You are on page 1of 6

2015 IEEE 54th Annual Conference on Decision and Control (CDC)

December 15-18, 2015. Osaka, Japan

Heave Spring and Ride Height Optimisation of a Formula One Car


Suspension System
Mehdi Imani Masouleh and David J. N. Limebeer

Abstract— The suspension of a Formula One racing car is A direct multiple shooting method [6] was used in [7]
modelled using kinematic analysis. A set of non-linear equations to solve the minimum lap time problem for a race car. The
are derived and then solved and fitted to a multivariate runtime for a full lap of Circuit de Catalunya-Barcelona on
polynomial. This meta-model maps the normal tyre forces and
heave spring stiffness to the roll and ride heights of the car. a Sun Spark station with six CPUs running at 336 MHz was
Numerical optimal control is used to calculate the minimum approximately 28 hours for the case which converged and 60
lap time trajectory for a vehicle model which includes the hours for cases which did not satisfy the tolerances within
suspension system characteristics. A parametric study is then maximum major iterations limit of 200.
conducted to determine the optimum suspension set-up. A direct collocation method based on the trapezoidal
I. INTRODUCTION rule was used to calculate the minimum lap-time for the
Barcelona circuit in [8]. The full lap simulation was solved
With increasing computational power and the introduc- in 15 minutes on a quad-core 3.4 GHz desktop station. In
tion of better numerical methods to tackle optimal control a later paper [9], the minimum lap time strategies for cars
problems, minimum lap time trajectory calculations have with energy recovery systems were studied. The optimal
become more important to racing teams. In some races even a control problem was solved using a variable-order adaptive
millisecond a lap can mean a different finishing grid position [10] pseudospectral method.
and hence the racing teams try to optimise every aspect of Suspensions allow the car to heave, pitch and roll relative
the car to suit the circuit as best as possible. to the road. They play a crucial role in the performance
Lap time analysis has traditionally been conducted using of Formula One cars where aerodynamic characteristics are
quasi-steady state methods [1]. Although these have proven sensitive to attitude and height of the chassis. Five link
to be a very useful tool in lap time analysis, they all have suspension quarter car models for road vehicles have been
an inherent limitation; they assume the lap can be broken considered in the literature [11]–[13]. In what follows we
into a series of steady state manoeuvres. This makes them will develop a quasi-static suspension model for a racing car
unsuitable for studying vehicle transient behaviour. Another with a double wishbone suspension. Such a model differs to
pitfall of such methods is that the racing line is typically that of road vehicles in that it employs heave springs and
required a priori. In order to alleviate these issues an optimal rockers which couple the left and right wheels.
control formulation can be used. Including the full dynamics of the suspension in the car
There are two general methods of solving optimal control model will result in impractical solution times, as fine meshes
problems numerically, direct and indirect methods. Indirect will be required to capture the fast dynamics of the stiff
methods involve deriving the conditions of optimality [2] and suspension. Furthermore, this approach will increase the
solving the consequent two point boundary value problem number of states which in turn will considerably inflate the
(TPBVP). This approach was taken by [3] where manoeu- number of decision variables in the NLP problem. Therefore,
vrability of motorcycles was studied and in [4] where the a quasi static model of the suspension will be considered
minimum lap time problem was solved for a motorcycle on and a meta-model will be used to act as a surrogate for
Ardia circuit. A downside of indirect methods is the need to the full kinematic model. This meta-model is then used to
symbolically derive the necessary conditions of optimality. determine the aerodynamic forces acting on the vehicle. A
Direct methods aim to discretise the problem and then full lap of Barcelona is then simulated using a pseudospectral
optimise [5] as opposed to indirect methods where optimality collocation method and a set of parameter optimisation
conditions are first derived and then solved by discretisation. studies are carried out to find the suspension set-ups which
Direct methods involve approximating the infinite dimen- result in optimal lap times.
sional problem into a finite dimensional problem and then This article is structured as follows: In Section II a kine-
applying Non-linear programming (NLP) techniques to solve matic suspension model is developed from first principles.
the optimal control problem. Such methods are typically Multivariate polynomials are then used to capture the main
divided into two general categories, shooting and collocation. behaviour of the suspension for fast calculations. Section III
This work was supported by the UK Engineering and Physical Sciences discusses the vehicle model which is used in the optimal
Research Council. control problem. The way in which the optimal control
Mehdi Imani Masouleh (mehdi.imanimasouleh@eng.ox.ac.uk) and problem is solved is covered in Section IV. The results of
David J. N. Limebeer (david.limebeer@eng.ox.ac.uk) are with the Depart-
ment of Engineering Science, University of Oxford, Parks Road, Oxford, minimum lap time simulations are then presented in Section
OX1 3PJ. V with the conclusion in Section VI.

