Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

SAE TECHNICAL
PAPER SERIES 2004-01-1372

A Load Torque Estimator

Marco Fam and Elbert Hendricks
Oersted-DTU, Automation, Technical University of Denmark

Reprinted From: Electronic Engine Controls 2004

(SP-1822)

2004 SAE World Congress

Detroit, Michigan
March 8-11, 2004

400 Commonwealth Drive, Warrendale, PA 15096-0001 U.S.A. Tel: (724) 776-4841 Fax: (724) 776-5760 Web: www.sae.org
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or
transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise,
without the prior written permission of SAE.

For permission and licensing requests contact:

SAE Permissions
400 Commonwealth Drive
Warrendale, PA 15096-0001-USA
Email: [email protected]
Fax: 724-772-4891
Tel: 724-772-4028

For multiple print copies contact:

SAE Customer Service

Tel: 877-606-7323 (inside USA and Canada)
Tel: 724-776-4970 (outside USA)
Fax: 724-776-1615
Email: [email protected]

ISBN 0-7680-1319-4
Copyright © 2004 SAE International

Positions and opinions advanced in this paper are those of the author(s) and not necessarily those of SAE.
The author is solely responsible for the content of the paper. A process is available by which discussions
will be printed with the paper if it is published in SAE Transactions.

Persons wishing to submit papers to be considered for presentation or publication by SAE should send the
manuscript or a 300 word abstract of a proposed manuscript to: Secretary, Engineering Meetings Board, SAE.

Printed in USA
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

2004-01-1372

A Load Torque Estimator

Marco Fam and Elbert Hendricks

Oersted·DTU, Automation, Technical University of Denmark

Copyright °
c 2004 SAE International

ABSTRACT also been developed for this application but such

sensors are too expensive for actual production
This paper deals with a Mean Value Engine Model vehicles. They are in general intended for laboratory
(MVEM) approach to dynamically estimating engine development uses [4].
load torque. The load torque estimator is made us-
ing parameter estimation and an extended Kalman fil-
ter. The load torque estimator can be used to simplify Apart from direct measurement, software engine
the control logic of modern engines as it can detect torque estimators suffer from the difficulty that they
load changes over time thereby accounting for wear must be calibrated and validated on dynamometers
and also differing accessory power consumption. The over a large range of barometric pressures, ambient
model can be implemented using a state equation for temperatures and accessory drive loads. This re-
the crank shaft speed, and a state estimating a con- quires a comprehensive and expensive experimental
stant describing the loading level. effort. Nevertheless torque control systems are an
important part of modern engine control systems as
INTRODUCTION they have been integrated into a large number of
useful and marketable subfunctions: traction control,
Recently engine torque based control systems have cruise control and transmission control.
been introduced into production by several manu-
factures, first by Bosch [1] and then by others, for
example Ricardo [2]. It is clear then that torque This paper does not deal with the estimation of en-
based systems are of increasing interest as the gine torque but with the complementary problem of
effect of engine torque is immediately obvious to a load torque estimation which has not yet been ad-
vehicle operator. Such systems translate the vehicle dressed in the literature. The basic idea is to use ob-
operator’s accelerator pedal demands into a torque server estimation techniques to derive the road and
requirement which is then fed to an electronic throttle road grade load information. If successful such an al-
which provides the necessary throttle air mass flow. gorithm could make possible more accurate dynamic
While they are based on engine models such as the engine torque control systems which would require a
MVEM (Mean Value Engine Model) they basically are minimum of calibration effort. This paper is based on
non-dynamic and have a feed forward configuration. a Masters thesis carried out at the Technical Univer-
The concept of torque is otherwise used to organize sity of Denmark (DTU) [5].
all of the functions of the ECU (Engine Control Unit).
Two other approaches to engine torque estimation MODEL
use ”Stochastic Estimation” and ”Frequency Analysis”
to give a cycle by cycle detailed picture of the engine This section presents the equations necessary to de-
torque output [3]. Special dedicated sensors have scribe a load torque estimator. First the well known

1
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

crankshaft speed state equation is given, afterwards tance.

a load model is presented which attempts to model
real driving conditions. A differential equation for the Phill : Power needed to climb hills.
load torque is derived and finally a loading parameter Paero : Power needed to overcome aerodynamic
model is given. drag.
Pacc : Power needed to accelerate.
CRANKSHAFT SPEED STATE EQUATION The
Paux : Power needed to supply secondary loads (al-
main differential equation for the load torque estimator
ternator, power steering, etc.).
is the crank shaft speed state equation. The deriva-
tion of the crank shaft speed state equation is based
on conservation of energy. Chemical energy is de- The individual contributions are given below and the
livered into the engine by fuel to create mechanical values used for the simulation model are for a typical
energy. It is this energy which is used to overcome small vehicle.
the friction losses, pumping loss, and load, and to ac-
celerate the crankshaft. The crankshaft speed equa- Proll = m · g · v · Crr · cos(θ) (3a)
tion can be found in [6] and is revised in [7] and [8]. Phill = m · g · v · sin(θ) (3b)
More details are provided in the appendix . Equation 3
Paero = 0.5 · A · v · Cd · ρ (3c)
(1) gives the differential equation for the crank shaft dv
speed. Pacc = m·v· (3d)
dt
Pl + Pb Hu ηi ṁf (t − τdn ) Paux = f (alternator, power steering, etc.)(3e)
ṅ = − + (1)
In In
µ ¶2 where m = 791 kg is the mass of the ve-
2π hicle, g is the gravitational acceleration in
where I = 1000 · (Iengine + Iload )
60 meters/sec2 , v is the speed of the car (meters/sec),
Here Hu is the lower heating value of the fuel, ηi is Crr = 0.0136 + 0.4 · 10−7 v 2 is the rolling resistance
the indicated efficiency and ṁf is the cylinder port coefficient (dependent on tire quality, tire pressure
fuel mass flow. The losses and load powers are Pf road surface, etc.), θ is the slope of the hill (radians),
(friction), Pp (pumping) and Pb (brake). The moment A = 1.72 m2 is the effective frontal area of the car,
of inertia I has been scaled so that n can have the Cd = 0.41 is the drag coefficient and f (·) is the
units of revolutions per minute (rpm). secondary power consumption in watts.

