Feature Engineering for Predictive Maintenance: Identifying Key Predictors of Machine Defects Using Machine Learning

In the modern industrial environments, the ability to predict equipment failure before it occurs is essential for minimizing downtime and maximizing operational efficiency. This research explores the use of feature engineering to identify key indicators of mechanical faults in a cement mill fan system. Vibration data were collected over 34 weeks from critical components of the fan and processed using several statistical techniques to extract relevant features. Various feature selection methods including Principal Component Analysis (PCA), Minimum Redundancy Maximum Relevance (mRMR), ReliefF, Chi-square, ANOVA, and Kruskal-Wallis were used to determine the most informative features. These features were then used to train and evaluate machine learning models, with neural networks demonstrating superior performance. Among all models, the neural network optimized with Chi-square-selected features achieved the highest classification accuracy, fastest prediction speed, and lowest misclassification cost. These results highlight the effectiveness of combining robust feature selection with deep learning methods for reliable fault detection and predictive maintenance in industrial systems.