Road Estimation and Fuel Optimal Control of an Off-Road Vehicle
J
¨
orgen Albrektsson
1,2
and Jan
˚
Aslund
1
1
Department of Electrical Engineering, Link
¨
oping University, SE-581 83 Link
¨
oping, Sweden
2
Volvo Construction Equipment, SE-631 85 Eskilstuna, Sweden
Keywords:
Off-road, Construction Equipment, Kalman Filters, Rolling Resistance, Optimal Control, Dynamic Program-
ming.
Abstract:
This paper explores the possibility to use optimal control to establish a Pareto front of fuel consumption vs
cycle time for a transport mission with an articulated hauler. The Pareto front can be utilised to optimise the
hauler transport mission on its own or as a part in a larger optimal control problem involving several con-
struction machines working together on a site transporting material at a set production rate. While rolling
resistance is a major energy consumer in an articulated hauler’s transport, the effect of varying rolling re-
sistance is included in the developed optimisation algorithm. A method utilising Extended Kalman Filter,
Rauch-Tung-Striebel smoothing and sensor fusion is formulated in order to calculate the road related data
needed in the optimisation algorithm. A potential fuel efficiency improvement, verified by computer simula-
tions, of up to 9% was found in the example transport mission where the optimal gear and speed trajectory
were followed instead of driving towards a mean speed target to achieve an equal cycle time for the transport
mission.
1 INTRODUCTION
In the construction industry today articulated haulers
are used in a vast amount of different transport mis-
sions due to its ability to work efficiently even at
tough (off-)road conditions. Commonly there is a set
production target [ton/h] for the transport mission, i.e.
the hauler should transport a certain amount of mate-
rial in a set time. This work explores the possibility
to use optimal control to establish a Pareto front of
fuel consumption vs cycle time for a transport mis-
sion with an articulated hauler. The Pareto front can
be utilised to optimise the transport mission on its
own or as a part in a larger optimal control problem
involving several construction machines. Predictive
cruise control is today a common feature in commer-
cial vehicles. Through calculation of an optimal ve-
hicle speed trajectory on a priori known road stretch
a reduction in fuel consumption and CO
2
emissions
can be achieved. The optimisation algorithm anal-
yses the road elevation to make use of the potential
energy that is released in down slopes, preserves the
energy as kinetic energy in the vehicle which is then
used when the road is going up again. In combina-
tion with optimising the choice of gear, enabling the
internal combustion engine (ICE) to work efficiently,
further fuel consumption reduction is possible. The
working conditions of an off-road vehicle differ in
some respects substantially compared to an on-road
vehicle. Off-road the fluctuation in elevation of the
road is steeper and more frequent and the road sur-
face is much rougher. The differences in the working
environment results in a much lower average speed
of an off-road vehicle compared to an on-road truck.
These conditions shifts the relation between energy
consumers making rolling resistance dominate over
air-drag off-road compared to the opposite in a typi-
cal on-road truck application (21st Century truck part-
nership, 2013). In (Fu and Bortolin, 2012) the possi-
bility of using Dynamic Programming is explored to
optimise a vehicle speed trajectory and to build a gear
shift strategy for an articulated hauler. This paper ex-
pands earlier research through also consider varying
rolling resistance in the optimisation algorithm and to
include an algorithm that estimates the rolling resis-
tance along the track. The paper is disposed as fol-
lows: Section 2 presents a vehicle model of an ar-
ticulated hauler, the model is essentially equal when
applied both in the map module and optimisation al-
gorithm. In Section 3 a method to estimate and record
the road characteristics is developed. Section 4 dis-
plays a method for optimal control of the hauler and
the method used to build the Pareto front. Section 5
presents results from tests and simulations.
58
Albrektsson, J. and Åslund, J.
Road Estimation and Fuel Optimal Control of an Off-Road Vehicle.
DOI: 10.5220/0006247200580067
In Proceedings of the 3rd International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2017), pages 58-67
ISBN: 978-989-758-242-4
Copyright © 2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
2 VEHICLE MODEL
In this section the driveline and the complete vehicle
are modelled. The models are used both in the map
module and in the optimisation algorithm. The dis-
played models are the ones used in the map module
while the end of the section explains the needed up-
dates when applied in the optimisation algorithm.
2.1 External Forces Acting on the
Hauler
There are several forces acting on a vehicle as it starts
to move, see i.e. (Guzzella and Sciaretta, 2013).
The main longitudinal forces acting on the articulated
hauler are displayed in Figure 1.
Figure 1: Longitudinal forces acting on an articulated
hauler.
In Figure 1 the following models / notation apply:
Vehicle Speed, v
Road Inclination, α
Tractive Force (retarding if negative), F
t
Aerodynamic Force, F
a
. The motion of an object
through the atmosphere gives rise to an aerodynamic
resisting force. For a vehicle F
a
is often modelled to
be dependent on the density of the surrounding air,
ρ
a
, the front area of the vehicle, A
f
, an aerodynamic
drag coefficient, c
d
and the velocity of the vehicle, v,
in square.
