Journal of Engineering Research and Reports

Background: The integration of smart technologies into modern facilities has underscored the need for proactive maintenance strategies to minimize unplanned equipment failures and enhance operational efficiency. Traditional maintenance approaches, including reactive and time-based preventive maintenance, often fall short in dynamic building environments. Predictive maintenance, driven by machine learning (ML) and Internet of Things (IoT) sensor data, offers a data-driven solution to anticipate equipment failures before they occur.

Methodology: This study proposes a comprehensive machine learning framework for predictive maintenance in smart facilities, evaluated using the ASHRAE Great Energy Predictor III dataset—a real-world benchmark containing operational data from diverse building systems. The framework compares both classical machine learning (Random Forest, XGBoost) and deep learning (LSTM) approaches to address different predictive maintenance scenarios. Data preprocessing included outlier removal, missing value imputation, feature engineering, and normalization. Model evaluation was conducted using precision, recall, F1-score, ROC-AUC, and inference time metrics. The system is designed for seamless integration with existing Computerized Maintenance Management Systems (CMMS) to ensure practical deployment.

Results: Among the models tested, the LSTM network achieved the highest predictive performance (F1-score: 0.89, ROC-AUC: 0.93), while XGBoost provided an optimal balance between accuracy (F1-score: 0.84) and computational efficiency (10ms inference time). Implementation resulted in a 40% reduction in maintenance response time, 25% cost savings, and 47% decrease in unplanned downtime. The framework demonstrates strong scalability across different facility types and equipment classes. A revised bar chart (Fig. 2) has been included with enhanced visual clarity through color-coded bars. Additionally, Appendix A outlines the practical steps for integrating the framework with CMMS platforms.

Conclusion: The developed machine learning framework effectively bridges the gap between research and practical implementation, offering a versatile solution for predictive maintenance in smart facilities. Its compatibility with CMMS platforms and demonstrated performance on real-world data (ASHRAE dataset) positions it as a viable tool for intelligent facility operations. Future work will focus on edge computing deployment, including strategies such as model quantization and device-level optimization, and expansion to additional industrial applications.