LOAD POWER MODEL Before the presentation of The acceleration power requirement is a measure
the load torque state equation some characteristics of the operator’s wish to increase the speed of the
of and comments on the choice of model should be vehicle. If the driver wants to increase the speed
made. The load power model is selected in such a then the power that the engine can deliver must
way that it is representative of a car and its driving be greater than all the external power sources that
characteristics. The purpose is to see how well an are mentioned above. In what follows it should be
observer can estimate the load torque or load power clear that the acceleration term is not used in the
given a realistic power demand. It is shown how a description since what is of interest here is a function
simple expression can be used to describe the power that describes the external loading of the engine.
demand so that it is usable for an observer setup.
Below the model and equations will be presented.
If the car is driven on a level road then what must
be taken into account is the rolling and aerodynamic
The power required from the engine in a vehicle is de- power consumption. Figure 1 shows four different
pendent on a number of factors. One way of describ- functions which are an attempt to describe the rolling
ing the power requirements is to use the equation. and aerodynamic power consumption Proad = Proll +
Paero (dashed line). Phill and Paux are not modelled
Pengine = Proll + Phill + Paero + Pacc + Paux (2) here as they are considered to be transient distur-
bances during normal operation. Since friction can-
This equation describes the traction force required by
not be described by a physically derived expression,
the vehicle under all possible driving conditions. The
polynomial fits are used here.
contributions are summarized below.
Equation (4) shows the polynomials drawn on the fig-
Proll : Power needed to overcome the rolling resis- ure.

2
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

10000
Proad equation derived for the load torque. This model
1. order fit
2. order fit
has been constructed to see whether or not a dy-
3. order fit
8000 k⋅ n
3 namic model of the load torque can be obtained.
In the MVEM literature differential equations have
been shown to work with good accuracy on several
6000
important estimation problems. For this reason it is
reasonable to attempt to construct a new differential
Power [W]

4000 equation to estimate the dynamic load torque.

2000
There is a very simple relationship between the load
power and load torque. This is given in the equation
0
below.

−2000 Pb = Qb ω = kQb n = kb n3 (6)

0 10 20 30 40 50 60 70 80 90
Vehicle Speed [km/h]
2π 2π
Figure 1: Models fitted to the rolling and aerodynamic Since ω = 60 n then k = 60 and Qb can be written as
power losses. Notice that the third order polynomial fit kb 2
is exact. The simple nonlinear fit is not quite as good Qb = n = f (kb , n) (7)
k
but is has only one fitting constant.
Equation (7) can be differentiated to obtain the re-
quired state equation. Here kb is also considered to
P1.order = 89.82n − 1283.29 (4a) be a time varying parameter which is why the differ-
2
P2.order = 1.19n − 17.18n + 319.93 (4b) entiation given in equation (8b) has the following form.
P3.order = 0.0088n3 + 2.5073 · 10−15 n2
+ 25.5929n + 1.8255 · 10−12 (4c) d(Qb ) ∂f (kb , n) ∂f (kb , n)
= k̇b + ṅ (8a)
Pnonlinear = 0.0132n 3
(4d) dt ∂kb ∂n
n2 kb
The third order polynomial can be seen to fit the Q̇b = k̇b + 2 nṅ (8b)
k k
physical model quite accurately. This is because the
equations involved are represented by a third and The crankshaft speed state equation (1) can be in-
first order expression whose parameters are basically serted in equation (8b) and the loss and load powers
independent of vehicle speed. The second order expressed as torques. This is summarized in equa-
expression is not too bad but the equation needs tion (9).
all three parameters in order to effectively describe µ ¶
n2 n Hu ηi ṁf (t − τdn )
the curve. The simple nonlinear fit represented by Q̇b = k̇b + 2kb −(Ql + Qb ) +
k I nk
Pnonlinear = kn3 is presented in order to see how
well it fits. Naturally this expression is not able to (9)
describe the curve but is however much simpler since
The parameter kb is modelled as a slowly varying con-
it has only one parameter.
stant and is included in the observer setup.

Since this work focusses on determining the load

SIMPLE LOAD TORQUE EQUATION A simpler so-
power dynamically the simple nonlinear model (equa-
lution to estimating the load torque is also proposed
tion (4d)) is used in order to keep the model simple.
here. The principle is very simple as it uses the model
that the load torque state equation was derived from.
In what follows the road load equation will have the
An estimate of the load torque can be read out of
following terms included and has been implemented
the model if the loading parameter kb and the engine
in the SIMULINKr model of the load power subsys-
speed, n, are estimated correctly.
tem.
P̂b k̂b n̂2
Pb = Proll + Phill + Paero + Paux (5) Q̂b = = (10)
k n̂ k

DERIVATION OF THE LOAD TORQUE STATE where kb and n are estimated using a constant gain
EQUATION This section describes a differential extended Kalman filter.

3
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

STATE EQUATIONS FOR THE OBSERVER SETUP mean value and variance:
To summarize the previous section the differential n o
equations presented above are collected here. E x(t0 ) = x0 (16)
n o
E [x(t0 ) − x0 ][x(t0 ) − x0 ]> = Q0 (17)

