Temporal Decision Trees: Model-based Diagnosis of Dynamic Systems On-Board

Abstract Machine learning has been used successfully as a tool to generate fault diagnostic rules from first principles. However, many diagnostic problems such as those involving temporal reasoning remain unsolvable due to the limitations of current learning algorithms. This paper describes an application of a new machine learning system, Golem, in the domain of generating temporal fault diagnostic rules for the power supply subsystem of a satellite. Temporal rules in the form of Horn (relational) clauses are successfully produced by exploiting the ability of Golem to discover relations from examples expressed as logical facts. The examples used to induce rules are generated by simulation from a qualitative model of the power supply subsystem. Results showed good diagnostic accuracy.