Machine Learning-Driven Predictive Maintenance in Smart Manufacturing

Unplanned equipment failures in smart manufacturing environments generate substantial direct costs through lost production, emergency maintenance labour, and spare parts procurement, as well as indirect costs through contractual penalties, customer attrition, and reputational damage. Predictive maintenance (PdM) — the paradigm of continuously monitoring equipment health indicators and generating maintenance interventions only when and where required — offers the prospect of 70–85% reduction in unexpected downtime relative to time-based preventive strategies, while simultaneously extending component service life and optimising technician scheduling. The proliferation of Industrial Internet of Things (IIoT) sensor infrastructure, edge computing platforms, and industrial communication standards has created the technical preconditions for pervasive, low-latency PdM implementation across manufacturing plants of diverse scales and sectors.This paper presents a comprehensive PdM framework integrating a novel hybrid Long Short-Term Memory — Gradient Boosting Machine (LSTM-GBM) architecture with an IoT-Edge-Cloud three-tier system design. The LSTM sub-component captures long-range temporal dependencies in multivariate sensor time series — including vibration, acoustic emission, current draw, temperature, and oil particle count — while the GBM sub-component provides rapid, high-precision fault classification on LSTM-generated feature embeddings. A dataset of 2.4 million sensor readings collected from CNC machining centres, industrial compressors, and conveyor systems across three manufacturing facilities over eighteen months was used for model training and validation. The proposed LSTM-GBM model achieved a fault detection accuracy of 94.7%, F1-score of 0.923, and AUC-ROC of 0.978 on the held-out test set, outperforming six benchmark methods including CNN-LSTM, standalone LSTM, standalone GBM, Random Forest, and SVM. Deployment across the three test facilities over a six-month operational trial demonstrated a 67.3% reduction in unplanned downtime, 31.4% decrease in maintenance costs, and a conservative projected annual saving of USD 2.84 million per facility, establishing a compelling operational and financial case for scaled IIoT-PdM implementation.