Pl + Pb Hu ηi ṁf (t − τdn )
ṅ = − + (11)
In In The main difference between a true extended Kalman
= f1 (α, ṁf , pi , Ti , n, kb ) filter and a constant gain extended Kalman filter is that
µ ¶ the system (14) is linearized at an appropriate operat-
n2 n Hu ηi ṁf (t − τdn )
Q̇b = k̇b + 2kb −(Ql + Qb ) + ing point and a constant gain calculated at this point.
k I nk
The choice of operating point depends on the system,
(12)
especially if the nonlinearities are distinct and not very
= f2 (α, ṁf , pi , Ti , n, k̇b , kb ) general. In other words the designer chooses a set of
k̇b = v1 (t), v1 (t) ∈ N (0, V1 (t)) (13) appropriate values of (x0 , u0 ) such that the linearized
= f3 (n), (constant driven by white noise) system gives the most general approximation to the
real system. A very familiar linear state space de-
The notation N (0, V1 (t)) means that the noise source scription arises since the system matrices are con-
is normally distributed with a zero mean with an inten- stant at a given operating point. The constant matri-
sity which is V1 ces are the Jacobian matrix of f with respect to x and
the Jacobian matrix of h with respect to x. This can
The equations are used in a constant gain extended be seen in equations (18a) and (18b).
Kalman filter. The next section shows how to set up
∂f (x(t), u(t), t)
the system equations so that the Kalman gain L can A = ∇x f¯¯x(t)=x0 = ¯
¯x(t)=x0 (18a)
be found using the proper noise assumptions. u(t)=u0 ∂x(t) ¯u(t)=u0
∂h(x(t), u(t), t)
EXTENDED KALMAN FILTERS AND PARAMETER C = ∇x h¯¯x(t)=x0 = ¯
¯x(t)=x0(18b)
u(t)=u0 ∂x(t) ¯
ESTIMATION u(t)=u0

CONSTANT GAIN EXTENDED KALMAN FILTER The matrices A, C, V1 and V2 are all constant due to
Extended Kalman filters have been shown to work the selection of a static operating point. This gives a
with good results in previous papers see e.g. [9] constant covariance matrix Q and a constant Kalman
and [10], which is why only a brief description of filter gain L. Q and L can be found by solving the
constant gain extended Kalman filters and parameter algebraic Riccati equation. L is then found using the
estimation will be included here. simple relationship given below.

0 = AQ + QA> + V1 − QC> V2−1 CQ, LTI (19)

The differential equations can generally be described
by equation (14). L = QC> V2−1 , LTI (20)
½
Sys. model ẋ(t) = f (x(t), u(t), t) + v1 (t) where LTI means the Linear Time Invariant case.
(14)
Meas. model y(t) = h(x(t), u(t), t) + v2 (t)
Finally the state estimation equation can be written as
where v1 (t) ∈ N (0, V1 (t)), v2 (t) ∈ N (0, V2 (t)), x ∈ in equation (21)
Rn , v1 ∈ Rn , y ∈ Rm , v2 ∈ Rm and m ≤ n. f :
Rn → Rn , h : Rm → Rm , v1 (t) is the state excitation ˙
x̂(t) = f (x̂(t), u(t), t) + L[y(t) − h(x̂(t), u(t), t)] (21)
noise and v2 (t) is the observation or measurement
noise. [v1 (t) v2 (t)]> are described as white noise
with intensities.
µ ¶
V1 (t) V12 (t) AUGMENTATION OF SYSTEM A nonlinear system
> , t ≥ t0 (15)
V21 (t) V2 (t) including an unknown parameter is given in equation
(22).
The noise sources are assumed uncorrelated 
>
(V12 (t) = V21 (t) = 0). Also it will be assumed that 
 ẋ(t) = f (x(t), u(t), θ(t), t) + v1 (t)
V2 (t) > 0 which means that there is white noise on all  y(t) = h(x(t), u(t), θ(t), t) + v2 (t)
the measurements. Additionally the initial state x(t0 ) (22)

 θ̇(t) = v3 (t)
is uncorrelated with v1 (t) and v2 (t) and has initially a 
where vi (t) ∈ N (0, Vi (t)) for i = {1, 2, 3}

4
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

Notice that the unknown parameter vector has been Crank Shaft Speed Sensor:
included in f (·) and h(·). Again it will be assumed
that v1 (t) and v2 (t) are the state and measurement n̂˙ = f1 (pi , Ti , n̂, α, ṁf ) + L11 (nm − n̂) (27)
noise and that they are uncorrelated white noise sig- ˙ ˙
Q̂b = f2 (α, ṁf , pi , Ti , n̂, k̂b , k̂b ) + L21 (nm − n̂) (28)
nals. The nonlinear system description can be lin-
˙
earized giving the following augmented state variable k̂b = f3 (n̂) + L31 (nm − n̂) (29)
x̄(t).
½ ¾ · ¸½ ¾ · ¸
ẋ(t) ∇x f ∇ θ f x(t) ∇u f The observer setup using a constant gain extended
= + u(t) Kalman filter can be seen in equation (30). This equa-
θ̇(t) 0 0 θ(t) 0
| {z } | {z } | {z } | {z } tion contains a constant Kalman gain matrix. By using
˙
x̄(t) Ā x̄(t) B̄
· ¸½ ¾ the theory on constant gain Kalman filters, augmen-
∇v 1 f 0 v1 (t) tation of systems and finding appropriate noise inten-
+ (23)
0 I v3 (t) sities a sub optimal Kalman gain can be found. An
example of this is also given in [9] where the stability
properties are also treated. Since only one measure-
½ ¾ ment is used L only has one column. If more mea-
£ ¤ x(t) surement signals are included the number of columns
y(t) = ∇x h> ∇θ h> + ∇u h> u(t)
| {z } θ(t) | {z } will likewise increase with the same number of mea-
| {z }
C̄
x̄(t)
D̄ surement signals.
>    
+ ∇v2 h v2 (t) (24) n̂˙ f1 (pi , Ti , n̂, α, ṁf , k̂b )
 ˙    £ ¤
The Ā, B̄, C̄ and D̄ matrixes can be identified directly  Q̂b  =  f2 (pi , Ti , n̂, α, k̂˙ b , k̂b ) + L3×1 nm − n̂ (30)
˙ f3 (n̂)
from the calculation of the Jacobian matrices. The fol- k̂b
lowing constants should likewise be defined in order
to make the presentation more compact and general.
µ ¶ SIMULATIONS AND RESULTS
(∇v1 f )V1 (∇v1 f )> 0
V̄1 (t) =
µ 0 V
¶3 This section shows the results based of the theory
(∇v1 f )V12 (∇v2 h)> (25) presented in the previous sections based on a simu-
V̄12 (t) =
0 lation study and also using real engine data taken on
V̄2 (t) = (∇v2 h)V2 (∇v2 h)> a dynamometer.
It is assumed that the process noise v1 (t) and the
noise that drives the parameter variation v3 (t) are un- An introduction to the variables used for the differ-
correlated. ent load torque curves is presented here in order for
the reader to see on what basis the comparisons are
Based on equations (22) - (25) a Kalman filter can be made. The curves all have a label or name tag which
written up in terms of the estimated system states and is presented in the equations below.
the unknown parameters.
( ) ½ ¾ Qbmeas = measured load torque (31a)
˙
x̂(t) f (x̂(t), u(t), θ̂(t), t)  ˙ 2 ˙
˙ = 
 Q̂b = n̂k k̂b³
θ̂(t) 0 


 k̂ Pl (pi ,n̂)+k̂b n̂3
 +2 kb n̂ − I n̂
+ L̄[y(t) − h(x̂(t), u(t), θ̂(t), t)] (26) ´
Qbhat = +
Hu ηi (n̂,pi )ṁf (t−τdn )

