A Train Control System Case Study in Model-Based Real Time System Desig
A Train Control System Case Study in Model-Based Real Time System Desig
A Train Control System Case Study in Model-Based Real Time System Desig
net/publication/220952907
A Train Control System Case Study in Model-Based Real Time System Desig.
CITATIONS READS
55 826
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Armin Zimmermann on 20 May 2014.
integrity control system and a computer that can control A Quality of Service parameter specification (maxi-
train speed. It communicates via GSM-R radio with base mum connection establishment delay etc.) is given in
transceiver stations (BTS), which are connected to base sta- the Euroradio form fit functional interface specifica-
tion controllers (BSC) by cable. The BSCs are communi- tion (FFFIS) [11].
cating with radio block centers via ISDN.
Allowed parameter ranges for some system design
Radio Block Centers (RBC) are the trackside part for ra- variables (like the minimum time between two subse-
dio at ETCS level 2 and 3. Their major functions include quent position reports sent by a train) are specified in
safe train separation based on safe allocation of routes by ERTMS Performance Requirements for Interoperabil-
regulation and interlocking. Position and integrity reports ity [12].
are sent by the trains periodically or upon request. Based
on this information and the train routes, safe track areas are Definitions of requirements of reliability, availability,
assigned to trains. This is done with so-called movement maintainability and safety (RAMS) as well as accept-
authority messages. able numbers of failures per passenger-kilometer due
The European Integrated Railway Radio Enhanced Net- to different reasons can be found in the ERTMS RAMS
work (EIRENE) project was started on behalf of the Eu- Specification [9].
ropean Railways to define a new digital radio standard for Some additional assumptions (mean time to complete
application on the European High Speed Rail System. The the on-board train integrity check etc.) have been taken
EIRENE System Requirements Specification [13] defines from a description of simulation experiments carried
the set of requirements which a railway radio system shall out by the German railways company [23].
comply with in order to ensure interoperability between na-
tional railways. GSM (Global System for Mobile Commu- Another detailed description of communication quality
nications) was chosen as the base technology because of of service parameters is provided in [16], serving as
availability and cost considerations. Additional functions acceptance criteria for future measurements and tests
which are tailored to the needs of railroad use (like area of actual ETCS communication setups.
addressing, automatic international roaming etc) have been
defined as Railway GSM (GSM-R [7]). For up-link and Results of such a quality of service test at a railway
down-link there are different frequency bands reserved for trial site are presented in [24], thus facilitating a com-
GSM-R around 900 MHz. parison with the original requirements. It turns out that
the QoS parameters are in the required range, although
The EURORADIO layer of the communication link often close to and even sometimes worse than the re-
specifies the Radio Communication System requirements quirements. In the following, however, we adopt worst
to the air gap interface between train and trackside equip- case assumptions based on the requirements, because
ment [11, 18]. The MORANE (Mobile Radio for Railway future implementations will be bound to perform as re-
Networks in Europe [21]) project was set up to specify, de- quired.
velop, test and validate prototypes of a new radio system.
Trial sites exist in France, Italy and Germany. Results of The communication link between train and RBC is al-
a quality of service test at one of these sites are presented ways connected in normal operation mode. In that situation
in [24]. the following failures may happen:
parameter ( ).
( / * + 2 3
offline + - - @ *
burst
tial transition loss is set accordingly. There is a certain
connected
lossindication handover amount of time needed to detect the communication loss,
which is required to be not greater than one second. This
indicate loss cellborder
is modeled by the deterministic transition indicate with
one as the fixed delay.
reconnect After being offline, the train communication system tries
to reestablish the link at once. The requirements specify
Figure 2. Failure and recovery model for GSM- that a connection attempt must be successful with proba-
R communication channel bility 99.9%, while in the remaining cases the establish-
ment is canceled after 7.5 seconds and retried. This be-
havior is modeled with immediate transitions carrying the
Figure 2 shows a deterministic and stochastic Petri success/fail probabilities estp and failp, and the deter-
net [2] model of the described behavior. The firing delays ministic transition fail with delay 7.5. Connection es-
and distributions have been chosen as follows. One unit of tablishment times are random, but required to be less than
model time means one second in reality. 5 seconds for 95% of the cases. The corresponding firing
Transition startburst models the beginning of a distribution of transition connect is thus exponential with
transmission error. It has an exponentially distributed fir- parameter 0.6.
ing delay because of the stochastic nature of transmission
The model shown in Figure 2 depicts states and state
errors. The corresponding firing time is comparable to a
transitions of the communication link. The initial state is
mean time to failure of the communication link due to trans-
connected. It is obvious that there will always be ex-
mission errors. The specification requires this value to be
actly one token in the model, letting the Petri net behave
greater than or equal to 7 seconds for 95% of all cases. From
like a state machine, and the reachability graph is isomor-
the density and distribution functions of the exponential dis-
phic to the net structure.