978-1-4799-7886-1/15/$31.00 ©2015 IEEE 165


II. S USPENSION respective mounting points.
Points P1 to P5 are fixed in car chassis and Q1 to Q6
A typical Formula One car consists of a double wishbone are fixed in the wheel carrier. Point P6 is a moving point
suspension mechanism with eight wishbones in total (two in the car body frame. The wheel carrier has a total of six
on each quarter). Fig. 1 illustrates the front left quarter of degrees freedom to begin with, however its movements are
such a suspension system, in a push rod configuration. In constrained with the presence of the six links. Since only
the following analysis we will make use of two coordinate P6 is free to move in the car body, each wheel carrier will
systems, OC xyz and OW xyz that are fixed in the car body only have one degree of freedom which can be seen as the
and the wheel carrier respectively. Throughout this paper the rotation angle of its corresponding rocker.
SAE coordinate system will be used in which the x-axis Since the distance between rod ends are fixed, these
points forwards, the y-axis points to the right and the z- kinematic constraints can be written as
axis points downwards as shown in Fig. 1. The direction of
rotations around each axis is determined using the right hand kPi − Qi k22 = li2 i = 1, · · · , 6 (1)
rule.
in which the lengths li are constant and Pi ’s and Qi ’s are
Rods 1-4 make up the wishbones, Rod 5 the steering
both described in car body reference frame. The wheel carrier
track rod and Rod 6 the suspension push rod. Wishbone
reference frame has a translational freedom with respect to
elements have one end fixed on the car body, we will refer
the car body frame given by
to these points as P1 , . . . , P4 . The other ends are fixed in
the wheel carrier, referred to as Q1 , . . . , Q4 . The steering  T
D= ∆x ∆y ∆z . (2)
track rod links the wheel carrier at Q5 to the steering rack
at P5 . The steering rack is linked to the steering wheel by It also has rotational degrees of freedom relative to the
a pinion which converts the steering wheel rotations into chassis frame which can be described in terms of the
relative lateral movements of the rack. In this analysis, we elementary rotations Rx (φ), Ry (θ) and Rz (ψ) such that the
will assume that the steering wheel rotation has a relatively total rotation is
small effect on suspension movement and hence treat P5 as
a static point on the car body. RT (ψ, φ, θ) = Rz (ψ)Rx (φ)Ry (θ) =
 
The push rod transfers the tyre loads to the heave spring cψ cθ − sψ sφ sθ −sψ cφ cψ sθ + sψ sφ cθ
and the shock absorber, which are both packaged inside the sψ cθ + cψ sφ sθ cψ cφ sψ sθ − cψ sφ cθ  (3)
nose of the car. The push rod links the wheel carrier at Q6 −cφ sθ sφ cφ cθ
to the rocker mounting point at P6 , which is a moving point
in which sφ and cφ are the sine and cosine, respectively, of
in the car body reference frame (refer to Fig. 2, shown as
φ. The same notation is used for θ and ψ. Therefore a point
P6L for the left rocker). All rods are considered to be rigid,
xwi that is fixed in wheel carrier frame can be described in
inextensible and massless links with lengths l1 to l6 . Ideal
car body coordinates by using
spherical joints are used to connect the rod ends to their
xci = RT (ψ, φ, θ)xwi + D (4)
Fig. 2 shows the rocker assembly. The rocker pivot is
RL3 dˆ3 mounted on a torsion bar at PRP (L and R subscripts denote
l3 left and right hand sides respectively) and rotates around the
Q3,4
l4
RL4 dˆ4 axis nR . One end of the rocker is connected to the push rod
at P6 and the other to the heave spring at PH . Besides the
x Q5
l5
RL5 dˆ5 coupling which is caused by the heave spring between left
Ow y RL6 dˆ6 and right rockers, an anti-roll bar is used to oppose excessive
z
l6 l1 RL1 dˆ1 roll. Heave spring stiffness is represented by Kh and the
Q1,2 RL2 dˆ2
l2 x
Q6 Kh

