possibilities of simple polynomial

models for exhaust gas sensors that were parameterized and

optimized by machine learning algorithms. e goal was to

t this model into area-wide used automotive control units.

Furthermore, the eciency impact of a new DFP with

such a model was studied, and its benets were shown.

For the upcoming regulations for OBM, the need for very

accurate models of engine emissions is increasing. Just as with

the models for air mass and intake pressure already shown

here, models for soot, hydrocarbons, carbon monoxide, form-

aldehyde, carbon dioxide, and oxygen can also becreated with

appropriate emission measurement technology. e problem

here is the measurement data collection of emissions without

existing regular sensors.

Recommendation

e most important recommendation to future implementa-

tions of machine-learned models into automotive control

units is to make use of the developments of processor units

that are optimized for the needs of parallelized structures.

ere is a huge potential to make use of machine learning in

the automotive sector, but the standard control units have too

many restrictions. e real benet from machine-learned

models is to have a high range of structural possibilities and

a high accuracy of mathematical functions (e.g., exponential,

logarithmic, tangential functions).

To enhance the accuracy of machine-learned emission

models, more attention should bepaid to the distribution

of the model features and the target value. The distribu-

tion of input values from sensors and especially from

actors are not always ideal for machine learning. In the

best case, every feature and the target value will beindi-

vidually transformed to get as near as possible to Gaussian

normal distribution.

Optimization of learning algorithms is the next going on

from this evaluation. It is very important to know well the

capabilities and limitations of each learning algorithm to

guarantee a stable model outside the trained data space.

Knowledge about the behavior of interpolation and extrapola-

tion in areas outside the known data range is essential for an

evaluation of the usability of the model in an

automotive application.

For emission models the use of polynomial structures is

relatively safe in the context of approval of emissions regula-

tion from the authorities. For other machine learning

approaches like neural networks (MLP in this evaluation),

there is no procedure at the moment to get an approval. Here

lies a major problem to use machine-learned models. A

process must beagreed upon with the authorities in order to

obtain approval for dierent soware structures. On the

developer side, it should beclaried in advance how such

structures can bevalidated and secured.

Acknowledgments

anks to DENSO for providing the evaluation results for the

DFP, AECC and IPA for providing the car to collect the data

and IAV for supporting this paper.

Contact Information

Marco Moser

Sr. Smart Data Analyst for OBD Functions

IAV GmbH, Carnotstr. 1, 10587 Berlin

Tel: +49(0)30 -399789176

marco.moser@iav.de

Steve Kipping

Function Developer OBD

IAV GmbH, Carnotstr. 1, 10587 Berlin

Tel: +49(0)30-399789247

steve.kipping@iav.de

Kazuhiro Higuchi

Powertrain System Development Div.

DENSO CORPORATION

1-1, Showa-cho, Kariya, Aichi 448-8661, Japan

Tel: +81(0)566-61-4441

kazuhiro.higuchi.j7r@jp.denso.com

Hiroyuki Hirayama

Processor Development Dept.

NSITEXE, Inc.

2-16-4, Kounan, Minato-ku, Tokyo, 108-0075, Japan

Tel: +81(0)3-5783-2415

hiroyuki_hirayama@nsitexe.co.jp

References

1. Huber Group, “Neuronale Netze in Automobilen sollen

NOX-Sensor vor Katalysator in SCR-Systemen ersetzen,”

2009, accessed May 2021, https://www.elektronikpraxis.

vogel.de/neuronale-netze-in-automobilen-sollen-nox-sensor-

vor-katalysator-in-scr-systemen-ersetzen-a-231895/.

2. Huber Group, “Neural Networks in the Automotive

Industry: Huber Group Is Working on Substituting a NOx

Sensor with a Neural Network/Learning Network to Replace

Physical Model,” 2010, accessed August 2021, https://www.

konstant.de/pressefach/huber-group/neura l-net works-in-

the-automotive-industry-huber-group-is-working-on-

substituting-a-nox-sensor-with-a-neural-network-learning-

network-to-replace-physical-model/.

Downloaded from SAE International, Friday, May 13, 2022

© 2021 Mos er, Kipping, Higu chi, Hirayama; Pu blished by SAE Inte rnational. Thi s Open Access art icle is publishe d under the terms of t he Creative Common s Attribution

License (http://creativecommons.org/licenses/by/4.0/), which permits distribution, and reproduction in any medium, provided that the original author(s) and the

source are credited.

Positions and opinions advanced in this work are those of the author(s) and not ne cessarily those of SAE International. Respo nsibility for the content of the work lies

solely with the author(s).

