University of Huddersfield Repository: Original Citation
University of Huddersfield Repository: Original Citation
University of Huddersfield Repository: Original Citation
Original Citation
Ward, C.P., Goodall, Roger M. and Dixon, R. (2011) Contact Force Estimation in the Railway
Vehicle Wheel-Rail Interface. In: 18th IFAC World Congress, 28th August - 2nd September 2011,
Milan, Italy.
• The authors, title and full bibliographic details is credited in any copy;
• A hyperlink and/or URL is included for the original metadata page; and
• The content is not changed in any way.
For more information, including our policy and submission procedure, please
contact the Repository Team at: [email protected].
http://eprints.hud.ac.uk/
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
Abstract: Increased patronage of railways in the UK in the past 20 years has put demands
on rolling stock to operate at peak availability with reduced time available for maintenance.
One possible tool to enable this is the use of real time fault detection and diagnosis on board
railway vehicles to detect faulty components and provide information about the current running
condition of the system. This paper discusses the development of one such technique for the
estimation of creep forces of the wheel-rail contact. Real time knowledge of which could be used
to predict wear of the wheel tread and rail head, predict the formation of rolling contact fatigue,
and identify any areas of low adhesion present on the network. The paper covers development
of a full vehicle nonlinear contact mechanics model, development of the Kalman-Bucy filter
estimation technique and how the technique will be developed and validated in the future.
of various creep forces at different parts on the vehicle IRF ψ̈RF =FLyRF RLxRF − FLxRF RLyRF
system, with the estimated lateral creep force of the trail- +FRyRF RRxRF − FRxRF RRyRF (6)
ing wheelset giving best correlation to the modelled creep +MsψRF + MgRF
force. Therefore this study is looking into the best candi-
date creep forces for detection on a more representative mRR ÿRR = FLyRR + FRyRR + FsyRR + FgRR (7)
system. IRR ψ̈RR =FLyRR RLxRR − FLxRR RLyRR
The current study considers a vehicle model with a full +FRyRR RRxRR − FRxRR RRyRR (8)
length body no longer constrained in yaw, plus two bogies +MsψRR + MgRR
with the accompanying four wheelsets, Figure 1. As with mF B ÿF B = − (FsyF F + FsyF R + FsyV F ) (9)
previous studies this model only considers the lateral and IF B ψ̈F B = −(MsψF F + MsψF R + MsyV F
yaw directions, as the longitudinal and vertical effects can (10)
be satisfactorily neglected, Wickens (2003). + L(FsyF F − FsyF R ))
mRB ÿRB = − (FsyRF + FsyRR + FsyV R ) (11)
IRB ψ̈RB = −(MsψRF + MsψRR + MsyV R
(12)
+ L(FsyRF − FsyRR ))
mV ÿV = FsyV F + FsyV R (13)
IV ψ̈V = MsψV F + MsψV R (14)
where Fijkl , Rijkl , Miψkl are the forces (creep, gravita-
tional and suspension), positions and moments, mkl is
the mass, Ikl is the moment of inertia, ykl is the lateral
position, ψkl is the yaw angle; where i =L(eft), R(ight),
s(uspension); j =x (longitudinal), y (lateral); k =F (ront
bogie), R(rear bogie), V (vehicle); l =F (ront wheelset),
R(rear wheelset), B(ogie)
The accompanying suspension forces and moments (for
small angles) for the primary and secondary suspension
are given by equations 15 to 26.
