Optimizing Predictive Maintenance: Addressing Imbalanced Data with Sampling Techniques and Stacking Models in Fault Prediction

In the modern industrial field, industrial machinery failures can bring high production losses and significant maintenance losses to enterprises. Predictive maintenance has become an important research direction to solve this problem, and fault prediction is the core link in predictive maintenance. Mastering fault prediction methods is crucial. This article is based on the AI4I 2020 predictive maintenance dataset of UC Irvine Machine Learning Repository. It mainly constructs a fault prediction model through decision trees, random forests, and MLPClassifier in machine learning, and explores the optimal model based on this dataset. On this basis, various methods such as random oversampling, random undersampling, SMOTE, etc. were used to optimize the fault prediction scheme for the imbalanced sample problem that occurred in the experiment. Some methods are effective, but the performance of the model is not significantly improved, so further explore the fault prediction scheme using Stacking to integrate multiple models, compare the evaluation results of single model and multiple models, and the scheme using Stacking to integrate decision trees, random forests and MLPClassifier models has the best effect in this experiment, with $\mathbf{F 1}$ score of $\mathbf{0. 8 0 9 8}$. This conclusion indicates the importance of using Stacking ensemble method to integrate multiple models in mechanical fault prediction. At the same time, it also verified the positive effect of various data augmentation methods on handling imbalanced data for fault prediction in the model. This study also provides a reference for future industrial machinery fault prediction.