F
a
=
1
2
ρ
a
· A
f
· c
d
· v
2
(1)
Rolling Resistance Force, F
r
. The rolling resistance
force can be modelled as
F
r
= c
r
· m
veh
· g ·cos(α) (2)
The rolling resistance coefficient, c
r
, depends on
many variables such as vehicle speed, tire pressure
and road surface conditions. In our case the rolling
resistance coefficient is one of the parameters that are
estimated in the map building process and variations
in these variables will indirectly affect the estimation
of c
r
.
Force Generated by Road Inclination, F
g
. As the
vehicle moves up-hill or down-hill the gravity will
create a retarding or accelerating force depending on
the angle of the slope, α. This force is modelled as
F
g
= m
veh
· g · sin(α) (3)
2.2 Drivetrain Model
The drivetrain in a Volvo articulated hauler is shown
in Figure 2. The main components in the hauler driv-
etrain are from left to right: internal combustion en-
gine, torque converter, gear box, drop box, central
gear, hub, brake and (wheel).
Figure 2: Drivetrain in a Volvo A40G Articulated hauler.
Figure 3: Drivetrain model.
Figure 3 displays how the model of the drivetrain
is built. The notation used in picture 3 is: T = torque
[Nm], ω = angular velocity [rad/s], gr = gear [-], J =
mass moment of inertia [ kgm
2
], i = gear ratio [-], η
= efficiency [ % ].
Internal Combustion Engine ( ICE ). When
used in the map module the measured torque gen-
erated by the internal combustion engine is an input
to the drivetrain model. The net generated torque,
T
engine
, is estimated by the engine management sys-
tem (EMS) and broadcasted on the vehicle CAN net-
work. No deeper description on how the torque is es-
timated by the EMS is considered to be needed at this
point. Newton’s second law of motion, with dots rep-
resenting the time derivatives, connects angular accel-
eration to torque,
J
engine
·
˙
ω
engine
= T
engine
T
TC
(4)
where J = inertia, ω
engine
= engine angular speed and
T
TC
is the torque at node torque converter.
Road Estimation and Fuel Optimal Control of an Off-Road Vehicle
59
Torque Converter ( TC ). A torque converter
with lock-up functionality is bolted on the crankshaft
of the ICE connecting the engine to the gear box. A
Volvo A40G hauler is predominantly running in lock-
up mode. Only at take-off and at very steep uphill
slopes the hauler runs in torque converter mode. Thus
the drivetrain model is limited to only consider lock-
up mode and the TC is modelled as a stiff axle with
additional inertia according to (5):
T
TC
= T
engine
ω
TC
= ω
engine
(5)
Gear Box ( gbx ). The Volvo A40G is equipped
with a Powertronic gear box with 9 forward gears and
4 reverse gears. The Powertronic gearbox is of plan-
etary type and gear shifts are made by means of en-
gaging and disengaging clutches. A major advantage
with this design is that gear shifts can be performed
without torque loss and the vehicle loses no kinetic
energy during a gear shift. During the gear shift a
change in efficiency of the gear box is likely but this
is not reflected in the gear box model when used in
the map module since it is judged to be of minor im-
portance when estimating the road grade. In the gear-
box there are three types of efficiency losses: friction,
viscous and a loss stemming from an oil pump in the
gearbox. The friction coefficient, µ
gbx
is gear depen-
dent and derived from measurements. The speed and
gear dependency of the viscous loss is interpolated
from measured curves. The loss from the oil pump
is modelled as a constant. The model of the gearbox
becomes:
T
gbx. f ric
= T
TC
· i
gbx
(gr) · µ
f ric
(gr)
T
gbx
=
T
TC
T
gbx.pump
· i
gbx
(gr)
|T
gbx.visc
(ω
TC
,gr)| |T
gbx. f ric
|
(6)
ω
gbx
=
ω
engine
i
gbx
(gr)
(7)
Drop Box ( dbx ). The dropbox is a gear box
placed in the middle of the hauler translating the
power downwards in the vehicles vertical direction
and splitting it to the front axle and rear axles. Sim-
ilar to the gearbox a friction loss and a viscous loss
applies. Equation (8) to (9) displays the model of the
drop box.
T
dbx. f ric
= (T
gbx
· i
dbx
|T
dbx.visc
ω
gbx
|) · µ
f ric
T
dbx
= T
gbx
· i
dbx
|T
dbx.visc
ω
gbx
| |T
dbx. f ric
|
(8)
ω
dbx
=
ω
gbx
i
dbx
(9)
Centre Gear ( cgr ). The denomination centre
gear is used for the differential gear in the middle of
the axle. The centre gear is modelled as:
T
cgr. f ric
= T
dbx
· µ
f ric
T
cgr
= (T
dbx
|T
cgr. f ric
|) · i
cgr
|T
cgr.visc
|
(10)
ω
cgr
=
ω
dbx
i
cgr
(11)
Hub Reduction ( hub ). In the hub of the wheel
a reduction gear and the brake is placed. The hub
reduction gear is modelled according to the equations
below.