FsyF F =ky1 yBF + ky1 LψBF − ky1 yF F
(15)
+ fy1 ẏBF + fy1 Lψ̇BF − fy1 ẏF F
MsψF F = kψ1 (ψBF − ψF F ) + fψ1 (ψ̇BF − ψ̇F F ) (16)
FsyF R =ky1 yBF − ky1 LψBF − ky1 yF R
(17)
+ fy1 ẏBF − fy1 Lψ̇BF − fy1 ẏF R
MsψF R = kψ1 (ψBF − ψF R ) + fψ1 (ψ̇BF − ψ̇F R ) (18)
FsyRF =ky1 yBR + ky1 LψBR − ky1 yRF
(19)
+ fy1 ẏBR + fy1 Lψ̇BR − fy1 ẏRF
MsψRF = kψ1 (ψBR − ψRF ) + fψ1 (ψ̇BR − ψ̇RF ) (20)
FsyRR =ky1 yBR − ky1 LψBR − ky1 yRR
Fig. 1. Vehicle system plan view model (21)
+ fy1 ẏBR − fy1 Lψ̇BR − fy1 ẏRR
MsψRR = kψ1 (ψBR − ψRR ) + fψ1 (ψ̇BR − ψ̇RR ) (22)
A detailed description of the development of railway vehi- FsyV F = − ky2 yV − fy2 ẏV + ky2 yBF + fy2 ẏBF
cle dynamics can be found in Garg and Dukkipati (1984), (23)
with the rigid body dynamics for the simulation model − ky2 cψV − fy2 cψ̇V
given below by equations 1 to 14. These equations encom- MsψV F = − ky2 c2 ψV − fy2 c2 ψ̇V + ky2 cyV + fy2 cẏV
pass lateral and yaw dynamics for the four wheelsets, two (24)
− ky2 cyBF − fy2 cẏBF
bogies and the vehicle body.
FsyV R = − ky2 yV − fy2 ẏV + ky2 yBR + fy2 ẏBR
mF F ÿF F = FLyF F + FRyF F + FsyF F + FgF F (1) (25)
+ ky2 cψV + fy2 cψ̇V
IF F ψ̈F F =FLyF F RLxF F − FLxF F RLyF F MsψV R = − ky2 c2 ψV − fy2 c2 ψ̇V − ky2 cyV − fy2 cẏV
+FRyF F RRxF F − FRxF F RRyF F (2) (26)
− ky2 cyBR − fy2 cẏBR
+MsψF F + MgF F where kmn and fmn are the suspension stiffness and
mF R ÿF R = FLyF R + FRyF R + FsyF R + FgF R (3) damper coefficients; with m = y(lateral) or ψ(yaw); n =
1(primary suspension), 2(secondary suspension).
IF R ψ̈F R =FLyF R RLxF R − FLxF R RLyF R
+FRyF R RRxF R − FRxF R RRyF R (4) Creep forces fundamentally provide the guidance mech-
anism for the wheelsets. These forces are generated in
+MsψF R + MgF R
reaction to the creeps (or slips) in the rolling contact of the
mRF ÿRF = FLyRF + FRyRF + FsyRF + FgRF (5) wheel-rail interface in normal running, these are relative
4399
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
velocities of the wheel and the rail in the contact area and (1967), the initial slope of the creep curve varies with
are defined as different adhesion levels, Pearce and Rose (1985), Harrison
wi
si = , i = x, y (27) and McCanney (2002). Four levels of adhesion are defined
V in this study as dry, wet, low and very low conditions.
where V is the forward velocity of the wheelset, wi is the The accompanying constants are given in Table 1 and the
creep (slip) velocity in the relevant direction (where x is creep curves are given in Figure 2. This varying slope
longitudinal direction and y is lateral direction), where this means that different adhesion levels should be feasibly
is defined as detected without the system becoming saturated. The
wi = Vw − Vr , i = x, y (28) effect of varying the adhesion levels on the running system
where Vw is the velocity of the wheel through the contact is shown in Figure 3. This shows the sum of the lateral
patch, and Vr is the velocity of the rail through the contact creep forces and gravitational stiffnesses for the rear bogies
patch. Creep generation is a highly nonlinear process, front wheelset and demonstrates that for the same system
however effects of hysteresis can be ignored due to this disturbance (i.e. the lateral position of the track), the
being a single direction rolling contact. creep forces generated reduces as the friction levels reduce,
Normal practice for wheel-rail contact modelling is to lin- meaning that detection of changes of adhesion level is
earise the creep forces generated in the model based upon feasible in practice.