 I n̂
where L̄ can be computed using the same technique 


 +L 21 (n m − n̂)
given in equations (19) and (20). 
solution to the diff. equation
n̂2
QbPbhat = k̂b (called simple solution) (31b)
OBSERVER CONFIGURATIONS BASED ON MEA- k ³ ´
 60 Pl (pi ,n̂) Hu ηi (n̂,pi )ṁf (t−τdn )
SUREMENTS Normally the choice of observers de-  2π n̂ − I n̂ + I n̂
pends on which sensors are available and/or neces- Qbhat
f orward = (31c)


sary. The observer to estimate the load torque has assuming, ṅ = 0
been made using a crank shaft speed sensor as this
was found to be sufficient. If the crank shaft speed Equation (31a) is the measurement of the load torque
sensor is used the system equations can have the fol- taken on the dynamometer. Equation (31b) is the so-
lowing innovations. lution to the differential equation using an observer

5
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

with the crankshaft speed as the innovation. Equation 27

Simulation study with constant throttle angle

Throttle angle [°]

Throttle angle [°]
(31b) is the solution to the simple load torque using 26

the observer estimate of k̂b and n̂. Finally equation 25

24
(31c) is the steady state solution to the load torque 23
0 10 20 30 40 50 60 70 80 90 100
state equation. It should be pointed out that the load

Load Torque [Nm]

36 Qb
Stateequation
torque measurement on the available dynamometer 34
Simple
32
is not fast enough to give a valid dynamic load indi- 30
28
cation. This is why only the steady state values are
0 10 20 30 40 50 60 70 80 90 100
used to verify the model. Specifically the feedforward 0.27

kb [kW/(krpm)3]
kb−estimated
solution (Qbhat
f orward ) is used to track the load torque.
0.26
0.25
0.24
0.23
0 10 20 30 40 50 60 70 80 90 100
r

err Load Torque [%]

SIMULINK MODEL The simulation presented 2
Stateequation
here has been made using an Adiabatic Mean Value 0
Simple

Engine Model (AMVEM) including a PI observer. The

model has been presented in [11] for further refer- −2
0 2 4 6 8 10 12 14 16 18 20
time [s]
ence. The AMVEM has been revised to include the
load model presented in the load power model sec- Figure 4: Steady state test with no throttle step. No-
tion. Figure 2 shows the implementation of the load tice that the observer was started in a different initial
model and the appendix gives the complete AMVEM condition than the system. The simple load torque ex-
observer implementation. It should be stressed that pression stabilizes first but has a larger steady state
even though the observer gives estimates of the man- error
ifold pressure and manifold temperature these are not
needed to estimate the load torque. Notice that the observer was started in a different ini-
tial condition than the system. The simple expression
f(u)
2 for the load torque stabilizes first but has a larger
P_roll

P_road
steady state error than the load torque state equation.
P_road m_car*g*u[1]*sin(u[2]*pi/180)

P_hill
2 THETA
The error can be minimized if the gain that drives the
4. gear
1
Pb
u[1]/1000

6
n_crank [rpm]−> wheel sped [km/h ]

v [m/s] v [km/h] n [krpm] parameter estimate is increased. This is not desir-

1/2*ro*Af*Cd*(u[1])^3

P_aero
u[1]/3.6 f(u) 1
n
able since larger gains lead to higher noise sensitivity.
1*u[1]

P_aux f(alternator, power stering, air condition system, ect.)

Figure 2: Road load power implementation

Transient Test with Auxiliary Power Steps The
The observer for estimating the load power is imple- SIMULINKr model was tested with different throttle
mented using a simple integration block seen in figure angle steps and auxiliary power step inputs. The
3. The innovation is here the difference between the purpose was to see how well the model can follow
models output of the crankshaft speed and the ob- dynamically the load torque. Figure 5 shows a
server estimate. simulation where different throttle angle inputs are
used. A load disturbance of 1000 W is applied from
kb2_hat t = 20 − 25 s. Notice the rapid detection from both
kb2_hat load torques equations. It is interesting to see how k̂b
1
1
1 changes in order to adapt to the new load setting.
s
kb2^ ddtkb2^
Integrator2

kb^
ddtkb^
From the simulation example it is shown that both
the simple load torque expression and the differential
Figure 3: Subsystem to estimate the loading param- equation can be used to find the load torque. How-
eter k̂b ever the system model has to be accurate to within
a few percent which means that careful calibration of
the crankshaft speed state equation is necessary oth-
Steady State Test Figure 4 shows the first test erwise the correct mean value for the load torque can
given no throttle input and no noise. not be found.

