Black-Box Modelling of Nonlinear Railway Vehicle Dynamics For Track Geometry Assessment Using Neural Networks
Black-Box Modelling of Nonlinear Railway Vehicle Dynamics For Track Geometry Assessment Using Neural Networks
Black-Box Modelling of Nonlinear Railway Vehicle Dynamics For Track Geometry Assessment Using Neural Networks
To cite this article: Sönke Kraft, Julien Causse & Aurélie Martinez (2018): Black-box modelling of
nonlinear railway vehicle dynamics for track geometry assessment using neural networks, Vehicle
System Dynamics, DOI: 10.1080/00423114.2018.1497186
Lines, Track and Environement Departement, SNCF Réseau, la Plaine St Denis Cedex, France
• Integration in the track geometry recording car: For the track geometry assessment,
a fast availability of the simulation results is of high importance. The integration of
the simulation in the measurement chain of the recording car would therefore be
advantageous.
• Model accuracy: The vehicle response-based track geometry assessment is based on the
quantitative results of the simulation. It is indispensable to guarantee a high quality of
the models. Therefore, the validation of the models with respect to measured vehicle
responses gains importance.
The multi-body simulation, even though well established in railway industry and
research, responds only partly to the requirements of the vehicle response-based track
geometry assessment. An integration of the multi-body simulation into the measurement
chain of the track geometry recording car is complicated since several input files to the
simulation have to be handled and an adequate computational environment is required.
Besides, the computational cost of multi-body models can be high.
Taking into account these drawbacks this work studies the use of black-box models
for the simulation of the vehicle dynamics. Black box models are described by a parame-
terised mathematical structure without any physical meaning. They do not require a costly
numerical integration process but can be applied such as a filter. The integration in the
measurement chain of the recording car is therefore much simpler. However, the main
issue of black-box models is their accuracy and the simulation uncertainty. It depends on
the capacity of the model structure to reproduce the dynamics of the vehicle and the selec-
tion of adequate input and output data sets for the training of the black-box model. The
consideration of the large variability of operation conditions in railway dynamics as well
as nonlinear effect is therefore a crucial issue.
Figure 1. Nonlinearities in lateral vehicle dynamics: wheel-rail contact, bumpstops, nonlinear charac-
teristics of rubber and airsprings.
rail profiles are nonlinear and relative movements between the wheel and the rail lead to
friction forces.
It is therefore convenient to separate vertical and lateral dynamics when analysing the
responses of the vehicle. In vertical direction, the trajectory of the wheel depends directly
on the longitudinal level without introducing additional degrees of freedom. Nonlineari-
ties in the primary and secondary suspensions have in general small impact. This is not
the case for the lateral direction. The wheelset moves relative to the track and the resulting
slip leads to friction forces which depend non-linearly on the slip. The contact proper-
ties between wheel and rail can change discontinuously due to flange contact and multiple
contacts as illustrated in Figure 1. Besides, many suspension elements have nonlinear char-
acteristics in lateral direction. Typically this is the case for air springs, rubber springs and
bump-stops as described in [4]. The consideration of nonlinear effects is therefore indis-
pensable. A detailed description of the dynamics of railway vehicles can be found in [5]
and [6].
The nonlinearities of the vehicle-track system can be taken into account in the multi-
body model. Wheel-rail contact models [5] consider non-linear kinematics and nonlinear
relationship between slip and friction coefficients. Suspension elements can be represented
by detailed physical models.
Multi-body models have the advantage that they are valid over the complete range of
running conditions. The same model can be used at different speeds, track geometry qual-
ities and track design parameters. The validation from measurements indicates that good
model qualities can be obtained [7]. However, the use of multi-body models for the track
geometry assessment application reveals several drawbacks:
Based on these drawbacks, this work studies the use of alternative modelling approaches
for the track assessment process.
