U.S. patent application number 10/296611 was filed with the patent office on 2004-02-12 for rail safety system.
Invention is credited to Seifert, Benedict.
Application Number | 20040026574 10/296611 |
Document ID | / |
Family ID | 9892203 |
Filed Date | 2004-02-12 |
United States Patent
Application |
20040026574 |
Kind Code |
A1 |
Seifert, Benedict |
February 12, 2004 |
Rail safety system
Abstract
A system which can easily and cheaply be installed to an
existing rail infrastructure is disclosed which seeks to avoid the
likelihood of train collisions. Each train travelling on a network
is provided with a means for determining its own position (such as
a GPS receiver) and a means for communicating with a central
controller. The train's position is communicated to the central
controller and the central controller uses this information to work
out where each train is located on the network. Velocity and
acceleration information are obtained by monitoring the train's
position over time and predictions are made as to the future
position of each of the trains on the network. If the future
position of any of the trains is determined to be in a range which
at least partially overlaps a range of future positions for another
train then a collision warning is sent to one or both trains. This
warning could trigger the train to stop automatically or for the
driver to apply the breaks. Successive re-start signals are then
sent so that the trains may continue on their journey without
likelihood of collision.
Inventors: |
Seifert, Benedict; (Oxford,
GB) |
Correspondence
Address: |
BOZICEVIC, FIELD & FRANCIS LLP
200 MIDDLEFIELD RD
SUITE 200
MENLO PARK
CA
94025
US
|
Family ID: |
9892203 |
Appl. No.: |
10/296611 |
Filed: |
July 16, 2003 |
PCT Filed: |
May 23, 2001 |
PCT NO: |
PCT/GB01/02311 |
Current U.S.
Class: |
246/5 |
Current CPC
Class: |
B61L 25/026 20130101;
B61L 27/0038 20130101; B61L 27/20 20220101; B61L 2205/04 20130101;
B61L 25/021 20130101; G01S 5/0072 20130101; B61L 2205/02 20130101;
B61L 25/025 20130101; B61L 23/34 20130101 |
Class at
Publication: |
246/5 |
International
Class: |
B61L 027/00 |
Foreign Application Data
Date |
Code |
Application Number |
May 23, 2000 |
GB |
0012519.5 |
Claims
1. A method of determining if a train collision is possible, said
method comprising the steps of: obtaining data relating to the
position of a train; transmitting said data to a controller;
receiving said data at said controller; performing a calculation at
said controller using said data to predict a position of said train
at a future time; and using said predicted future position to
determine whether any action needs to be taken.
2. A method according to claim 1, wherein said calculation includes
deriving differenced position vectors from said position data, said
differenced position vectors having vector coordinates representing
the difference between two consecutive coordinates in a vector of
said position data.
3. A method according to claim 1 or 2, wherein said calculation
includes referring to a database of stored data relating to
previous train journeys and identifying in said database vectors
similar to a present journey vector, wherein said present journey
vector is either said position data or differenced position vectors
derived from said position data, said differenced position vectors
having vector coordinates representing the difference between two
consecutive coordinates in a vector of said position data.
4. A method according to claim 3, wherein said database is a
database of stored differenced position vectors.
5. A method according to claim 3, wherein said database is a
database of stored position vectors.
6. A method according to any one of claims 3 to 5 wherein said step
of identifying similar vectors comprises evaluating whether the
magnitude of a further vector obtained by subtracting a potentially
similar vector in said database from said present journey vector is
less than some prescribed positive number.
7. A method according to any one of claims 3-6, wherein said
calculation includes extracting from said database predictive data,
said predictive data being data recorded a particular period of
time after said data identified as similar.
8. A method according to claim 7, further comprising the step of
adding the same number to each coordinate of said extracted
predictive data vector so as to translate it by the amount of the
number added.
9. A method according to claim 8, wherein said number added
corresponds to the current train position.
10. A method according to any one of claims 7 to 9, wherein a set
of predictive data is extracted, said set having a maximum value
and a minimum value for any given particular period of time.
11. A method according to claim 10, wherein said minimum and
maximum values represent a range of possible train positions at
said future time.
12. A method according to any preceding claim, wherein said steps
of obtaining data relating to the position of the train and
transmitting said data to said controller are carried out a
plurality of times for a plurality of respective trains and said
controller determines if a collision is possible based on a
prediction of a future position of said plurality of trains.
13. A method according to claim 12, wherein said step of predicting
a future train position is a step in which a plurality of intervals
of future possible train positions for a plurality of respective
trains is predicted in accordance with the method of claim 8; and
said step of determining if a collision is possible comprises
determining whether at least some of said plurality of predicted
intervals of possible train positions for said plurality of trains
overlap.
14. A method according to any preceding claim, further comprising
the step of: transmitting a warning signal to said train if it is
determined that a collision is possible.
15. A method according to any preceding claim, said method further
comprising: transmitting a stop message to said train if it is
determined that a collision is possible.
16. A method according to claim 15, further comprising the step of:
automatically stopping said train when said stop message is
received from said controller.
17. A method according to any preceding claim, wherein said step of
determining a position of the train uses signals transmitted from
GPS satellites.
18. A method according to any preceding claim, wherein said
transmitting and receiving steps are carried out using an
electronic communications network.
19. A method according to any preceding claim, wherein said
transmitting and receiving steps are carried out using Internet
protocols within a extranet structure.
20. A method according to any preceding claim, wherein said step of
transmitting said data to said controller is carried out using a
mobile telephone.
21. A method according to any preceding claim, wherein said
obtained data relating to a position of a train is converted to
position data along a real line model.
22. A method according to claim 21, wherein said calculation
performed at said controller predicts a future position, or range
of future positions, along said real line model.
23. A method according to claim 22, wherein said predicted
position, or said range of predicted positions, is converted to
data representing a position, or range of positions, on said track
network.
24. Apparatus for determining the possibility of a train collision,
said apparatus comprising first means situated on said train and
second means situated remote from said train; said first means
comprising: position determining means for determining the position
of said train; communication means operatively linked to said
position determining means for transmitting the position determined
by said position determining means; said second means comprising:
receiving means for receiving said determined position from said
communication means; prediction means for predicting a future
position of said train based on said received determined position;
and collision determining means for determining whether a collision
is possible based on the result of said prediction means.
25. Apparatus according to claim 24, wherein said second means
further comprises a database of stored data relating to previous
train journeys.
26. Apparatus according to claim 25, wherein said prediction means
comprises: identifying means for identifying stored data which is
similar to said determined position or is similar to data derived
from said determined position.
27. Apparatus according to claim 26, wherein said prediction means
further comprises: extracting means for extracting predictive data,
said predictive data having originally been obtained at a future
time to said identifying stored data.
28. Apparatus according to claim 27, wherein said extracting means
is for extracting a set of predictive data, said set having a
minimum and maximum value.
29. Apparatus according to claim 28, wherein said set defines a
range of predicted future positions for said train.
30. Apparatus according to claim 29, wherein said collision
determining means uses a plurality of said sets of predictive data
obtained for different trains and determines a collision is
possible when at least two of said sets overlap.
31. Apparatus according to any one of claims 24 to 30, wherein said
position determining means comprises a GPS receiver.
32. Apparatus according to any one of claims 24 to 31, wherein said
communication means comprises a mobile telephone.
33. Apparatus according to any one of claims 24 to 32, wherein said
communication means comprises a portable computer having a
modem.
