Probabilistic Safety Assessment

Probabilistic risk assessment (PRA) is a systematic and comprehensive methodology to evaluate risks associated with a complex engineered technological entity (such as an airliner or a nuclear power plant) or the effects of stressors on the environment (Probabilistic Environmental Risk Assessment - PERA) for example.

Risk in a PRA is defined as a feasible detrimental outcome of an activity or action. In a PRA, risk is characterized by two quantities:

Consequences are expressed numerically (e.g., the number of people potentially hurt or killed) and their likelihoods of occurrence are expressed as probabilities or frequencies (i.e., the number of occurrences or the probability of occurrence per unit time). The total risk is the expected loss: the sum of the products of the consequences multiplied by their probabilities.

The spectrum of risks across classes of events are also of concern, and are usually controlled in licensing processes – it would be of concern if rare but high consequence events were found to dominate the overall risk, particularly as these risk assessments are very sensitive to assumptions (how rare is a high consequence event?).

Probabilistic Risk Assessment usually answers three basic questions:

Two common methods of answering this last question are event tree analysis and fault tree analysis – for explanations of these, see safety engineering.

In addition to the above methods, PRA studies require special but often very important analysis tools like human reliability analysis (HRA) and common-cause-failure analysis (CCF). HRA deals with methods for modeling human error while CCF deals with methods for evaluating the effect of inter-system and intra-system dependencies which tend to cause simultaneous failures and thus significant increase in overall risk.

One point of possible objection interests the uncertainties associated with a PSA. The PSA (Probabilistic Safety Assessment) has often no associated uncertainty, though in metrology any measure shall be related to a secondary measurement uncertainty, and in the same way any mean frequency number for a random variable shall be examined with the dispersion inside the set of data.

...
Wikipedia