Predictive Maintenance dengan Klasifikasi OnevsRest dan Random Search

Abstract

This study develops a predictive model for machine maintenance using machine learning, focusing on optimizing the Random Forest algorithm with a One-vs-Rest (OvR) approach. The research is motivated by the industrial need for accurate predictive maintenance systems to anticipate machine failures early, addressing the limitations of currently dominant reactive and preventive maintenance approaches. The main challenge lies in significant class imbalance, where the "No Failure" class dominates 80% of the dataset while other failure classes comprise only 2-5%. The research methodology includes data preprocessing, exploratory data analysis, hyperparameter optimization using Random Search, and model evaluation with various metrics. Results show the model achieves 99% accuracy and a macro ROC-AUC of 0.9408, but fails to predict minority classes like Tool Wear Failure (TWF) and Random Failures (RNF) with 0.00 precision and recall. These findings confirm the negative impact of class imbalance on model performance. The study makes important contributions by identifying Random Forest's limitations for imbalanced data while recommending specialized handling techniques for future research. The results provide a foundation for developing more robust predictive maintenance systems in industrial applications.