Towards a new fault diagnosis system for electric machines based on dynamic probabilistic models

This paper presents a new approach to diagnose faults in electrical systems based on probabilistic modelling and machine learning techniques. Our framework consist of two phases: an approximated diagnosis on the first phase and a refined diagnosis on the second phase. On the first phase the system behavior is modelled with a dynamic Bayesian network that generates a subset of most likely faulty components. In this phase the structure and parameters of the dynamic Bayesian network are learned off-line from raw data (discrete and continuous). On the second phase a particle filter algorithm is used to monitor suspicious components and extract the faulty components. The feasibility of this approach has been tested in a simulation environment using several interconnected electrical machines.