Figure 2. Analysis of nonlinearities from track geometry – vehicle response transfer functions: lon-
gitudinal level – vertical bogie acceleration for two sections 1: 10–15 km and 2: 50–55 km (a) in the
spectrogram of the transfer functions (b). Alignment – lateral bogie acceleration for two sections 1 and
2 (c) in the spectrogram of the transfer functions (d).
analysis of the SISO transfer functions in Figure 2, using spectrograms and local trans-
fer functions computed at different positions along the track, reveals significant non-
linear effects in lateral direction. While the vertical transfer functions show only small
variations (Figure 2(a,b)), lateral transfer functions change significantly along the track
(Figure 2(c,d)) indicating nonlinear behaviour.
The results of the analysis of the SISO transfer functions and their effect on the total vehi-
cle response are summarised in Figure 3, showing the impact of nonlinearities on lateral
vehicle dynamics.
Figure 3. Analysis of the SISO functions between track geometry parameters and vehicle responses and
their effect on the lateral (a) and vertical (b) bogie acceleration.
4.1.1. Input and output data for model training and validation
All black-box models require the definition of the input and output variables for the train-
ing and validation of the models. The excitation of the vehicle is given by the track design
parameters and the irregularities of the track geometry (Table 1). In order to remove the
effect of speed changes on the vehicle dynamics, the black-box models are trained and
validated for a defined constant speed.
For the vehicle responses, vertical and lateral bogie accelerations are considered. In
curves, the lateral accelerations are composed of two parts: the dynamic response of the
vehicle due to the excitation by track irregularities and the centrifugal accelerations which
are a function of the curvature, the cant and the vehicle speed as well as the static stiffness of
the suspension. They do not depend on the dynamics of the vehicle and can be computed
using either a linear transfer function model of order 1 by defining only the first order term
in the denominator polynomial (1 pole and no zeros) (Figure 4(a)), or the analytical for-
mula for centrifugal accelerations with a coefficient representing the static stiffness of the
suspension (Figure 4(b)).
Therefore, for the simulation of the dynamic response of the vehicle using black-box
models the quasi-static part (0–0.2 Hz) is removed.
For the output data, the results of a reference simulation obtained by a validated multi-
body model (configuration A) or measured vehicle responses (configuration B) can be used
as illustrated in Figure 5.
Figure 4. Simulation of the quasi-static lateral bogie acceleration using a first order transfer function (a)
and an analytical formula (b).
Figure 5. Training (a) and validation by computing the simulation error E (b) of the black-box model
using either a multi-body simulation [configuration A] or measured vehicle responses [configuration B]
as reference.
are perfectly synchronised and can be used directly for training and validation of the black-
box model. Furthermore, the simulation can be performed easily for different operation
conditions thus allowing the creation of representative training and validation data.
Configuration B: Training using measured vehicle responses: The use of measured vehi-
cle responses is advantageous. First, the construction of a multi-body model is not
required. Secondly, the black-box model is trained on measured vehicle data without
adding additional uncertainty due to the multi-body model. However, often measured
vehicle responses are not available for all running conditions of interest. Besides, for the
training of the black-box model, the spatial synchronisation between the track geometry
data and the vehicle response has to be very accurate.
The synchronisation, illustrated in Figure 6, takes advantage of the excellent correlation
between the vertical track irregularities (longitudinal level) and the vertical displacements
of the wheelsets and bogies obtained by integrating twice the measured accelerations.
Using the cross-correlation function, the vehicle responses and the track geometry can
be synchronised with high precision.
VEHICLE SYSTEM DYNAMICS 9
Figure 6. Synchronisation process between the track geometry and the measured accelerations (a) and
a section of the longitudinal level and the synchronised axle-box displacements (b).
For a good model performance, the training data have to be representative of the
running conditions which the vehicle will meet under operation. The dynamic vehicle
responses depend on the track irregularities and the curvature and cant combinations.
The latter determine the uncompensated lateral accelerations and the running behaviour
in curves. The contact in the wheel flange and large displacements in the suspension can
lead to nonlinear effects.