QGC Oc y
PHL PHR
z
MTL
FTL Ktl Ktr
P6L P6R
PRPR
Fig. 1: Front-left quarter of a typical Formula One car sus- PRPL
pension in a push rod configuration. The OC xyz coordinate FPL FPR
system is fixed in the car body, while OW xyz is fixed in the
wheel carrier with its origin at the wheel’s geometric centre. Fig. 2: Push-rod and rocker set-up. The heave spring has
The reaction force magnitudes are given by RLi with their stiffness Kh . Right and left hand quantities are labelled with
directions given by dˆi for i = 1, . . . , 6. subscripts R and L respectively.

166
torsion bar stiffness by Kt . total set of unknown forces, three more equations are re-
The mounting points on the rocker move as it rotates quired which can be obtained by balancing moments around
around rotation axis nR . To describe movements of these Q1 . For the left hand wheel this results in
points in terms of rocker rotation angle α and rotation axis
MT L + d1GC × FT L + d1W × FLg + RL2 (dˆ2 × d12 )
nR we define the rotation matrix R = R(nR , α):
+ RL3 (dˆ3 × d13 ) + RL4 (dˆ4 × d14 ) (11)
R(nR , α) = cos(α)I + (1 − cos(α))nR nTR + sin(α)S(nR )
(5) + RL5 (dˆ5 × d15 ) + RL6 (dˆ6 × d16 ) = 0.
in which the skew-symmetric matrix S(nR ) is defined by
  In (11) MT L is the external torque acting on the tyre. The
0 −nz ny d1i represents the displacement vector pointing from Q1 to
S(nR ) =  nz 0 −nx  . (6) Qi . d1W is the vector pointing from Q1 to the geometric
−ny nx 0 centre of the wheel and d1GC points from Q1 to the ground
contact point.
Balancing moments around the pin joint PRP L gives
In order to calculate d1GC the geometry of the ground
(FP L (R(nRL , αL )(P6L − PRP L )) × dˆ6 ).nRL contact point needs to first be considered. If we let êw to be
+ (Fheave (R(nRL , αL )(PHL − PRP L )) × dˆhs ).nRL (7) the wheel spindle axis in wheel carrier frame then we can
− Ktl αL − Karb (αL + αR ) = 0 define êcw to be the spindle axis in car body frame given by

in which FP L is the left push rod force, nRL is the left êcw = Rz (ψ)Rx (φ)Ry (θ)êw (12)
rocker pivot rotation axis, dˆ6 is the unit vector aligned with Provided that the chassis pitch and roll relative to the road
the push rod, dˆhs is the unit vector aligned with the heave is ‘small’, then unit road-normal expressed in the chassis
spring, Fheave is the heave spring force, Karb is the anti- frame can be approximated by n̂z = [0 0 1]T . With this
roll bar stiffness, ‘.’ designates the scalar product and × the assumption, the vector pointing from the wheel centre to the
cross product. ground contact point expressed in the car body frame can be
Formula One cars typically employ heave springs stiff- written as
nesses which are non-linear functions of variations in spring êcw × (êcw × n̂z )
length. The heave spring force Fheave can therefore be rw = − Rf w (FT L ) (13)
kêcw × (êcw × n̂z )k2
expressed as
Fheave = kh (∆L)∆L (8) and finally d1GC can be expressed as
in which ∆L = |PHR − PHL | − L0 with L0 being the d1GC = rw + d1W . (14)
unloaded length.
The next step is to carry out the force analysis on the Since Formula One car suspensions are very stiff, the tyre
assembly. With the aid of Fig. 1 the force equilibrium squash can have a sizeable impact on the movements of the
equation can be written as chassis. Using a polynomial fitted to experimental data, the
loaded tyre radius for the front wheels is given by

