Transcript OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS

7TH FRAMEWORK
PROGRAMME FOR RESEARCH
FP7-SST-2012-RTD-1
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Overview of WP4:
Development of Fuzzy and Computational
Intelligence based models for
maintenance management
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Goal:
Develop intelligent models to represent railway
infrastructures, maintenance, management and
traffic processes
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
TASK 4.1: Development of fuzzy models for railway infrastructures and components
WP1 and WP2
Components
of railway
Infrastructure
Geometric
auscultations
Maintenance
data
Data Mining Techniques
Data preprocessing
Model selection
Model identification
Model fine tuning
Model validation
Fuzzy
and CI
Models
Partners involved: UGR (ES), CARTIF (ES), OSTFALIA (DE), EVOLEO (PT), SINTEF (NO).
Output: D4.1 Report on fuzzy systems built for railway infrastructure component modeling
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
TASK 4.2: Development of fuzzy models for maintenance, management and traffic
processes
WP1 and WP2
Maintenance
processes
Maintenance
operations
(work orders)
Traffic
Data Mining Techniques
Data preprocessing
Dynamic models
Ensembles
CRISP-DM
FRBS
SVM
ANN
Partners involved: UGR (ES), CARTIF (ES), OSTFALIA (DE), EVOLEO (PT).
Output: D4.2 Report on fuzzy systems built for maintenance processes modeling
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
TASK 4.3: Knowledge Extraction from Experts
Knowledge Extraction
K. Representation
K. Acquisition
K. Validation
Knowledge
base
Partners involved: UGR (ES), CARTIF (ES), OSTFALIA (DE), EVOLEO (PT), SINTEF (NO).
Output: D4.3 Report on knowledge extraction from experts and combination with data-driven
models
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
TASK 4.4: Cooperation/Fusion of expert knowledge and data-driven models
Expert
Knowledge base
Knowledge
aggregation and
fusion
Combined
Knowledge base
Data-Driven
Models
Partners involved: UGR (ES), CARTIF (ES), OSTFALIA (DE), EVOLEO (PT).
Output: D4.4 (Combined with Task 4.5)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
TASK 4.5: Development of multicriteria decision-making
Input
Data
Criteria
Multiple objective
optimization
Maintenance
Decisions
Models and
Knowledge
Partners involved: UGR (ES), CARTIF (ES), OSTFALIA (DE), EVOLEO (PT).
Output: D4.4 Report on multi-criteria decision making and multi-objective optimization
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Deliverable 4.1: Report on fuzzy systems built for
railway infrastructure component modeling
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Introduction
• Use machine learning methods learn a functional
relationship between features and targets
Possible inputs in OPTIRAIL
Feature
1
Feature
3
...
Target
Feature
1
Target
Feature
2
Instance1
Instance2
Instance3
• Historical geometrical condition data
Feature
2
Input
Output
...
• Infrastructure data (asset characteristics), such as sleeper type and curvature
• Available work order data
Possible outputs in OPTIRAIL
• Predictions of geometrical condition data (min, max, mean, sd)
• Thresholded predictions  prediction of need for interventions
• Prediction of work orders ( D4.2)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Overview of processing steps
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Alignment of geometrical inspection data
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Examples
Methods for data alignment
• Some methods take objects such as bridges and crossings into account
• Our approach is based on the correlation of excerpts/snippets of the curvature of
the measurements
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Some remarks
• The offset is by no means constant, but varies quite significantly due to the way
the position of the train is determined (number of wheel rotations)
• Hence: The offset between the measurements is determined every, say, 1km,
i.e. at discrete points, and linear interpolation is used in between
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Dynamic Segmentation
The idea is to use the available asset characteristics
to determine homogeneous track sections.
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Predictive modeling of deterioration
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Predictive modeling of deterioration
Two deterioration models for TQIs can be used:
• Linear deterioration model
Q(t) = Q0 + tb
• Exponential deterioration model (derived from the observation that the
deterioration of track quality is proportional to the current quality)
Q(t) = Q0 + exp(tb)
•
is the track quality at time t = 0 (immediately after a work order)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
•
In the example of the Swedish Q-value, if no
maintenance is performed, the quality gets worse,
and the Q-value decreases
•
Predicting future geometrical inspection data is
„simply“ fitting the (exponential) model
•
When work along the track is performed,
the quality increases, with a jump, and the
parameters of the (exponential) model may
change
•
The time series is cut into pieces by the work orders.
