Hybrid AI‑Powered Real-Time Predictive Maintenance for Railway Wheelsets and Bogies

Abstract

This thesis presents a Hybrid AI-based predictive maintenance system specifically designed to monitor railway wheels and bogies. The implemented solution uses real-time sensor data such as vibrations, wheel-rail forces, angles, and weather conditions to detect anomalies and predict maintenance requirements in advance. The presented system applies a hybrid anomaly detection framework consisting of Sigma-based statistical analysis, Isolation Forest, and Autoencoder methods (Context-Aware Anomaly Detection (CAAD)), with adaptive and accurate alerts for anomalies.

A CNN-LSTM-based predictive model identifies both short-term and long-term trends in sensor readings to accurately forecast future failures. In addition, an RL agent that employs the PPO algorithm optimizes maintenance choices, balancing cost savings against operational reliability.

The system is calibrated to operate on vast quantities of real-world data from the Haugfjell rail monitoring station. The architecture aspires to overcome traditional approaches, forecasting gains in fault detection (target F1 ≈ 0.81), reducing maintenance expenditures (≈30\%), and increasing system availability (≈22\%), relative to theoretical standards and model characteristics.