A Computational Intelligence Approach To Railway Track Intervention Planning

1
A Computational Intelligence Approach To Railway
Track Intervention Planning

• Derek Bartram
• Supervisors
• Dr. M. Burrow
• Prof. X. Yao

2
Contents

• Current Technologies
• Structure
• Problems / Issues
• Project
• Aims
• Comparison Of Methodologies
• Assumptions
• Tasks
• Task Details (including progress / results)
• Problems / Issues
• Conclusions

3
Existing Technologies Decision Support Systems

• What
• Where
• When
• Why
• Expert System
• Knowledge Base
• Fact Base
• Inference Engine
• Diagnostic

4
Existing Technologies Decision Support Systems

• Knowledge Base
• Geometry Data
• Work History Data
• Location Data
• Component Information
• Curve Data
• Line load / speed

5
Existing Technologies Decision Support Systems
Knowledge Base Segmentation

6
Existing Technologies Decision Support Systems

• Rule Base
• Intervention limits
• Deterioration Models
• Rules

if (ballast thickness lt 30mm) then repair
7
Existing Technologies Decision Support Systems
Rule Base Deterioration models

8
Existing Technologies Decision Support Systems

• Rule Base Rules
• Division of problem based on components

9
Existing Technologies Decision Support Systems

• Inference Engine
• Simplifies rules
• Applys to knowledge

10
Existing Technologies Decision Support Systems

• Diagnostic
• Where
• When
• What

• Why
• Confidence

11
Decision Support Systems Problems / Issues

• Expert system only as good as the rule base
• Simplified models
• Possible rule / intervention flaws
• Large track segments

12
Project Aims

• Improved deterioration modelling
• Improved intervention planning
• Improved localised fault detection
• Improved total life-cycle costing

13
Methodology Comparison Static Vs Dynamic
Solutions

• Static solution
• Guaranteed good behaviour initially
• Never improves
• Dynamic solution
• Initial behaviour potentially bad
• Requires high quality existing dataset
• Improves with time

14
Methodology Comparison Problem Division

• Existing solution
• Divide Per Component
• Proposed solution
• Divide Per Failure Type

15
Assumptions (1)

• The various possible faults for track are
  identifiable by unique combinations of track
  component deterioration

16
Assumptions (2)

• For each type of failure, the solution to the
  problem is not related to other failure types

17
Assumptions (3)

• Once a track sections starts failing with a
  particular failure type, it will continue to fail
  with the same failure type

18
Solution Tasks

• Data Preparation
• Classify the various failure types
• Produce a deterioration model for each failure
  type
• Determine best intervention for each failure type
• Provide a mechanism for classifying new
  unclassified track sections

19
Data Processing Segmentation

• Fixed non-uniform segmentation

20
Data Processing Produce Data Runs

• A set of data all geometry and work data for a
  particular segment of track
• A run of data all geometry data for a
  particular segment of track between two
  successive work interventions

21
Data Processing Make Values Relative

• Need to be able to directly compare values

22
Data Processing Handle Missing Values

• Remove Rows
• Remove Columns
• Fill with value
• Average
• 0
• Extrapolation

23
Produce Data Runs Filling Missing Values

• Extrapolation technique

24
Problems / Issues

• Large number of missing values in geometry data
• Inconsistent / missing? work history data
• Data anomalies

25
Problems / Issues Data Anomalies

• Expected deterioration does exist

26
Problems / Issues Data Anomalies

• Too high for recording noise?

27
Determine Failure Types

• Number of failure types unknown
• Data Mining
• Rival Penalised Competitive Learning (RPCL)
• Performed on last data recording per run

28
Determine Failure TypesClustering On Non-Uniform
Shapes
29
Determine Failure TypesRPCL

• Initial number of clusters
• Learning rate
• De-learning rate (ratio to learning rate)

30
Determine Failure Types
31
Deterioration Modelling

• Curve fitting (simple model)
• 3rd Order curve
• Genetic Algorithm to learn curve parameters

32
Deterioration Modelling Enhanced Simple Model

• Extension to simple modelling using extra
  parameters into function
• Component age
• Cumulative load
• Effective load
• Needs further research

33
Deterioration Modelling Genetic Algorithm Model

• Genetic Algorithm
• Produce curve function
• Train curve parameters to data

34
Work Determination

• For each run in failure type
• Calculate fitness of subsequent intervention
• Calculate average of fitness's for each
  intervention type
• Choose intervention with best average fitness

35
Work Determination Fitness metric

• Length of time before next intervention
• Next failure type
• Integration of costing data

36
(No Transcript)
37
Classifying New Track Sections

• We know sets of individual data points and
  associated failure types
• Assumption 3 Failure type does not change until
  intervention
• Region based classification

38
Classifying New Track Sections

• Classifying blue points
• Red points are other failure types
• Define region around points
• Optional Pruning
• Bad performance just after intervention work

39
Classification based on previous work / failure
history
40
Conclusions

• Long term improvements over static solutions
• Deterioration models
• Intervention planning
• Costing

41
Thank you for listening

• Questions?

42
(No Transcript)