Advanced hyperparameter optimization and adaptive synthetic sampling in machine learning for predictive maintenance of industrial machinery

Given the evolution of industrial machines, Predictive Maintenance (PdM) requires innovative solutions as the most advanced means of operating efficiency and cost-effectiveness enhancement. This research presents a comprehensive framework that integrates advanced hyperparameter optimization with Adaptive Synthetic Sampling (ADASYN) to solve the PdM problems in Machine Learning (ML) applications. In hyperparameter tuning, the Tree-structured Parzen Estimator (TPE) can apply an efficient method to dynamically select hyperparameters, while Optuna minimizes computational burden and enhances prediction accuracy. ADASYN addresses the problem of skewed datasets by improving the minority class representation in the training dataset and simultaneously facilitating better predictions. An extremely imbalanced industrial dataset is assessed using ۱۶ ML models, including encoder and ensemble models like LightGBM and Random Forest. The results are much better than boosting-based models, like LightGBM (۹۹.۲۶% accuracy and ۹۹.۲۶% F۱ score), highlighting their adaptability to complex datasets with variable feature interaction. Data balancing techniques and robust hyperparameter tuning significantly address most challenges posed by the noisy, imbalanced, and large-scale industrial data. This study introduces a novel unified framework integrating Optuna and ADASYN for PdM—an approach that has not been previously explored in existing literature. Previous research has used either optimization or resampling, but this work shows how their combination improves minority class sensitivity, reduces overfitting, and increases generalization in complex industrial settings. Meanwhile, the importance of advancing PdM methodologies is emphasized, and the possibility of a milestone in accurate, scalable, and interpretable scaling of ML models in dynamic industrial environments is investigated. Future research should advance predictive capability by domain-related feature engineering, real-time model testing, and hybrid methods. Intelligently combining predictive theory with field applicability, the proposed method can potentially bridge field applicability with the ability to deliver dependable, efficient industrial utility across sectors like manufacturing, transportation, and healthcare equipment maintenance.

کلیدواژه ها:

Predictive Maintenance ، Machine Learning ، hyperparameter optimization ، Adaptive Synthetic Sampling