34. Apparatus according to any one of claims 24 to 33, further
comprising means at said second means for transmitting a message to
said first means.
35. Apparatus according to claim 34, wherein said message comprises
a warning message.
36. Apparatus according to claim 34, wherein said message comprises
a stop message.
37. Apparatus according to claim 24 to 36, further comprising means
at said second means for transmitting a re-start signal when it is
determined that it is safe for a stopped train to start.
38. Apparatus according to claim 37, further comprising means at
said first means for re-starting a stopped train upon receipt of a
received re-start signal.
39. Train safety means for use with the method according to any one
of claims 1 to 23, said train safety means comprising: position
determining means; communication means operatively linked to said
position determining means for transmitting a position so
determined.
40. A controller for use with the method of any one of claims 1 to
23, said controller comprising: receiving means for receiving data
relating to a determined train position; prediction means for
predicting a future train position based on said received train
position data; and collision determining means for determining
whether a collision is possible based on said predicted future
position.
41. Use of a mobile telephone for transmitting data relating to a
determined position of a train and for receiving data relating to a
possibility of a collision.
42. Use of a computer system for receiving data relating to a
determined position of a train, for predicting a future position of
said train using said received positional data, for predicting the
possibility of a collision based on said predicted future position
and for transmitting messages to said train if the possibility of a
collision is predicted.
43. A memory means storing computer readable instructions which,
when connected to a suitable computer, is operable to cause said
computer to carry out the following method: receive data from a
plurality of trains, said data relating to a determined position of
said respective train; accessing a database available to said
computer, said database containing information about the train
network on which said train is located; calculating, using said
received position data and said network database, the position of
each train on a track of said network; predicting the position of
each train at a point in time in the future; and determining if a
collision between two or more of said plurality of trains is
possible using said predicted future positions.
44. A method substantially as hereinbefore described with reference
to any one of FIGS. 1 to 9 of the accompanying drawings.
45. Apparatus constructed and arranged substantially as
hereinbefore described with reference to any one of FIGS. 1 to 9 of
the accompanying drawings.
Description
[0001] The present invention generally relates to the field of
safety systems for use with trains and other rail-based
vehicles.
[0002] More particularly, the present invention relates to a safety
system using mobile computerised communication technology so as to
avoid train crashes and other accidents.
[0003] The present state of the art in rail safety systems is known
as Automatic Train Protection (ATP). In an ATP System, information
is transmitted to the train cab electronically. In particular, the
speed limit of the section in which a train is travelling and
usually the speed limit of the next section are each transmitted to
each train. In its most sophisticated form, the ATP system will be
interlocked with the braking system of the train so as to invoke a
braking application if the driver exceeds the speed limit for the
section. The ATP system consists of a series of speed bands which
appear behind the first train on a line so that the speed of a
second train on the line reduces in steps as it approaches. These
speed bands move with the first train (either smoothly or in
"jumps" as the first train moves between block sections) and the
permitted speed increases from zero in the band directly behind the
train to full speed at some distance further behind the train. The
size of the bands and the speeds are chosen in accordance with the
safe braking distance of any train. The ATP system assumes that all
trains have the same braking distance which is not the case on
standard mainline railways. On such railways, there are many
different types of trains which all have different braking
characteristics. The main drawback of ATP is that to install it
requires updating of the train and rail infrastructure, resulting
in a very high cost. The equipment required for ATP is also
vulnerable to bad weather, electronic interference, damage,
vandalism and theft.
[0004] WO 99/52091 discloses a system for alerting the crew of a
train of the proximity of other trains. Simply, each train
transmits its own position and is able to receive the transmitted
positions of other nearby trains. These positions are displayed to
the driver who can take appropriate action if required. Thus, the
system is still prone to driver error and is not useful in the
vicinity of a station where traffic is so dense that the knowledge
that other trains are in the vicinity is of no relevance. The
onboard train computers of this system do not themselves have a
methodology for predicting a collision but simply organise
descriptive information for the driver.
[0005] It is an objective of the present invention to provide a
system that can be installed in an existing rail infrastructure at
a lower cost than ATP. Furthermore, it is desired to provide a
system which can automatically predict a collision and take
automatic action to prevent such a collision.
[0006] Accordingly, the present invention provides a method of
determining if a train collision is possible, said method
comprising the steps of obtaining data relating to the position of
a train; transmitting said data to a controller; receiving said
data at said controller; performing a calculation at said
controller using said data to predict a position of said train at a
future time; and using said predicted future position to determine
whether any action needs to be taken to collide with another train
or other object.
[0007] Further, the present invention also provides apparatus for
determining the possibility of a train collision, said apparatus
comprising first means situated on said train and second means
situated remote from said train; said first means comprising:
position determining means for determining the position of said
train;
[0008] communication means operatively linked to said position
determining means for transmitting the position determined by said
position determining means; said second means comprising: receiving
means for receiving said determined position from said
communication means; prediction means for predicting a future
position of said train based on said received determined position;
and collision determining means for determining whether a collision
is possible based on the result of said prediction means.
[0009] Furthermore, the present invention also provides a use of a
mobile telephone for transmitting data relating to a determined
position of a train and for receiving data relating to a
possibility of a collision.
[0010] Furthermore, the present invention also provides a use of a
computer system for receiving data relating to a determined
position of a train, for predicting a future position of said train
using said received positional data, for predicting the possibility
of a collision based on said predicted future position and for
transmitting messages to said train if the possibility of a
collision is predicted.
[0011] The present invention also provides a memory means storing
computer readable instructions which, when connected to a suitable
computer, is operable to cause said computer to carry out the
following method:
[0012] receive data from a plurality of trains, said data relating
to a determined position of said respective train;
[0013] accessing a database available to said computer, said
database containing information about the train network on which
said train is located;
[0014] calculating, using said received position data and said
network database, the position of each train on a track of said
network;
[0015] predicting the position of each train at a point in time in
the future; and
[0016] determining if a collision between two or more of said
plurality of trains is possible using said predicted future
positions.
[0017] In a preferred embodiment of the invention, differenced
position vectors are derived from the obtained position data, these
differenced position vectors having vector coordinates representing
the difference between two consecutive coordinates of a vector of
the position data. Preferably, the differenced position vector
derived from the position data is compared with a series of stored
differenced position vectors in a data base. Similar stored
differenced position vectors are then identified. Alternatively,
the position data itself may be compared with other position data
stored in a database. Preferably, two vectors are determined to be
similar if the magnitude of a third vector obtained by subtracting
the stored vector from the present vector is less than some
prescribed number.
[0018] Preferably, the calculation also includes extracting
predicted data from the database, this predicted data being data
recorded at a particular period of time after the data which has
been identified as similar. In order to translate the predicted
data, which might be in the form of differenced position vectors,
to useful data, it may be necessary to add an offset to all the
predicted data so that a predicted position of the train on the
track is obtained.
[0019] The advantage of the using differenced position vectors
rather than simple position vectors themselves is that the
differenced positioned vectors are independent of the train's
absolute position. Thus, differenced positioned vectors on previous
train journeys on other tracks for example can be stored in the
database and used to obtain a prediction for the present train
journey.
[0020] Typically, the predicted data will comprise a number of
vectors extracted from the database which are similar to the
present journey vector. This range of extracted vectors will have a
maximum value and a minimum value for any given particular period
of time. These maximum and minimum values may be used to give a
range of possible train positions at this future time.
