Fault Diagnosis in Rotating Machinery Using Rough Set Theory and ROSETTA

This report outlines an approach to use rough set theory and the ROSETTA system in the field of fault diagnosis in rotating machinery. The complexity of rotating machinery makes fault diagnosis a difficult task. Several computer based methods exist using mathematical modelling together with fuzzy logic, neural networks, faults matrices, machine specific experience and simulation. The ROSETTA system takes a machine learning approach to fault diagnosis by inducing rules on the basis of measured data classified by a domain expert. Data used in this work was collected from 15 different measurement points on a large diesel engine. Six different machine states were recognised, five fault states and one normal state. Our primary goal was twofold. First, to establish whether machine specific knowledge could be used indirectly in order to reduce the amount of data one has to consider when inducing rules in ROSETTA. Second, to learn how much data is needed to induce effective rules in ROSETTA. Our analysis shows that machine specific knowledge not only can be used to reduce the amount of data one has to consider, but also to obtain considerably better results. ROSETTA is proven to be a powerful tool for fault diagnosis, comfortably exceeding what domain experts consider to be good results.