The training data should therefore include all relevant curvature, cant and track quality
combinations. In practice this is difficult, in particular for vehicles running on many con-
ventional lines, since the variability of the track conditions is very high: Figure 7(a) shows
the distribution of curvature and cant combinations for 23 line sections of the French net-
work computed per section of 200 m length. In Figure 7(b) the cumulative distributions of
the maximal values for alignment are compared for these line sections.
10 S. KRAFT ET AL.
Figure 7. Combinations of curvature and cant per section of 200 m for 23 different conventional
lines (a), distribution of standard deviations of alignment per section of 200 m (b) and curvature/cant
combinations of high-speed lines (c).
In return, on high-speed lines (Figure 7(c)) the variability of the track design values and
the track quality level is much smaller, making the selection of representative training data
less laborious.
Transfer function (TF): When computing the zero points of the numerator and denom-
inator polynomials of the ARX model, the transfer function can be expressed by the poles
and zeros.
• Open-loop parallel training: Since the vehicle responses are available during the training
process, the feedback of the neural network output can be replaced by the known vehicle
response. This allows a fast training using back-propagation.
• Two-step training using open-loop and closed-loop configurations: The training of the
recurrent network is performed in two steps as illustrated in Figure 8. In order to ini-
tialise the network parameters the fast parallel training is used first. Then the open-loop
network is transformed in the closed-loop configuration followed by a second training.
This approach gives better results in some cases but leads to higher computational cost.
12 S. KRAFT ET AL.
Figure 8. Model initialisation and training using combined parallel training (open-loop) and series
parallel training (closed-loop).
Figure 9. Two-step modelling: linear transfer function model as additional input to the nonlinear
recurrent neural network model.
4.2.3. Two-step modelling: linear transfer function and nonlinear neural network
model
If the system behaviour is characterised by a combination of linear and nonlinear operation
ranges, the use of a two-step modelling approach can be advantageous [10]. In the first
step, a linear transfer function model is optimised from the input-output data set. The
modelling takes advantage of the good training properties of linear systems. In order to
model nonlinear operation ranges, the simulated response of the linear model is used as an
additional input to the non-linear recurrent neural network model as illustrated in Figure 9.
The performance of the combined model is at least as good as the performance of the linear
model. The nonlinear recurrent neural network can be trained specifically for nonlinear
operation conditions.
Figure 10. Training and validation section of the black-box model for one line.
• Vehicle type: high-speed train with two traction units and eight coaches, three-car
regional trainset,
• Line type: high-speed line, conventional lines with different speed classes,
• Running conditions: straight track, curved track.
A representative section of the line is used for training. The validation of the identified
black-box model is realised on the complete line as illustrated in Figure 10.
The performance of the black-box model is evaluated using several criteria:
Transfer function model: For the simulation of the vertical bogie acceleration, very good
model performances are obtained for all vehicles. The validation for the regional train-
set on the conventional line in Figure 11(a) shows an excellent correlation of the time
14 S. KRAFT ET AL.
Figure 11. Transfer function model for simulating the vertical bogie acceleration. Comparison of time
signals (a), misfit function values per section with curvature (b) and correlation of maximal values
per section transfer function – multi-body model (c) for a regional train on a conventional line at V
= 160 km/h.
data signals leading to a mean misfit function value of 36%. The analysis of the mis-
fit function per section together with the curvature in Figure 11(b) and the correlation
between maximal values per section obtained for the transfer function and the multi-
body model in Figure 11(c) indicates that the model performance does not depend on
the running conditions.
However, for the simulation of the lateral bogie acceleration the transfer function model
gives insufficient results. Figure 12 shows the validation for the high-speed train on the con-
ventional line. While good model performances are obtained on straight track, the vehicle
response is strongly underestimated in curves.
