Survival-analysis-guided machine learning for failure and remaining-useful-life prediction in railway compressor

Predictive maintenance requires reliable failure detection and remaining useful life (RUL) estimation under conditions of noisy sensor data, censored observations, and limited failure labeling. This work presents an interpretable predictive maintenance framework in which a Cox proportional hazards model is used to identify the operational variables most strongly associated with failure risk and to guide the construction of data-driven predictive models. The framework is validated on the real-world MetroPT-3 railway air-compressor dataset, using failure dates reported by the operator to derive both the binary failure-event label and a proxy RUL target. Machine learning and deep learning approaches, including neural networks, tree-based models, and sequential architectures, are then evaluated for failure event prediction and Proxy RUL estimation. The results show that Cox-guided variable selection enables effective dimensionality reduction while preserving high failure detection performance, particularly in terms of sensitivity and specificity. Sequential models prove well suited for failure event prediction, whereas nonlinear tree-based models are more robust for Proxy RUL estimation due to the non-stationary and discontinuous nature of the target variable. The main contribution is a unified, leakage-controlled pipeline that couples Cox-PH survival-based feature selection with a chronological evaluation protocol, showing when statistical guidance helps and when it does not across two distinct maintenance tasks. Overall, the study highlights the practical value of combining statistical explainability and machine learning performance to support robust and interpretable predictive maintenance decision-making in industrial contexts.