[0021] In a preferred embodiment of the present invention, the
above method is carried out for a number of trains on the network
so that a range of potential future positions is obtained for each
train. The step of determining if a collision is possible
preferably comprises determining whether at least some of the
plurality of predicted intervals of possible train positions for
the plurality of trains overlap. Remedial action can be taken if
any overlapping is found.
[0022] Preferably, in accordance with the present invention, the
obtained data relating to the position of the train is converted to
position data along a real line model. The calculation of predicted
train positions is then carried out along the real line model and
the results are changed back to data representing actual positions
along the track network. The has the advantage that the method of
prediction can be kept separate from the physical topology of the
track network, which is dictated by the position of the points
etc.
[0023] The present invention will be more fully described, by way
of non-limitative example only, with reference to the accompanying
drawings, in which:
[0024] FIG. 1 shows schematically the main features of an
embodiment of the present invention;
[0025] FIG. 2 shows schematically a model of a simple rail network
having three stations and two tracks;
[0026] FIG. 3 shows schematically a model of another simple rail
network having four stations and two tracks;
[0027] FIG. 4 shows schematically a model of yet another simple
rail network having six stations and three tracks;
[0028] FIG. 5 shows schematically a model of a network similar to
the network of FIG. 3 which is used to describe the train future
position prediction method of the present invention;
[0029] FIG. 6 shows schematically the relationship between a
modelled track network and the concept of a "real line";
[0030] FIG. 7 shows schematically an inverse relationship between a
real line model and a track network model;
[0031] FIG. 8 shows a graph of velocity versus time for a
particular train; and
[0032] FIG. 9 shows a corresponding graph of distance versus time
for the same train.
[0033] The present invention comprises monitoring at some central
location at least the position of the active trains on a particular
network. This is achieved by providing means, on each train, for
detecting the position of the train and for communicating this
position to the central location. Appropriate means are provided at
the central location for analysing the positions of the trains and
predicting if a collision is possible or likely. The means at the
central location are operable to communicate to at least some of
the trains on the network a message that a collision is possible or
likely so that appropriate action may be taken.
[0034] FIG. 1 shows a schematic illustration of an embodiment of
the invention. In FIG. 1, only three trains are shown and these are
intended to represent all of the active trains on a particular
defined network. A network may be defined arbitrarily as a
particular group of tracks. Alternatively (and more usually), it
may be defined as a group of tracks upon which the trains run
wherein the tracks are chosen because there is at least a
possibility of a train on one track being able to collide with a
train on another, or the same track.
[0035] Each train is provided with a communication device
represented in FIG. 1 by a mobile telephone. The mobile telephone
is capable of communicating with a central controller via a
communications network, which could include, for example, a
cellular telephone network. Preferably, the mobile telephone is
operable to exchange data directly with the central controller in
the form of digital messages. Each train is also provided with
means for determining its own position either relative to some
known position or absolute with respect to some defined axis.
Preferably, these means comprise a GPS (Global Positioning System)
receiver. The means for determining the train's position are
operable to communicate with the mobile telephone so that the
position of the train may be transmitted to the central controller.
In this way, the position of every train on the network is known
by, or is able to be calculated by, the central controller. The
central controller can then make a prediction as to whether it is
possible that a collision may occur. This prediction may be based,
for example, on whether a train comes within a certain distance of
another train on the network.
[0036] The means for communicating with the central controller may
take any suitable form. If a mobile telephone is used, it is
preferably a "smart mobile phone" which is able to handle data and
has computer connectivity. Alternatively, a DPA (Digital Personal
Assistant) with integrated modem or a mobile computer attached to a
digital mobile phone may be used. All of these devices allow the
transmission of digital data over the telephone network which is
advantageous to the present invention.
[0037] The data may be transmitted over the telephone network using
the GSM (Global System for Mobile Communications) standard or WAP
(Wireless Application Protocol) which is the standard for
connecting mobile phones and other devices to the Internet.
[0038] Further information on suitable devices may be found in a
paper entitled "Mobile Commerce Report" by Durlacher Research
Limited, London 1999.
[0039] Preferably the known global positioning system is used as
the means to determine the position of the train in space. These
means may simply comprise a receiver for receiving the GPS signals
from the GPS satellites so that these received signals may be
transmitted to the central controller via the communication means.
The central controller may then decode the signals to provide the
location of the train. Alternatively, the GPS receiver on each
train can be operable to decode the GPS signals itself so that the
communication means can transmit to the controller data
representing the position of the train in space. Preferably, the
GPS receiver comprises a commercially available receiver which can
output a position on the surface of the earth as a data string.
[0040] Other technology for calculating one's position is known and
may also be used. For example, a signal transmitted by the
communication means may be received by a number of nearby base
stations and the time of receipt compared so that the position of
the train may be triangulated.
[0041] The central controller preferably consists of a computer
system linked to the telephone network by modem or other suitable
means. The computer system has access to a database which stores a
geometric model of the rail network. It is therefore possible for
the central controller to construct a model of the rail system
using the geometric rail network database and the received train
positions. The trains may therefore be tracked as they move along
the tracks of the network. The central controller preferably also
has access to a database holding information on the expected
trajectory of a train. For example, it is not uncommon for two
stations to be connected by two tracks with sections at various
points along the track at which the train may change track (known
as "points" or "switches"). The expected trajectory of the train is
information on which track the train is expected to be on at
various distances between the two stations. For example, a specific
train could be on track 1 for 3 miles, then cross over to track 2
for 27 miles, then to track 1 again for 34 miles. This information
is constructed by referring to the present state of the points in
the network (for more information on what is meant by the "state of
the points", see later). This additional information database
allows the central controller to determine which track a particular
train is on when the positional information transmitted by the
train is not accurate enough to be able to distinguish between
adjacent tracks. This additional information is not, however,
necessary if the positional data is extremely accurate (so that the
track that the train is on can be determined by comparing the
positional data with the model of the rail network).
[0042] The present commercial version of the GPS system has an
accuracy of 100 metres, although planned new commercial versions of
the system will be accurate to within 4 metres. That is to say, the
actual point in space of the receiver could be up to 100 metres
away from the measured point. Thus, it is necessary for the central
controller to construct a disc of diameter 200 metres the centre of
which is the measured position. The actual position of the train is
then known to be somewhere in the disc. This disc is then
superimposed over the model of the rail network and the expected
trajectory data is used to identify which track the train is on. A
section of track may then be identified (maximum length=200 metres)
upon which the head of the train is located.
[0043] The error in the commercial GPS signal is a random one and
as such it is possible to reduce or eliminate it using standard
noise elimination procedures. More accuracy can be obtained if a
nearby base station (whose position is known) also receives the GPS
satellite signals so that the error (which is an offset error) may
be calculated and deducted from the measurements.
[0044] Preferably, the communication means and central controller
communicate with one another via a secure extranet. An extranet is
a physically secure network using IP (Internet Protocol) technical
standards but being physically detached from the open Internet, for
the purposes of security and speed of transmission. In such a case,
the central controller would have the function of a web server
providing Internet facilities. Each communication means located on
respective trains would be operable to access a particular web site
provided by the web server in order to upload the positional
information. Information may also be downloaded from the web site,
but this is optional.
