Chapter 1 · Lesson 1.1
What a fault tree assumes
A fault tree is an excellent tool for combining causes. But it rests on assumptions that no longer hold as soon as the system is repaired, backed up or degraded in a particular order. This chapter identifies those assumptions on a case that serves as the running thread of the whole course.
- State the hidden assumptions of a fault tree
- Recognise five situations where it gives a wrong or unverifiable result
- Place dynamic trees, Markov chains, Petri nets and simulation
- Choose a model according to the size of the system and the decision-maker’s need
The running case: a loading bay
A (fictional) depot for flammable liquids loads tanker trucks. The loading bay comprises:
- a storage tank that feeds the loading arm through two pumps: P1 in service, P2 on cold standby (stopped as long as P1 is running). P2 starts automatically on failure of P1 and may fail to start;
- a single maintenance crew, which repairs one pump at a time;
- a high-level shutdown function for the truck (detector, logic, shutdown valve), tested periodically, which stops loading in the event of overfilling;
- a neighbouring bay, close by, which could be hit by a pool fire.
Two undesired events are tracked: the unavailability of loading (an operating issue) and the overflow of the tank (a safety issue: pool, ignition, possible domino effect).
| Data (example assumption) | Value |
|---|---|
| Failure rate of a pump in operation, λ | 2·10-3 h-1 (MTTF of 500 h) |
| Repair rate, μ (a single repairer) | 5·10-2 h-1 (MTTR of 20 h) |
| Probability that the standby pump fails to start, γ | 0.05 |
| Shutdown function: λDU detector, logic, valve | 2·10-6, 1·10-6, 2·10-6 h-1 |
| Interval between two tests of the shutdown function, T | 8,760 h (one year) |
| Common-cause factor of two detectors, β | 10% |
| Demands on the shutdown function (overfilling not stopped by the operator) | 0.2 per year |
| Cascade after overflow: bund failure, ignition, exposure of the neighbouring bay, escalation | 0.1; 0.1; 0.5; 0.2 |
These values do not describe any real installation: they are used to calculate and compare. They are reused as they stand in the chapters, the exercises and the quizzes.
The “loading unavailable” tree
The first idea is a single-gate tree: loading is unavailable if P1 and P2 are both down. With the probability that a pump is down at a given instant, q = λ/(λ+μ) = 0.038:
This figure seems reasonable. Yet it is wrong, and above all it explains nothing of what happens in the installation. The following lessons show why.
Four implicit assumptions
The basic events are independent: the failure of P1 changes nothing in the probability of failure of P2. Yet a single crew does the repairs, and P2 does not behave like P1.
The AND gate does not care about the order of failures. Yet “P1 then P2” and “P2 then P1” do not have the same consequences when P2 is on standby.
Repair is hidden in q = λ/(λ+μ), as if each pump had its own crew. The maintenance queue does not exist.
The result is an average probability at an arbitrary instant. We know neither when, nor for how long, nor after what the undesired state occurs.
A fault tree assumes basic events that are independent, with no order, no repair queue and no time. When these assumptions cease to be true, its result is no longer reliable, and we do not know in which direction it is wrong.