Influence of Dataset Structuring on Condition Monitoring of a Rotating System by Machine Learning

Resumo

Condition monitoring consists of constant data acquisition from a machine of interest to determine its
operational condition, and also to give reasonable predictions regarding its behavior over time. Considering that
vibration generated by a machine carries information about internal conditions and is sensitive to structural
changes, vibration analysis can be employed to detect faulty components. As some defects have known vibrational
responses (“vibrational signatures”), it is possible to infer the type of defect by analyzing the vibration signal
characteristics. An algorithm capable of automatically doing this type of analysis could potentially prevent
financial or physical harm. In this context, the present study focuses on preprocessing vibrational response data
related to induced defects in a rotating system, extracting features of interest, using a machine learning classifier
to identify common problems, and segregating troublesome conditions from expected normal operation ones. The
processed data was obtained from the Machine Fault Dataset (MaFaulDa/UFRJ). The obtained results show the
influence of dataset structuring on the algorithm generalization capability, revealing that bigger datasets do not
always lead to superior performance and that an increase in the amount of attributes is not always the most
interesting choice.