Explainable AI-Driven Predictive Maintenance Framework for Industrial Equipment in Industry 4.0 - Enabled Smart Factories

Abstract

Predictive maintenance has now become one of the most important strategies of enhancing the workability of the system as well as cutting down unplanned downtime in the complex industrial systems. GRID deviation is one of the essential measurements of the imbalance and stress factors in a power grid operation, which requires precise predictability and timely interventions. This paper suggests a full machine learning predictive maintenance system to predict grid deviation using high-scale operation records. The framework combines systematic feature engineering, ensemble learning models, and explainable artificial intelligence methods to conquer the issues of non-linear behavior, data imbalance, and rare extreme events. It used a dataset of 267, 324 observations, including the operational, maintenance, generation, and time variables. The exploratory analysis showed that grid deviation is a highly skewed, heavy-tailed, variable, which has some high-magnitude events, which highlights the constraints of the conventional linear modeling methods. Several of the ensemble learning models were developed and tested within a similar experimental setup, namely, Random Forest, Gradient Boosting Regressor, and Light Gradient Boosting Machine (LightGBM). Comparative findings indicate that each of the models has good predictive performance; nevertheless, LightGBM has better predictive ability on outliers, robustness, and efficiency. In order to make it more transparent and practical, model explainability was implemented with the help of Shapley Additive Explanations (SHAP). The explainability analysis also found the generation alignment variables, which include actual and programmed generation, and station-level statistical factors to be the most influential in predicting the grid deviation. Besides giving high prediction accuracy, the proposed framework also gives actionable insights to grid operators and maintenance planners. In general, the research paper evidences the efficiency of combined using advanced machine learning and explainability methods of predictive maintenance in large-scale power systems.