Integrating Artificial Intelligence and IoT for Predictive Maintenance in Smart Manufacturing Systems

Abstract

Downtime is one of the most expensive events in today's manufacturing and downtime on a critical line can run the organization a ten of thousands of dollars per hour in lost production, lost product and expedited repair. The traditional maintenance strategies (reactive maintenance after failure and preventive maintenance on fixed calendar schedule) are not effective enough to satisfy the reliability, cost and sustainability requirements of production environments in Industry 4.0. This paper introduces a single Artificial Intelligence and Internet of Things (AI-IoT) framework for predictive maintenance (PdM) in smart manufacturing systems, which is capable of continuously monitoring the health of equipment, diagnosing incipient failure, predicting Remaining Useful Life (RUL), and generating timely, actionable maintenance decisions before failure. The proposed framework consists of a multi-modal IoT sensing layer, an edge-cloud hybrid computing architecture, a feature-engineering pipeline optimized for rotating and reciprocating machinery, and a hybrid deep-learning core, which features the Convolutional Neural Network (CNN) for spatial feature extraction, the Long Short-Term Memory (LSTM) network for temporal degradation modeling, and the attention-weighted fusion layer for joint fault classification and RUL regression. A rule and model hybrid decision support layer converts model results into prioritized maintenance work orders with the consideration of resources. The framework is tested against two widely used public test datasets: the NASA C-MAPSS turbofan degradation dataset and the Case Western University University (CWRU) bearing fault dataset, as well as on a simulated multi-machine factory-floor testbed that simulates a representative discrete-manufacturing production line. Experimental results demonstrate that the proposed hybrid CNN-LSTM-Attention model can achieve 98.1% fault-classification accuracy and reduce the error of RUL prediction (RMSE) by 17.6% compared with the baseline model (LSTM) and 24.3% compared with the baseline model (Random Forest), while maintaining the inference latency under 45 mS per prediction window at the edge. A cost-benefit analysis performed on the factory floor case study simulated results of an estimated 32–38% decrease in unplanned downtime and 18–25% decrease in overall factory maintenance cost compared to a preventive-maintenance baseline, which will need to be verified through field deployment. The paper also explores the architectural, organizational and data-governance considerations that must be taken into account for scaling such a framework and outlines open challenges, such as cross-machine generalization, label scarcity for rare failure modes and integrating with existing Manufacturing Execution Systems (MES) and Computerized Maintenance Management Systems (CMMS), which will motivate future work.