Kalker coefficients, Kalker (1967). Due to the importance
Model parameter Dry Wet Low Very Low
here of modelling the non-linear adhesion characteristics
kA 1.00 1.00 1.00 1.00
up to and beyond the creep saturation, use is made of the ks 0.40 0.40 0.40 0.40
contact force model developed in Polach (2005). This is µ0 0.55 0.30 0.06 0.03
essentially a practical curve fitting mechanism, where the A 0.40 0.40 0.40 0.40
creep force (excluding spin effects) are calculated as B 0.60 0.20 0.20 0.10
Table 1. Polach contact model parameters
2Qµ
F = + arctan (29)
π 1 + 2
where Q is the wheel load, with
2 Cπa2 b 0.5
= s (30) 0.45
DRY
3 Qµ WET
LOW
where C is the proportionality
coefficient of the contact 0.4 VERY LOW
sheer stiffness N/m3 . Kalker coefficients can be used for 0.35
COEFF. OF FRICTION
0.05
The forces Fx , Fy in the longitudinal and lateral directions
are 0
si 0 0.002 0.004 0.006
CREEPAGE
0.008 0.01
Fi = F , i = x, y (33)
s
and the adhesion coefficients Fig. 2. Creep curves for varying adhesion conditions
Fi
fi = , i = x, y (34)
Q
The friction coefficients rely upon the slip velocity, where 1500
µ = µ0 (1 − A) e−Bw + A (35)
Sum of Creep and Gravitational Forces (N)
where −1500
10 11 12 13 14 15
kA + kS Time (s)
k= (38)
2
Experimentation has shown that, contrary to expectation Fig. 3. Lateral creep forces for varying adhesion conditions
from theoretical models such as that of Kalker, Kalker
4400
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
3. CREEP FORCE ESTIMATION TECHNIQUE where K is the ‘Kalman gain’, P is the error covariance
which in this continuous case the initial conditions of which
This study applies a Kalman-Bucy filter technique (Kalman require setting.
(1960)) as developed for the half vehicle model in Charles The second part is updating the estimates. The estimated
et al. (2008), to the full vehicle model of the previous state and output are then calculated simultaneously as
section. As mentioned earlier this method attempts to
estimate the total reactive creep and gravitational forces ŷ = Ck x̂ + Dk u (53)
generated at the wheel rail interface. The size of these
creep forces can then be analysed to determine the level of
x̂˙ = Ak x̂ + Bk u + K (y − ŷ) (54)
adhesion present, and the impact of the rail vehicle upon
the track infrastructure. where ŷ is the estimated output and x̂ is the estimated
state.
The Kalman-Bucy filter model uses simplified versions of
the system equations 1 to 8. As previous studies high- In this example, it is assumed that there is no system
lighted, the detection technique used here can not dif- input, and the filter becomes output only, therefore Ck =
ferentiate between left and right creep forces, and the Dk = 0. Where the state vector is defined as
gravitational stiffness present. Therefore these are brought
together as one state to be estimated, these are now x = [yF F ẏF F ψF F ψ̇F F yF R ẏF R ψF R ψ̇F R · · ·
mF F ÿF F = FF F + FsyF F (39) · · · yBF ẏBF ψBF ψ̇BF yRF ẏRF · · ·
IF F ψ̈F F = MF F + MsψF F (40) · · · ψRF ψ̇RF yRR ẏRR ψRR ψ̇RR · · ·
(55)
mF R ÿF R = FF R + FsyF R (41) · · · yBR ẏBR ψBR ψ̇BR yV ẏV · · ·
IF R ψ̈F R = MF R + MsψF R (42) · · · ψV ψ̇V FF F FF R FRF FRR · · ·
mRF ÿRF = FRF + FsyRF (43) · · · MF F MF R MRF MRR ]T
IRF ψ̈RF = MRF + MsψRF (44) The primary tuning parameter here is the Q matrix,
mRR ÿRR = FRR + FsyRR (45) defined as a diagonal matrix, where
IRR ψ̈RR = MRR + MsψRR (46) Q = diag[1 1 1 1 1 1 1 1 1 1 1 1 1 1 · · ·
where for the purposes of the filter, the following assump- · · ·1 1 1 1 1 1 1 1 1 1 1 1 1 1 · · · (56)
tions are made
· · ·1e9 1e9 1e9 1e9 1e9 1e9 1e9 1e9 ]
ḞF F = ḞF R = ḞRF = ḞRR = 0 (47)
The high values associated with the last eight positions
ṀF F = ṀF R = ṀRF = ṀRR = 0 (48) in the matrix assign uncertainty to the assumptions of
For this proof of concept work it is assumed that all of equations 47 and 48, allowing the filter to adapt the state
the states can be measured. Practically this will not be estimates to the creep force levels.
