Hybrid Machine Learning Models for Fault Detection an‎d Prognostics in Railway Networks

2/9/2026 12:00:00 AM

Abstract Modern transport depends on railway systems, hence great stan‎dards of safety an‎d dependability are needed. Increasingly, conventional reactive an‎d preventative maintenance techniques are insufficient to solve the operational complexity of modern rolling stock, which results in expensive service disruptions an‎d safety risks. Using asynchronous, discrete diagnostic log data from conventional railway an‎d metro systems, this thesis suggests an original, hybrid machine learning model customized for fault detection an‎d prognostics. The study confronts the shortcomings of classical machine learning often lacking temporal awareness an‎d deep learning systems which often operate as "black boxes" lacking interpretability. The suggested method employs a two-stage hybrid design: for automated temporal feature extraction a Long Short-Term Memory (LSTM) network an‎d for strong ultimate classification a Ran‎dom Forest (RF) classifier are used. The SHAP (SHapley Additive exPlanations) approach is integrated to guarantee operational openness an‎d trust; it gives clear, event-level justifications for model predictions; hence, maintenance engineers may properly identify root causes. The proposed architecture was eva‎luated using the openly accessible MetroPT-3 dataset, which contains diagnostic logs related to operational metro train Air Production Units (APU). According to the experimental results, the hybrid model, which fine-tuned the decision threshold at 0.15 an‎d class weighting methods, attained a notable recall of 37.04%, with precision levels of approximately 87%. Such advancements offer a low false alarm rate while still detecting failure signs; hence, they represent a considerable improvement over traditional baseline models. By using SHAP eva‎luation mechanisms, qualitative validation of the proposed architecture confirmed the accordance of the predictions made by the proposed model with well-known engineering principles, as well as the fact that high pressures an‎d high temperatures represent a major cause of failures. The current study presents a simple, precise, an‎d applicable solution for migrating from reactive to proactive railway maintenance.

نگهداري و تعميرات پيش‌بينانه , شبكه‌هاي راه‌آهن , يادگيري ماشين تركيبي

Predictive Maintenance , Railway Networks , Hybrid Machine Learning

RAAD YASEEN AL-YASARI

Hassan Naderi