| dc.description.abstract | Remaining useful life (RUL) of an equipment or system is a prognostic value that depends on data gathered from multiple and diverse sources. Moreover, assumed for the sake of the present study as a binary classification problem, the probability of failure of any system is usually very much smaller than that of the same system to be in normal operating conditions. The imbalanced outcome (largely much more 'normal' than 'failure' states) at any time results from the combined values of a large set of features, some related to one another, some redundant, and most quite noisy. Previewing the development and requirements of a robust framework, it is advocated that by using Python libraries, those difficulties can be dealt with. In the present Chapter, DOROTHEA, a dataset from UCI library with a hundred thousand of sparse anonymized (i.e. unrecognizable labels) binary features and imbalanced binary classes are analyzed. For that, an ipython (jupyter) notebook, pandas are used to import the data set, then some exploratory analysis and feature engineering are performed and several estimators (classifiers) obtained from scikit-learn library are applied. It is demonstrated that global accuracy does not work for this case, since the minority class is easily overlooked by the algorithms. Therefore, receiver operating characteristics (ROC), Precision-Recall curves and respective area under curve (AUCs) evaluated from each estimator or ensemble, as well as some simple statistics, using three hybrid methods, that are, a mix of filter, embedded and wrapper methods, feature selection strategies, were compared. | |
