Académique Documents
Professionnel Documents
Culture Documents
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 2, JUNE 2012
I. I NTRODUCTION
Fig. 1.
S=
(VTrain VWheel )
.
VTrain
(1)
NOORI AND JENAB: INTELLIGENT TRACTION CONTROL FOR SPEED SENSOR VEHICLE IN TRANSIT SYSTEM
681
Fig. 3. (a) Pulse rise/fall using the sensors potential (in volts). (b) Pulse
rise/fall using the sensors current (in milliamperes).
Currently, most of existing traction control systems are vehicle oriented and are the application of vehicle mechanics and/or
vehicle dynamics, in which existing rail traction controllers
use either pure mechanical components or electromechanical
devices to detect and maintain the traction conditions in the
component level [8][12].
In some approaches, the traction was considered as an application of electrical propulsion, along with pulsewidth modulation in [13] and [14], control, and torque mitigation [15]. These
components can be controlled through the logic circuit, track
circuit, and power lines [16][20]. In addition, the electrical
and power applications of the traction can be sensor based and
are controlled by an embedded computer. These systems are
considered hard real time and work in very low application
cycles (e.g., 10 ms); there is no time for traction compensation
due to the short application cycle and noisy nature of the
environment [21].
Embedded traction control processors are mainly independent, with no intrusive interaction with the central computer
due to the fast action module and slow communication backbone. Power traction control is necessary for railway signaling;
however, according to the pertaining literature, there is still no
connection through the integrated system under automatic train
protection (ATP) and automatic train operation (ATO) [6].
Modern traction control has become very popular over the
past few years due to the ability of new methods to address uncertainty, learning ability, and adaptability [22][26].
Since slip/slide is an uncertain phenomenon- and time-varying
process, intelligent logic is a good choice for slip/slide
control. An artificial intelligent method was introduced in [27]
for processing, simulation of the traction control model, and
682
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 2, JUNE 2012
NOORI AND JENAB: INTELLIGENT TRACTION CONTROL FOR SPEED SENSOR VEHICLE IN TRANSIT SYSTEM
683
Fig. 6.
(6)
In addition, risk and loss function is another term to calculate the cost of decision making and expected loss (risk).
Let {1 , 2 , . . . , c } be the set of classes (state of natures),
and let {1 , 2 , . . . , a } be the set of possible actions to be
taken. Then, (i |j ) is the loss function of action (1 ) when
the state of nature is (j ). As a result, Bayesian risk can be
formulated by
c
R (i | X) =
(i | j ) P (j | X) i.
(7)
j=1
Now, the decision can be formulated for minimum conditional risk that is given by
Decide on j :
if R (j | X) = min {R (1 | X), R(2 | X), . . . , R (a | X)}
(8)
where i = 1, 2 . . . a, and j = 1, 2 . . . c.
(2)
p(X | j )P (j )
.
p(X)
(3)
(X
)
. (4)
2
(2)d/2 ||1/2
684
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 2, JUNE 2012
Fig. 7.
( wl )
.
(pw ppr )
(12)
NOORI AND JENAB: INTELLIGENT TRACTION CONTROL FOR SPEED SENSOR VEHICLE IN TRANSIT SYSTEM
Fig. 8.
685
(14)
Total
(j ) =
Oi
(15)
Total
v(j ) =
[Oi (j )]2
Total
(16)
P (X | j )P (j )
.
P (X)
(18)
686
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 2, JUNE 2012
applied for all possible conditions; however, the classification is based on the greatest value of all posterior
probabilities. On the other hand, the greatest posterior
probability is the classified traction for that input data.
Repeat this process until the end of the input stream, or
stop the application in case of failure or implausible conditions. In case of the equal posterior probability or other
issues such as noise and lack of data, which might happen
in some application cycles, the error and risk of classification need to be calculated, and the classification would
be based on the lowest error and/or lowest risk of action.