FT L + FLg + RL1 dˆ1 + RL2 dˆ2 + RL3 dˆ3 Rf w = 326.0 + 4.00 × 10−3 FT L + 4.075 × 10−8 FT2 L (15)

+ RL4 dˆ4 + RL5 dˆ5 + FP L dˆ6 = 0 (9) in which FT L is the normal tyre load expressed in Newtons
and unit of Rf w is in millimetres.
in which FT L is the external tyre force applied at the
To summarise the suspension analysis, the equations (1)
ground contact point and FLg is the force exerted on the
constitute the kinematic displacements of the links and de-
centre of wheel carrier due to earth’s gravity and is given by
termine the rotation and displacement of the wheel carriers.
 T Equation (7) determine the rocker angle. The six equations
FLg = 0 0 mL g . (10)
defined in (9) and (11) allow the reaction forces in the
If the chassis roll and pitch are not negligible, the tyre links to be calculated. Therefore, there are thirteen non-linear
force FT L and the gravitational force FLg will have to equations on each side and in total twenty six equations
transformed from inertial frame to car body frame. Here, which need to be solved simultaneously (since the equations
we can assume the two coordinates are aligned so FT L are not independent) for each axle in response to the exerted
and FLg can be assumed to be in car body frame. RLi is tyre force and torque. To this end, a Newton based nonlinear
the magnitude of the reaction force which acts at Qi in dˆi equation solver such as fsolve implemented in MATLABTM
direction and FP L is the magnitude of push rod force in dˆ6 can be used to determine the twenty six unknown variables.
direction acting on the rocker mounting point. This vector The result of solving these equations for the case where
equation provides three more scalar equations in terms of the right tyre is unloaded and the left tyre has only a vertical
left hand reaction force magnitudes. force applied is shown in Fig. 3. One can see as the load
Equations (9) contains six unknown force magnitudes but on the left tyre increases, the left wheel cambers and toes
only provides three scalar equations. In order to resolve the towards the car and also goes up in negative z direction.

167
−4
x 10
2.64 10 10
20
Kh=500 kN/m
9 5
2.62 15 K =750 kN/m
Kh=500 kN/m h

Tyre Vertical Displacement (mm)


8 Kh=1000 kN/m
0 10
2.6 K =1250 kN/m
h
Camber Angle (deg)

7
5

Toe Angle (deg)


−5 Kh=1500 kN/m

Ride Height (mm)


2.58 6
0 Kh=1500 kN/m
−10
2.56 5
−5
−15
4
−10
2.54
−20
3
−15
2.52 −25
2
−20

2.5 1 −30
−15 −10 −5 0 −15 −10 −5 0 −15 −10 −5 0 −25
Tyre Force (kN) Tyre Force (kN) Tyre Force (kN)
−30
−7 −6 −5 −4 −3 −2 −1 0
Vertical Tire Forces (kN)
Fig. 3: Variations in the front-left wheel carrier camber
angle, toe angle, and the tyre ground contact point vertical
displacement as a function of tyre vertical load. Fig. 4: Front axle height ride in mm for different values of
heave spring stiffness as the tyre loads are increased equally.