6
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

Simulation study with varying throttle angle

small errors. At the first throttle step the differential
Throttle angle [°]

50

40
equation has a much faster response and overshoots
30 Throttle angle [°] the feed forward solution. The differential equation
20 never recovers as the steady state value stabilizes at
0 10 20 30 40 50 60 70 80 90 100
around 100 Nm. QbPbhat on the other hand gives a
Load Torque [Nm]

80

60 much better estimate. Compared to the feed forward

40 Qb solution the agreement is very good. Once the crank
Stateequation
20 Simple shaft speed stabilizes (around 2.5 s) the error is no
0 10 20 30 40 50 60 70 80 90 100
more than 3 Nm. On tip-out QbPbhat never stabilizes
kb [kW/(krpm)3]

kb−estimated
0.3 but this is because the crank shaft speed is not con-
0.25 stant which is confirmed on the graph below the load
0.2
torques. At around t = 8 s the crank shaft speed is
0 10 20 30 40 50 60 70 80 90 100
again constant implying that Qbhat f orward and QbPbhat
err Load Torque [%]

4
2 should be the same which is also the case.
0
−2
−4 Stateequation
−6 Simple Notice that the crank shaft speed is being estimated
0 10 20 30 40 50
time [s]
60 70 80 90 100
correctly for the entire test except during fast tip-ins.
Figure 5: Transient test with with a 1000 W load dis- On figure 7 it is clear that these excursions are on the
turbance at t = 20 − 25 s.. Notice how k̂b changes order of about 4% and can also easily be seen on the
to adapt to the load disturbance. The maximum error crank shaft speed curves.
is about -6% for the simple load torque equation and
about -5.5% for the state equation. The steady state
error is however much better for the crankshaft speed The next measurement estimation study shows the
state equation. case when kb is not estimated but a constant kb value
is used in the crank shaft speed state equation and
load torques equations. The test is otherwise exactly
REAL ENGINE DATA The following tests have been the same as that shown in figure 6 and 7.
made in the constant speed mode on the dynamome-
ter. Figures 8 and 9 show the results gathered from the
test. Notice that the load torque estimation is not
even approximately close to working now, either for
1. 2000 rpm and throttle step from 20◦ − 40◦ , esti- the simple solution nor the state equation. QbPbhat
mating kb can be seen to be moving up and down as it naturally
follows the small variations in the crank shaft speed
2. 2000 rpm and throttle step from 20◦ − 40◦ , not
estimate. Again this observation is as expected since
estimating kb
both load torque equations and crank shaft speed are
3. 3000 rpm and throttle step from 20◦ − 60◦ , esti- very dependent on the estimate of kb .
mating kb
The error plots for the crank shaft speed suggest that
the estimation is not as good as earlier. The mean
Notice that the dynamometer in these tests has a very error deviation including transient operation has in-
slow response time (time constant of about 1.5 sec- creased from −0.0143 % to −0.77 % and the standard
onds) and should only be considered correct at the deviation has also increased from 1.024 % to 1.174 %.
beginning of the tests where the engine is in steady This can also easily be verified by looking at the crank
state operation. shaft speed estimate on figure 8 as the measured and
estimated curve do not lie on top of each other.

Testing ”Constant” Engine Speed (TR2024.mat)

Figures 6 and 7 show the test results when measure- Testing ”Constant” Engine Speed (TR3026.mat)
ments are taken when the crankshaft speed is held Figures 10 and 11 show the test results when mea-
approximately constant near 2000 rpm, and throttle surements were taken in constant speed mode (3000
angle steps between 20 ◦ and 40 ◦ . rpm) and throttle angle steps between 20 ◦ to 60 ◦ .

kb was started with the correct value so that both Similar to the previous test (TR3024.mat) the param-
QbPbhat , Qhat
b are at the correct starting value. No- eter estimate converges rapidly. After the first tip-in
tice that Qbhat meas
f orward , Qb are also right on top of QbPbhat finds the correct load torque level based on
each other and fit the estimated load torques with very the solution to Qbhat
f orward . At t ∈ [10; 12] s the crank

7
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

shaft speed is again held at a constant speed and CONCLUSION

QbPbhat is again able to find the correct mean value.
It was investigated if a dynamic load torque esti-
The differential equation is again not able to find the mate could be found using a simulation model and
correct mean value not even when the system is op- on a dynamometer. The simulation study made
erating in steady state. in SIMULINKr showed that the differential equation
could better capture the load disturbances but oth-
Notice that the crank shaft speed is varying up to 500 erwise was very similar to the simple expression put
rpm after tip-out. Therefore it is difficult to compare forward. The simple expression showed using real
the estimated load torque during this period since vehicle data that a steady state estimate of the load
neither Qbhat
f orward or Qb
meas
are showing the correct torque could be found. This was concluded based on
load torque. the measurements taken in the constant speed mode.
The dynamic response was difficult to judge due to
the lack of bandwidth in the available dynamometer
measurement system. However it has been shown
that using a simple load model the observer system
is capable of estimating the load torque when the in-
novation from the crank shaft speed sensor is used.
µ = −0.081051 [%] & σ = 8.1372 [%]
Using File TR2024.mat (iii = 4, jjj = 1) 1400
ThrottleAirMassFlow = algebraric, pobs = on, Tobs = on, nobs = on, kbobs = on, integrators = on 60
1200
30 dmatdtmeas Error dmatdt [%] 40
1000

No. of samples
dmatdt [g/s]

dmatdtAlgebraric
20 800
20
0 600
10
−20 400

0 1 2 3 4 5 6 7 8 9 −40 200

0
0 2 4 6 8 −100 −50 0 50 100
0.9
pmeas
p [bar]

0.8 man
hat
µ = −0.051717 [%] & σ = 0.69468 [%]
pman 1000
i

0.7
4
0.6 800

No. of samples
0 1 2 3 4 5 6 7 8 9 2
Error pi [%]

600

340 0
meas 400
330 Tman
hat
T −2
man
T [K]

320 200
i

310 −4
0
300 0 2 4 6 8 −4 −2 0 2 4 6

0 1 2 3 4 5 6 7 8 9 µ = −0.026458 [%] & σ = 1.0245 [%]

1200
4
80
Load Torque [Nm]

2 1000
No. of samples

60 meas 0 800
Error Ti [%]

Qb
hat −2
Qbhat
40 Qb 600
Pbhat −4
Qbhat
20 forward
−6 400
0 1 2 3 4 5 6 7 8 9 −8 200
−10
2.2 0
nmeas 0 2 4 6 8 −10 −5 0 5
2.1 nhat
n [krpm]

2 µ = −0.014275 [%] & σ = 1.0233 [%]

1000
1.9 6
1.8 4 800
0 1 2 3 4 5 6 7 8 9
No. of samples

2
Error n [%]

600
2 0
khat [kW/krpm3]

−2 400
1.5
−4
200
kbhat −6
1
b

0
0.5 0 2 4 6 8 −4 −2 0 2 4 6
0 1 2 3 4 5 6 7 8 9 time [s] error [%]
time [s]
Figure 7: Percentage error in ṁat , pi , Ti and n
Figure 6: Test made in constant speed mode using (TR2024.mat). Figures on the left show the percent-
file TR2024.mat. The test shows how kb is estimated age errors, while figures on the right show bar plots of
and adapts to the load change. the percentage errors.