[0045] Operationally linked to the web server is a control server
which carries out the various calculations and predictions
required. In practice, the web server and control server preferably
comprise a single computer system. It is the control server that
has access to the database storing the geometric model of the
network and the optional database storing the expected train
trajectory information.
[0046] Preferably, communication will be such that train positional
information is transmitted from each train to the web server at
regular time intervals, for example once every second. This
regularity could be fixed for the whole system or variable
depending on various parameters of the rail network system. For
example, information could be transmitted more regularly when the
train is travelling faster than when it is stationary. As another
example, information could be transmitted more regularly in busy
areas where collisions are more likely to occur than in areas
comprising of a single track in open countryside.
[0047] The structure of the database storing the geometric model of
the rail network will now be described. This description is by way
of example only and any way of storing the rail network data so
that the positions of trains can be tracked and predictions can be
made regarding collisions is encompassed by the present
invention.
[0048] As an example, FIG. 2 shows a simple model for a two-track
railway system with three stations. The stations are labelled A, B
and C and the tracks are labelled t1 and t2. Track 1 (t1) runs
between station A and station B and track 2 (t2) runs between
station B and station C. Track t1 is 1000 m long and track t2 is
500 m long. In this simple case the model defining the tracks would
solely consist of the set {t1, t2} where t1 and t2 are defined as
follows:
t1={A, B, [0,1000]}
t2={B, C, [0,500]}
[0049] Thus, the model holds information about which stations each
track connects and the distance of track there is between the
stations. It is to be noted that the notation assumes that
measurements commence from the first mentioned station to the
second mentioned station. Thus, 0 m along track 1 is station A and
1000 m along track 1 is station B. A train travelling from B to A
on track 1 would be travelling in the negative direction.
[0050] The position of the trains themselves are modelled as
information about which track they are on and the distance along
the track they are on. For example, a train on track t1 halfway
between stations A, B (and therefore 500 m from station A) would be
modelled as:
train1=(t1, 500)
[0051] This train position information would be linked with
information regarding the time at which the position was measured.
The information would therefore change with time as the train moves
so that if the train was travelling from A to B the number "500" in
the above notation would increase until it reaches "1000" (at which
point the train is at station B).
[0052] If sophisticated predictions about the train motion are to
be made, then it is necessary that information about where tracks
meet or cross each other is stored. At such locations, it is
generally possible for a train travelling on one track to either
change tracks or carry on on the same track. A simple example is
shown with reference to FIG. 3. In FIG. 3, stations A and B are
connected by track t1 and stations C and D are connected by track
t2. The tracks interconnect at a distance of p1 from station A
along t1 and at a distance of p2 from station C along t2. A train
travelling along track t1 in the direction of station B can either
carry on down track t1 when it reaches Point1 or change tracks to
track t2 towards station D. This railway point Point1 is modelled
as:
Point1={t1, t2, (p1, p2)}
[0053] A more complicated point system is described with reference
to FIG. 4. In this example, three tracks are shown t1, t2 and t3. A
train travelling from A to B on track t1 can either continue on
track t1 or change to track t3 in the direction of station F. In
the process of changing the train crosses t2 but it is not able to
join track t2. Here, the point (labelled Point2) would be modelled
as:
Point2={t1, t2, t3, (p1, p2, p3)}
[0054] Where p1, p2 and p3 denote the distance from the three
stations A, C and E along the tracks t1, t2 and t3 respectively at
which a train would be able to change over to track t1 to track t3
crossing track t2 on the way.
[0055] The above describes how the position of the tracks, trains
and points is modelled. For a full model of the network system, it
is necessary also to determine the state of the various components
of the network. The state of the rail system would in general be
modelled by the following set of data:
[0056] 1) The state of all railway points indicating which track a
train would end up on if it approached a point on a certain
track.
[0057] 2) The position of all active trains.
[0058] Regarding the state of the points FIG. 3 shows a point
system connecting track t1 and track t2 and which is referred to as
Point1. The state of Point1 would be described by referring to what
would happen to a train approaching Point1 on track t1 from station
A. In this example, it is clear that the state of Point1 is
described by a simple piece of information saying whether a train
would stay on t1, or would change to track t2. In general, the
state of the points is described by the following notation:
{Point1, t1(A.fwdarw.B).fwdarw.t2(C.fwdarw.D)}
[0059] The above expression indicates that a train travelling on
track t1 in the direction of A to B would change to track t2 in a
direction toward station D. The points also work in the opposite
sense such that a train travelling from D to C along track t2 would
change to track t1 in the direction towards A. In other words the
above expression is equivalent to:
{Point1, t2(D.fwdarw.C).fwdarw.t1, (B.fwdarw.A)}
[0060] In the alternative state of the points a train travelling on
track t1 from A to B would remain on track t1.
[0061] Point1 described above and illustrated in FIG. 3 does not
influence a train travelling from B or from C towards Point1
because in direction of travel the track does not diverge so there
is no choice of possible routes. To enable a train travelling from
B or from C to change track, another point would have to be
provided and its state of switching track or not would be
separately defined.
[0062] Regarding the state of the active trains, it is noted that
an active train is one that is either moving or standing between
two stations on one of the tracks. A train in a depot is not
active. The state of the train is in general described by three
data:
[0063] a) position of the train at present (time=now), and at a
number of equally spaced points in the recent past (now-1 time
unit, now-2 time units, . . . , now-n time units, etc).
[0064] b) velocity of the train, described by a set of recent
points in time and the corresponding positions of the train,
and
[0065] c) acceleration of the train described by a set of recent
points in time and corresponding positions.
[0066] These data can be stored as a table of the present and
delayed train positions. These data can be used to obtain a table
of differenced train positions as will be later described. In fact,
only one table of data is required to be measured because velocity
and acceleration data can be determined from the position data by
appropriate differencing algorithms.
[0067] Once the state of the train network system has been
ascertained and the present and delayed train positions, along with
details on the state of the points which aid in predicting the
trajectory of the trains is known, it is possible to make
predictions as to the possibility or likelihood of a collision.
[0068] The predictions made can be used for purposes other than
determining the possibility of a collision. For example, the
velocity of the train can be predicted according to the method of
the present invention and it can be determined whether the train is
likely to exceed a certain speed limit for the piece of track that
it is predicted to be on. In this case and also in the case of
possible collisions, remedial action would be taken if an incident
is predicted which would preferably be in the form of sending a
message to the train to slow down in the form of a warning to the
driver or an actual remote control message interlocked with the
train braking system.
[0069] The central control server is operable to make a set of
predictions of the future state of all active trains based on the
present state of all active trains and the position of the points.
Preferably, a special prediction method known as "ensemble
prediction" is used. The prediction method uses the knowledge of
previous positions of the same train, or perhaps trains of similar
characteristics over the same or similar areas of track in order to
yield a range of points at which the train could be located in some
future interval of time and the collision detection algorithm
checks to see if this range overlaps with the range of points
calculated for any other train on the network at the same time
point. Such prediction methods are disclosed in "Time Series:
Forecasts in the Future and Understanding the Past", proceedings of
the NATO Advanced research workshop on comparative time series by
Weigend and Gershenfeld (Editor).
[0070] Traditional methods of prediction, based on either ordinary
differential equations (smooth deterministic systems) or linear
statistical techniques (for probabilistic systems) are not ideal
for predicting collisions on rail networks. The preferred method of
the present invention relies on a system in which it is not
necessary to know the explicit form of the equations of motion.
