A DECISION SUPPORT SYSTEM BASED ON NEURO-FUZZY

SYSTEM FOR RAILROAD MAINTENANCE PLANNING

Michele Ottomanelli

Dept. of Environmental Engineering and Sustainable Development, Technical University of Bari

Viale del Turismo, 8 – 74100 Taranto, Italy

Mauro Dell’Orco, Domenico Sassanelli

Dept. of Highways and Transportation, Technical University of Bari, Italy

Via Orabona, 4 – 70125 Bari, Italy

Keywords: Decision Support Systems, Neuro-Fuzzy, Railroad Maintenance

Abstract: Optimization of Life Cycle Cost (LCC) in railroad m

aintenance, is one of the main goals of the railways

managers. In order to achieve the best balance between safety and operating costs, “on condition”

maintenance is more and more used; that is, a maintenance intervention is planned only when and where

necessary. Nowadays, the conditions of railways are monitored by means of special diagnostic trains: these

trains, such as Archimede, the diagnostic train of the Italian National Railways, allow to observe every 50

cm dozens of rail track characteristic attributes simultaneously. Therefore, in order to plan an effective on

condition maintenance, managers have a large amount of data to be analyzed through an appropriate

Decision Support System (DSS). However, even the most up-to-date DSSs have some drawbacks: first of

all, they are based on a binary logic with rigid thresholds, restricting their flexibility in use; additionally,

they adopt considerable simplifications in the rail track deterioration model. In this paper, we present a DSS

able to overcome these drawbacks. It is based on fuzzy logic and it is able to handle thresholds expressed as

a range, an approximate number or even a verbal value. Moreover, through artificial neural networks it is

possible to obtain more likely the rail track deterioration models. The proposed model can analyze the data

available for a given portion of rail-track and then it plans the maintenance, optimizing the available

resources.

1 INTRODUCTION

This study is addressed to tamping operation in

railways maintenance. It is compacting and forcing

the ballast against the rails and sleepers. Such an

operation is frequently carried out by railways

companies, with the exception of the new slab

tracks, where the ballast is substituted by concrete or

asphalt slab.

The aim of the tamping is to improve geometrical

arameters of the railway track such as alignment,

longitudinal level, super-elevation, gauge and

buckling, to reach a higher safety level of the

railway.

A variety of mechanical (a

utomated) tamping device

are available: they are able to tamp ballast under 2 or

3 sleepers at the same time, thus it is possible to

tamp up to 2200 meters of track per hour.

We have to point out that frequently repeated

amping operations could cause negative effect for

the ballast. In particular, with respect to the effect on

the ballast aging, each tamping cycle is equivalent to

a 20 Megatons (MGT) of cumulate traffic, due to

increasing of the finer material percentage.

During the life cycle, a given track shows three

ifferent phases with respect to the capacity of

preserving its original geometrical and

morphological characteristics.

For example, let us assume as track quality index a

arameter that shows the deterioration of the

geometrical layout, such as the standard deviation

(SD) of the track alignment. In Figure 1 such a

parameter has been connected to the traffic,

Ottomanelli M., Dell’Orco M. and Sassanelli D. (2005).

A DECISION SUPPORT SYSTEM BASED ON NEURO-FUZZY SYSTEM FOR RAILROAD MAINTENANCE PLANNING.

In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 43-49

DOI: 10.5220/0002524500430049

 SciTePress

expressed in MGT of trains load, accumulated from

the track operation starting. The aging curve is made

up of three sections: two are curvilinear and one

usually is quasi-linear. The aging curve represents

the behaviour of the track when no maintenance

operations are carried out.

The first portion (a) of the curve is the youth phase

of the track and it shows a quick increase of the SD,

due to an early track settling: the higher SD value is

the lower is the ballast compactness. It is very

difficult to foresee how long the youth phase lasts

since it depends on a number of factors. Moreover,

such a phase could be affected also by the part of the

railway track we are not dealing with.