8
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

Using File TR2024.mat (iii = 4, jjj = 1) µ = −0.70291 [%] & σ = 8.1847 [%]
ThrottleAirMassFlow = algebraric, pobs = on, Tobs = on, nobs = on, kbobs = off, integrators = on 1400
60
30 1200
dmatdtmeas
40
dmatdt [g/s]

Error dmatdt [%]

dmatdt 1000

No. of samples
Algebraric
20 20 800
0 600
10
−20 400
0 1 2 3 4 5 6 7 8 9
−40 200

0
0.9 0 2 4 6 8 −100 −50 0 50 100
pmeas
pi [bar]

0.8 man
phat
man µ = −0.051518 [%] & σ = 0.69547 [%]
0.7 1000
4
0.6 800
0 1 2 3 4 5 6 7 8 9

No. of samples
2

Error pi [%]
600
340
0
Tmeas 400
330 man
That
man
Ti [K]

320 −2
200
310
−4
300 0
0 2 4 6 8 −4 −2 0 2 4 6
0 1 2 3 4 5 6 7 8 9 µ = −0.025952 [%] & σ = 1.0283 [%]
1200
4
Load Torque [Nm]

35 2 1000

No. of samples
0 800
Error Ti [%]

30
Qbmeas
−2
Qbhat 600
25 Qbhat
Pbhat −4
Qbhat 400
forward −6
0 1 2 3 4 5 6 7 8 9
−8 200
2.2 −10
0
2.1 0 2 4 6 8 −10 −5 0 5
n [krpm]

2
nmeas µ = −0.77233 [%] & σ = 1.1736 [%]
1.9 1000
nhat
6
1.8
0 1 2 3 4 5 6 7 8 9 4 800
No. of samples

2
Error n [%]

600
0.82 0
kb [kW/krpm ]
3

hat
0.81 kb −2
400
0.8 −4

0.79 −6 200
hat

−8
0.78
0
0 1 2 3 4 5 6 7 8 9 0 2 4 6 8 −4 −2 0 2 4 6
time [s] time [s] error [%]

Figure 8: Test made in constant speed mode using Figure 9: Percentage error in ṁat , pi , Ti and n
file TR2024.mat (const. kb ). The test shows that the (TR2024.mat). Figures on the left show the percent-
estimation of kb is necessary in order to estimate the age errors, while figures on the right show bar plots
load torque. Notice now n also is degraded when a of the percentage errors. Notice how the crankshaft
constant kb value is used. speed error has increased due to using the constant
value of kb .

9
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

Using File TR3026.mat (iii = 11, jjj = 1)

µ = 0.23805 [%] & σ = 10.065 [%]
ThrottleAirMassFlow = algebraric, pobs = on, Tobs = on, nobs = on, kbobs = on, integrators = on
2500
60

2000
dmatdt [g/s]

50

Error dmatdt [%]

40

No. of samples
dmatdt
meas
dmatdtAlgebraric 1500
20 0
1000

0 2 4 6 8 10 12 14 16
−50 500
1
0
pmeas
man 0 5 10 15 −100 −50 0 50 100
0.8 phat
pi [bar]

man

µ = −0.087441 [%] & σ = 0.83429 [%]

0.6 2500
6
0.4 4 2000
0 2 4 6 8 10 12 14 16

No. of samples
2

Error pi [%]
1500
340 0
Tmeas
man
That −2 1000
man
Ti [K]

320
−4
500
−6
300
−8 0
0 5 10 15 −10 −5 0 5 10
0 2 4 6 8 10 12 14 16
µ = −0.045359 [%] & σ = 0.99593 [%]
100 2500
Load Torque [Nm]

5
2000

No. of samples
50
Error Ti [%]

0 1500
Qbmeas
Qbhat
hat
0 QbPbhat 1000
hat −5
0 2 4 6 8 Qb
10 forward 12 14 16
500
3.2 −10
0
3 0 5 10 15 −15 −10 −5 0 5 10
n [krpm]

2.8
nmeas µ = 0.0073844 [%] & σ = 0.65985 [%]
2.6 nhat 2500
4
0 2 4 6 8 10 12 14 16 2000
No. of samples

2
Error n [%]

1500
1
[kW/krpm3]

0
1000
0.5 −2
kbhat 500
khat
b

−4
0 0
0 2 4 6 8 10 12 14 16 0 5 10 15 −10 −5 0 5 10
time [s] time [s] error [%]

Figure 10: Test made in constant speed mode using Figure 11: Percentage error in ṁat , pi , Ti and n
file TR3026.mat. The test is made using a different (TR3026.mat). Figures on the left show the percent-
operating point and throttle movement. Again kb is age errors, while figures on the right show bar plots
estimated and adapts to the load change applied by of the percentage errors. Notice how the well the es-
the dynamometer. Notice how the crankshaft speed timates of the crank shaft speed is.
is estimated correctly except during the rapid throttle
tip in and tip out.

10
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

References [11] Elbert Hendricks, Donn Engler and Marco Fam.

A Generic Mean Value Engine Model for Spark
[1] Gerhart, J., Benninger, N. and Hess, W. Torque- Ignition Engines. Proc. SIMS-2000 Conf.,
Based System Structure of an Electronic En- September 2000.
gine Management System as a New Basis for
Drive Train Systems (Bosch ME7 Systems). 6.
Aachener Kolloqium, Fahrzu- und Motorentech-
nik, 20.-22 October 1997. Aachen Germany.

[2] N. Heintz, M. Mews, G. Stier, A. J. Beaumont

and A. D. Noble. An Approach to Torque-Based [12] J.,B., Heywood. Internal Combustion Engine
Engine Management Systems. SAE Tecnical Pa- Fundamentals. 1988.
per, 01(0269), 2001.