A. Suspension Meta-model
Including a suspension model of the type described so far for a range of left and right tyre forces and a set of varying
would significantly increase the solution times in context of spring stiffnesses. The solution dataset can then be fitted
optimal control calculations. Therefore we will instead use to a multivariate polynomial using a linear least squares
a meta-model to replace the full model by making use of algorithm [14]. Table I gives the coefficients associated with
multivariate polynomial functions. A map of this kind will the multivariate polynomial descriptions of the front and rear
keep the solution accurate to a high degree whilst decreasing axle ride height and roll angles.
calculation times. Even though the polynomial is of quadratic order, the
The most important factor which influences the suspension maximum error for the front axle ride height across all mesh
movements of a Formula One car is the vertical tyre force. points was calculated to be 0.44 mm and 0.047o for the roll
In order to simplify matters further, the tyre moments MT L angle. Similar results were obtained for the rear axle.
and MT R , along with the x and y-axis components of FT L
and FT R are set to zero. Fig. 4 shows the front ride heights III. C AR M ODEL
when equal vertical tyre forces are applied on each wheel
A vehicle model with yaw, lateral and longitudinal degrees
for a range of heave spring stiffness values. It can be seen
of freedom based on [8] will be used here. This car model
that the ride height increases (in SAE coordinates) with tyre
includes non-linear tyres and a limited slip differential.
load, bringing the car closer to the road. The static front ride
However, instead of using constant drag and downforce
height of the vehicle is defined by hf 0 = −14 mm with a
coefficients, a more sophisticated aerodynamic map shown
rear ride height of hr0 = −74.1 mm.
in Table II will be used. This map is responsive to the front
and rear ride heights, side-slip (β), steering (δ) and roll (φ)
- 1 Fl Fr Kh Fl2
angles. The ride heights and roll angle are fed into this
hf -2.53e1 -3.70e0 -3.70e0 -1.19e-3 -1.07e-1
aerodynamic map by making use of the suspension meta-
φf 6.56e-16 8.57e-3 -8.57e-3 -6.2e-19 2.88e-4 model developed in Section II.
hr -9.07e1 -6.90e0 -6.90e0 -1.83e-2 -9.71e-2
φr -3.95e-16 1.11e-2 -1.11e-2 7.2e-19 6.88e-5 The front and rear aerodynamic downforce terms are given
by
- Fl Fr Fl Kh Fr2 Fr Kh Kh2 1
hf 1.22e-1 9.77e-4 -1.07e-1 9.77e-4 3.13e-6
f
Faz = CLf ρ A u2 , (16)
2
φf 6.49e-18 2.24e-7 -2.88e-4 -2.24e-7 1.8e-22
hr 9.43e-2 4.04e-3 -9.71e-2 4.04e-3 3.01e-5 r 1
Faz = CLr ρ A u2 (17)
φr 1.91e-18 2.28e-7 -6.88e-5 -2.28e-7 -3.5e-22 2
respectively and the drag force is given by
TABLE I: Multivariate polynomial coefficients of the front
1
and rear axle ride height and roll angle. The left and right Fax = − CD ρ A u2 . (18)
wheel normal loads Fl and Fr , respectively, are given in kN. 2
Heave spring stiffness must be given in kN/m. hf and hr are CD is the drag coefficient, while CLf and CLr are the front
in mm and φf and φr are in radians. and rear axle downforce coefficients respectively. The front
and rear downforce are applied at the centre of the front and
The ride height and roll angle of each axle can be solved rear axles. The drag has a negative sign, since it acts in the

168
negative x-axis direction and is applied at the vehicle’s mass control problem is then represented by an NLP problem
centre. whose cost function is obtained by approximating the cost
In [8] a suspension related roll moment distribution factor functional using LGR quadrature. This NLP problem is
was used to resolve the load transfer equations (see equation subject to the transcribed algebraic constraints and path
(26) in [8]). Since a suspension model was developed here constraints of the original problem imposed at the collocation
we will replace this equation with a constraint that keeps the points.
front and rear axle roll angles equal. This will ensure that the GPOPS-II uses a segment and polynomial degree adap-
tyre loads are distributed in accordance with the suspension tation scheme [10] so that error reduction can still be
specifications. This can be written as achieved in the presence of non-smooth problem features.
The transcribed NLP problem is typically very large but
φr = φf (19) sparse. The IPOPT [16] software library (based on interior
point methods) was used here to solve the NLP problem.
IV. O PTIMAL C ONTROL Automatic Differentiation was used to provide IPOPT with
fast and accurate first and second order derivatives.
In order to calculate the minimum lap time we will set
up an optimal control problem. The states of this problem V. R ESULTS
are the vehicle perpendicular distance to the centreline, the
A full lap simulation of Circuit de Catalunya-Barcelona
angle with respect to the track, the longitudinal and lateral
was considered for a car with the suspension meta-model
velocities, and the yaw rate. The controls are the vehicle
from Table I and the aero-map from Table II. A map of the
steering angle, the longitudinal slip and the normal load of
circuit together with the calculated racing line for the car in
the four wheels. The static parameters considered here are the
its baseline set-up is shown in Fig 5. The corner distances
suspension ride height and heave spring stiffness offsets. The
from the start-finish (SF) line are marked on the figure. A
aim is to find the controls and values of the static parameters
series of parameter optimisation cases were then studied with
which minimise the lap time given by
the results summarised in Table III.
Zsf
1000
J= Sf (s) ds (20) 3900
4050