The simulation error of the linear model serves as an indicator of the degree of nonlin-
earity of the system. The misfit function as a function of the curvature reveals nonlinear
behaviour in curves (Figure 12(c)). Furthermore, the correlation analysis between max-
imal values per section obtained from the transfer function and the multi-body model
(Figure 12(d)) shows that the simulation error increases at high vehicle response ampli-
tudes. As curves, they lead to nonlinear effects in the vehicle behaviour.
Wiener Hammerstein: The validation of the Wiener Hammerstein model for the high-
speed train on the conventional line indicates a significant degradation of the model
performance in curves. Misfit function values between 100% and 500% are observed in
VEHICLE SYSTEM DYNAMICS 15
Figure 12. Transfer function model for simulating the lateral bogie acceleration in a high-speed train on
a conventional line at 160 km/h. Comparison with the multi-body simulation on straight track and curve
(a), misfit function values per section and curvature (b), misfit function as a function of the curvature (c)
and correlation between maximum values transfer function – multi-body model (d).
Figure 13. Wiener–Hammerstein model: Misfit function and curvature for the simulation of the lateral
bogie acceleration on a conventional line at 160 km/h using different static nonlinearity models (a) and
misfit function and curvature for the simulation of the lateral acceleration on a high-speed line (b).
Figure 13(a). The application of different static nonlinearity types (piece-wise, sigimoid,
etc.) in the Wiener–Hammerstein models does not lead to an improvement of the model
accuracy in curve.
In return, when applied to the high-speed line at 300 km/h, good model performances
are obtained with the multi-body model as reference (Figure 13(b)). This improvement can
16 S. KRAFT ET AL.
Figure 14. Recurrent neural network (NARX) model: Identification of an adequate model structure by
parametric analysis (a), comparison of the misfit function for Wiener-Hammerstein and neural network
models for the lateral bogie acceleration of a high-speed train on a conventional line (b) and misfit
function for a neural network model of a high-speed train on a high-speed line (c).
be explained with the larger radii and better track quality therefore reducing the effects of
nonlinearities in curves.
Recurrent neural network (NARX): The optimal structure of the recurrent neural net-
work is identified using a parametric analysis. As illustrated in Figure 14(a) the global misfit
function is computed while increasing the numbers of neurons, layers and delays. Each
configuration is trained five times, considering the effect of different model initialisation
on the local optimisation.
The neural network model of the high-speed train on the conventional line is vali-
dated and compared to the Wiener–Hammerstein models (Figure 14(b)). A significantly
improved model performance in curve is obtained. However, for some curve sections
underestimations of the vehicle response with misfit function values up to 200% are
observed. Interestingly, for linear operation conditions on straight track the performance of
the neural network is very good (Misfit smaller than 50%) but worse than the performance
of linear models.
The validation of the neural network model of the high-speed train on the high-
speed line in Figure 14(c) indicates a very good model performance in straight track
and curve.
VEHICLE SYSTEM DYNAMICS 17
Figure 15. Two-step modelling approach: Misfit function per section and curvature (a) and misfit func-
tion as a function of the curvature (b) for the linear transfer function model, the recurrent neural network
and the two-step modelling approach for simulating the high-speed train on the conventional line.
Two-step modelling: Transfer function + Recurrent neural network: The good perfor-
mances of linear models on straight track and of recurrent neural networks in curve and at
high vehicle response amplitudes are combined by using the two-step modelling approach.
Figure 15 shows the misfit function for simulating the high-speed train on the conventional
line computed per section and as a function of the curvature for the linear transfer func-
tion model, the recurrent neural network and the two-step modelling approach combining
them both. The best performance is obtained using the two-step modelling.
Figure 16. Recurrent neural network NARX: Validation for a high-speed train on a high speed line (a),
simulation and measurement time signals (b), correlation of simulated and measured maximal values
per section (c) and misfit function as a function of curvature (d).