2) Error classification is used when the posterior probability
does not give the certain classification (either the posterior probability normal condition is equal to slip condition
or the posterior probability normal condition is equal to
slide condition). The error of classification for each state
of nature for the given delta speed and train is calculated
by (6). Overall, the model decides on the state of nature
with minimum error of decision, which is formulated by
P (error | X) = min {P (1 |X), P (2 |X), P (3 |X)} . (19)
There is no possibility that either all posterior probabilities are equal or posterior probability slip is equal to
slide. In such case, the data input is discarded and waiting
for the next cycle input (see Fig. 8).
3) Risk is used when posterior probability does not give the
precise decision (the same applied conditions that kick
in the error calculation). There is a risk associated with
each decision, and loss factors can be assigned to each
decision action on the state of natures (i.e., slip, normal,
and slide). These loss factors are used to calculate risk by
using (7) and (8). The classification is made based on the
lowest risk of action on the state of natures for the given
delta speed and train speed. The risks of action in (8) are
given by
R(i | X) = [(i | 1 )P (1 | X)]+[(i | 2 )P (2 | X)]+
(20)
where i = 1, 2, 3, . . . , a. In addition, the decision rules in
(7) are formulated by
Decide on 1 :
if R(1 | X) = min [R(1 | X), R(2 | X), R(3 | X), . . .]
(21)
Decide on 2 :
if R(2 | X) = min [R(1 | X), R(2 | X), R(3 | X), . . .]
(22)
Decide on 3 :
if R(3 | X) = min [R(1 | X), R(2 | X), R(3 | X), . . .]
(23)
respectively.
D. Safety and Plausibility Step
This step is to make sure that the traction classification is
safe and plausible among the sensors. Basically, the detected
traction for speed sensors 1 and 2 should be the same or within
the defined tolerance (see Fig. 9). In addition, the acceleration
Fig. 9.
NOORI AND JENAB: INTELLIGENT TRACTION CONTROL FOR SPEED SENSOR VEHICLE IN TRANSIT SYSTEM
687
classified traction with the traditional tractions. Patter recognition is done during the engineering phase, and the technical
team needs to know all the aspects of the project from vehicle
characteristics to guideway topology. Pattern classification and
plausibility phase are done in real time and can be loaded in
the target system, depending on the application and configuration. The succeeding sections are to discuss the effectiveness,
performance, and results of the implemented model.
A. Training Data Discussion
A training data set is simulated and compiled from the field
data. Field data need many filtration and deletion due to the
nature of field data. For instance, in the real field data, a train
can sit in the line for hours, which will cost multimegabyte data
with no use; in addition, some of the logs may have no slip/slide
condition and thus are not good candidates for training. The
training data set needs to have all traction conditions and all
vehicle behavioral characteristics; this can be achieved during
the system-engineering phase of the project and would alleviate
the later cost of field tests. In this paper, the VisSim application
is used for simulation of the stop-and-go motion patterns,
which include all possible traction conditions during a course of
daily operation. The training sample includes engineered train
characteristics such as maximum acceleration/deceleration and
braking curve to comply with the actual condition in the field.
Acceleration and deceleration are the key players in the traction because, generally, the slip (spin) happens in acceleration
during the traction and the slide happens in the deceleration
during the braking. Weather conditions (snow and rain) contribute in the traction condition, as such; the wheel adhesion to
the rail is very dependent on the dryness and wetness of the
road. However, in the training samples, the weather condition
is ignored.
For the input analysis, the wheel speed in the training set is
analyzed by BestFit software to find out the best probabilistic
distribution that matches the curve. The distribution chart in
Fig. 10 for wheel speed data samples shows the goodness-of-
688
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 2, JUNE 2012
rate for finite separable elements [43]), which will decrease the
variance and estimation as it gets closer to density. However,
for the sake of practicality, in this paper, 2000 samples are
introduced for training purposes that includes all applicable
traction conditions and vehicle behavior based on the field
data and simulated/engineered data. This number of record
(2000) is sufficiently enough to prove the practicality and
precision of the proposed model. Although we used this number
of training samples, for real application, a higher number of
training records is suggested. For other projects, the number of
records may go up to millions, depending on the application
and required conditions.
