In Industry 4.0 manufacturing environments, un-expected machine failures lead to substantial economic losses and operational disruptions. Traditional maintenance approaches often prove inadequate for modern industrial systems. This research introduces a novel framework combining ensemble machine learning with explainability techniques to predict equip-ment failures. We evaluated multiple classification algorithms including Random Forest, SVM, XGBoost, LightGBM, and a Stacking approach using the AI4I 2020 industrial dataset containing 10,000 instances across 14 attributes. The dataset exhibits significant class imbalance with only 339 failure cases against 9,661 normal operations. We employed Synthetic Minor-ity Oversampling Technique to balance the training data. Our LightGBM implementation demonstrated superior performance, achieving 97.85% classification accuracy and an F1-measure of 0.7514, surpassing baseline Random Forest results (F1: 0.7027) by 5.71%. To enhance model transparency, we incorporated SHAP value analysis, which revealed that Torque measurements (SHAP: 1.5035), Tool Wear duration (SHAP: 0.8527), and Heat Dissi-pation indicators (SHAP: 0.7035) serve as the most significant failure predictors. These findings enable maintenance personnel to prioritize monitoring of critical operational parameters. We validated practical applicability through a web-based prediction system offering real-time failure forecasting with interpretable explanations. Our approach advances Sustainable Development Goal 9 by facilitating data-driven maintenance decisions that reduce industrial waste and enhance operational sustainability.

Index Terms—Predictive Maintenance, Ensemble Learning, LightGBM, XGBoost, SMOTE, SHAP, Explainable AI, Industry 4.0, Machine Learning, Industrial IoT