Inductive learning applied to diagnostics

Abstract

The author describes how inductive learning techniques which give an expert system the ability to gain most, or all, of its knowledge from actual field experience can be applied. J.R. Quinlan's ID3 algorithm is used as a fundamental building block. This algorithm induces a decision tree from a set of training examples and uses this tree to classify future cases. Extensions were made for handling unknown test values and uncertain class assignment for the training examples. By treating fault-isolation sessions in the field as training examples, the expert system can develop its own optimized decision tree and improve its performance as it gains experience. The result is lower mean time to repair (MTTR) and retest-okay (RTOK) rates, with development cost being approximately that of a traditional test program set. Because the system tracks real-world failure rates, probability values can be assigned to fault callouts, thus reducing the amount of ambiguity in a group. This allows the maintenance technician to determine the actual cause of failure more often when presented with an ambiguity list. >