Autoencoder based feature reduction analysis applied to electromechanical systems condition monitoring

Condition monitoring in electromechanical systems represents, currently, one of the most critical challenges dealing with the advancement and modernization in intelligent manufacturing. In this regard, machine learning based algorithms widely applied in other technological fields are being considered now to face the automatic feature extraction on the electric machine monitoring. In this study, a monitoring scheme is considered for faults detection performance evaluation, where vibrations signal under different fault conditions are acquired. Thus, the common electric machine monitoring framework, that is, a set of features estimated from a limited number of measurements, is considered in front of the three main dimensionality reduction approaches: principal component analysis, linear discriminant analysis and auto-encoder based. Performance of the corresponding approaches are studied and discussed experimentally. It is revealed that, although scheme based on auto-encoder provides enhanced diagnosis results, it is still necessary to carry out a detailed study on the automatic extraction capabilities of important features for the detection of faults.