T
hub. f ric
= T
cgr
· µ
f ric
T
hub
= (T
cgr
|T
hub. f ric
|) · i
hub
|T
hub.visc
(ω
cgr
)|
(12)
ω
hub
=
ω
cgr
i
hub
(13)
Wheel Brake ( brake ).
The wheel brake is mounted in the hub. A speed
dependent viscous loss applies to the brake indepen-
dent if the wheel brakes are applied or not. If the
wheel brakes are in operation the hydraulic brake
pressure is measured and broadcasted via CAN. The
wheel brake is modelled according to (14) to (16),
T
brake
= P
brake
· A
brakepiston
· µ
brake
(ω
hub
)
· N
f ric.disc
· r
f ric.disc
(14)
T
wheel
= T
hub
|T
break.visc
(ω
hub
)| T
brake
(15)
ω
wheel
= ω
hub
(16)
where P
brake
= brake pressure, A
brakepiston
= area of
brake piston, µ
brake
= speed dependent brake disc fric-
tion coefficient, N
f ric.disc
= number of brake friction
discs and r
f ric.disc
= average brake friction disc radius.
A threshold is set to the brake pressure to avoid mea-
surement ripple, thus, when P
brake
200kPa, T
brake
is
set to 0.
Wheels ( Wheel ). At the wheels the torque gen-
erated by the ICE and the applied brake force is trans-
lated into a traction (or retardation) force, F
t
, see
equation (17). At the wheels, the rotational inertia is
also translated into a equivalent mass, equation (18),
F
t
=
T
wheel
r
wheel
(17)
m
rot.TC
= (J
engine
+ J
TC
+ J
gbxin
)
· (i
gbx
(gr) · i
dbx
· i
cgr
· i
hub
)
2
/r
2
wheel
m
rot.gbx
= (J
gbxout
+ J
dbxin
)
· (i
dbx
· i
cgr
· i
hub
)
2
/r
2
wheel
m
rot.dbx
= (J
dbxout
+ J
cgrin
) · (i
cgr
· i
hub
)
2
/r
2
wheel
m
rot.cgr
= (J
cgrout
+ J
hubin
) · (i
hub
)
2
/r
2
wheel
m
rot.wheel
= (J
hubout
+ J
brake
+ J
wheel
)/r
2
wheel
(18)
where r
wheel
is the radius of the wheel. The equivalent
rotational mass is summarised with the vehicle mass,
m
veh
, to a total mass denominated m
tot
.
m
tot
= m
veh
+ m
rot.TC
+ m
rot.gbx
+ m
rot.dbx
+ m
rot.cgr
+ m
rot.wheel
(19)
VEHITS 2017 - 3rd International Conference on Vehicle Technology and Intelligent Transport Systems
60
2.3 Complete Vehicle Model
Combining the drivetrain model, the external forces
and Newton’s second law of motion enables an ex-
pression for the longitudinal dynamics of the hauler.
In continuous time the expression is:
m
tot
(gr) ·
d
dt
v(t) = F
t
(t)F
a
(t)F
r
(t)F
g
(t) (20)
where gr = gear and t = time.
2.4 Model Alteration for Optimisation
The complete vehicle model is to a large extent reused
when applied in the optimal control algorithm. Fol-
lowing the energy flow with the aim to calculate the
engine speed and torque corresponding to a step in
position and states, the structure of Section 2.1 to 2.3
is reused but in reversed order. When the ICE speed
and torque is known it is translated into a fuel flow.
Wheel Brake and Hub Reduction
(torque/speed node cgr). Differing from the
wheel brake model in Section 2.2 there is no need to
calculate the brake force when calculating the fuel
cost. If the step in energy state (speed decrease)
is negative and large enough so that application of
the wheel brakes is necessary to achieve balance in
equation (42), this will render in a loss of kinetic
energy but since the ICE will operate without adding
positive torque, no fuel will be injected and thus the
fuel cost is 0.
Torque Converter and Auxiliary Equipment
(torque/speed node engine). As the torque converter
is modelled only to be operating in lock-up mode, the
hauler never comes to full stop and consequently the
energy state is always > 0 in the optimised speed tra-
jectory. In practise this has limited impact since the
very high combined gear ratio in a hauler enables very
low vehicle speeds and since it is only at the end of
the cycle v = 0 is desired. At the engine speed/torque
node the loss from the ICE’s auxiliary equipment (al-
ternator, fan, etc) is added by means of a look-up ta-
ble, equation (21). This is not necessary when the
model is used for road estimation since the measured
torque signal from the engine ECU includes the aux-
iliary equipment loss.
