Banca de QUALIFICAÇÃO: MAILSON RIBEIRO SANTOS

Online and Offline Approaches to Fault Detection, Classification and Estimation in Dynamical Systems

Fault Detection and Classification, Online Approach, Offline Approach, Feature Selection.

In this work, we present two new approaches to perform the detection, classification, and severity estimation of faults in dynamic systems. The first proposal we use when the dataset is fully available a priori, this proposal is characterized by using offline training, composed of the phases of feature extraction; feature selection; detection, classification, and severity estimation of fault.  We perform the extraction of several statistical features in the time domain, then we propose to use techniques such as principal component analysis, autoencoder networks, and explainable artificial intelligence to perform the selection of features, finally we use the Support Vector Machine technique in the phases detection, classification, and estimation of fault severity. The second proposal should be used when the data are obtained continuously, in other words, the dataset is not fully available a priori, this approach is composed of the same phases as the first approach, however, we use online techniques. We will calculate the features recursively, we will apply an online feature selection technique such as incremental principal component analysis, and evolving autoencoder. In the phase of fault detection, we use the Typicality and Eccentricity Data Analysis (TEDA) algorithm, samples identified as faults are applied to the AutoCloud algorithm to determine the classification and severity estimate of the fault. To validate the proposed approaches, we applied the methodologies to the dataset of faults in electric motor bearings.