ISSN 0148-7191

12 MACHINE-LEARNED EMISSION MODEL FOR DIESEL EXHAUST ON-BOARD DIAGNOSTICS

3. Li, Q., Qiao, F., and Yu, L., “A Machine Learning Approach

for Light-Duty Vehicle Idling Emission Estimation Based on

Real Driving and Environmental Information,” Environ

Pollut Climate Change 1 (2016): 106.

4. BOSCH, “Modeling a Physical System for System Control and

Calibration,” accessed May 2021, https://www.bosch-ai.com/

research/research-applications/ai-based-dynamic-modeling/.

5. BOSCH, “NOx sensor – Regulation and diagnosis of

cataly tic nitrogen oxide reduction,” accessed August 2021,

https://ww w.bosch-mobility-solutions.com/en/solutions/

exhaust-gas-treatment/nox-sensor/.

6. Rudlo, J., Leroy, T., Ngo, C., and Denys, F., “Schätzung der

NOX-Emissionen moderner Dieselmotoren unter realen

Fahrbedingungen,” MTZ- Motortechnische Zeitschri 80,

no. 2 (2019): 58-62.

7. Lumpp, B.C., “Echtzeitfähige Stickoxidmodellierung zur

Integration im Steuergerät eines Nutzfahrzeug-

Dieselmotors,” Technischen Universität München

(Technical University Munich), 2011, http://nbn-resolving.

de/urn/resolver.pl?urn:nbn:de:bvb:91-

di ss-2 0111212-1083664 -1-3.

8. Arsie, I., Cricchio, A., De Cesare, M., Lazzarini, F. et al.,

“Neural Network Models for Virtual Sensing of NOx

Emissions in Automotive Diesel Engines with Least Square-

Based Adaptation,” Control Engineering Practice 61 (2017):

11-20 , https://doi.org/10.1016/j.conengprac.2017.01.005.

9. Sequenz, H. and Isermann, R., “Emission Model Structures

for an Implementation on Engine Control Units,” IFAC

Proceedings Volumes 44, no. 1 (2011): 11851-11856, https://

doi.or g/10.3182 /2 0110828- 6 -IT-10 02.03131.

10. Cython, “Faster Code via Static Typing,” accessed August

2021, https://cython.readthedocs.io/en/latest/src/quickstart/

cythonize.html.

11. Demuynck, J., Favre, C., Bosteels, D., Randlshofer, G. etal.,

AECC, IPA, IAV, “Integrated Diesel System Achieving Ultra-

Low Urban and Motorway NOX Emissions on the Road,” in

Wiener Motorensymposium, Vienna, 2019.

12. Rezaei, R., Dinkelacker, F., Tilch, B., Delebinski, T. et al.,

“Phenomenological Modeling of Combustion and NOX

Emissions Using Detailed Tabulated Chemistry Methods in

Diesel Engines,” International Journal of Engine Research 17

(2016): 846-856.

13. Todo, Y. et al., “Development of High Accuracy and Quick

Light-o NOx Sensor,” SAE Technical Paper 2018 -01-0334 ,

2018, https://doi.org/10.4271/2018-01-0334.

14. Guyon, I. and Elissee, A., “An Introduction to Variable and

Feature Selection,” J. Mach. Learn. Res. 3 (20 03): 1157-1182 .

15. Zheng, A. and Casari, A., Feature Engineering for Machine

Learning: Principles and Techniques for Data Scientists

(Germany: O’Reilly Media, 2018)

16. Scikit-Learn, “User Guide,” https://scikit-learn.org/stable/

user_guide.html, accessed May 2021.

17. ENVIRON International Corporation, “Humidity and

Temperature Correction Factors for NOx Emissions from

Diesel Engines,” Final Report, 2003.

18. McCormick, R.L., Graboski, M.S., Newlin, A.W., and Ross, J.D.,

“Eect of Humidity on Heavy-Duty Transient Emissions from

Diesel and Natural Gas Engines at High Altitude,” Journal of the

Air & Waste Management Association 47, no. 7 (1997): 784-791,

htt ps://doi.org/10.1080/10 473289.1997.10463937.

19. Liu, Y., Wang, Z., Li, L., Wan, K. et al., “Reaction Mechanism

Reduction for Ozone-Enhanced CH4/Air Combustion by a

Combination of Directed Relation Graph with Error

Propagation, Sensitivity Analysis and Quasi-Steady State

Assumption,” Energies 11 (2018): 1470, https://doi.

org/10. 3390/en11061470.

Downloaded from SAE International, Friday, May 13, 2022