The second phase, usually known as intermediate

phase, is represented by a less than linear curve part

(b). This phase starts when the track settling is near

to be completed and, consequently, the track aging

ratio can be assumed as constant. The intermediate

phase lasts for most of the track lifetime; for this

reason and for sake of computing simplicity, a lot of

algorithms consider only this part of the track

lifetime.

The last phase (old phase) is representative of the

tracks near to the end of the life-time. This curve

section (c) shows a quick aging of the track: the

value of SD increases in approximately as in

exponential way. Generally, in order to preserve

railway safety through proper maintenance

operations, this phase should be avoided.

Dashed line in figure 1 represents the lower bound

of SD, that is the maximum improvement that can be

obtained for the track through tamping operations.

From the same figure it is possible to note that

effectiveness of tamping decreases as the aging of

the track increases; thus, starting from a certain time,

it should be more economical to renew ballast, rails

or sleepers than carrying out tamping operations

since they would become more frequent and less

effective.

Existing Decision Support Software (DSS), such as

Ecotrack, usually has the following drawbacks:

a) the phases a and c are not taken into account,

thus the aging function is assumed to be linear

for the whole life-time of the track;

b) the considered aging function is assumed to be

the same for the whole rail track;

c) such models assume as intervention thresholds

some rigid and crisp values.

Actually, the approximation imbedded in the point

1) is not suitable from the safety standpoint. In fact,

because of several different reasons such as bad

maintenance programs or problems concerning local

conditions of the superstructure or roadbed, the track

deterioration could be quicker than linear. In this

case, the thresholds should be crossed sooner than

the DSS foresaw.

In figure 2 such a situation is depicted referring to a

200m long track segment. Track alignment was

observed each 50 cm, then 400 observations are

Figure 1: Aging Function

ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

Actual trend

Above-threshold

values

T1 T2

available for each segments: on the basis of the

observed values, the DSS determines the Standard

Deviation (SD). The ordinates represent the value of

SD for the alignment with respect to time. The

vertical lines T1, T2 e T3 represents the tamping that

have been carried out in the last years. The

horizontal dashed line represents the threshold value

suggested by the European Rail Research Institute

(ERRI) for the acceptability of the observed

parameter.

From the figure 2, it is possible to see that linear

extrapolation of SD can mislead, since in case the

track-segment is in its old phase -as frequently

happens in reality, the date for tamping maintenance

would be later than necessary.

Concerning the statement 2), it has to be pointed out

that different track-segments usually have different

geometrical deterioration trend and different wear

and tear of the rails, even under the same load

conditions.

European railways companies have widely verified

this issue, thus the general three-phases shape of the

aging function (figure 1) should to be adjusted for

each track-segment in order to take into account

local characteristics of the track.

The issue in the statement 3) has got conceptual

origin. Traditional DSS’s are quite rigid, since they

assume crisp values for the thresholds: as the

observed parameter exceed the threshold, the

decision rule suggests maintenance. An example of

this kind of rule could be:

IF (SDalignment > 1.4) THEN tamping

Figure 2: Example of incorrect extrapolation

1997

1999

Threshold value

2001

Linear trend

Reference time

T1 T2

(<0)

19981997

Over threshold values

1999

days

SD (>0)

20012000

threshold

days

2002 2003

∆SD

∆days

Figure 3: Scheme of the variation of quality indexes.

A DECISION SUPPORT SYSTEM BASED ON NEURO-FUZZY SYSTEM FOR RAILROAD MAINTENANCE

PLANNING

This approach does not allow to take into account

how much the threshold has been exceeded; in fact,

such a DSS gives the same importance to very

different values of the SD, for example 1.4 and 3.0.

