QUALITATIVE REASONINGS WITH DEEP-LEVEL MECHANISM MODELS FOR DIAGNOSES OF DEPENDENT FAILURES (ARTIFICIAL INTELLIGENCE)

Traditional studies on mechanism diagnoses have been based on the "single failure" assumption even though multiple failures are main concerns in the real world. In this research, we concentrate on a subclass of multiple failures, called dependent failures, where a primary failure may induce subsequent secondary failure(s). The dependent-failure case is important because the probability of its occurrence is the same as that of single failure, yet it leads to multiple failures with potentially catastrophic results.;Our approach is unique in three aspects. First, owing to the explicit failure knowledge in state-transition models, this approach allows dependent failures to be treated as rationalizable consequences of a single failure rather than just a case of multiple failures, and thus reduces the computational complexity involved in performing diagnoses. Second, the approach adopts the concept of qualitative diagnosis under which quantitative time-observations of sensory inputs are abstracted into sequences of symptom-events. By matching symptoms with model-predictions at qualitative-event level, the diagnosis system is relieved of its total dependency on the quantitative precisions of mechanism models as well as sensor data. Third, the time nature of this approach enables the transient behaviors of a post-failure mechanism to be used as diagnostic symptoms, and allows predictions of failure-consequences before the full development of the symptoms.;While a straightforward rule-based expert system exhibits impressive performance owing to its ability to heuristically initiate failure hypotheses, it can not further disambiguate among failure possibilities for lack of deep-level domain knowledge. The adopted expert diagnosis approach is founded on a reasoning process, called the predictive analysis, which enables the computer to understand time-elapsed mechanism behaviors through deep-level mechanism models. By incorporating explicit time-related knowledge about failure causalities into component state-transition models, our approach is able to justify or dispute each failure hypothesis by qualitatively analyzing its time-elapsed multiple-failure consequence.