s0 4000
3500
3650 4200
where Sf is the reciprocal of the vehicle speed and s0 and 800 3350
sf are the start and end distances. The problem is subject to
state dynamics of the form
dx(s) 600
= f (x(s), u(s)) (21)
ds
Y−position (m)

2750
given in Equations (15), (16), (20), (21) and (22) of [8] SF

and path constraints given in Equations (23), (24), (25), (34), 400
(35) of [8] and (19). The vehicle perpendicular distance to
the centreline is also bounded to not exceed the track width. 1650
The pseudospectral numerical optimal control solver 200 2420
GPOPS-II [15] based on the Legendre-Gauss-Radau (LGR)
collocation scheme was used here to solve the minimum lap
2250
time problem. In this scheme the states are approximated
0 2000
using Lagrange polynomials as basis functions. The state
dynamics are then transcribed into algebraic constraints using 725
1050
800
the derivative of the approximating polynomial. The optimal
−200
−500 0 500
X−position (m)
CLf CLr CD Input
1.2000 0.9494 0.9455 Nominal Value Fig. 5: Racing line with distances from Start/Finish (SF) line
10.0000 1.0000 0.1000 hf (m) in (m) for Circuit de Catalunya-Barcelona. The suspension
-5.0000 -22.2222 -0.1000 hr (m) and aerodynamic parameters are set to their base-case values.
0 -222.2222 0 h2r (m2 )
-5.1294 -10.2588 0 β 2 (rad2 ) In case A, the vehicle is considered in its baseline set-
-0.1641 -0.1641 0 δ 2 (rad2 ) up with no parameter optimisation. The calculated lap time
-8.2070 -32.8281 0 φ2 (rad2 ) was 84.59 s. In case B, the front and rear ride height were
allowed to have an offset of between ±10 mm and ±30 mm
TABLE II: Down force and drag coefficients map. respectively. The optimum set-up was shown to increase

169
100
the front and rear ride heights by 7 mm and 12.6 mm (in Baseline
Optimised
SAE coordinates this brings the car closer to the road). This Speed
80
resulted in a lap time reduction of 30 ms. 1.7
In case C, the front and rear heave springs were allowed

Front Downforce Coefficients


1.65
to vary between ±500 kN/m of their nominal value. The 60
1.6

Speed (m/s)
optimiser stiffened the rear axle by the maximum allowed
and softened the front by 500 kN/m. This is due to the 1.55 40