T
Engine
= T
TC
+ T
aux.equip
(ω
TC
) (21)
Internal Combustion Engine (ICE). The inter-
nal combustion engine model in the optimisation al-
gorithm consists of a measured look up table with en-
gine speed, ω
Engine
, and engine torque, T
Engine
as in-
put and the fuel mass flow, ˙m
f
as output. The cost in
fuel mass, m
f
[kg], for one step in distance with corre-
sponding changes of states is evaluated using equation
(45).
3 ESTIMATION OF ROAD
INCLINATION AND ROLLING
RESISTANCE
Section 3 outlines a method for collecting the road re-
lated data that is needed in the optimisation algorithm.
The intention is to use sensors available in a standard
articulated hauler complemented with a commercially
available GPS. The data is collected and processed in
a Matlab algorithm and the road is stored in a map
like format with latitude and longitude coordinates as
identification points. The main parameters that are
identified in the algorithm are: latitude (ϕ), longitude
(λ), altitude (z), mean vehicle speed (v), road incli-
nation (α), vehicle heading (β), rolling resistance co-
efficient (c
r
), speed limit (v
max
) and travelling direc-
tion (Dir.). Of the quantities stored in the map, only
ϕ,λ,α,c
r
and v
max
are used in the optimisation algo-
rithm presented in Section 4. A method to estimate
the road inclination for on-road commercial vehicles
is exhaustively described in (Sahlholm, 2011) and has
been further enhanced for off-road vehicles and to in-
clude rolling resistance in the work of (Almes
˚
aker,
2010) and (Saaf and Hana, 2011). The method used
utilizes an extended Kalman filter (EKF) to work as
an observer for the unmeasured parameter rolling re-
sistance and also to help removing potential bias error
that develops when only using an inclination sensor to
measure the road inclination (Sahlholm, 2011), p.80-
88. In the estimation model, Section 3.2.6, the vehicle
model described in Section 2 and the road model de-
scribed in Section 3.2.5 are combined to generate the
quantities stored in the road map.
3.1 Map Building Process
The intention with the proposed map building process
is that the operator initially travels the track between
the loading and unloading sites 2 times to initiate the
map. The map data is updated off-line after each fin-
ished run which is a feasible scenario on a construc-
tion site since the distance between the loading and
unloading sites normally are within 2 km (approx. 6
min travel). On a high level, the map building process
can be described according to the steps below:
1. Operator drives the track between the loading and
unloading site forth and back as fast (but safe) as
possible while necessary sensor data is recorded.
2. The direction of the travel is detected and the data
updated accordingly.
3. The collected data is processed in the map-
building algorithm according to:
(a) Calculation of applied brake force.
Road Estimation and Fuel Optimal Control of an Off-Road Vehicle
61
(b) Translation of ICE torque into force at wheels.
(c) Geographic and vehicle dependent data (mea-
sured and calculated) are merged in an Ex-
tended Kalman Filter (EKF).
(d) Smoothing of estimates with Rauch-Tung-
Striebel algorithm to remove potential lag.
(e) Merge the estimates into a map utilising a fu-
sion algorithm.
4. The highest recorded speed at each coordinate is
used to set the max speed limit.
As additional runs are travelled during production
new data is merged into the map after each run, im-
proving the quality of the map.
3.2 Sensor and Data Fusion
The proposed map building algorithm utilises data
recorded from an external GPS: ϕ, λ,z,β and the vehi-
cle CAN: v, α, vehicle articulation (Φ), engine torque
(T
engine
). The GPS sensor used in the tests was a
Garmin GPS18x OEM. The accuracy according to the
manufacturer (Garmin international inc., 2011) is for
position: < 15m, 95 % typical and for velocity: 0.1
knot RMS steady state in GPS Standard Positioning
Service mode and position: < 3m, 95 % typical and
velocity: 0.1 knot RMS steady state in WAAS mode.
No specific data on the accuracy of the altitude signal
is given.
3.2.1 Time vs Spatial Sampling
The predominant way to describe a road is through
map coordinates. Using distance rather than time as
the independent variable facilitates the fusion of data
from several different runs along the road since dif-
ferent runs may have been travelled in opposing di-
rection and at different speeds. To shift to distance as
the independent variable the following conversion is
used in the vehicle’s longitudinal model.
dv
dt
=
dv
ds
ds
dt
= v
dv
ds
dv
ds
=
1
v
dv
dt
,v 6= 0 (22)
3.2.2 Sensor Fusion
Several different methods for sensor fusion are avail-
able, see e.g. (Gustafsson, 2012). The Kalman filter
and the extended Kalman filter (EKF) are well-known
mathematical methods for sensor fusion that also en-
ables the possibility to estimate the states of a pro-
cess. This is a valuable feature since the rolling re-
sistance is not directly measurable with the standard
mounted sensors on an articulated hauler. While the
Kalman filter method only are able to estimate the
states of a linear process the extended Kalman filter
method gives the possibility to estimate the states in a
non-linear process (Welch and Bishop, 2006). Based
on the findings in (Sahlholm, 2011) and (Almes
˚
aker,
2010) the extended Kalman filter was chosen as the
method for sensor fusion in the map building pro-
cess. The use of the extended Kalman filter in the
proposed map building process follows to a large ex-
tent the guidance given in (Welch and Bishop, 2006).