When a system has both deterministic and random components, it is
in principle impossible to predict its precise future position.
However, it is possible to predict an interval (or ensemble) in
which the system will fall within certain probability bounds. These
probability bounds can and should be made so small that they
represent a virtual certainty that no collision will occur on the
basis of the assumption that the underlying model of the train
(i.e. the physical characteristics of the trains and their
behaviour observed over long periods of time) is fundamentally
accurate. The predicted interval has a size which is a function of
the amplitude and other characteristics of the noise in the system.
This method is suitable to the present invention because (a) it is
not necessary to know the differential equations describing ideal
or actual train motion and (b) the method for obtaining the trains
position contains noise, not least due to the random offset
inherent in the GPS signal.
[0071] The method yields an interval of possible values for the
future position of the system. Thus, the method does not yield a
precise future prediction but rather a whole interval of such
positions.
[0072] The method of prediction later described relates to a
certain fundamental theorem due to Takens (see F. Takens, "On the
Numerical Determination of the Dimension of an Attractor"). This
theorem applies to all systems which can be described in principle
by knowledge of some finite set of observations all in some finite
"observation space". In other words, the theorem can be applied to
systems which in principle have a finite number of dimensions and
are such that all observations lie in some compact subset of that
space. The theorem states that it suffices to produce (almost) any
sufficiently large set of observations about the system at time t
in order to predict the system behaviour at some future time as
well as if one had known the original state of the system.
[0073] It is therefore possible to use a one dimensional time
series of observations (for example relating to the positions of a
train) to "re-construct" a system equivalent to the actual system
simply by using the observation at the present time and a certain
number of observations in the recent past.
[0074] This general approach makes it possible to use any method of
functional approximation to obtain a computational scheme for
getting the prediction. Possible techniques for this are local
linear predictions, neural networks (see James Freeman, "Simulating
Neural Networks with Mathematica", Addison-Wesley 1994), splines
(see W. Press et al, "Numerical Recipes in C", Cambridge University
Press, 1998), wavelets (see Y. Meyer, "Wavelets and Applications",
Springer 1992).
[0075] This ensemble prediction spans an entire interval of
relevant time points [0,dt], where dt is some time interval chosen
by the operator. By this, it is meant that for each interval of
time between 0 and dt there is calculated a corresponding interval
of train positions. Taking a very simple example, if an interval of
time of 5 minutes is taken (dt=5) and the train is travelling at
constant speed equal to 30 m/s on a single track, the corresponding
predicted interval of position (with a prediction based on future
constant velocity being used) is [0, 9000 m]. This is because the
train will travel 9000 m after 5 minutes at 30 m/s.
[0076] If the prediction model predicts that two trains will have
the same, or an otherwise overlapping position interval on the same
track for the same time period then a collision is predicted.
Optionally, a prediction of near misses, (i.e. closely adjacent
position intervals) may also be made so that a factor of safety can
be built into the system. Clearly the longer the time intervals
chosen, the larger the position intervals and the more chance of
overlap. However, longer time intervals give rise to less accurate
position intervals since any prediction algorithm is less accurate
for events further in the future.
[0077] The prediction model used can take any suitable form. For
example (as in the simple example above), it can be assumed that
the train travels at its present velocity at a trajectory dictated
by the rail points without accelerating or decelerating. This will
yield a position interval of the train for the chosen time
interval, the position intervals of all trains in the same time
interval being compared to check for overlapping.
[0078] An upper bound prediction may be made by assuming all trains
travel at their maximum velocity. This will give a position
interval that is longer, or the same length as, the correct
interval if it were calculated using the exact train speed and
acceleration.
[0079] FIG. 5 illustrates a situation where a train travels on
track t1 from station B toward station A and changes at the point
onto track t2 in the direction of station D. A time interval dt is
allocated and it is assumed that the train travels at constant
speed V for the whole of this period. The position of the train
when the prediction is made is (t1, p) i.e. a distance p from
station A and the position of the point is defined as:
Point1={t1, t2, (r, q)}
[0080] where r and q are the distances of Point1 along tracks t1
and t2 respectively.
[0081] If the train stays solely on track t1 for the interval dt
(i.e. (p-r)/V.ltoreq.dt), then the position interval of the train
will be:
train1=[p,(p-vdt)]
[0082] If (p-r)/V>dt, the train will change tracks (assuming the
state of Point 1 is {Point1, t1(B.fwdarw.A).fwdarw.t2(C.fwdarw.D)}
and the position interval will then be:
train1={[p, r, t1], [q, q+V(dt-(p-r)/V), t2]}
[0083] Taking the example of a train travelling at 30 m/s 2.7 km
from Point1 with dt=3 minutes and assuming p=8 km and q=4 km. Here,
(p-r)=4 km and V=30 m/s, thus (p-r)/V=2700/30=90 s=1.5 minutes
which is <dt. Thus, the train will change tracks. The interval
for train1 will therefore be:
train1={[8000, 5300, t1], [4000, 4000+30(180-(2700/30)),
t2]}={[8000, 5300, t1], [4000, 6700, t2]}
[0084] The interval of train1 for the next 3 minutes is therefore
from 8 km to 5.3 km on track t1 and from 4 km to 6.7 km on track
t2.
[0085] The preferred method considers the train's trajectory as
being on a real line, labelled by the tracks on which the train is
travelling (as determined by the position of the points). The
preferred prediction algorithm uses an ensemble prediction method
to yield a range of possible positions on this real line. Actual
positions may then be determined using an inverse mapping function.
As an alternative to ensemble prediction, it is possible to use a
simple system of ordinary differential equations to perform the
necessary integrations and work out which track the train would be
travelling on taking into account the positions of the points. The
predicted motion of the train may be obtained using estimated
position, velocity and acceleration only in practice. Methods for
solving such differential equations are provided in "Differential
Equations, Dynamical Systems, and Linear Algebra" by Hirsch and
Smale, 1974.
[0086] The concept of a real line will be explained with reference
to FIG. 6 of the accompanying drawings. On the left of FIG. 6 is
shown the modelled track layout comprising tracks and points. It is
assumed that a particular train travels firstly on track t1 for 10
km before changing to track t2 at Point1. The train travels for a
further 12 km before changing to track t3 at Point2. After
travelling 22 km, the train changes to track t4 at Point3 whereupon
it travels a further 20 km. The transformation labelled F in FIG. 6
is used to convert this model into a more abstract model wherein
the train travels on a single track for 64 km. An inverse
transformation F.sup.-1 exists which maps the real line back onto
the track network. The transformation F and its inverse F.sup.-1
are used to simplify the prediction algorithm. Thus, before
prediction is attempted, the state of the various relevant points
is ascertained and the trajectory of the train is calculated on the
track network model. The present position of the train is
determined and this position is mapped onto the real line model
using transformation F. For example, suppose the train is located
at track t2 7 km from Point1. This would be mapped as 17 km (7
km+10 km) along the real line model. This is shown in FIG. 7. Once
the mapping of train position has been made, a time interval is
chosen and a corresponding position interval is calculated. Suppose
the time interval chosen is [5, 10] (i.e. the interval of time
between (now+5 minutes) and (now+10 minutes). If a simple
prediction based on present train speed is used and the speed is 10
m/s, then the position interval obtained will be [3 km, 6 km] (i.e.
the train will be between 3 and 6 km away during the time interval
chosen). This position interval is shown in FIG. 7.
