Aurorisk FR

Chapter 1 · Lesson 1.2

Five situations that defeat it

On the loading bay, five ordinary situations make the fault tree insufficient. Each calls for a different family of models.

1. Passive redundancy: cold standby

P2 is stopped as long as P1 is running: it cannot fail in operation during that time. In the tree, however, P2 has the same probability of failure as P1, at any instant. The logic “P2 only ages from the failure of P1” is a temporal dependency: it requires a state model.

2. Repair and the queue

When both pumps are down, the single crew repairs one and the other waits. The downtime of the second depends on the first. With one crew per pump, the unavailability of loading would be 1.7·10-3; with a single crew, it is 3.45·10-3, twice as high. The tree does not know the difference.

3. Failure on demand

P2 may refuse to start when P1 fails (probability γ = 0.05). This event is not a rate per hour: it is a probability attached to an event (the failure of P1). It can be added to the tree as a basic event, but it combines with the dynamics of the rest, not just with a gate.

4. The order of failures

If P2 fails before P1, it is on cold standby: it could not have failed while running (it was not running). If it fails after, it is a failure in service. The two sequences do not have the same probability. An AND gate does not distinguish these orders; you need a priority gate (PAND) or a state model.

5. Cascade propagation

An overflow can produce a pool, then a fire, then involve a neighbouring bay. Each step depends on the previous one and on the barriers that did or did not work. Here time and order are the data of the problem: a tree only represents the combination of causes of a single final event.

What dynamic calculation gives

Chapter 2 calculates the unavailability of loading with a Markov model that represents cold standby, repair with a single repairer and failure to start. The figures:

ModelUnavailability of loadingHours per year
Single AND-gate tree (two independent pumps, repaired separately)1.5·10-313
Tree with the failure to start of P2 added as a basic event3.3·10-329
Markov: cold standby, one repairer, failure to start3.45·10-330

The second tree is close to the Markov result (3.5% below), but for the wrong reasons: it still assumes that P2 can be down before the demand with the same probability as P1, and that the two repairs are independent. It remains optimistic, all the more so as repair is slow: 8% below for an MTTR of 50 h, 13% for an MTTR of 100 h. The single AND-gate tree underestimates the unavailability by more than half.

Warning

Agreement within a few per cent between two models does not validate the simpler one. You must understand which assumptions each makes, and what happens if the parameters change.

Key takeaway

Cold standby, shared repairer, failure on demand, order of failures and cascade propagation are five common situations where a fault tree is wrong or cannot answer. For the loading bay, the single AND-gate tree announces 13 hours of unavailability per year, the state model 30.