3.2.3 Smoothing
To compensate for filtering delay and to include later
measurements in the estimate for each data point a
smoother is applied after each run. The Rauch-Tung-
Striebel (RTS) smoother (Rauch et al., 1965) is an ef-
ficient two-pass algorithm for fixed interval smooth-
ing. The use of the RTS smoother is possible since
the intention is to update the map only after a com-
plete run along the road.
3.2.4 Data Fusion
A general data fusion method is used to merge data
from different runs along the road. The data fusion
method is described in (Gustafsson, 2012), p.30. Af-
ter the data has been merged, the data is stored in the
map for each coordinate pair along with the covari-
ance matrix.
3.2.5 Road Model
In the proposed road model the identification points
are separated with a nominal distance s, as described
in Section 3.2.7. Out of the 7 appended road pa-
rameters only the correlation between road altitude,
z, and road inclination angle, α, is modelled as the
other parameters are either measured or observed in
the Kalman filter. The correlation between road alti-
tude and road inclination angle is modelled as
dz
ds
= sin(α(s)) (23)
3.2.6 Estimation Model
This section describes how the road parameter esti-
mation is made and the models used in the estimation
process.
Extended Kalman Filter ( EKF ) and Smoothing.
The states to be estimated presented in continuous
time are displayed in (24).
ˆx (t) = [ϕ(t) λ(t) z (t) v (t) α (t) β(t) c
r
(t)]
T
(24)
VEHITS 2017 - 3rd International Conference on Vehicle Technology and Intelligent Transport Systems
62
The explanation of the parameters are found in the
beginning of this section. As described in 3.2.1 spa-
tial samples are used instead of continuous time in the
model. To shift to distance as the independent vari-
able equation (22) is used. Equation (24) is translated
into discrete notation, see equation (25), where k rep-
resents the index of the location.
ˆx
k
= [ϕ
k
λ
k
z
k
v
k
α
k
β
k
c
r.k
]
T
(25)
Time update (a priori estimate).
1. Define two distances, one in meters and one in
degrees:
s
m.k
= ˆv
k
· T
s
s
deg.k
=
s
m.k
r
earth
·
180
π
(26)
2. Project the state ahead (state equations).
ˆx
k
=
ϕ
k1
+ s
deg.k1
cos(α
k1
)cos(β
k1
)
λ
k1
+ s
deg.k1
cos(α
k1
)sin(β
k1
)
z
k1
+ s
m.k1
sin(α
k1
)
v
k1
+
s
m.k1
v
k1
F
t.k1
F
a.k1
F
g.k1
F
r.k1
m
tot
α
k1
β
k1
+ s
m.k1
cos(α.k1)
r
turn.k1
c
r.k1
(27)
With:
r
turn.k1
= l
1
cot(Φ
k1
) +
l
2
sin(Φ
k1
)
(28)
where r
turn
is the turning radius of the vehicle, Φ is the
articulation angle, l
1
and l
2
distances between axles
and articulation point (front / rear).
3. Project the error covariance ahead
Define the Jacobian: A[i,j] = df[i]/dx[j] and
project the error covariance:
P
k
= A · P
k1
· A
T
+ Q (29)
4. Measurement update (a priori estimate) Define
the measurement vector:
y
k
=
ϕ
k.gps
λ
k.gps
z
k.gps
v
k.CAN
α
k.CAN
β
k.gps
T
(30)
Measurement equation:
y
k
= H · x
k
+ e
k
(31)
where the H matrix is:
H =
I
6
h
7
= 0
(32)
Calculate the Kalman gain:
K
k
= P
k
H
T
(HP
k
H
T
+ R)
1
(33)
5. Update estimates with measurement
ˆx
k
= ˆx
k
+ K
k
(y
k
H ˆx
k
) (34)
6. Update error covariance
P
k
= (I K
k
H)P
k
(35)
7. Save ˆx
k
, P
k
, ˆx
k
and P
k
at each coordinate [k]
to be used in smoothing process.