[0087] This position interval is then mapped back onto the track
network model using transformation F.sup.-1. It can then be seen
that this position interval maps onto the last 2 km of track t2 and
the first 1 km of track t3.
[0088] The advantage of mapping the model to a real-line model is
that the methods of ensemble prediction can be applied to motion on
a real line, with the results then re-converted to the actual model
of the railway system using a mapping. The prediction method itself
need not be concerned with which track the train is on and the
state of any points since this is all taken care of by the
transformation mapping F.
[0089] The above method describes how to obtain a set of points on
the rail track system which cover the possible positions of the
train during the time period between the time at which the
prediction was made (0 minutes) and dt time units further on. This
would be available for each train in the system and the control
server simply checks for any overlap between the sets corresponding
to any two or more trains.
[0090] In the case that an overlap is predicted, the server could
communicate to one or any number of the trains to stop via the
communication means. It is then possible to monitor the position
data received to see whether the orders are being followed or not
and to check the updated prediction collision. After the trains
have stopped, and a prediction of collision no longer exists, it is
preferable to send a "re-start signal" to each of the trains
involved in the predicted collision so that the trains may continue
on their journey. Preferably, these re-start signals are issued
successively to ensure that a collision is not predicted due to the
trains starting on their way simultaneously. It is preferable that
one train is started first and is allowed to pass the predicted
collision interval before the other train is started.
[0091] The choice of time period dt must be such that any remedial
action is possible within the time period dt. For example, dt
should allow time for transmission of information to and from the
communication means, time for the driver to react to any stop
instruction and time for the train to come to a complete halt. A
margin of error should also be allowed to increase the safety of
the system.
[0092] It is preferable for the system to be used in a way in which
the time period dt can be dynamically set, so that the crude
ensemble prediction, using the physical maximal speed limit of the
train is not used when this would be impractical, for instance in
the vicinity of a station. In fact, the time period dt may be
determined by the speed at which trains are travelling which is
itself a factor in determining the stopping time of the train.
[0093] The ensemble prediction method is preferred because it
accounts for systems in which one does not know the "laws of
motion" of the system under study. This prediction method uses past
time and position histories to yield predictions for an interval of
positions of the trains as a function of an interval of times.
[0094] The method uses a database of train positions built up over
a period of time. More specifically, the method makes use of a
database of differenced train positions (which can be derived from
a database of absolute train positions). A "differenced train
position" is a vector describing the amounts of track covered by a
train in a series of certain time intervals. This vector can be
easily derived from another vector which describes the absolute
position of the train as it changes with time. For example, assume
the following vector describes the absolute positions of the train
(measured from an absolute datum) such that each entry in the
vector represents the train's absolute position at one second
intervals:
(0, 1, 2, 4, 7, 11, 13, 14, 14)
[0095] This is the vector of absolute train position. A vector of
differenced train position may be obtained by taking the difference
between adjacent terms like such:
([1-0], [2-1], [4-2], [7-4], [11-7], [13-11], [14-13], [14-14])
[0096] In other words:
(1, 1, 2, 3, 4, 2, 1, 0)
[0097] Since the difference in time between each successive point
of the absolute train position vector is constant, the entries of
the differenced train position vector are equal to the train's
average speed in some particular time period.
[0098] As already mentioned, the controller of the present
invention has access to a database of differenced train position
vectors representing many different journeys on the tracks of the
network. This database is referred to during the prediction
process. It is to be noted that the database need not necessarily
hold actual differenced train position values itself, any values
(such as data of absolute train position) which allow differenced
train position vectors to be extracted, may be used.
[0099] The method of predicting the train's position at time t+dt
which is made at time t will be more precisely described below.
[0100] Firstly, a sampling frequency f is fixed. This frequency
dictates how often the train is polled by the central controller to
obtain its position.
[0101] This frequency may be fixed for the system at, say, 1 second
or it can be variable depending upon the position or type of train
or even external parameters.
[0102] One would also need to establish an "embedding dimension" d.
This dimension d dictates how many samples are looked at at a time
by the controller. The process of creating a d-dimensional series
from a 1-dimensional series is referred to as "phase space
reconstruction" or "embedding".
[0103] Thus, for any particular train, at any particular time, the
controller would be in possession of a d-dimensional vector P(t)
observed and recorded at time t. This would take the form:
P(t)=(x(t), x(t-f), x(t-2f), . . . x(t-(d-1)f))
[0104] where x(t) represents the train's absolute position along
the track at time t, x(t-f) represents the absolute position one
sample earlier and so on.
[0105] Thus, for example, when f=2 minutes, d=4 and t=6 pm, one
would obtain:
P(6pm)=(x(6pm), x(5.58pm), x(5.56pm), x(5.54pm))
[0106] In other words, the controller would be in possession of the
train's position at the present time and at three successive two
minute intervals prior to the present time. This vector is often
called the "d-dimensional time series", the "time series in
embedding space" or the "embedded time series". All of these terms
are equivalent and the entries in the vector P(t) are actually
distance measurements.
[0107] The train obtains its own position preferably using a GPS
receiver and transmits this to the central controller. Each train
in the system would perform this procedure so that the controller
is able to store a d-dimensional vector for time t (the present
time) for each train in a dynamic array. The sampling frequency may
be chosen depending on whether one wishes to obtain a fine
prediction (eg across 3 minute prediction horizon) or a course
prediction (eg across a half hour prediction horizon). In these
cases, the frequency might be one second or twenty seconds
respectively. The precise optimal choice of such frequency is a
matter of trial and error. The "prediction horizon" is the period
of time between the present time and the time for which a
prediction is required.
[0108] In the following explanation, the train's position is
measured along a one dimensional track. This is made possible using
the mapping F already described. It should be borne in mind,
however, that the train may be physically travelling on any
particular track and may change track several times on its journey.
The mapping F allows one to condense this physical complication
into a simple one dimensional real line model.
[0109] A trivial example of an embedded time series will now be
given. Assume a train is travelling from x=0 at a constant speed of
1 metre per second and has its position (in metres) sampled every
second. A vector of sampled positions taken after 10 seconds would
be as follows:
(10, 9, 8, 7, 6, 5, 4, 3, 2, 1)
[0110] These are simply the positions of the train at the present
time and at previous 1 second intervals before. The most recent
data is given first. The differenced position vector at time t is
in general given by the following formula:
([x(t)-x(t-1)], [x(t-1)-x(t-2)], . . . , [x(t-n+1)-x(t-n)])
[0111] where n is one less than the total number of samples taken.
In the present case, this expression has the value:
(1, 1, 1, 1, 1, 1, 1, 1, 1)
[0112] The d-dimensional differenced embedded time series is
obtained by taking the first d values of this list. For example,
the 3-dimensional differenced embedded time series for the time
(t-1) is;
(1, 1, 1)
[0113] In this example, all the 3-dimensional embedded differenced
time series for all times comprise 3 one's. For completeness they
are listed below;
now: (1, 1, 1)
now-1: (1, 1, 1)
now-2: (1, 1, 1)
now-3: (1, 1, 1)
now-4: (1, 1, 1)
now-5: (1, 1, 1)
now-6: (1, 1, 1)
[0114] Such d-dimensional differenced embedded time series vectors
are used because it is recognised that it is the step-by-step
changes in position of the train and not absolute positions of the
train which are important in terms of being able to predict future
train positions. For example, if a train travels along a certain
speed trajectory, it does not matter whether it does this in London
or Birmingham if the track quality is similar. The presently
described prediction algorithm thus only looks at the changes in
train position and not the absolute train position. The absolute
train position may become important when it is known that the track
quality varies over the network. This will be described in more
detail later.
