SMOTE-Enhanced Predictive Maintenance of Hydraulic Systems using Random Forest and Logistic Regression

The industry’s hydraulic system defects often result in excessive repairs and unexpected downtime. One of the most difficult parts of predictive maintenance is dealing with an imbalance in the data. In most cases, such as my research, we have a lack of failure cases compared to a sizeable number of nominal or normal operating records. This ultimately causes difficulty for a model in detecting rare faults like we have with hydraulic systems. This current work tackles the problem by utilizing Random Forest and Logistic Regression models along with the Synthetic Minority Oversampling Technique (SMOTE) to have a more balanced dataset. Each model was evaluated on accuracy, precision, recall, F1-score, and ROC-AUC. Overall, the results show that SMOTE enabled the model to identify faults in a more successful manner, while especially benefiting rare failure cases. Of the studied models, Random Forest with SMOTE proved to be the most successful, giving a better performance with higher accuracy and fewer errors. This enables a better approach to hydraulic system maintenance and limit unexpected breakdowns.