A Hybrid Stacked-Ensemble Deep Learning Framework for Machine Learning-Based Predictive Maintenance in Smart Manufacturing Operations

Abstract

Unplanned equipment downtime remains one of the most persistent sources of productivity loss in discrete and process manufacturing, with industry surveys attributing substantial annual revenue loss to reactive maintenance practices. While machine learning (ML) has been widely proposed as a substitute for time-based and reactive maintenance strategies, most published predictive maintenance (PdM) studies rely on a single learning algorithm and report performance on heavily engineered, often near-noise-free benchmarks, leaving open questions about robustness under class imbalance and about the transparency of model decisions to maintenance engineers. This study proposes SE-PdM, a hybrid stacked-ensemble framework that combines two heterogeneous base learners — a Random Forest (RF) classifier and an Extreme Gradient Boosting (XGBoost) classifier — with a one-dimensional Convolutional Neural Network coupled to a Bidirectional Long Short-Term Memory network (1D-CNN-BiLSTM) operating on sliding-window sensor sequences, integrated through a logistic-regression meta-learner. A strict class imbalance of failure modes is dealt with through a hybrid SMOTE-Tomek resampling strategy within training folds while a SHAP value is computed to explain the decisions of the ensemble model to the plant engineers. The framework is tested on the AI4I 2020 predictive maintenance benchmark (10,000 operating records, five failure modes), stratified 5-fold cross validation. Experimental results indicate that SE-PdM is superior to all base learners and to the following deep-learning based baselines for both macro-averaged F1-score and ROC-AUC and that the most important features for the prediction of failure risk are always torque, tool wear and the interaction between the rotational speed and the process temperature, irrespective of the base learner. This paper makes the following three key contributions: (i) a tabular sensor data representation that incorporates a tree-based and sequence-based representation; (ii) a training protocol that does not suffer from resampling leakage between cross validation folds, a common, but underreported pitfall in past PdM research; (iii) an interpretability layer that maps model output into actionable maintenance triggers. The proposed model provides an operational example of explainable predictive maintenance pipelines in smart manufacturing with the consideration of imbalance.