The performances of the analysed black-box models are summarised in Table 2 for the
high-speed train. It gives the value of the mean-square misfit function over the complete
validation section (mean) and the 99% percentile of the distribution of the misfit function
values per 300 m section (max).
Among the analysed nonlinear models, the two-step modelling approach using recur-
rent neural networks combined with linear transfer function models gives the best
Table 2. Comparison of black-box model performances for the simulation of a high-speed train on a
conventional and a high-speed line.
Model validation: Mean-square misfit function [%]
Type Model mean max (99%) mean max (99%) mean max (99%)
Linear Transfer Function (TF) 64 757 21 112 21 101
Nonlinear Wiener–Hammerstein 59 460 18 118 19 83
NARX network 73 192 36 112 26 98
Linear + nonlinear TF + NARX network 46 136 24 92 23 86
VEHICLE SYSTEM DYNAMICS 19
performances. However, it is observed that the model performance degrades in curve and
at high response amplitudes. Section 5 therefore focuses on the training on representative
track data and the consideration of strongly varying running conditions.
and the network weight and delays are optimised in parallel using global optimisation.
For this work, the simulated annealing algorithm has been applied to the model train-
ing and the performance has been compared with the results obtained from repeated
local training algorithms. The simulated annealing allows obtaining the best result of
5 local trainings using only one optimisation. However, the computation time until the
convergence increases significantly.
• Simulation performance in curves and transitions: The high variability of the curva-
ture/cant combinations in curves in combination with nonlinear effects of the vehicle
leads to a degradation of the model performance in curve.
• Simulation performance at high-amplitude vehicle responses: For the track geometry
assessment, the high-amplitude responses of the vehicle are of principal interest. They
represent a risk for safety and comfort and shall be modelled with a good accuracy. How-
ever, due to their rare appearance, high-amplitude responses can be underrepresented
or even missing in the training data and consequently be prone to important simulation
errors.
• Training based on measurement and multi-body simulation data: The use of measured
acceleration for the training and validation on a high-speed line showed very good
results in Section 4. It is therefore interesting to compare the performances of the neural
network and the multi-body model with respect to measurements.
Curves Curves represent a principal source for nonlinear behaviour and degraded simula-
tion performance. Therefore the effect of varying cant values at constant curvature has been
studied. The neural network is trained on a curve with a given curvature-cant combination
using real track geometry data and validated on the same track section. The obtained model
performance is compared to the validation results obtained on the same track section but
with modified cant values.
The results in Figure 17 indicate that the correct representation of all existing curvature-
cant configurations is an important requirement for good model performance. While a
sufficient model performance is obtained for the original curvature-cant configuration,
the model performance degrades significantly for increased or decreased cant values. This
might be explained with the uncompensated lateral accelerations which depend directly
on the cant deficiency and determine the dynamic behaviour in curve. In particular, the
change between positive and negative cant deficiency leads to the different guidance of the
wheelset in curve.
High amplitude responses In order to improve the simulation results at high amplitude
responses, track sections with high response amplitudes are artificially over-represented in
the training data. After the classification of all vehicle responses according to their ampli-
tude and the corresponding curvature (Figure 18(a)), the sections including high ampli-
tude and/or high curvature values are selected for the model training. The comparison of
VEHICLE SYSTEM DYNAMICS 21
Figure 17. Validation of a neuronal network model on a curve with the real cant value used for training
and increased cant values.
the selected training sections (Figure 18(b)) with the distribution of the curvature/cant
(Figure 18(c)) and the curvature/vehicle amplitude (Figure 18(d)) combinations of the
complete data set allows verifying the selection. All curvature-cant combinations of the
line should be represented and high amplitude responses emphasised.
Additionally, the training data is weighted as a function of the vehicle response ampli-
tude using a normalised amplification factor.
Based on the requirements outlined above, different modelling approaches have been
applied. They are summarised in Table 3 and outlined in the following. The aim is to obtain
optimal model performance at all operation conditions. Besides, the complexity and the
computational cost of the model training and the simulations should be as small as possible.
