Analysing Aviation Accidents using WB-Analysis An Application for Multimodal Reasoning

We describe our ongoing work in accident analysis. Accident reports should tell us at least what the accident was and what the critical events were. A third requirement they should fulll is to explain these events (see below) and their sequence (!temporal reasoning). Explanation concerns causes (!causal reasoning), human intentions, purposes, capabilities and behavior (so-called human factors). Causality also involves the unfolding of events in time (Explaining Failure in Tense Logic, RVS-RR-96-13). We include social factors-obligations and the regulatory environment-amongst the human factors (!deontic reasoning). Formal reasoning about accidents involves addressing three basic issues:{ 1. We must determine what to reason about (what's in the universe?). To do this in a formally correct way, we describe the world in the ontology of TLA, Lamport's Temporal Logic of Actions, because this ontology has shown itself suucient for describing the temporal behavior of artifacts (process algebra & Petri net semantics use a similar ontology). 2. In principle, this should involve us in clarifying the relation between set theory (or other formal data structuring) and`the world'. But we leave this to philosophers of mathematics. We're more interested in the accidents. 3. We must determine what kind of reasoning is involved. Since a narrative is involved, tense logic (with the Kripke semantics) is an appropriate reasoning tool, given that the ontology has been declared suitable. We show that the ontology of TLA is suucient for (A) description of machine behavior, (B) formulation of accident histories, and (C) determination of sequences of states leading to an accident. States are individuated by the collection of state predicates which are true in that state. This makes states into types. Particular occurrences of states can be identiied by means such as timestamps or positions in the causal chain. States furthermore may have a duration. In contrast, events are particulars, representing speciic changes in state. An individual event cannot recur, but its type (a TLA action) can be instantiated more than once. There are also processes, state/event mixes of bounded duration which describe undiierentiated actions. Non-events, the non-occurrence of awaited events, are also important (see below). On closer inspection the logical operators in TLA are insuucient for adequate reasoning about accidents, so we extend TLA by introducing a temporal point of view (tpov), additional temporal operators ((P ],hP i,,F],hFi) and additional rules (RVS-RR-96-13). We show that 1