In other words, it gives the same importance when

the SD of the parameter exceeds “very slightly” or

“heavily” the given threshold.

low-medium

low

medium

medium-high

2 METHODOLOGY

In order to overcome some of the mentioned

drawbacks of traditional DSS, in this paper we

propose a new methodology based on an Adaptive

Network-based Fuzzy Inference System (ANFIS).

Using the proposed ANFIS algorithm, it is possible

to calibrate the Membership Functions (MF) of the

Fuzzy Inference System (FIS), with reference to a

given track-segment. Subsequently, the results

obtained by the proposed model have been

compared to those obtained by using Ecotrack.

Figure 5: Membership function for ∆SDvertical

To test the robustness of the two DSS, we also

simulate some rough errors in data observation.

The FIS has been applied to 200 m long track-

segment. In figure 3 is the scheme of changes in

time of the track-segment quality indexes. The

proposed Fuzzy Inference System has been specified

with three input elements and one output element; in

particular as input elements we have considered the

SD of alignment, the SD of vertical level and the

number of days past from last tamping. As output

the FIS provides with an estimate of the date for

tamping works on the track-segment.

The Standard Deviations of the two input parameters

have been represented through the following five

Gaussian Membership Functions (MF):

low, low-medium, medium, medium-high, high.

The output is made up of singletons, indicating the

number of days, starting from the date of the

analysis, before the next tamping.

The time period “Ddays” is the number of days

between the tamping carried out straight before each

measurement and the date of the survey P, while

DSD

alignment

and DSD

vertical

are respectively the

differences between given threshold values and the

SD of alignment and the SD of longitudinal level. It

is evident that, if the threshold value is exceeded, the

differences DSD reach negative values.

Such a FIS has been calibrated through an ANFIS

algorithm: all the MFs have been calibrated by

training the Artificial Neural Network (ANN)

relevant to the FIS. The structure of the adaptive

ANN is chosen by the analyst, which chooses also

the number of MF to be associated to each input and

output. The ANN represents an analytical tool able

to learn and emulate the behaviour of a real system.

The ANN for a given time interval operates side by

side with the real system (training phase) and

modifies its characteristics until the error done is

minimised. The characteristics of the ANN of the

given ANFIS, such as input and output nodes and

the number hidden layers are relevant to the FIS.

Given a certain number of input−output pairs, the

error ε is the difference between the output value Y*

determined by the ANN and the “true” output value

Y (so-called “pattern”):

low-medium

low

medium medium-high high

Figure 4: Membership function for ∆SDalign.

ε = |Y* -Y| (1)

The characteristics of an adaptive ANN are the node

functions, which correspond to the MF of the FIS.

Those functions are to be calibrated during the

training phase. The node functions correspond to the

Membership Functions of the FIS referring to; such

a functions are calibrated during the training phase.

3 MODEL SPECIFICATION AND

CALIBRATION

To carry out the specification and calibration of the

proposed model, a 200 m long track-segment has

ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

been considered. The following features were

available for this segment:

– maximum speed of the railway line 140 km/h;

– laying date for all components of the track, like

ballast, rails, sleepers and fastenings, 1/1/1985;

– during the period February 1992 - July 2002, ten

measurements a year for SD alignment and SD

vertical level have been carried out;

– in the same period, three tamping have been

carried out, namely in January ‘92, August ‘96,

October 2001;

– the period from 1/1/2003 to 31/12/2007 has been

considered as planning period.

On the basis of the International Railways Union

(UIC) rules, the maximum acceptable value for SD

alignment is 1.4, while for SD vertical level is 2. In

this paper, three input and one output has been

chosen for the FIS; the output is the date of

intervention, while the input are ∆SD

alignment

∆SD

vertical

, ∆days, where:

– ∆SDalignment = 1.4 – measured SD alignment;

– ∆SDvertical = 2.1 – measured SD vertical

level;

– ∆days are the days past from the last tamping.

A set of 107 pairs of input- output vectors has been

used; in the table 1 a sample of the input database is

reported.