Figure 18. Selection of training data based on the defect amplitude and the curvature (a), selected
training sections (b), validation and estimation data for the cant (c) and the acceleration amplitudes (d)
as a function of the curvature.
Table 3. Analysed approaches for the training and application of neural network models.
Approaches for the training and application of neural network models
Training
Number of conditions
Type models Curvature Cant Track quality Remark
General model 1 model random random random simple simulation,
difficult training due
to highvariability of
conditions
weighted random weighted
Multi-model n model defined range random random less variability in training data,
complex simulation
defined range defined range random high number of models,
difficulty to create training
data
Line-model 1 model per line data line data current quality all design conditions covered,
line(group) limited track quality range
weighted line weighted line weighted all design conditions covered,
data data quality weighting of high amplitude
defects and curvature
degraded training on artificially degraded
quality track quality
VEHICLE SYSTEM DYNAMICS 23
Figure 19. Multi-model approach using several models as a function of the curve radius.
several neuronal network models are applied in parallel as a function of the curvature and
cant parameters. The approach is illustrated in Figure 19 for a curvature-triggered multi-
model. Since the curve radius is directly available during the measurement run, a switch
allows applying the adequate neural network model at any time.
As outlined before, the consideration of the curvature for model training is not suffi-
cient. The simulation requires representative curvature-cant combinations. Considering
the large variability of cant at constant curvature on conventional lines, extensive training
data is required.
Table 4. Training and validation of neuronal networks for good, degraded and
artificially degraded track geometry qualities.
Track quality effect on training and validation (misfit function [%])
Validation
Figure 20. Power spectral densities of the track geometry (cross level) before maintenance, after
maintenance and for the artificially degraded track.
5.3. Results
The proposed modelling approaches have been applied to a representative number of lines
including both high-speed and conventional lines. The simulations are performed for the
VEHICLE SYSTEM DYNAMICS 25
Figure 21. Curvature – cant characteristics for three conventional lines (a), Validation on line 1 for the
line-specific model (mean misfit 32%) and the General model (mean misfit 40%) (b) of the high-speed
train.
lateral bogie accelerations of a high-speed train, a regional trainset and an electric 4-axle
locomotive.
On conventional lines the training and validation of the neural network models is
performed using the results of multi-body simulations. For high-speed lines the neural
network model has been trained using measured accelerations.
Multi-model: Neural network models (NARX) have been trained at 160 km/h for specific
curvature ranges using data of several lines. The validation results for these models are
compared with the performance of the Line-model on line 1 in Figure 22 for the curvature
ranges 0.45–0.55 and 0.9–1.1 1/km.
As expected, the curvature-specific model gives the best performance at the curvature
for which it has been trained. However, the model performance cannot be improved further
with respect to the Line-model. This might be explained by the relatively small amount of
training data at specific curvatures used until now and the fact that the distribution of the
cant is not considered.
26 S. KRAFT ET AL.
Figure 22. Model performance (misfit function) as a function of the curvature. Comparison of the line
model and curvature-specific model (0.5 and 1 1/km) for the high-speed train.
Line-model: The comparison of the modelling process and the validation results reveals
that line-specific models are the most convenient approach. Considering the different track
design and track quality characteristics of conventional lines, line models allow the fast
training based on relative short but representative training data sets of one line. Besides,
since track data is stocked and assessed with respect to the line, the use of Line-models is
convenient from the practical point of view.
Figure 23 shows the validation results for a two-step Line-model (linear transfer func-
tion + neural network NARX) of the high-speed train, the regional train and the electric
locomotive at 160 km/h. The best simulation performance is obtained for the high-speed
train. The regional train is still acceptable while for the locomotive misfit function val-
ues up to 400% are obtained indicating insufficient model quality. These differences can
be explained by the different mechanical structures of the vehicles and the effect of
nonlinearities in the model.