possible, so later stages of the work will look into reducing Figure 4 show results of the filter output compared to the
the number of states to be measured, and gauging how modelled total lateral creep forces for the rear wheelset of
many can be estimated without detrimentally affecting the the front bogie, by inspection this shows good correlation
creep force estimation quality. The Kalman-Bucy filter, for the specific estimate. The simulation parameters are
Grewal and Andrews (2001), is a very well known state for dry friction and the model velocity is 20 m/s, with
space method, where the state equation is defined as suspension parameters given in Table 2. The coefficient of
ẋ = Ak x + Bk u + z (49) determination (R2 ) values for the four lateral creep forces
where x is the state vector, ẋ is the rate of change of the from the simulations with different levels of adhesion are
state vector, z is the Gaussian noise source on each of the shown in Table 3, where this is calculated as
state vectors, Ak is the state matrix and Bk is the input σ 2 ((t))
matrix. The output equation for the system is defined as R2 = 1 − (57)
σ 2 (y(t))
y = Ck x + Dk u + v (50)
where σ 2 ((t)) is the variance of the residuals and σ 2 (y(t))
where y is the output vector, v is the Gaussian noise on
is the variance of the output, Ljung (1999). This shows
the output vector, Ck is the output matrix and Dk is the
that there is some reduction in the quality of the creep
input matrix.
force estimation as the friction level reduces. This is due to
Design choices are made by selecting covariance matrices the nonlinear effects of the creep force saturation becoming
of the state and the input. These define how much noise more prominent as the adhesion level reduces. The best
is present in the system, therefore affecting the output of estimates are achieved overall are for the front bogie rear
the system due to the model measurements not exactly wheelset creep forces. The estimator can also be shown
matching the system. to track adhesion changes in real time, Figure 5. This
Figure demonstrates a simple interpolation scheme of the
The filter algorithm can be split into two sections. The first adhesion conditions in the simulation model, and how the
section calculates how much to adapt the filter to changes estimator successfully track the changes to the creep force
in the system. level. This is the early stages of this investigation and
T −1
K = P CK R (51) further work will look into how these estimates are affected
Ṗ = Ak P + P ATk − KRK T + Q (52) by removing some of the measured signals.
4401
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
Adhesion level
Parameter Description Value Units dry
fy1 primary lateral damper 0 Ns/m
Adhesion condition
CoE wet
fψ1 primary yaw damper 0 Nms
CoE low
fy2 secondary lateral 0.06e6 Ns/m
damper CoE vlow
0 5 10 15 20 25 30
Ib bogie yaw inertia 3500 kgm2 Time (sec)
Iv vehicle yaw inertia 30000 kgm2 x 10
4 Creep force and gravitational stiffness
1
Iw wheelset yaw inertia 700 kgm2
4402
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
and yaw dynamics of the system. This was performed Polach, O. (2005). Creep forces in simulations of traction
as previous studies highlighted that some of the lateral vehicles running on adhesion limit. Wear, 258(1), 992–
creep forces and creep moments gave a better estimation 1000.