Table 1: Example of the input database

DATE SD

align.

vert.

∆SD

align

∆SD

verti

∆days

15/1/92 1,23 1,78 0,17 0,32

27/1/92 TAMPING

15/2/92 0,88 1,06 0,52 1,04 18

15/3/92 0,87 1,08 0,53 1,02 48

15/4/92 0,89 1,08 0,51 1,02 78

15/5/92 0,89 1,12 0,51 0,98 108

15/7/92 0,91 1,12 0,49 0,98 168

15/8/92 0,89 1,16 0,51 0,94 198

15/9/92 0,89 1,14 0,51 0,96 228

.... .... .... .... .... ....

15/4/01 1,26 2,38 0,14 -0,28 1667

15/5/01 1,53 2,53 -0,13 -0,43 1697

15/7/01 1,52 2,5 -0,12 -0,4 1757

15/8/01 1,5 2,52 -0,1 -0,42 1787

15/9/01 1,54 2,59 -0,14 -0,49 1817

11/10/01 TAMPING

15/10/01 1 1,65 0,4 0,45 4

15/11/01 1,03 1,65 0,37 0,45 34

15/1/02 1,05 1,68 0,35 0,42 94

.. .. .. .. .. ..

15/5/02 1,09 1,78 0,31 0,32 214

15/7/02 1,11 1,84 0,29 0,26 274

This set of input vectors has been divided into two

groups:

– training vectors, used for training a neural

network, that will subsequently calibrate the

MF’s;

– checking vectors, used to check the model.

The MF’s used in our case are gaussian curves,

characterized by mean and standard deviation; the

output is a singleton.

The training results consist in calibrated MFs both

for input and output, as well as the rules of the

inference engine. In the following figures 4 and 5

the MF’s for ∆SD

alignment

and ∆SD

vertical

, respectively,

are reported. Of course, when both ∆SD

alignment

and

∆SD

vertical

are 0, the thresholds are reached; then, the

pair [0 0] as input allows to forecast when these

thresholds will be reached.

Note that not necessarily both thresholds will be

reached at the same moment. On the contrary, highly

likely this situation will never happen.

The proposed FIS uses the logical OR to get the

lowest value ∆SD as a precautionary condition. In

figure 6 are the rules of inference system obtained

by ANFIS.

4 ROBUSTNESS OF THE METHOD

TEST

A glaring mistake in measurement has been

simulated: one of the measured values has been put

over threshold, keeping hold other values. Table 2

shows the results of tests for different location of the

error; in particular, in the test 4 an error in the

second-last measure of SD of alignment has been

simulated. It is easy to see that the influence of a

measurement error on the FIS forecast is very low,

not greater than 8%. The reason is that the system

decision is based not only on a unique peak value,

but on an overall analysis of the trend, over time, of

the track parameters.

Table 2: Results of the tests

A DECISION SUPPORT SYSTEM BASED ON NEURO-FUZZY SYSTEM FOR RAILROAD MAINTENANCE

PLANNING

5 CONCLUSIONS

Main advantages in using such a kind of FIS are

flexibility, since thresholds are no longer crisp, and

versatility, since the forecast is carried out for each

specific track-segment and the method allows to take

into account the company policy as for maintenance.

Moreover, it is possible to use the expert knowledge,

without predefined mathematical models: in this

sense, the fuzzy inference is close to reasoning way

of railway officials and technicians.

Tamping 11 October 2001

The proposed model is also able to recognize glaring

mistakes in measurement, since it analyzes the

overall behavior of the track on the basis of the

whole body of training data. It also allows to

overcome some drawbacks of the binary logic. In

fact, without verifying the correctness of the

parameter value, according to binary logic approach

even if one value of the parameter exceed the

threshold then maintenance operations have to be

carried out.