One such piece is called a deterioration branch
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
An example from the data
sigH = TQI used in OPTIRAIL
(maximum of the sd of the long. levelling
of left and right rail)
Q-value
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Machine learning models
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Use nonlinear regression methods (FRBS, neural networks, SVR, random forest, etc.)
to establish a relationship between asset characteristics and the parameters Q0 and b
of the deterioration model for the TQI
Sleeper
type
Rail
type
Traffic
data
Curvat
ure
...
Q0
b
Asset1
Asset2
Asset3
...
Input
Output
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Experiments for Sweden (track 118 of the Iron Ore Line)
• Non-constant attributes are used
as input
• One model for Q0, one for b
• RMSE to assess quality of models
Q0
b (exp)
#rules/units
ANFIS
1.598
0.000499
104/104
DENFIS
2.229
0.000493
7/8
GFS.MEMETIC
0.594
0.000514
35/35
Random Forest
0.578
0.000349
500/500
SVR
0.628
0.000386
-
MLP
0.661
0.000422
3/3
• Random forest performs well
• Resulting models are not straightforwardly interpretable
• sigH is usually between 0 (perfect) and 3 (maintenance threshold)
• With an error around 0.6, we see that the approach is feasible, but has a high
error, due to uncertainty in the data, data quality, etc.
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Conclusions D4.1
• A methodology for deterioration modelling was developed
• Goes beyond current state of the art
• Data has high amounts of uncertainty  models are currently not very accurate
Possible solutions:
• Acquire more data (monthly (geometric) inspections)
• Get more information about the track (ballast type, drainage , subsoil type, etc.)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Deliverable 4.2: Report on fuzzy systems built for
maintenance processes modeling
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Predict future auscultation results, and predict from auscultations the work to do
• Historical geometrical
auscultation data D4.1
• Infrastructure data
Future geometrical auscultation data
predictive model
Future work
D4.2 orders
predictive model, or existing
(expert knowledge) model
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Two approaches for D4.2
Expert knowledge driven:
• We predict the auscultation, the expert tells us what has to be done.
• Simple version: Use the thresholds from D1.1
 This only tells us that something has to be done, but not what
• Probably more expert knowledge needed to distinguish operations
Data driven:
• Use historical work orders to learn the condition in which the asset was directly
before the work order
• Problem: How to include policies, such as thresholds?
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Expert knowledge driven modeling of work orders:
• Idea: After having predicted an auscultation, apply the process currently
implemented by the railway administrator to get work orders from the
auscultations
• Detailed expert knowledge is needed for this approach (see also D4.3)
• We collected information from Spain (and some information from Sweden). E.g.:
Parameters
Maintenance action
Longitudinal level
Automatic Tamping
Alignment
Automatic Tamping
Cant
Automatic
Tamping/
Ballast renewal
Gauge
Fastening renewal
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Data driven modeling of work orders:
Longitu …
dinal
left D1
Alignm
ent D1
…
Twist9
m
Work to do
Output
Asset1
Nothing
Asset2
Tamping
Asset3
...
Input
Sleeper
replacement
…
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Data driven modeling of work orders:
Problems:
• Is the information in the data?
• We need historical work orders and geometrical
data that fit together, also, we need sufficient amounts of work orders for every
type of work to be done. Example: With 5 historic sleeper renewals difficult to
build a model for this type of operation
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Results Sweden:
Error
Recall
Precision
Random Forest
0.59%
100.00%
91.18%
SVM
10.30%
99.56%
37.25%
FRBCS.W
6.09%
0.44%
100.00%
GFS.GCCL
6.12%
0.00%
N/A
•
Good results especially for Random Forest:
•
100% recall (this means that all tamping work orders are correctly classified)
•
91% precision (which means that when the classifier determines that a tamping should be
performed, this is correct in 91% of the cases)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Modeling of the effect of a maintenance operation,
Lifecycle modelling
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
We need a TQI which adequately assesses deterioration behavior through time, not at a single
time point. With an exponential deterioration model, we have:
At time t=0:
So, Q‘ as the product of Q0 and b can be used as a quality measure (the tangent to the TQI)
We use the following formula to model the effect of a tamping operation (i, i+1):
I.e., a tamping operation will worsen the quality by a constant c.
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Lifecycle modelling:
•
We consider the following asset:
Radius
Speed
Class
1
•
Det.
Branch Sleeper Age
Rail Age
Num.
