Multiple decision trees to diagnose a transient state of dynamic systems. Application to a DC motor.

Abstract

In this paper, a novel methodology is proposed to diagnose a transient state of a dynamic system us­ing supervised learning. lt is composed by two steps: one off-line process and another on-Iine pro­cess. The off-line phase begins gathering data from the system, both when it is running free of fault and when the system is running in each fault mode. Also, it is possible to generate these data from Monte Cario simulations of a system model. A seg­mentation and normalization algorithm is used to reduce the large amount of gathered data. The fi­nal step of the off-line process is the generation of a decision tree by a classification tool. The on-lme process of the methodology consists in evaluating a new reading of the system sensors with the gen­erated decision trees. The system diagnosis is the result of this evaluation which has a linear compu­tational cost due to the simplicity of the decision trees. In arder to improve diagnosability problems of this methodology, it is proposed a new solution in this work. Instead of generating only one de­cision tree, a different decision tree is generated for each fault mode and free of fault mode. Therefore multiple possibilities of diagnosis can be offered for a given behaviour of dynamic system. Meth­odology has been applied to diagnose a DC motor. Eight different faults have been considered and the results have been discussed including diagnosabil­ity conflicts.