[3] Lee, B., Rizzoni, G., Guezennec, Y., Soliman,

A., Cavalletti, and Waters, J.,. Engine Control
Using Torque Estimation. SAE Tecnical Paper,
(01-0995), 2001.
NOMENCLATURE
[4] Corsetti, A., O’Connell, G., and Watkins, K., . In-
Vehicle Engine Torque Model Validation. SAE Abbreviations Description
Tecnical Paper, (01-1143), 2002. AMVEM Adiabatic Mean Value En-
gine Model
[5] Marco Fam. Observer Based Torque Estimation CGEKF Constant Gain Extended
of an SI Engine. Masters Thesis, DTU, Technical Kalman Filter
University of Denmark, DTU, Institute of Automa- EGR Exhaust Gas Recircula-
tion, IAU, 2800 Lyngby, March 2002. tion
EKF Extended Kalman Filter
[6] Elbert Hendricks and Spencer C. Sorensen. LTI Linear Time Invariant
Mean Value Modelling of Spark Ignition En- MAF Mass Air Flow
gines. SAE Tecnical Paper, (900616):15, Febru- MVEM Mean Value Engine
ary 1990. Model
PI Proportional Integral
[7] Elbert Hendricks, Alain Chevalier, Michael RPM Rounds Pr. Minute
Jensen, Spencer C. Sorensen, Dave Trumpy and SI Spark Ignition
Joe Asik. Modelling of the Intake Manifold Fill-
ing Dynamics. SAE Technical Paper, 0037:25,
February 1996. International Congress and Ex- Greek Symbols Description Unit
position. α throttle angle [◦ ]
α0 closed throttle angle, [◦ ]
[8] Christian Winge Vigild and Michael Struwe. Reg- 15
ulering af en SI-motors luft/brændstof-forhold β β = β1 (α)β2 (pi )
(Robust H∞ Control of an SI Engines, Air/Fuel ηi indicated efficiency [−]
Ratio). Masters Thesis, DTU, Technical Univer- ηvol volumetric efficiency [−]
sity of Denmark, DTU, Institute of Automation, κ specific heat in dry air [−]
IAU, 2800 Lyngby, August 1997. cp
cv = 1.4
µ mean value of per- [%]
[9] Per B. Jensen, Mads B. Olsen, Jannik Poulsen,
centage error
Elbert Hendricks, Michael Fons and Christian
ω angular speed of [ rad
s ]
Jepsen. A New Family of Nonlinear Observers
crankshaft
for SI Engine A/Fuel Ratio Control. SAE Techni- kg
cal Paper, February 1997. ρi air density [m3]

σ standard deviation of [%]

[10] Alain Chevalier, Christian Winge Vigild and El- percentage error
bert Hendricks. Predicting the Port Air Mass θ crank shaft angle [◦]
Flow of SI Engines in Air/Fuel Ratio Control θM BT ignition timing at opti- [◦]
Applications. Society of Automotive Engineers, mal engine torque
01(0260):28, 2000. ωs sample frequency [ rad
s ]

11
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

Constants Description Value Unit Variables Description Unit

Ai cross sec- 1.3 · 10−3 [m2 ] Pacc power needed to accelerate [kW ]
tional area Paero power for aerodynamic drag [kW ]
of intake Paux power from auxiliary devises [kW ]
manifold Pb load power on engine [kW ]
Hu lower heat- 4.3 · 104 [ kJ
kg ] Pengine power delivered by engine [kW ]
ing value for Pf riction , Pf friction losses [kW ]
fuel Phill power to climb hills [kW ]
2
·min2 ·m2
I scaled mo- 5.2638 [ g·rad s2 ] Plosses , Pl total losses [kW ]
ment of Pload , Pb load power [kW ]
inertia Ppumping , Pp pumping losses [kW ]
Iac actual mo- 0.48 [kg ·m2 ] Proll power for rolling resistance [kW ]
ment of n crank shaft speed [krpm]
inertia Qb load torque on engine [N m]
2π si (n) fitting parameter for ηvol -
k change from 60
rpm → rad s
approximation
li distance be- 14 · 10−2 [m] Ta , Tf ilt ambient temperature [K]
tween injec- TEGR EGR temperature [K]
tor and in- Ti intake manifold temperature [K]
take valve yi (n) fitting parameter for ηvol -
Lth stoichiometric 14.67 approximation
air/fuel-ratio
mat0 , mat1 fitting con- DETAILS OF THE CRANKSHAFT SPEED MODEL
stants used
in ṁat The crankshaft speed state equation presented in the
pc critical pres- 0.4125 [bar] paper is repeated here and expressions are given for
sure (two its submodels and parameters.
flow path
model)
3 Pl + Pb Hu ηi ṁf (t − τdn )
R 287.09·10−5 [ bar·m ṅ = − + (32)
kg·K ]
gas constant In In
ncyl number of 4 µ ¶2
2π
cylinders I = 1000 · (Iengine + Iload )
Vd displacement 1.275 [l] 60
volume
Vi volume 3.825 · 10−3 [m3 ] The moment of inertia I has been scaled so that n
of intake can be described in terms of rpm. In equation (32)
manifold it can be seen that there has been introduced a time
delay τdn . The physical interpretation comes from the
fact that there is a delay from fuel injection until the
actual torque response can be seen. The injection-
Variables Description Unit torque time delay is given in equation (33)
β1 (α) effective open throttle area
β2 (pi ) air mass flow dependency on µ ¶
60 1 ρi li Ai
manifold pressure τdn = 1+ + (33)
kb kW n ncyl ṁax
loading parameter [ krpm 3]
kg
ṁat air mass past throttle plate [s]
ṁap [ kg where ρi is the density of the air in the intake manifold,
air mass at intake valve s ]
kg ncyl is the number of cylinders, li is the effective dis-
ṁEGR EGR mass flow [s]
tance between the fuel injector(s) and the intake valve
ṁf fuel injected at the intake [ kg
s ] and Ai is the effective cross sectional area of the in-
valves
take manifold. ṁax either represents the air mass flow
ṁi change of air mass in mani- [ kg
s ] past the throttle plate ṁat (CFI strategy) or the port air
fold mass flow ṁap (EFI strategy) depending on the type
pa ambient pressure [bar] of injection system used.
pi absolute intake manifold [bar]
pressure The power losses in equation (32) can be written as