tribution
Because in every marking there is at most one transition
with non-exponentially distributed firing delay enabled, the
and
model can be numerically analyzed with standard DSPN
we can calculate the necessary parameter value: algorithms [6]. Because of the state machine structure it
# % '
would also be possible to exchange all deterministic tran-
( * + * * - / /
and -
burst
>
+ / * 3 @ *
handover
>
3 + 2 * - @ *
lossindication
>
B
+ - 3 @ *
recover fail
establish
B
+ 3 2 * @ *
estfail
>
+ * * @ *
state probability vector. It could then be analyzed as a sim- plexity and model accuracy. We decided to condense the
ple GSPN [1]. failure model into a two-state system with the basic states
Numerical analysis of the example is computationally in- ok and failed. The corresponding stochastic Petri net
expensive due to its small state space. Despite of the “stiff- (GSPN) is shown in Figure 4.
ness” of the problem (e.g. firing rates of transitions end- The question is then how to specify the transition firing
burst and loss differ by eight orders of magnitude) the rates to minimize the approximation error. The main char-
exact solution is a matter of seconds. A simulation with acteristic of the failure model is the mean availability, which
proper confidence interval control would take quite some shall be equal in the exact and condensed model. Thus the
time because of the mentioned rare events. probability of having one token in place ok needs to be
Table 3 shows the results of the numerical analysis. The 0.99166.
connection is working with a probability of 99.166%. This Even with a correct availability an error can be intro-
is much worse than the required availability of 99.95% as duced by selecting a wrong speed of state changes between
specified in [11]. This requirement is commented to be ok and failed. If the speed would be too high, no cor-
a coverage requirement, although we see from the model rect packet transmission is possible, because a certain undis-
evaluation that already the allowed handover downtimes vi- turbed time (given by the packet length and transmission
olate this requirement. bit rate) is always necessary. The second restriction which
In fact, handovers account for more than 70% of the we impose on the condensed model is thus to keep the
overall unavailability. To avoid their impact on the com- mean sojourn time in state ok exactly as it was in the full
munication link, there are discussions about installing two model. This time is the reciprocal value of the sum of all
independent GSM-R devices in each train. For instance in transition firing rates going out from the state, in our case
;
! " !
been duplicated for a higher reliability and availability. The With these two restrictions the transition rates can be eas-
connection to the next BTS can then be carried out when the ily calculated. Let denote the transition rates and the &
train gets close to the cell border, thus avoiding any offline state probability vector in steady-state.
state due to handovers. Then
* + 2 2
(
&
! '
& *
+ "
! ' + + ! -
ing block mode (ETCS level 3) under communication fail- sojourn time: *
/ + - -
! -
and thus
>
>
RBC and processed there, which takes a maximum of 0.5 and the minimum theoretical distance for
+
/ * *
sec. The resulting information is sent to the following is thus . This simple consideration already
* * *
Train2, telling him either that everything is fine to go on shows that the common term of “driving in braking dis-
driving (by sending a new movement authority packet) or tance” with ETCS is misleading, because even if everything
that an emergency braking is necessary immediately. would run perfectly trains cannot get closer than 4km.
However, if a communication packet is lost on either Figure 6 shows a Petri net model for the above explained
the communication up-link (Train1 RBC) or down-link
behavior. The upper part models the generation of the po-
(RBC Train2), Train2 needs to decide on its own at what
ter the last integrity check has been completed, a part of waiting
the train’s carriages are lost from the main train and stop
where they are. Another danger for the following train is lossUp ok lossDown Timeout
an accident after sending the position/integrity report. The recover fail
failed
ETCS system requirements specification thus states that “In stopped
moving block operation the movement authority shall never
exceed the min safe rear end of the preceding train” [10]. Resetup
for
2 + 3 * + 3 / * + /
for
* + * 3 - * + / + / /
for
> >
* + * * + / / + 3 /
References
*
otherwise
For the performance evaluation of the model the nu- [1] M. Ajmone Marsan, G. Balbo, G. Conte, S. Donatelli, and
G. Franceschinis. Modelling with Generalized Stochastic
meric analysis cannot be used, because the restriction of not
Petri Nets. Series in parallel computing. John Wiley and
more than one enabled non-exponential transition per mark- Sons, 1995.
ing is violated. Switching to an underlying discrete time [2] M. Ajmone Marsan and G. Chiola. On Petri nets with de-
scale [27] does not help either, because then the state space terministic and exponentially distributed firing times. In
becomes so large that it cannot be handled by the available G. Rozenberg, editor, Advances in Petri Nets 1987, volume
computing hardware. This is due to the fact that there are 266 of Lecture Notes in Computer Science, pages 132–145.
small delays and small differences between delays, necessi- Springer Verlag, 1987.
tating a very small underlying time step, leading to a state [3] G. Bucci and E. Vicario. Compositional validation of
space explosion. time-critical systems using communicating time Petri nets.
Thus simulation was the only choice, but has its own IEEE Transactions on Software Engineering, 21(12):969–
problems. For a distance the deadline is 6
3 * *
992, 1995.
[4] A. Chiappini, A. Cimatti, C. Porzia, G. Rotondo, R. Sebas-
seconds. In that case the model evaluation shows that the
tiani, P. Traverso, and A. Villafiorita. Formal specification
train is stopped with a probability of 94%. Even for a dis-
and development of a safety-critical train management sys-
tance and a resulting deadline of 12 seconds, the
3 * * *