B. Input Data Discussion
Fig. 13.
The input data set is collected from a field test that includes
2000 records. These records are sorted into four columns of
data, i.e., speed sensor 1 signal (wheel speed 1), speed sensor
2 signal (wheel speed 2), train speed (witness data), and actual
traction condition (witness data). Each record of data represents
data for an application cycle. As such, the number of inputs
does not affect the classification result. This number may go
up to millions, depending on the hours of operations and
application conditions.
Since the models are implemented and tested in the laboratory, the input data injected into the model (record by record)
and the classification results (record by record) are collected accordingly. The implemented model runs through all the records
to classify the traction for each record (each row). The given
train speed and the given traction condition are for witness and
validation purposes only. In fact, the model calculates the wheel
speed and train speed, and cross compares for plausibility, as
discussed in Section IV-C.
The model classifies the data input, regardless of the number
of records; however, for sake of practicality, the authors run the
program with 2000 input records, and it proves the effectiveness
and precision of the model. The model is able to classify the
traction based on the vehicle pattern in case of noise, but
the traditional sensor-based traction control failed to classify
the condition (unknown). In most of the cases, due to the safety
criticality of the rail application, the lack of robustness and
healthiness of the input data may cause application halt.
C. Result Discussion
The output of the model is traction classification, along
with plausibility check among the sensors (speed sensors 1
and 2, and the accelerometer). As a result of the classification
exercise and as a part of the validation, the model output
is compared to the field test traction result to validate the
results, as described in the previous section. For the sake of
comparison, the proposed model and traditional model run on
the same data collected from the field and compare the results.
As a result of this comparison analysis, the intelligent traction
classification is improved by 25% over the traditional results.
Improvement of 25% is conducted from the counting of the
number of misclassifications in the field test log versus the
classification output. Therefore, the intelligent model demonstrates better classification, more precise decision on noise, and
possible missing data, along with low cost of error and risk of
misclassification.
In fact, the model works with the data, which has neither
signal nor noise that the traditional traction classifier cannot
process. The traditional sensor-based model is easily misled by
noise or mistaken as input, but the intelligent model ignores
those noisy data and follows the vehicle motion pattern. Fig. 13
shows a snapshot of the 100 cycles for illustration purposes
only and can be extended to the rest of the cycles but cannot be
displayed in the chart due to lack of space. The perforated line
represents traditional classification, and the solid line represents
Bayesian classification. The traction results (Unknown = 0,
Slip = 1, Normal = 2, and Slide = 3) are on the y-axis, and
the records are on the x-axis. The traditional method has
many unknown tractions, along with many misclassifications,
compared with the proposed model. Traditional data come from
the field test log, and it is known that the field test suffers from
missing data due to noise and misclassification.
VI. C ONCLUSION
The proposed model has focused on the speed sensor input
(wheel speed and train speed) to detect the traction condition in
an intelligent format. The intelligent model has been derived
from the Bayesian decision theory, which was implemented
in the form of pattern recognition of traction conditions. The
practice of implementation has shown that the proposed model
can be a suitable candidate for such a system due to fact that
the number of traction conditions is limited to the few motion
conditions, which is very manageable for computation.
Although the model itself is independent from the data, for
the sake of accuracy of the patterns, we need to train the system
with an ample number of training samples. Due to the high
rate of convergence and fast learning rate, the model works
with a limited number (i.e., 2000) of training samples. The
proposed model is relatively simpler than the traditional system
and is more precise in terms of traction classification. In fact,
the training sample should include all aspects of the traction
conditions and characteristics, which need to be engineered
prior to the design of such system.
The computer-based intelligent train traction safety model
based on Bayesian decision theory has less error and less risk
of misclassification, compared with a similar system that is
NOORI AND JENAB: INTELLIGENT TRACTION CONTROL FOR SPEED SENSOR VEHICLE IN TRANSIT SYSTEM
689
690
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 2, JUNE 2012