12
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

empirical regression expressions. ADIABATIC MEAN VALUE ENGINE MODEL IN-

CLUDING A PI - OBSERVER
Pl (n, pi ) = Pf (n) + Pp (n, pi )
= n(kf 0 + kf 1 n + kf 2 n2 ) Figure 12 shows the AMVEM including a PI observer.
+ n(kp0 + kp1 n)pi (34)
The load power has traditionally been expressed as Random
Number Throttle Position

in equation (35) Random

Number1
Random
Number2
Ti
2
Ti2
Ti2
inlet temp 2 Ti2 1
1

k b n3
Pi2

Pb (n) = (35) 1 a1 a a n
inlet pressure 2
pi2

3
pi2

pi2
a1
Crank speed 2 n2
Throttle position n2
Qb2

where kb is the loading parameter, which determines

wave generator
4 n2
Torque Qb2
dmfdt from state equation Qb2
Qb2curve1
AMVEM

the power or torque in a given operating point. 5

K
3
3
3
3
Qb2

Kalman Gain 3
L

The constants for the power losses in equation (34) k1

can be seen below and have been found to fit the ex- k2
Gain
Demux
3
1
s
3

Gain1 Integrator3

perimental engine.
k3

Gain2

kf 0 = 1.673 , a
dmfdt
k1
k2
Ti^

Pi2^

kf 1 = 0.272 , kf 2 = 0.0135 (36) k3

k4
ddtkb2^
OBSERVER
n^

Qb^ Qb2curvehat

kp0 = −0.969 kp1 = 0.206

Qb^

, Demux
5

The indicated efficiency is written as a product of 4

terms in equation (37)
Figure 12: Block diagram of the complete model
ηi (n, θ, pi , λ) = ηin (n) · ηiθ (θ, n, pi ) · ηip (pi ) · ηiλ (λ)(37) used in the project
where the terms are written in decreasing order of
importance. For the experimental engine they have
been found to be:
NONLINEAR DIFFERENTIAL EQUATIONS FOR PI ,
ηin (n) = 0.558(1 − 0.392n−0.360 ) (38) TI , N , QB , KB The nonlinear equations for the
2
ηip (pi ) = 0.9301 + 0.2154pi − 0.1657pi (39) model are given below.
κR
ṗi = (−ṁap Ti + ṁEGR TEGR + ṁat Ta )
 Vi
 −1.299 + 3.599λ − 1.332λ2 , if λ≤1 = f1 (pi , Ti , n, α)
ηiλ (λ) = (40) ·
 RTi
−0.0205 + 1.741λ − 0.745λ2 , if λ>1 Ṫi = −ṁap (κ − 1)Ti + ṁat (κTa − Ti )
pi Vi
¸
ηiθ (θ, n, pi ) = 0.7 + 0.0240(θ − θmbt (n, pi )) + ṁEGR (κTEGR − Ti )
− 0.00048(θ − θmbt (n, pi ))2 (41) =
f2 (pi , Ti , n, α)
where the Maximum Brake Torque (MBT) spark tim- Pl + Pb Hu ηi ṁf (t − τdn )
ṅ = − +
ing for the experimental engine is given by equation In In
42. µ ¶2
2π
I = 1000 · (Iengine + Iload )
n47 = if n < 4.8 then 4.7n else 4.7 · 4.8 60
θ1 = 47.31pi + 2.6214 + n47 = f3 (pi , Ti , n, α, ṁf , kb )
µ ¶
n2 n Hu ηi ṁf (t − τdn )
θ2 = −56.55pi + 57.34 + n47 (42) Q̇b = k̇b + 2kb −(Ql + Qb ) +
k I nk
θ12 = min(θ1 , θ2 )
= f4 (pi , Ti , n, α, ṁf , k̇b , kb )
θmbt (n, pi ) = min(θ12 , 45) k̇b = v1 (t), v1 (t) ∈ N (0, V1 (t))
The indicated efficiency has been expressed in table = f3 (n)
form as it was found to more accurately describe the
engine. However for the purpose of system analy-
sis the equations presented above are better suited THROTTLE AIR MASS FLOW ṀAT The throttle air
and the equations given are also implemented in the mass flow has been derived in [6] and modified in [7]
SIMULINKr model. and is presented here using the last reference.

13
 Downloaded from SAE International by Stony Brook Univ, Sunday, August 12, 2018

The air mass flow past the throttle plate has been
modelled as two physically separated parallel isen-
tropic flows.
pa
ṁat = mat1 √ β1 (α)β2 (pr ) + mat0 (43)
Ta

where β1 and β2 are given as

α02
β1 (α) = 1 − cos(α) − (44)
2!

 s
 µ ¶2

 p −p
 1 − 1−pc
r c
, if pr ≥ pc
β2 (pr ) = (45)




1 , if pr < p c

where α0 is given by the closed throttle angle mea-

sured on the engine. In this case the throttle is
α2 15◦ π
closed at 15◦ which gives 2!0 = 2!180 ◦ = 0.0343 and

pc = 0.4125.

PORT AIR MASS FLOW ṀAP The port air mass

flow ṁap to the engine has been modelled as an al-
gebraic equation. The port air mass flow is known in
the literature [7], [12] as the speed density equation
and it is given in equation (46).

Vd
ṁap = ηvol pi n (46)
120RTi
Vd is the displacement volume of the engine and ηvol
is the volumetric efficiency. The volumetric efficiency
is a measure of the amount of fresh air drawn into
the engine. On SI-engines without turbocharger the
volumetric efficiency is not able to exceed 1. In [7]
it has been shown that ηvol pi can be modelled as in
equation (47).

ηvol pi = si (n)pi + yi (n) (47)

The two terms si (n) and yi (n) have a weak depen-

dence of the crank shaft speed and are around 0.98
for si and −0.1 to −0.2 for yi .

14