Comparison of machine learning approaches by using oversampling techniques on imbalanced datasets

Industrial maintenance covers all the technical and administrative operations to extend the lifetime of the equipments and avoid from unplanned system failures in order to ensure the manufacturing facilities uninterruptedly. The cost of unplanned breakdowns and accidents caused by insufficient maintenance operations brings serious risks for the industrial organizations. Although many facilities try to manage these risks with traditional maintenance approaches, their success may be limited. Companies that update and advance their maintenance strategies in the way of technological developments have the opportunity to manage related risks and losses more effectively. The analysis of data collected from sensors and equipment, allowing for the prediction of potential failures before they occur, and the maintenance activities conducted based on these predictions, is called predictive maintenance. With the developments in the Internet of Things (IoT) and the integration of cyber-physical systems, it has become easier to collect data in real time via industrial equipment. Analytic insights based on the processing of this data with artificial intelligence algorithms have brought a new dimension to predictive maintenance strategies. When we aim to devise an analytical model based on prediction for problems such as predictive maintenance and fault detection, we are faced with a dataset with an imbalanced class distribution, due to the nature of these events. The number of observations representing the non-faulty condition usually far exceeds the observations representing the faulty condition. Classifying imbalanced datasets using artificial intelligence algorithms poses significant challenges, as algorithms tend to overfit with the information representing the non-faulty condition, where there is a larger number of observations. This issue makes it difficult to detect failures using algorithms in real-life applications. In this study, artificial intelligence algorithms were used to perform a classification task on two different datasets representing data collected from equipment in an industrial environment. To prevent the problem of overfitting on both datasets with imbalanced class data distributions, various oversampling methods and hybrid methods were applied on the training data to balance the datasets. Classification was performed using models created with standalone machine learning algorithms, ensemble learning algorithms, and deep learning algorithms, using the balanced datasets. The performance of the models was evaluated in terms of the Cohen Kappa score, F1 score, recall, and accuracy. Ensemble learning algorithms carried out higher performance than the other algorithms in both datasets. The effect of the data balancing methods on the performance of ensemble learning algorithms was also analysed. The performance of the ensemble learning models varied in the datasets balanced using different methods, but generally performed better in datasets balanced using random sampling. In this thesis, the parameters affecting model performance for the classification task in imbalanced datasets were presented with a multidimensional perspective.