title: Best-fit dynamic behaviour: can fault tree upgrades be machine-learnt?
keywords: Fault Tree Analysis, Artificial Intelligence, Fitness Optimisation
topics: Algorithms and Data Structures , Dependability, security and performance
committee: Doina Bucur ,
Carlos Esteban Budde


Fault Tree Analysis is a world-standard for safety and reliability assessment of complex machines, due to its versatility, simplicity, and extensibility. Expert engineers build a fault tree (FT) of a system component to describe its failure mechanisms. The FT is refined until the component’s resilience metrics, such as its mean time to failure, can be predicted based on available data or technical information, such as the component’s material and the average/expected system load.

When building and refining the FT, engineers traditionally assume static behaviour, e.g. “the short-circuit is caused by degraded insulation and excessive rain,” even though dynamic behaviour may the most accurate predictor: “the short-circuit is caused by excessive rain after degraded insulation (and no maintenance in-between).”


Design, implement, and evaluate an Artificial Intelligence algorithm, or other heuristics, for this problem. The input is a static FT and a set of failure traces (time series data) from the system. The algorithm must determine if the failures described by the traces are better modelled by a modified FT whose gates are dynamic. More specifically, considering for instance an n-ary AND gate in the FT, switch it for an n-ary Priority-AND, and evaluate whether the modified FT more closely resembles the system failures observed in the data traces.

For example, the static FT in Roadtrip has two gates that can be better modeled as dynamic gates: The AND-gate at the top can be replaced by a Priority-AND (modeling that a trip by car does not fail if the car fails before the phone fails, as the driver can call for assistance), and the tires are better modeled using so called "spare gates" describing that a spare tire does not fail when it is not in use (see Roatrip (dynamic)).


Research aims

  • Determine which dynamic gates are good upgrade candidates for the traditional AND/OR/VOT static gates.
  • Assess the resulting failure-prediction quality of the upgraded FT.
  • Try to address the (combinatorial) problem of upgrading subsets of gates in the FT.
  • Study the occurrence of simultaneous failures, which would introduce “functional dependencies” as new dynamic gates in the FT.


  1. The Faul Tree survey (Digital version available here)
  2. Dynamic Fault Trees (Digital version available here)
  3. Machine-learning Fault Trees (Digital version available here)

Additional Resources

  1. Roadtrip
  2. Roadtrip (dynamic)