Novel condition monitoring approach based on hybrid feature extraction and neural network for assessing multiple faults in electromechanical systems

New challenges involve the development of new condition monitoring approaches to avoid unexpected downtimes and to ensure the availability of machines during operating working conditions. The feature calculation from vibrations and stator currents is one of the most common an important signal processing included in condition monitoring strategies; however, the calculation of features from only one signal alone can only detect some specific faults. Thus, disadvantages are presented if multiple faults are addressed. Aiming to avoid this issue, in this work is proposed a novel condition monitoring approach based on a hybrid feature calculation of statistical features from the available vibrations and stator current signals. Thus, the characterization of the available signals is performed by estimating a hybrid set of features, then, through the Linear Discriminant Analysis, such hybrid set of features is subjected to a dimensionality reduction procedure resulting into a 2-dimensional space. Finally, the assessment and identification of multiple faulty conditions are carried out through a Neural Network. The effectiveness of the proposed approach is validated by its application to two different experimental test benches, which makes the proposed approach feasible to be applied in industrial processes.