Bayesian Modelling of Subway Train Door Failures: Reliability Estimates and Operational Impact Scenarios

Subway maintenance requires tools that relate component reliability to the consequences of operational unavailability. This study developed and applied a Bayesian network framework integrating failure estimates, maintenance costs and scenario-based operational impacts for 22 door-system components of one subway operator. Three-parameter Weibull distributions fitted to maintenance records were combined with operational interference probabilities elicited from specialists, used as conditional inputs whose predictive capacity was not validated. Expected failure counts were compared retrospectively with 2019 records. The model estimated 822.6 failures against 940 recorded, an aggregate underestimation of 12.5%, with a mean absolute error of 11.93 failures and a weighted absolute percentage error of 27.9%; 14 of the 22 observed counts fell within 95% Poisson prediction intervals, and dispersion exceeded the Poisson assumption; a previous-year forecast reached a WAPE of 36.5% and a three-year mean, 27.0%. Expected canceled trips per failure, recovered from the implemented network, ranged from 0.87 for depot detections to 6.06 for severe failures detected on the commercial route. Under joint perturbations of up to 50% in these values, the ordering of interventions by investment per avoided trip remained unchanged, whereas the comparison between modernization and a combined lubrication and depot-inspection intervention was not robust. Canceled trips and the monetary proxy were not validated against observed disruptions. The contribution is an exploratory decision-support structure whose managerial use depends on documented parameters and prospective comparison with operational outcomes.