[0115] In general, the ensemble prediction method comprises
obtaining a d-dimensional differenced embedded time series vector
for a train which is called the "predictee". The database of
previous differenced embedded time series vectors is then searched
for vectors already observed in the past which are similar to the
predictee. The degree of similarity can be defined in numerous ways
but one convenient way is to extract all the previous vectors
having coordinates which are within a certain number of metres of
the coordinates of the predictee.
[0116] This degree of similarity is defined by an "observation
diameter epsilon." This parameter defines how close the word
"close" means. This is because the prediction method consists in
looking up, in a database of past d-dimensional observations of
previous train's differenced positions, any data points which are
"close to" the present data point. Intuitively, a small epsilon
means that there is a more stringent criterion for defining two
points as "similar" or "close to" each other. Hence, a small
epsilon means that only points which are practically the same will
be classed as such. On the other hand, a small epsilon means that
fewer data points are likely to be found near any other given point
and hence the forecast will be less reliable from a statistical
point of view. The choice of epsilon is a practical trade off
determined by trial and error. For example, suppose the following
vector is the predictee:
(1, 1, 1)
[0117] Suppose the database holds the following three vectors:
(0.6, 1.5, 1.2)
(0.1, 0.8, 0.9)
(1.2, 1.8, 0.9)
[0118] Now suppose the observation diameter epsilon is 0.7. The
method comprises checking for stored vectors having vector entries
within 0.7 of the corresponding vector entries of the predictee. If
such a vector is found, it is considered to be "similar" or "close"
to the predictee. In this example, only the vector (0.6, 1.5, 1.2)
will be considered "close" because the others contain data entries
differing from the corresponding data entry in the predictee by a
value exceeding the observation diameter epsilon.
[0119] In general, the vectors retrieved from the database will be
smaller portions of larger vectors representing an entire previous
train journey. For example, the vector extracted from the database
in the above example may be part of a larger journey having a
complete differenced embedded time series as such:
(2.2, 2.5, 1.9, 1.4, 0.6, 1.5, 1.2)
[0120] A crude prediction of the train position one second in the
future would then be given by simply looking at the differenced
data stored in the extracted data item. The entry one second
further on from the data item which corresponds to the last data
item of the predictee is looked at. In other words, since the
predictee corresponds to the part of the vector (0.6, 1.5, 1.2)
which was extracted from the database, a prediction can be
determined by looking at the next data item in the future, in this
case 1.4. The differenced time series for the present train would
then be presumed to be:
(1.4, 1, 1, 1)
[0121] which translates to an absolute train position vector
of:
(11.4, 10, 9, 8, 7)
[0122] in other words, the train's position would be predicted to
be 11.4 metres from the absolute datum at one second in the future.
In this example the ensemble reduces to a point because only one
data vector was extracted from the database in this particular
example
[0123] Of course, this example is very simple in order to aid in
the understanding of the invention. In reality, a value for d of
more than 3 would be chosen and many vectors would be extracted
from the database that are close to the d-dimensional predictee.
The next entry from each of these vectors would be looked at and
the highest and lowest would be identified so as to obtain a range
of possible train positions in the future. This is explained in
more detail below.
[0124] It is also to be noted that it is not necessary that only
the next entry in the extracted data vector is looked at. An entry
any number of samples forward in time could be examined. For
example, if it is necessary to obtain a prediction of the train's
position in 4 seconds time using the example above, the next four
values would be taken from the extracted data item and added to the
predictee so as to obtain:
(2.2, 2.5, 1.9, 1.4, 1, 1, 1)
[0125] This gives a predicted train position of
(10+1.4+1.9+2.5+2.2)=18 metres. In a similar way, the train's
position after 3 seconds, 2 seconds etc can be obtained. The above
method gives the train's position after 3 seconds as 15.8 metres
and after 2 seconds as 13.3 metres. Of course, the further away in
time that a prediction is, the less accurate it will be, since
there will be more time for the train in question to do something
differently to the trains whose position created the stored
database data. However, if the database is large enough so as to
store the data corresponding to a great many journeys, many points
will be extracted as being similar to the predictee and a range of
possible positions can be given for the future position of the
train with a relative degree of certainty.
[0126] The present invention will now be described by way of a
specific example. Firstly, consider the velocity profile shown in
FIG. 8. As can be seen, the train accelerates to 2 metres per
second in the first three seconds, travels constantly at this speed
for two seconds and then decelerates to 1 metre per second over the
next 2 seconds where it stays for a further second. FIG. 9 is a
graph showing the absolute position of the train as it moves.
Suppose we examine what happens when it is required to make a
prediction of the train's position in one second at a time of t=8
seconds. At this point in time, the past history of train positions
is given as:
(11, 10, 8.75, 7, 5, 3, 1.33, 0.33)
[0127] The corresponding differenced time series vector will
be:
(1, 1.25, 1.75, 2, 2, 1.66, 1)
[0128] It can be seen from FIG. 7 that this vector corresponds
closely to the average speed of the train at a point midway between
adjacent data entries.
[0129] An embedding dimension d of 5 will be chosen for this
example. Thus, the 5-dimensional differenced embedded time series
vector is:
(1, 1.25, 1.75, 2, 2)
[0130] This vector is the predictee used in the prediction
process.
[0131] We now assume we have a database of past trajectories which
not only correspond to a succession of time points similar to the
succession of time points in the predictee (ie every one second)
but also comprise additional differenced position data for time
points in the future. Our assumed database is represented in the
matrix below:
1 t t-1 t-2 t-3 t-4 t + 1 0.3081 0.7572 1.6665 1.8042 1.0224 0.9738
0.8606 1.2352 1.1094 1.2623 1.9732 1.2378 0.1067 0.4861 1.1673
1.3146 1.0329 0.7897 0.2228 0.6396 0.7614 1.9517 1.0146 1.1353
0.0875 0.5845 1.2877 1.9517 1.5396 0.5400 0.7106 0.5549 1.4907
1.2868 1.2796 0.6067 0.3777 0.2602 1.5976 1.7967 1.1807 1.2816
0.4615 1.0598 1.1505 1.7077 1.9087 0.8332 0.1159 0.6344 1.7036
1.0481 1.8310 0.5133 0.3886 0.4027 1.6359 1.3508 1.8852 0.8666
0.3168 1.1167 1.2859 1.9287 1.4188 0.7740 0.7447 1.0115 1.7340
1.6153 1.2427 0.7648 0.5919 1.0543 1.2378 1.2867 1.1326 0.5982
0.0052 0.3833 1.2642 1.0967 1.9797 0.7032 0.0559 1.2092 1.4643
1.3227 1.4138 1.2750
[0132] The above matrix represents the entire database of stored
points. The last column of the matrix shows the points stored for
the time slot after the time slot represented by the first column
(which corresponds to the most recent time slot in the
predictee).
