Abstract
The emergence of advance sensor data has led to considerable advancements in predictive maintenance (PdM) within metro systems, whereas energy and telecommunication infrastructures remain constrained by inadequately labelled fault data. This study examines the feasibility of transferring predictive maintenance knowledge from metro vehicles to data-scarce infrastructure domains using a transfer-learning framework. The MetroPT dataset was utilised and modelled using a supervised learning approach. The results show high in-domain accuracy of 98.8%, while the telecom baseline model achieved an accuracy of 0.9778 but exhibited low anomaly recall (0.125) reflecting the challenges associated with limited labelled fault data. Furthermore, the metro-trained model failed to generalise to telecom data under a source-only transfer-learning strategy, resulting in negative transfer and an AUC of 0.500. These findings provide initial empirical evidence that direct model transfer between metro and telecom infrastructures is ineffective without appropriate domain alignment. The study therefore highlights the need for advanced adaptation techniques and domain-invariant representation learning to enable cross-sector PdM deployment in critical infrastructures.