William Marsh

1
Using Bayesian Networks to Model Accident
Causation in the UK Railway Industry

• William Marsh
• Risk Assessment and Decision Analysis Group
• Department of Computer Science
• Queen Mary, University of London
• George Bearfield
• Transport Safety and Reliability
• Atkins Rail, London

2
Outline

• Signals Passed at Danger (SPADs)
• Organisational Accidents
• Bayesian Networks
• Building a BN for SPADs
• Conclusions

3
Signals Passed At Danger
Southall
Ladbroke Grove
4
Signals Passed At Danger

• Train has passed a stop signal without
  authority
• Incident on 27/3/03 at Southampton
• 360 yard overrun
• affected by low sunlight
• driver read adjacent signal
• signal is approached on a curve
• wrong signal into the drivers direct line of
  sight for a short time

5
Waterloo
Southampton
From Railway Safety Assessment of Railtracks
Response to Improvement Notice I/RIS/991007/2
Covering the Top 22 Signals Passed Most Often
at Danger HSE, 2002
6
Organisational Accidents

• Operator errors have organisational causes
• gradual relaxation of alertness
• pressure to increase efficiency

Increasing Resistance
Increasing Vulnerability
Currents acting within the Safety Space
7
Organisational Causes of SPADs
Within the workforce there is a perception that
emphasis on performance has affected attitudes to
safety.
Ladbroke Grove report
the industry is generally poor at identifying
organisational issues that may underpin SPAD
incidents

• Infrastructure multi-SPAD signals
• Driver training and timetable pressure

8
Bayesian Network
9
Organisational Model

• Actors in the organisation (idea from Rasmussens
  AcciMap)

• Responsibilities of actors
• Interactions between actors

10
BN Variables from Attributes
quality
pressure
experience alertness
visibility curve
traffic

• Actors and interactions can have attributes

11
SPAD Scenarios

• Each SPAD scenario modelled as a BN
• events
• influences attributes of driver, infrastructure,
  
• Scenario model merged

12
SPAD Scenario
Influence
Event
13
Expert Judgement

• Strength of probabilistic influences judged by
  experts
• Modify network structure
• Build probability tables
• Aggregated data
• SPAD frequencies
• Used to validate judgements
• Status
• Not yet completed!

14
Using the Causal Model

• Assess frequency / risk
• Where are SPADs likely?
• Monitor organisational changes
• Use audit results
• Select interventions
• How can the frequency of SPADs be reduced?

15
Summary

• Integrated causal model of SPADs
• Organisational influences
• Event sequence
• Bayesian networks
• Generalise other probabilistic modelling
• Future challenges
• Use