[0133] The extraction process comprises choosing an observation
diameter epsilon which determines how close stored data points must
be in order for them to be used. In this example we will use an
observation diameter epsilon of 0.79. All stored vectors whose
individual entries differ from the corresponding individual entries
of the predictee by a value exceeding 0.79 are discounted. Applying
this in the present case leaves the following matrix:
2 t t-1 t-2 t-3 t-4 t + 1 0 0 0 0 0 0 0.8606 1.2352 1.1094 1.2623
1.9732 1.2378 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.7106 0.5549
1.4907 1.2868 1.2796 0.6067 0 0 0 0 0 0 0.4615 1.0598 1.1505 1.7077
1.9087 0.8332 0 0 0 0 0 0 0 0 0 0 0 0 0.3168 1.1167 1.2859 1.9287
1.4188 0.7740 0.7447 1.0115 1.7340 1.6153 1.2427 0.7648 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0
[0134] In order to obtain a prediction, the column representing
"t+1" is interrogated. The ensemble prediction is made by looking
at the range of values occurring in this column of the matrix. In
this case the values lie in the interval [0.6067-1.2378]. The
controller is therefore able to conclude that, on the basis of past
data, the value for the future position of the train, in one
second's time will be given by the interval between 0.6067 and
1.2378. This corresponds to a possible range of absolute positions
of [11.6067-12.2378]. This interval is shown graphically in FIG.
9.
[0135] If necessary it is possible to make a point prediction
rather than an ensemble prediction by taking an average of all of
the future predictions given by each data vector extracted from the
database. In the above example, this point would be the average of
1.2378, 0.6067, 0.8332, 0.7740 and 0.7648 which is 0.8433.
Alternatively, a point prediction can be obtained by taking the
simple average of the lowest and highest figure obtained. In the
above example this point prediction would be the average of 1.2378
and 0.6067 which is 0.9223.
[0136] It can be seen, therefore, that the above method can be used
for each train on the network to predict a possible range of
positions at the certain time point anywhere in the future. These
ranges of 1-dimensional positions can then be converted to a range
of physical positions using the mapping f.sup.-1 as shown in FIG.
7. It can then be ascertained whether any of the predicted physical
positions for one train overlap with any of the predicted physical
positions for another train on the network. If this is the case,
appropriate action can be taken to remedy the situation. For
example, one or both trains could be ordered to slow down so that
its trajectory changes and the predicted range of physical
positions changes accordingly.
[0137] The method of searching the database may advantageously take
the form of cross-correlating the predictee with the entire
database. The resulting correlation data can be examined to see
which parts of the database have the highest correlation with the
predictee and these can be extracted. Then, predictions can be
based on data extracted from those parts of the database which
represent the appropriate forward-in-time position for each
identified point of the database. Another method to compare the
predictee with the database is to firstly compare the first
coordinate of the predictee with the entire database of stored
coordinates. Then stored coordinates within the observation
diameter epsilon of the coordinates of the predictee can be
identified. Next, the second coordinate of the predictee is
compared with the coordinates immediately following the identified
coordinates. This carries on until all the coordinates of the
predictee have been compared. The method identifies stored
coordinate strings having coordinates that are each within the
observation diameter epsilon of the respective predictee
coordinates.
[0138] The method of ensemble prediction may be thought of as a way
to use prior knowledge of how trains have reacted in the past in
order to predict any train's actions in the future. The method of
extracting similar vectors from a database may be likened to
comparing the present trains velocity curve with a set of stored
velocity curves. Any stored velocity curves which are similar to
the velocity curve of the train under consideration will be
examined and a range of velocity data at a point of time in the
future can be obtained. The use of an observation diameter epsilon
in matching velocity trajectories is only one way of sorting the
database. Any known curve matching method could be used to the same
effect, as long as such a method is based on a topology derived
from the usual metric on Euclidean spaces.
[0139] A refinement of the above system is to store in the central
data base a track classification which can be used in the
prediction process. For example, part of the track might be very
old, and uneven, and hence one would expect the train to move
differently, brake more slowly, etc. This can be reflected in the
prediction algorithm by defining grades of track. For example,
portions of track could be designated Grade A, Grade B, etc and the
database used for predicting the motion of the train might draw
only on past data points drawn from the same type of track. In
other words, different databases would be built up for different
track grades. If it is determined that a train is travelling on a
certain grade of track, only previous trajectories recorded for the
same type of track are used to provide the prediction.
[0140] This may be done by coding the predictee and stored vectors
as above, with an extra type classification of the track as a last
coordinate. The definition of distance may be changed by adding an
infinite term to any pair of data points with different type
classification coordinates. Adding such an infinite term would
effectively disqualify the data from being considered.
[0141] As well as quality of track, further coordinates such as the
distance of the train from the next railway point could be
considered. These may be relevant for the prediction because it can
be assumed that part of the deterministic data that determines the
train's progress is the information about when it needs to cross a
point or what the quality of line is, which is associated with
friction and possible instability of motion.
[0142] A further refinement is to label certain parts of the track
in terms of a speed limit. The speed limits for a particular track
would be an extra coordinate in the track definition vector, t1.
The ensemble prediction method can be used to predict the velocity
of the train in the future (as shown in FIG. 8) and it can be
determined whether this velocity is likely or possible to exceed
the track speed limit at the part of the track from which the
prediction is made. If so, appropriate remedial action can be
taken.
[0143] There is a choice as to how to build up the empirical
database used for predictions. For example, the database could be
quite segregated so that a certain train travelling on a certain
track quality will have its trajectory compared only with
trajectories obtained from similar trains travelling on similar
tracks. Alternatively, one can use data obtained from a whole host
of trains on a whole series of tracks when making any particular
prediction. It is an empirical question as to how precisely one
should segregate the data so as to "learn" a good prediction
method.
[0144] The above described method works well both when movement is
easily predictable and when it is more erratic. It is expected that
the predicted intervals will be very accurate when the method is
applied to a real train network.
[0145] As can be seen from the above, the present invention
provides a system which uses relatively simple and commonly known
components to avoid train collisions. Mobile telephones and GPS
receivers have been commercially available for some time and the
cost of such devices is small compared to the cost of installing
ATP. By using a central server which monitors the position of all
the trains, intelligent action can be taken to minimise the
possibility of collisions and to minimise disruption to the
service. Furthermore, the system can be completely automated by
operationally linking the communication means with the train's
power system so that the train may be stopped by remote
control.
[0146] It is to be noted that the present invention does not rely
at all on the existing signal system which is known to be
unreliable and prone to fault. Since there are substantially no
mechanical moving parts involved, the present system is extremely
reliable and requires little maintenance once set up.
[0147] The present invention could be used to replace the present
signalling system and conventional signalling could be phased out
as the invention is more widely used.
[0148] The instruction to stop issued to the communication means in
the event that a collision is predicted could be relayed via aural
or visual means on the train and may be issued for the attention of
the driver or another person situated in the train cabin whose job
it is to monitor the mobile telephone acting as the communication
means. The instruction could be issued to all drivers involved in a
predicted collision, or just a subset of all drivers. Independent
alerts in the manner of the present invention markedly decrease the
likelihood of an actual crash since it is generally necessary that
at least two of the drivers warned must ignore the warning for a
crash to occur (which is the square of the probability of one
driver ignoring a warning), and in any event the present invention
can provide for an automated link between the controller and the
braking system of the train. This contrasts with previous systems
(such as ATP) wherein only one driver is warned that a collision is
possible.
[0149] A further advantage of the invention is that it can be
implemented quickly without a need to modify the train or track
infrastructure.
* * * * *