8. Initiate smoothing with the last predicted val-
ues ( ˆx
N+1|N
) and last predicted covariance matrix
(P
N+1|N
), where N is the total number of measured
data points. Run smoothing backwards along the
track. Kalman smoothing gain:
K
s
k
= P
k|k
+ A
T
P
−−1
k+1|k
(36)
Smoothed estimates:
ˆx
s
k|N
= ˆx
k|k
+ K
s
k
( ˆx
s
k+1|N
ˆx
k+1|k
) (37)
Smoothed error covariance matrix
P
s
k|N
= P
k|k
(P
s
k+1|N
P
k+1|k
)K
sT
k
(38)
3.2.7 Fusion of Map Data
A reference track is chosen and split into 6m long
sections. The knot points are identified through the
ϕ and λ coordinates and the corresponding states are
appended. When the reference map is compared with
a recorded track the search area of new measurements
is limited to points which have the same heading,
|
β
|
15
and to a rectangular area that is ±1.5m in the
heading direction and ±8m orthogonal to the heading.
If driven in reversed direction, the sign of α and the
heading is switched (180 deg). Out of the points in
the new track that is in the search area, the point that
is closest to the reference point in the horizontal plane
is chosen. The tracks are merged into the stored map
through fusion of independent estimates as described
in (Gustafsson, 2012), p.30. The states in the map is
calculated according to equation (39).
P
f
k
= ((P
1
k
)
1
+ (P
2
k
)
1
)
1
ˆx
f
k
= P
f
k
· ((P
1
k
)
1
ˆx
1
k
+ (P
2
k
)
1
ˆx
2
k
)
(39)
4 OPTIMAL CONTROL OF AN
ARTICULATED HAULER
In this section a method to control the velocity and
gear shift of an articulated hauler as it travels along a
road with varying inclination and surface conditions
is developed. The target is to derive a Pareto front
of minimum fuel consumption vs cycle time. Input
to the optimisation algorithm is machine data and the
road dependent data developed in Section 3.
Road Estimation and Fuel Optimal Control of an Off-Road Vehicle
63
4.1 Objective
The objective is to transport material at a set produc-
tion rate [ton/hour], which easily translates into cycle
time, while minimising fuel consumption. A Pareto
front is built through running the optimisation proce-
dure a number of times with different cycle time tar-
gets achieving a set of discrete cycle time - min fuel
consumption points. Denominating each individual
optimisation procedure with i, the objective becomes:
minimise M
i
i = 1, ...,n (P1)
s.t. t
i
where M
i
= fuel consumption in cycle i and t
i
= cy-
cle time in cycle i. Dynamic programming is used as
method for the optimisation and to avoid the Curse
of dimensionality ((Bellman, 1961)) the approach of
(Monastyrsky and Golownykh, 1993) and (Hellstr
¨
om
et al., 2010) is used, i.e. the trip time is added to the
criteria in (P1) which becomes:
minimise M
i
+ β
i
t i = 1, ...,n (P2)
The trade-off between fuel consumption and cycle
time is represented by the scalar coefficient β. The n
number of discrete points in the Pareto front is estab-
lished through varying β n times. The lower limit of
the cycle time in the Pareto front is found through set-
ting β high enough to reach maximum speed limit and
the upper limit is found through setting β low enough
so no further fuel consumption decrease is found.
4.2 Dynamic Programming
Dynamic Programming (DP), developed in the 1950’s
by Richard Bellman, is a well known algorithm to
solve optimal control problems. Considering road
topology and rolling resistance as a priori known dis-
turbances (by means of the earlier described map-
module) and since dimension is small, DP suits the
optimal control problem at hand well. While the DP
algorithm is not described in-depth here, the reader
is referred to (Bellman and Dreyfus, 1962) and e.g.
(Guzzella and Sciaretta, 2013).
4.2.1 State Space
While it is the cycle time / fuel consumption trade off
that is the main objective a natural choice for the first
state variable would be vehicle speed. However, fol-
lowing the findings in (Hellstr
¨
om et al., 2010), hav-
ing energy as the state variable damps the oscilla-
tory behaviour of the control while using the preferred
Euler forward method for discretisation. Thus, en-
ergy is chosen as the first state variable. The sec-
ond state variable is gear number rendering in the
state vector: x
k
= [ e
k
gr
k
]
T
, where e = energy and
gr = gear number. Denominating the control vari-
ables u, the control vector is u
k
=
u
e.k
u
gr.k
T
=
[ e
k+1
e
k
gr
k+1
gr
k
]
T
.
4.2.2 Control Constraints
Since the proposed drivetrain model is limited to lock-
up mode, see Section 2.4, a min limit for the velocity
needs to be set. The max limit of the speed is an input
to the optimisation procedure from the map module.