110 km/h
0
18 years
18 years
1.731
We lower the nominal track speed to 90km/h. The model of D4.1 predicts values of 1.467
and 0.0001485 for Q0 and b, respectively
•
0.0002113
We set the threshold of triggering a work order to 3
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Lifecycle modelling:
Black curve: lifecycle model with
110km/h nominal track speed,
remaining life: 33.48 years
Red curve: lifecycle model with
90km/h nominal track speed,
remaining total life: 44.23 years
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Conclusions D4.2:
•
A methodology for modelling maintenance decisions from current auscultations was
developed
•
This can be done data-driven and/or with expert knowledge
•
We showed that the data-driven approach works well for the case of Sweden and tamping,
and a random forest can classify tamping vs. no-tamping reliably
•
We investigated the effect of tamping on deterioration behaviour, and applied/developed a
model, using a constant change in Q0‘=Q0*b for exponential deterioration.
•
We also did some modelling of traffic data from Norway, and from work orders of Spain (not
shown in this presentation)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Deliverable 4.3: Report on knowledge extraction process
from experts and
combination with data-driven fuzzy models
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
A questionnaire was developed
Aims:
• Gather expert knowledge regarding infrastructure deterioration and maintenance
operations
• Which variables/asset characteristics have an influence on infrastructure
deterioration? How big is this influence?
• How does the geometrical measurement determine a maintenance decision?
Which other data is necessary for a maintenance decision?
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Results:
•
•
We collected 26 answers:
Country
Number of answers
Spain (SP)
17 (VIAS 16, ADIF 1)
Poland (PO)
3
Norway (NO)
2
England (EN)
1
Germany (GE)
2
Austria (AU)
1
Problems: The answers differ considerably between the countries, and for all countries
except Spain, not enough answers are available to make a within-country analysis
•
The only distinction that is done is Spain vs the rest (we didn‘t take into account that this
may underrepresent the answer from ADIF)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Results (summarized):
• quality and condition of the ballast, the track load and the type of traffic, i.e.
freight traffic and passenger traffic strongly influence track deterioration
• subsoil and drainage is also considered important for track deterioration, but in
many cases, information along the whole track is not available
• deterioration of the track in curves is generally higher than compared to straight
tracks
• The answer for the shortest tamping interval still considered feasible varies from
20 days to 12 months
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Results (summarized):
• Decision for performing tamping is based on auscultation data
• Experts think that there exists an optimal interval between consecutive tamping
operations in the sense that the useful track life is largest.
• Rail substitution can be based on the geometrical inspection data only in a
limited way. Instead, other factors such as the rail condition (obtained from visual
inspection) should be taken into account.
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Results (summarized):
•
There is no consensus whether or not rail grinding should be based on the geometrical
inspection data. Additionally, visual and ultrasonic inspections should be considered.
•
The replacement of sleepers can partially be based on the geometrical inspection data,
especially on the gauge. Moreover, also the condition of the sleepers and of the fastenings,
which can be determined by visual inspections, as well as the age of the components,
should be taken into account.
•
Finally, for ballast cleaning, there is no consensus among the experts which geometrical
variables (besides the longitudinal level of the left and right rail) should be considered for
ballast cleaning. Instead, information on the condition of the ballast and its contamination
(determined by visual inspections) should be considered.
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Deliverable 4.4: Report on multi-criteria decision
making and multi-objective optimization
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Main objectives
• track maintenance cost
• the availability/capacity
• the safety/quality of the track
Other important factors
• Constraint or objective?
• Planning horizon (3,5,30 years)
• Granularity of planning in time (daily, monthly, trimestral)
• Granularity of planning in space (track sections, track length)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Objective: Minimize maintenance cost
• Maintenance operation cost (MOC) in framework from Lulea:
i sums over the track sections and j over the time intervals in the planning horizon.
The variable r is a discount rate.
: material cost for the maintenance operation in €/km
: average time to perform the maintenance operation (MO) on the ith track section in hours/kilometre
: total length of maintenance section in kilometres
: average labour cost in €/hour
: equipment cost for the maintenance operation in €/hour
: cumulative load / time (in MGT or years)
: interval for the maintenance operation of the ith track section in MGT (or years)
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Objective: Minimize maintenance cost
• The costs can be described with parameters that can later be easily changed by
each railway operator.
