Conceptual framework for predictive maintenance in urban mass transport systems based on signal processing and generative artificial intelligence: From fault prognostics to maintenance decision modelling

Urban mass transport systems require high availability, controlled life-cycle cost, and strict safety performance. This study develops and validates a predictive maintenance framework that links signal processing, deep generative modelling, and reliability-based decision analysis. Multi-sensor data from a tram traction system, including vibration sampled at 10 kHz and thermal and current signals sampled at 1 Hz , were processed using short-time Fourier transform and wavelet packet decomposition. Statistical and spectral features were extracted from segmented operating windows. A variational autoencoder and a generative adversarial network were trained on healthy-state data to model latent degradation behaviour. Reconstruction error and latent deviation were used to derive degradation indicators, which were embedded into a proportional hazards model for remaining useful life estimation. The approach was evaluated on five years of operational data from a fleet of 40 trams. Detection accuracy increased from 0.87 with a support vector machine baseline to 0.94 with the generative adversarial network. Mean absolute error of remaining useful life prediction decreased from 18 days to 9 days. Monte Carlo propagation of latent uncertainty produced 90 percent calibrated prediction intervals. Fleet-level simulation over one operational year showed reductions in corrective interventions and expected failure cost when compared with time-based maintenance. The results demonstrate that generative prognostic outputs can be formally linked to reliability functions and risk-based maintenance planning. The framework additionally integrates prognostic outputs into reliability-centered maintenance and risk-based scheduling models, clarifying the role of generative models in maintenance engineering and outlining future research directions for ESREL 2026.