Consequently the vehicle speed is limited to
v
min
v v
max
(40)
Due to limitations in the gearbox a limit on gear
steps is introduced i.e. the maximal number of gear
shifts is 2 (both up and down shift).
gr
k
2 gr
k+1
gr
k
+ 2 (41)
4.3 Dynamic Model
The vehicle model in Section 2, with alterations de-
scribed in Section 2.4, is used. Thus the complete
vehicle model is the same as in equation (20). Trans-
lated into spatial coordinates and reformulated into
terms of energy (20) becomes:
de
ds
= F
t
F
a
F
r
F
g
(42)
4.4 Discretisation
The optimisation problem is solved numerically and
must be discretised. The data from the map module
is discrete and split into N steps of length h such that
the total distance of the transport mission, S, equals
S =
N
k=1
h
k
(43)
Utilising Euler forward method to discretise equa-
tion (42), the discretised complete vehicle model is
written
e
k+1
e
k
h
k
= F
t.k
F
a.k
F
r.k
F
g.k
(44)
Similarly the fuel mass flow ˙m
f
is transformed
into spatial representation using equation (22) and
then discretised with the Euler method.
m
f .k+1
= m
f .k
+
h
k
v
k
˙m
f .k
(45)
VEHITS 2017 - 3rd International Conference on Vehicle Technology and Intelligent Transport Systems
64
4.5 Cost Function
The cost function is a central part of the DP algo-
rithm and in the work at hand based on calculating the
equivalent fuel cost, m
f
, for bringing the vehicle from
one position on the road to the next position. Dur-
ing the transition both states, i.e. the kinetic energy
(speed) and the gear, may change. A time penalty, in-
troduced in Section 4.1 and the cost for changing gear
described below, are added to the cost function.
ζ = m
f
+ βt + m
f .gs
(46)
Differing from the drivetrain model in the map
module, see 2.2, the efficiency loss in the gearbox at
gear shifts is accounted for in the optimisation algo-
rithm. The cost of a gear shift is approximately equal
to the work that is lost speeding up or slowing down
the engine to meet the next gear. In the model, this is
implemented as the fuel flow needed to accelerate or
decelerate the engine inertia plus the inertia of com-
ponents up to the point where the gear is engaged.
The fuel flow is multiplied with the time of the gear
shift resulting in a fuel mass penalty, m
f .gs
.
˙
ω
gs
=
ω
engine.k
(gr) ω
engine.k+1
(gr)
t
gs
(47)
T
gs
=
˙
ω
gs
J (48)
m
f .gs
= ˙m
f .gs
(ω
engine.k
,T
gs
) · t
gs
(49)
5 RESULT
To test the map module a 1.2km long gravel road
was travelled with an articulated hauler 3 times in
each direction while needed sensor data was recorded.
The data was processed off-line in a Matlab script
designed according to the method described in Sec-
tion 3. A comparison between measured / estimated
parameters and the resulting fusioned data stored in
the map is shown in Figure 4 to 7. As seen in Fig-
ure 4 and 5 the latitude - longitude and especially the
altitude signal from the GPS have rather poor accu-
racy. In the vertical plane the spread in the GPS sig-
nal is approximately 10m. Also the signal from the
road angle sensor is wildly fluctuating in real working
condition. Except for the endpoints of the track, the
use of sensor fusion and fusion of data from several
runs averages the spread in the individual measure-
ments to make a uniform estimation of the road. At
the endpoints of the track (i.e. the loading/unloading
area) there is much larger spread in the hauler’s move-
ments (position, speed, heading, articulation etc.) and
the method has some difficulties in finding reference
measurements rendering in some unexpected fluctua-
tion in the stored map. A solution to this would be
to define a loading respectively unloading area, e.g.
implemented as a circle, and then only store map data
between the periphery of the two circles.
Figure 4: Measured and resulting latitude and longitude
coordinates stored in the map. (axes normalised)
Figure 5: Measured and resulting altitude stored in the
map.
Figure 6: Measured and resulting road angle stored in the
map.
Road Estimation and Fuel Optimal Control of an Off-Road Vehicle
65
Figure 7: Individual estimations and fusioned rolling
resistance stored in the map.
A test to see how well the method estimates the
rolling resistance was performed. Since it is very
difficult to measure the rolling resistance of an ar-
ticulated hauler in actual working conditions, the
measured GPS data was combined with a pre-set
rolling resistance, c
r
= 3%, and then the hauler drive
was simulated with a Volvo in-house developed soft-
ware to generate the engine torque and speed signals
needed in the map module. The map module was run
on the created data and a comparison of the estimated
rolling resistance and the pre-set is shown in Figure 8.
Even if only one drive is used as input, due to diffi-
culties in syncing measured and simulated data, (the
estimation is enhanced if several turns are driven) the
rolling resistance is well represented after some ini-
tialisation time.
Figure 8: Verification of rolling resistance estimation.
Utilising the created map as input to the opti-
mal control module a speed and gear shift optimisa-
tion can be performed. Figure 9 to 10 displays op-
timal speed and gear-shift trajectories when the time
penalty is set to β = 0.007g/s. In the figures a com-
parison is made to speed and gear trajectories as sim-
ulated with an in-house Volvo tool when the speed is
limited to 30/km/h which gives a similar cycle time.
Figure 9: Vehicle speed trajectory.
Figure 10: Gear trajectory.