than others, and this needed to be investigated in a more Vasic, G., Franklin, F., and Kapoor, A. (2003). New
representative model of a rail vehicle system. rail materials and coatings. Technical Report
RRUK/A2/1, University of Sheffield, prepared for
It then discussed the Kalman-Bucy filter method used
RSSB. http://portal.railresearch.org.uk/RRUK/Shared
to estimate the lateral creep forces and the yaw creep
%20Documents/rssba2a.pdf.
moments of the four wheelset in the model, and how these
Ward, C., Goodall, R., and Dixon, R. (2010a). Wheel-
have given successful estimations in simulations.
rail profile condition monitoring. In Proceedings of the
Finally the future direction of the project was highlighted UKACC control conference, Coventry.
with further development potential of the algorithm given, Ward, C., Weston, P., Stewart, E., Li, H., Goodall,
and an explanation of the methods of validation that R., Roberts, C., Mei, T., Charles, G., and Dixon, R.
will be used, by data generated in multi-bodied dynamics (2010b). Condition monitoring opportunities using ve-
simulation software and through track testing. Suggestions hicle based sensors. In press: IMechE proceedings, Part
were also given for other applications of the creep force F: Rail and Rapid Transit.
estimations. Wickens, A. (2003). Fundamentals of Rail Vehicle Dynam-
ics: Guidance and Stability. Swets and Zeitlinger, first
ACKNOWLEDGEMENTS edition.
Xia, F., Cole, C., and Wolfs, P. (2008). Grey box-based
The authors would like to thank Rail Research United inverse wagon model to predict wheel-rail contact forces
Kingdom (RRUK) and the Engineering and Physical Sci- from measured wagon body responses. Vehicle System
ence Research Council (EPSRC) who funded this research. Dynamics, 46(Supplement), 469–479.
REFERENCES
Bombardier (2010). Orbita - predictive asset
management, the future of fleet maintenance.
http://www.bombardier.com/en/transportation/
accessed 7th April 2010.
Charles, G., Goodall, R., and Dixon, R. (2008). Model-
based condition monitoring at the wheel-rail interface.
Vehicle System Dynamics, 46(1), 415–430.
Garg, V. and Dukkipati, R. (1984). Dynamics of Railway
Vehicle Systems. Academic Press, first edition.
Grewal, M. and Andrews, A. (2001). Kalman Filter-
ing: Theory and Practice Using MATLAB. Wiley-
Interscience Publications, second edition.
Harrison, H. and McCanney, T. (2002). Recent devel-
opments in coefficient of friction measurements at the
rail/wheel interface. Wear, 253(1), 114–123.
Hussain, I. and Mei, T. (2010). Multi kalman filtering
approach for estimation of wheel-rail contact condi-
tions. In Proceedings of the UKACC control conference,
Coventry.
Kalker, J. (1967). On the Rolling Contact of Two Elastic
Bodies in the Presence of Dry Friction. Ph.D. thesis,
Delft University of Technology, Delft, Netherlands.
Kalman, R. (1960). A new approach to linear filtering and
prediction. Transactions of ASME - Journal of Basic
Engineering, 35–45.
Li, P., Goodall, R., Weston, P., Ling, C., Goodman,
C., and Roberts, C. (2006). Estimation of railway
vehicle suspension parameters for condition monitoring.
Control Engineering Practice, 15:43-55.
Ljung, A. (1999). System Identification, Theory for the
User. Prentice Hall, second edition.
Mei, T. and Li, H. (2008). Measurement of vehicle ground
speed using bogie based inertial sensors. IMECHE
proceedings, Part F - Rail and Rapid Transit, 222(2),
107–116.
Pearce, T. and Rose, K. (1985). Measured force-creep rela-
tionships and their use in vehicle response calculations.
In Proceedings of the IAVSD 9th Symposium, Linkoping.
4403