Figure 23. Line-model for three vehicles (high-speed train, regional train and electrical locomotive).
Misfit function per section with curvature (a) and misfit function as a function of the curvature (b).
VEHICLE SYSTEM DYNAMICS 27
• High-speed lines are characterised by much larger curve radii and good track quality.
The operation conditions are therefore less variable and consequently nonlinearities of
the vehicle less emphasised.
• The neural network model can be trained and validated from measured acceleration
data. This has several advantages: The creation of the neural network model does not
Figure 24. Comparison of the measured and simulated lateral bogie accelerations per section for the
multi-body and the neural network model (a) and cumulative distribution of the misfit function values
(b) and relative amplitude errors for responses above the threshold (c).
28 S. KRAFT ET AL.
require a validated multi-body model for creating the training and validation data.
Additional uncertainty due to the multi-body model is avoided. By using only neu-
ral network models, the modelling and simulation process on high-speed lines can
be significantly simplified. Finally, it is possible to validate both the multi-body and
the neural network model with respect to the measurement and to compare their
performance.
Two-step neural network models (linear transfer function model + recurrent neu-
ral network) have been trained for 300 km/h on several high-speed lines. The valida-
tion with respect to measured lateral bogie accelerations gives very good results on all
lines. Furthermore, the comparison of the performances of the neural network model
and the multi-body model in Figure 24 reveals, that better results can be obtained
with the neural network. Both models are compared section-wise with the measured
lateral bogie acceleration. For each section, the mean square misfit function, the coher-
ence function and the power spectral density (PSD) are computed (Figure 24(a)). For
a generalised comparison, the cumulative distribution of the misfit function values of
all sections is computed in Figure 24(b). Better results are observed for the neural
network model.
In the vehicle response-based track assessment the exceedance of vehicle response
thresholds is used as a criterion for the track defect identification. The simulation accu-
racy at high response amplitudes is of particular interest. Therefore the relative amplitude
error EA rel = (Asimu /Ameas )% between the simulation Asimu and the measurement Ameas
has been computed for vehicle responses above a threshold which corresponds to the 95%
percentile value of the distribution of all vehicle responses. The comparison of the error
distributions for the multi-body model and the neural network model in Figure 24(c)
indicates better results for the neural network model.
the latter. Considering this result as well as the simple model creation based on train-
ing from measurement data and the fast simulation, the black-box modelling approach
is advantageous.
However, the results also show that the performance of the black-box model depends
significantly on the modelled vehicle, the characteristics of the line and the available train-
ing data. The good simulation performances on high-speed lines are not obtained on
conventional lines with smaller radii and much higher variability in track characteristics.
Nonlinear effects in curve and at high vehicle response amplitudes lead to a degradation
of the simulation result. Due to insufficient measurement data, the training has to be per-
formed based on multi-body simulation results. Some vehicles as the electrical locomotive
are characterised by highly nonlinear behaviour in curves. Therefore on some conventional
lines multi-body models are preferable. Future work will aim at improving the black-
box modelling on conventional lines by realising sufficient and precise measurements on
different vehicles which allow successful training of the black box models.
Disclosure statement
No potential conflict of interest was reported by the authors.
References
[1] Coudert F, Causse J, Kraft S. Optimization of track maintenance standards based on vehicle
simulation. Proceedings of the 10th World Congress on Railway Research; 2013 Nov 24–27;
Sydney, Australia; 2013.
[2] Dynotrain. D2.6 – Final report on track geometry, WP2 – Track Geometry Quality, Dynotrain
project; 2013.
[3] CEN. EN 13848-1:2003+A1:2008, Railway applications – Track – Track geometry quality –
Characterisation of track geometry, CEN/TC 256; 2008.
[4] Bruni S, Vinolas J, Berg M, et al. Modelling of suspension components in a rail vehicle dynamics
context. Veh Syst Dyn. 2011;49(7):1021–1072.