As seen in Figure 9 in the optimal trajectory the
hauler picks-up speed in the down slopes generating
kinetic energy which is utilised when the road goes
up. Figure 10 shows that when the gear shift is opti-
mised higher gear is consistently used enabling lower
engine speeds and reducing fuel consumption. The
combination of optimal speed and gear saves approx.
9 % fuel compared to when a fixed target speed is set.
The comparison of how the fuel consumption devel-
ops along the track is shown in Figure 11.
Figure 11: Accumulated fuel consumption (normalised).
Following the method described in Section 4.1 a
Pareto front trading fuel consumption against cycle
time is built, see Figure 12. The graph shows a clear
increase in fuel consumption if the cycle time is de-
creased below approx. 200s.
VEHITS 2017 - 3rd International Conference on Vehicle Technology and Intelligent Transport Systems
66
Figure 12: Pareto front showing trade-off (normalised) fuel
consumption vs cycle time.
6 CONCLUSION
A method to generate a map which includes parame-
ters important for optimisation in off-road conditions,
such as road inclination and rolling resistance, has
been developed. The method utilises sensors mounted
as standard on a Volvo articulated hauler combined
with a commercially available GPS sensor. The main
algorithms used is an extended Kalman filter to merge
sensors, a RTS smoother to remove the filter lag and
a data fusion algorithm to merge data from several
runs along the road. The map is used as input to the
optimal control problem to minimise fuel consump-
tion in a articulated hauler transport mission towards
a set cycle time. A Dynamic Programming algorithm
is developed to solve the optimal control problem.
In the DP algorithm, optimal vehicle speed and gear
shift trajectories are computed enabling the hauler to
make best use of its kinetic energy and to consistently
choose high gears to enable low engine speed, min-
imising fuel consumption. In the test case, a poten-
tial reduction of fuel consumption of up to 9%, ver-
ified by computer simulations, is shown when com-
pared to a simulation of the same transport mission
where a fixed mean speed target is used to achieve an
equal cycle time. The proposed optimisation method
is utilised to create a Pareto front of fuel consump-
tion vs cycle time for the transport mission, which can
be used for the hauler transport itself or when solv-
ing a larger optimal control problem involving several
construction machines working together on a trans-
port mission. Different means to get the articulated
hauler to follow the optimal control trajectories are
plausible. One is to implement a human machine in-
terface (HMI) instructing the driver to follow the op-
timal speed trajectory, a second could be to design a
cruise control software that controls speed and gear
shifts (under the operators supervision) and in an au-
tonomous hauler the system could be integrated into
the control system of the machine.
ACKNOWLEDGEMENTS
This research is supported by FFI - Strategic Vehicle
Research and Innovation.
REFERENCES
21st Century truck partnership (2013). 21st cen-
tury truck partnership and technical white papers.
www.energy.gov/sites/prod/files/2014/02/f8/21ctp
roadmap white papers 2013.pdf.
Almes
˚
aker, B. (2010). Iterative map building for gear shift
decision. Master’s thesis, Uppsala University.
Bellman, R. (1961). Adaptive control process. Princeton
University Press.
Bellman, R. E. and Dreyfus, S. E. (1962). Applied dynamic
programming. Princeton University Press.
Fu, J. and Bortolin, G. (2012). Gear shift optimization
for off-road construction vehicles. In Procedia - so-
cial and behavioral science, volume 54. SCIENCEDI-
RECT.
Garmin international inc. (2011). Gps
18x technical specifications.
http://static.garmin.com/pumac/GPS 18x Tech
Specs.pdf.
Gustafsson, F. (2012). Statistical sensor fusion. Studentlit-
teratur AB, 2:1 edition.
Guzzella, L. and Sciaretta, A. (2013). Vehicle propulsion
systems. Springer-Verlag, 3 edition.
Hellstr
¨
om, E.,
˚
Aslund, J., and Nielsen, L. (2010). Design
of an efficient algorithm for fuel-optimal look-ahead
control. Control Engineering Practice, 18(11):1318–
1327.
Monastyrsky, V. V. and Golownykh, I. M. (1993). Rapid
computations of optimal control for vehicles. Trans-
portation Research, 27B(3):219–227.
Rauch, H. E., Striebel, C. T., and Tung, F. (1965). Maxi-
mum likelihood estimates of linear dynamic systems.
AIAA Journal, 3(8):1445–1450.
Saaf, M. and Hana, A. (2011). Map building and gear shift
optimization for articulated haulers. Master’s thesis,
M
¨
alardalen University.
Sahlholm, P. (2011). Distributed road grade estimation for
heavy duty vehicles. PhD thesis, Royal Institute of
Technology, Stockholm.
Welch, G. and Bishop, G. (2006). An introduction to the
kalman filter. Technical report, University of North
Carolina at Chapel Hill.
Road Estimation and Fuel Optimal Control of an Off-Road Vehicle
67