• Example costs used by Lulea:
Constraints regarding maintenance costs
• One tamping machine is available for the considered track (otherwise external
factors have to be taken into account)
• Constraints regarding time of day: Maintenance window of approx. 5h, which
translates into maximal tamping distance per day
• Constraints regarding time of year
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Objective: Availability and capacity
The objective has the following aspects:
• capacity of the track
• punctuality of trains
• penalties if capacity and punctuality goals are not met
Capacity loss is the sum of delays and non-availability of the track due to maintenance.
Train delays occur due to parts of the track that cannot be used with their nominal track speed
as their quality is not sufficient.
Changes in speed should be minimal (as they waste resources and cause more maintenance
necessity)
Capacity may often be a constraint and not an objective
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Objective: Safety and quality of the track
• maximize safety and ride comfort
• minimize costs caused by damage to trains and track components due to bad
track condition
• minimize penalties that may result from these
• Cost is proportional to how much measured values lie above the thresholds
 Both safety and ride comfort are difficult to measure
 Safety is not optimized but guaranteed, so this is modeled as a constraint and
not an objective
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
• Cost (tamping and renewal cost) is defined by:
• Cost is subject to minimization:
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Calculation of train delays:
• Maximal admissible speed can be obtained from (predicted) track quality using,
e.g., EN-13848-5:
• Delay is to be minimized:
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
In this approach, the constraints handle important parts:
• Ensuring that the speed the trains go with is admissible given the tamping and
renewal actions:
• Machine limits (of overall tamping and renewal that can be performed):
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
• Degradation model from D4.1:
• Tamping effect model from D4.2: Derivative of sigH is constant over tamping
operations:
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
• Renewal effect model: Both
and
are set to values that can be
considered “as good as new“
• Renewal is always programmed in a whole section k
• Tamping and renewal are exclusive. Renewal has higher priority
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Initialization of the solutions
•
Problem has very high dimensionality and complexity
•
An intelligent initialization can be performed and is necessary
 Program operations in first trimester where necessary
 Calculate the remaining tamping capability up to the established threshold. Generate a random
number r between 0 and this remainder.
 Sort the track segments that have no tamping scheduled according to decreasing quality. Include
tamping along this list up to the computed limit.
 From this list, choose for each segment randomly if tamping will be performed or not.
 Do the same for renewal
 Apply deterioration and effect models, and begin from start iteratively for all trimesters
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Parameters for the Sweden case study:
•
Sections ij are statically segmented, 5km each
•
Segments k are dynamic segmentation from D4.1
•
•
Values for “as good as new“ (see also D4.2):
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Results
•
500,000 evaluations of the fitness function, 104 solutions in the initial population
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Results (2)
•
AMOSA yields better results than NSGA2: the problem is complex and high dimensional, and
the initialization already uses a lot of problem-specific knowledge. AMOSA favors local
search instead of exploration
•
Best solutions with AMOSA:
Cost (€)
Delay
Tampings
Renewals
(hours)
5689494
137.98
665
8
5690557
126.63
667
8
6286985
126.63
623
9
6289973
126.14
625
9
6293171
119.84
630
9
6293924
118.34
630
9
…
…
…
…
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Conclusions D4.4:
•
Possibilities for optimization have been analysed in depth, and a framework was developed
•
We implemented a state-of-the-art model: Multi-objective optimization minimizing cost and
train delays, using the degradation model from D4.1 and the tamping effect model from
D4.2
•
Complexity and high dimensionality is a big problem
•
Intelligent initialization helps to cope with this problem
•
However, with current resources, simplifications have to be made (regarding granularity of
planning in time and space)
•
Scheduling is another problem not touched here. However, this is necessary to define
realistic benefits regarding transport and fixed costs
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
More conclusions:
--> Track is so bad that tamping doesn't lift it anymore above threshold
--> Renewal has to be programmed
--> Important how cost of tamping and renewal relate to each other. Currently, 20 tampings
cost the same as one renewal
--> With a planning horizon of 3 years, the situation that a renewal saves cost will not occur
--> With longer horizons, complexity even bigger, and deterioration model unreliable
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
Conclusions
OPTIRAIL WORKSHOP · OCTOBER 23, 2014 · BRUSSELS
• We have developed methodologies for modeling of infrastructure and
maintenance operations data, and we have shown how a predictive maintenance
plan could be generated
• We have shown how data and expert knowledge can be used to achieve the
OPTIRAIL goals, i.e., predict maintenance operations.
• We have adapted and implemented two multiple-objective algorithms for
maintenance decision making