fact that the front lift coefficient increases linearly with ride 1.5
height (see Table II) and hence more downforce can be 20
1.45
generated by operating the front of the car closer to the road.
In all the cases investigated here, the front ride height of the 1.4
0
car was constrained to not exceed 15 mm. This prevented 1.35
the car from getting too close to the road surface which
20
could potentially cause excessive wear on the under-body 0 500 1000 1500 2000 2500 3000 3500 4000 4500
Distance from SF line (m)
plank of the car. This explains why the optimiser does not
increase the front ride height by the maximum allowed. The Fig. 6: Front downforce coefficients comparison for the
rear downforce coefficient is a quadratic function of the rear baseline car (red) and the vehicle with optimised ride height
height ride and by hardening the rear, the car is kept closer and heave spring stiffness offsets (case D, shown in blue).
to its optimum operating point for a larger proportion of the Speed of the baseline car is shown in black.
lap.
In case D, both the heave spring stiffnesses and ride
heights for front and rear axles are allowed to be optimised R EFERENCES
by the same amount as in case B and C. The results show
[1] W. F. Milliken and D. L. Milliken, Race car Vehicle dynamics. SAE,
that ride heights are increased by 9.3 mm and 7.7 mm at the 1995.
front and rear respectively and both axles are stiffened by [2] D. E. Kirk, Optimal Control Theory. Dover Publication, 1970.
the maximum value of 500 kN/m. The difference between [3] V. Cossalter, M. Da Lio, R. Lot, and L. Fabbri, “A general method
for evaluation of vehicle manoeuverability with special empashis on
the front downforce coefficients for the baseline and case motorcycles,” Vehicle System Dynamics, vol. 31, pp. 113–135, 1999.
D is shown in Fig. 6. The speed profile of the baseline [4] E. Bertolazzi, F. Biral, and M. Da Lio, “Symbolic-numeric indirect
car is shown on the same figure in black. One can see method for solving optimal control problems for large multibody
system,” Mutlibody Systems Dynamics, vol. 13, pp. 233–252, 2005.
that the optimised car benefits from higher front downforce [5] J. T.Betts, Practical Methods for Optimal Control and Estimation
coefficients for the majority of the lap. It is evident that Using Nonlinear Programming. Society for Industrial and Applied
at high speeds when downforce levels are significant, the Mathematics, 2009.
[6] P. J. Enright and B. A. Conway, “Discrete approximations to optimal
front ride height increases and supports further gains in Clf ’s. trajectories using direct transcription and nonlinear programming,”
The total reduction in lap time for case D was 60 ms. The Journal of Guidance, Control, and Dynamics, vol. 15, pp. 994–1002,
simulation time for case D was less than 10 minutes on an 1992.
[7] D.Casanova, “On minimum time vehicle manoeuvring: The theoret-
8 core 3.4 GHz machine. ical optimal lap,” Ph.D. dissertation, Cranfield University, School of
Mechanical Engineering, 2000.
VI. C ONCLUSION [8] G. Perantoni and D. J. Limebeer, “Optimal control for a formula one
car with variable parameters,” Vehicle System Dynamics, vol. 52, pp.
The work presented here demonstrates how meta- 653–678, 2014.
modelling can be used to include sophisticated model be- [9] D. Limebeer and G. Perantoni, “Optimal control of formula one car
haviour into racing car dynamic lap time simulations without energy recovery systems,” International Journal of Control, 2014.
[10] C. L. Darby, W. W. Hager, and A. V. Rao, “An hp-adaptive pseu-
losing tractability. A suspension model was developed and dospectral method for solving optimal control problems,” OPTIMAL
a multivariate polynomial was used to capture its relevant CONTROL APPLICATIONS AND METHODS, vol. 32, pp. 476–502,
characteristics. A pseudospectral collocation method was 2011.
[11] H. M. Hiller and S. Frik, “Road vehicle benchmark 2 five-link
then used to simulate a full lap of the Barcelona circuit and suspension,” Vehicle System Dynamics, vol. 22, pp. 254–262, 1993.
time optimal values of heave spring stiffness and ride height [12] J. Knapczyk and S. Dzierżek, “Displacement and force analysis of five-
set-ups were found. rod suspension with flexible joints,” Journal of Mechanical Design,
vol. 117, no. 4, pp. 532–538, 1995.
[13] P. Simionescu and D. Beale, “Synthesis and analysis of the five-link
rear suspension system used in automobiles,” Mechanism and Machine
Case ∆RHf ∆RHr ∆Khf ∆Khr Lap Time
Theory, vol. 37, p. 815832, 2002.
A - - - - 84.59 [14] S. M.Kay, Fundamentals of Statistical Signal Processing: Estimation
B 7.0 12.6 - - 84.29 Theory. Prentice-Hall, 1993.
C - - -500 500 84.21 [15] M. A. Patterson and A. V. Rao, GPOPS - II Version 1.0: A General-
D 9.3 7.7 500 500 83.99 Purpose MATLAB Toolbox for Solving Optimal Control Problems
Using the Radau Pseudospectral Method, University of Florida, 2013.
[16] L. Biegler and V. Zavala, “Large-scale nonlinear programming using
TABLE III: Lap times for various optimisation cases. Ride ipopt: An integrating framework for enterprise-wide dynamic opti-
height offsets are in (mm), heave spring stiffness offsets are mization,” Computers and Chemical Engineering, vol. 33, pp. 575–
in (kN/m) and lap times are in (s). 582, 2008.

170

You might also like