Abstract:
Shallow knowledge-based fault detection and diagnosis may be carried out using production rules. If the knowledge is provided in the form of linguistic rules then fuzzy logic is a suitable inference mechanism. However, the conclusions drawn can communicate either the likelihood of a fault nor the certainty of the. diagnosis. This is of consequence if information used to make diagnoses is unavailable or unreliable. The mathematical theory of evidence provides a solution that utilises bodies of evidence rather than fuzzy sets. Measures reflecting the amount of evidence supporting and contradicting a conclusion can be derived. Structural similarities with fuzzy models are maintained but methods of combining evidence are more completely specified. Evidential reasoning has been applied to anaesthesia monitoring in the form of SENTINEL. SENTINEL aims to diagnose seven problems that sometimes occur during general anaesthesia. It makes use of a temporal pattern matching system known as fuzzy trend templates, which are capable of providing bodies of evidence about complex, temporal relationships between events on multiple time-series. They have a transparent, graphical interpretation. SENTINEL has been tested retrospectively on more than 250 hours of data collected during actual surgical procedures at a general hospital. The data represents surgery on adult patients; no particular type of surgery was selected. Performance is significantly better than traditional anaesthesia alarms, with up to 95% sensitivity and 90% specificity.