Machine learning-based predictive maintenance of an industrial thermal plant

The transition toward Industry 4.0 has positioned Predictive Maintenance as a critical strategy for minimizing unplanned downtime and optimizing asset lifecycle management. This thesis develops an end-to-end data-driven predictive maintenance system for an industrial thermal plant, utilizing high-frequency telemetry data acquired via an IoT supervision platform. The primary objective is to overcome the limitations of traditional reactive PLC alarms by forecasting thermodynamic and mechanical faults with a five-hour prognostic horizon. A rigorous data engineering pipeline was designed to transform raw sensor streams into an augmented feature space, integrating physical domain knowledge and temporal rolling statistics to capture early-stage multivariate degradation. Following a comparative evaluation of supervised and unsupervised machine learning architectures, a cost-sensitive binary Random Forest classifier was identified as the optimal solution. The evaluated model achieved a Recall of 84\% and a Precision of 93\%, successfully anticipating the vast majority of plant shutdowns. Finally, the validated model was industrialized as an autonomous, scheduled inference task. To ensure operator trust and system interpretability, the pipeline was augmented with SHAP (SHapley Additive exPlanations) values, providing real-time, sensor-level root-cause analysis for every generated predictive alarm. The resulting architecture delivers a robust, scalable, and explainable prognostic solution ready for active industrial deployment.