Aurorisk FR

Chapter 1 · Lesson 1.2

Accidents without a failed component

Several well-documented accidents happened while every element worked as intended, or nearly so. Three examples help identify the mechanisms that the STPA method seeks to make visible.

The three cases are described from public investigation reports; the details are simplified: read the reports themselves before citing these cases.

Mars Polar Lander (1999)

The probe was meant to shut down its engines on touching the ground. According to the investigation board’s analysis, the landing-leg sensors produced a contact signal when the legs deployed; the software took it for a landing and shut the engines down some forty metres above the ground. This is the most probable cause retained by the JPL review board (2000). Every element did what it was designed to do: it was the requirement that ignored this spurious signal.

Texas City (2005)

During the restart of a refinery unit, a column was filled beyond the intended level, then flammable liquid overflowed. The investigations (including the CSB’s) point to unreliable level instruments, procedures not followed and organisational decisions: the unit was restarted although the level transmitter and the redundant alarm were known to be faulty. The explosion killed 15 people. The flaw is spread between instruments, operators and management.

Boeing 737 MAX (2018-2019)

An automatic flight-control function relied on a single angle-of-attack sensor; an erroneous value could trigger a repeated action on the stabiliser, which the crew, who had not been told that this function (MCAS) existed, did not always know how to interpret. Two accidents (Lion Air in 2018, Ethiopian Airlines in 2019) killed 346 people. The software executed its logic; the flaw lies in the design assumption and in the information given to the pilots.

What these cases have in common

Caution

These examples do not say that component failures do not exist: they show that failures are not the only route to an accident. A complete analysis deals with them too (see lesson 5-2 on the complementarity of methods).

Key takeaways
  • An accident can result from interactions without any component failing.
  • The three cases share an inappropriate command based on a wrong process model.