Test

otes

date of

measure

SD align

SD vert.

days to

intervention

∆days %

1 reference 15-july-02 1,11 1,84 713 0

2 error on last measure 15-july-02 1,06 1,74 736 3.22

15-febr-02 1,02 2,11

error on fifth-last

measure

15-july-02 1,11 1,84

760 6.59

15-may-02 1,09 2,11

error on second-last

measure

15-july-02 1,06 1,84

765 7.29

5 error on last measure 15-july-02 1,11 2,11 718 0.70

An important difference between the two approaches

is relevant to the way of taking into account the two

considered parameters. In fact, Ecotrack considers

SD of alignment

and SD of longitudinal level

separately; then by linear extrapolation forecasts the

dates on which parameters threshold will

respectively be reached and assumes as tamping date

the closest one by applying a OR (logic) rule.

On the contrary, the FIS model consider both the

parameters at the same time: it is based on a

systemic and multicriteria approach to the

parameters values analysis. Thus, a parameters

correlation is implicitly assumed. Actually, in order

∆

SDalign=0

∆

SDvertical=0

forecast = 826

Figure 6: Rules of the fuzzy inference system

ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

to held higher safety level each fuzzy rule assumes

the lowest membership degree of the two parameter

by means of an AND (logic) rule.

For both the approaches a large amount of track

technical and geometrical qualitative and

quantitative data is needed and usually most of the

railways company do not have enough data.

Ecotrack assumes the available data within the

software database while the proposed FIS system the

data are used during the training phase of the ANN

and to validate the calibrated model.

Some problem to reach an effective and consistent

implementation of such a fuzzy system may occur

when there is a lack of data or when the analyst

assumes incorrect data as reliable. Actually, under

these assumptions rail track analysis is difficult

using traditional software too.

Further research will be devoted to improve the

proposed DSS. In particular, we will deal with:

– define a hierarchically higher analysis level in

order to gather homogeneous maintenance works

on adjacent track-segments; to this purpose we

are defining a Subtractive Clustering method;

– introducing new rules in order to allow the DSS

to carry out analysis when insufficient or

incorrect data are available;

– defining time-function of the track quality

indexes in order to determine a track

deterioration model closer to the theoretical one.

REFERENCES

Dell'Orco, M., Ottomanelli, M., Pace, P., Pascoschi, G.,

2002. Optimisation of Railway Maintenance Planning

Using Intelligent Decision Support Tools. In Proc. of

the 13

Mini-EURO Conf. on "Handling Uncertainty

in the Analysis of Traffic and Transportation Systems".

Polytechnic of Bari, 10-13, June 2002. pp. 218-223.

Bari, Italy

Esveld C., 2001. Modern Railway Track, MRT

Productions

Leung, M., Chuah, K.B., Rao Tummala, V.M., 1998. A

Knowledge–based System for Identifying Potential

Project Risks. Omega, n°5, 623 – 638

Jovanovic S., Esveld C., 2001. ECOTRACK: An objective

condition-based decision Support system for long-term

track M&R Planning directed towards reduction of

Life Cycle Costs. 7th International Heavy Haul

Conference, Brisbane, Australia

Zadeh, L.A.,1973. “Outline of a new Approach to the

Analysis of a complex Systems and Decision

Processes”. IEEE Transaction on Systems, Man and

Cybernetics, 3, pag. 28-44

Zimmermann, H. J., 1991. Fuzzy set theory and its

applications. Kluwer Academic Publisher, Dordrecht

Zoeteman, A., 2001. Maintaining the infrastructure of the

conventional railway network in the Netherlands,

State-of-the-art Report. May 2001

Miwa, M., Ishikawa, Oyama T., 2001. Modeling the

Transition Process of Railway Track Irregularity and

its Application to the Optimal Decision-Making for

Multiple Tie Tamper Operations. Railway Engineering

2000, London, July 2000

A DECISION SUPPORT SYSTEM BASED ON NEURO-FUZZY SYSTEM FOR RAILROAD MAINTENANCE

PLANNING