Combining machine learning methods and data augmentation for misaligned journal bearings design

Abstract

Shaft misalignment is one of the most common defects observed in rotating systems and has a substantial effect on dynamic behaviour, stability, and lifetime. Aim of this study is the binary identification of misalignment using five Machine Learning techniques: Logistic Regression, K-Nearest Neighbours, Support Vector Machines, Decision Tree and Random Forest. Nevertheless, the limited quantity of provided data points coupled with the substantial imbalance between the aligned and misaligned cases necessitated the implementation of oversampling and data augmentation methods. The utilization of SMOTE-LOF for oversampling the minority class, alongside the adoption of a Conditional Tabular GAN for the generation of synthetic data points yielded substantial outcomes. The application of SMOTE with the Local Outlier Factor on the original dataset achieves the oversampling of the minority class by using additional synthetic data, while the LOF factor overcome the noise problem. After the SMOTE-LOF implementation, new synthetic samples of the minority class are added to the dataset, eliminating the imbalance, however the ‘cleaned’ dataset’s small number of samples could lead to poor performance of the Machine Learning techniques. Hence, using a Generative Adversarial Network to expand the dataset is requisite, and since the experimental data is tabular, utilizing a Conditional Tabular Generative Adversarial Network is ideal for this case of study. The results of the current study elucidate the dataset that, among the augmented datasets, exhibits the best quality score when contrasted with the original dataset. Furthermore, the dataset that performs the best on the Machine Learning algorithms is determined.