Design and Implementation of an IoT-Driven Predictive Maintenance System Using Machine Learning for Industrial Equipment

Unplanned equipment downtime remains a major source of production loss, safety risk, and unnecessary maintenance cost in industrial settings, particularly where maintenance practice remains reactive or fixed-schedule rather than condition-based. This study presents the design, implementation, and validation of an IoT-driven predictive maintenance (PdM) system for realtime monitoring and Time-To-Failure (TTF) prediction of industrial equipment. The system is structured around a five-layer architecture data acquisition, transmission, storage, processing/analytics, and visualization/alerting built using a Python-based sensor simulation, the InfluxDB time-series database, Grafana for dashboarding, and Telegram for automated alerting. Due to practical constraints on accessing live industrial machinery, synthetic time-series sensor data (temperature, vibration, current) was generated to simulate realistic equipment degradation, while a public Kaggle predictive-maintenance dataset supplied historical, labeled degradation data for model training. Three machine learning models Long Short-Term Memory (LSTM) networks, Random Forest, and XGBoost were developed and compared for TTF prediction. LSTM achieved the lowest Mean Absolute Error (MAE = 39.59 seconds, R² = 0.9600), outperforming Random Forest (MAE = 63.03 s, R² = 0.9884) and XGBoost (MAE = 59.91 s, R² = 0.9894) on absolute prediction accuracy, despite the tree-based models showing marginally higher R² scores. Feature importance analysis was consistent across all three models: temperature emerged as the dominant predictor of equipment failure, ahead of vibration and current. The trained LSTM model was integrated into a real-time inference pipeline that queried live sensor data every 10 seconds, generated continuous TTF forecasts, and dispatched automated Telegram alerts once predicted TTF fell below a critical threshold, while a Grafana dashboard provided continuous visual monitoring. These results validate a practical, scalable, and reproducible framework for IoT-based predictive maintenance that can be deployed with modest computational resources, and demonstrate the specific value of LSTM networks for precise failure-time estimation in industrial condition monitoring