[5] Iwnicki S. Handbook of railway vehicle dynamics. Boca Raton: CRC Press; 2006.
[6] Knothe M, Stichel A. Schienenfahrzeugdynamik. Berlin: Springer Verlag; 2003. German.
[7] Kraft S, Causse J, Coudert F. An approach for the validation of railway vehicle models based
on on-track measurements. Veh Syst Dyn. 2014;52(Supp 1):125–141.
[8] Ljung L. System identification: theory for the user. 2nd ed. Upper Saddle River (NJ): Prentice
Hall PTR; 1999.
[9] Walter E, Pronzato L. Identification of parametric models from experimental data. London:
Springer Verlag; 1997.
[10] Sjoberg S, Zhang Q, Ljung L, et al. Nonlinear black-box modeling in system identification: a
unified overview. Automatica. 1995;31(12):1691–1724.
[11] Luber B, Haigermoser A, Grabner G. Track geometry evaluation method based on vehicle
response prediction. Veh Syst Dyn. 2010;48(Suppl 2010):157–173.
[12] Mathworks. Toolbox system identification – user’s guide. Natick (MA): Mathworks; 2013.
[13] Dreyfus G. Neural networks. Berlin: Springer Verlag; 2005.
[14] Galushkin A. Neural networks theory. Berlin: Springer Verlag; 2007.
[15] van de Ven PWJ, Johansen TA, Sorensen AJ, et al. Neural network augmented identification of
underwater vehicle models. Control Eng Pract. 2007;15(6):715–725.
[16] Matusko J, Petrovic I, Peric N. Neural network based tire/road friction force estimation. Eng
Appl Artif Intell. 2008;21:442–456.
[17] Tan Y, Saif M. Neural-networks-based nonlinear dynamic modeling for automotive engines.
Neurocomputing. 2000;30:129–142.
30 S. KRAFT ET AL.
[18] Li D, Meddah A, Hass K. Relating track geometry to vehicle performance using neural network
approach. Proc Inst Mech Eng J Rail Rapid Transit. 2006;220(3):273–281.
[19] Nefti S, Oussalah M. A neural network approach for railway safety prediction. IEEE Inter Conf
Syst Man Cybern. 2004;3915–3920.
[20] Iwnicki S, Parkinson H, Stow J. Assessing railway vehicle derailment potential using Neural
Networks. The Rail Technology Unit, Manchester Metropolitan University; 2016.
[21] Mathworks. Toolbox neural networks – user’s guide. Natick (MA): Mathworks; 2013.
[22] Reed R. Pruning algorithms – a survey. IEEE Trans Neural Netw. 1993;4(5):740–747.
[23] Castellano G, Fanelli AM, Pelillo M. An iterative pruning algorithm for feedforward neural
networks. IEEE Trans Neural Netw. 1997;8(3):519–531.
[24] Loghmanian SMR, Jamaluddin H, Ahmad R, et al. Structure optimization of neural network
for dynamic system modelling using multi-objective genetic algorithm. Neural Comput Appl.
2012;21(6):1281–1295.
[25] Deb K., Pratap A., Agarwal S., et al. A fast and elitist multiobjective genetic algorithm: NSAG-II.
IEEE Trans Evol Comput. 2002;6(2):182–197.
[26] Paul S. Novel hybrid compact genetic algorithm for simultaneous structure and parame-
ter learning of neural networks. Proceedings of the 2012 IEEE Congress on Evolutionary
Computation; 2012 Jun 10–15; Brisbane, QLD, Australia; 2012.
[27] Worden K., Wong C.X., Parlitz U., et al. Identification of pre-sliding and sliding friction
dynamics: grey box and black-box models. Mech Syst Signal Process. 2007;21:514–534.
[28] Xie WF, Zhu YQ, Zhao ZY, et al. Nonlinear system identification using optimized dynamic
neural network. Neurocomputing. 2009;72:3277–3287.