A documentation-based Modelica model of a metro air production unit: preregistered validation and Kalman-filter leak warning on MetroPT

Open predictive-maintenance data of the Porto metro (MetroPT) have been analysed almost only with data-driven methods. We asked whether a physics-based model of the air production unit, built from its documentation, describes and predicts its operation and warns of failures at least two hours ahead, as the operator requires. Hypotheses, data splits, thresholds and predictions were registered before computation. The object is the roof-mounted unit of a Eurotram car: an oil-injected rotary vane compressor with load–idle–stop control, a two-tower dryer and two 60 l reservoirs. It was recorded every 10 s from February to August 2020 (MetroPT-3), with five failures by loss of air in the maintenance log. Our model-synthesis method produced a Modelica model from a passport of the unit; the model passed all 141 behavioural acceptance checks without manual changes. Calibrated on two months, the model was tested on the next five: it predicts the next load phase with a median error of 5.7 % and a mean absolute error of 23.7 s, better than regression (26.4 s) and persistence (38.2 s); the cooling of the receiver after loading is physical (RMSE −52 %), and the delivery stays within ±10 % in all 23 weeks; the registered description of the cycle and the one-node oil model fail. The registered physical failure indicators responded to the failures but warned none: their thresholds were set by single calibration outliers. We then developed a predictor from the mass balance of the model: the consumption share over the last recorded hour. On the same unit it warned three of four failures 3–161 h ahead at 0.88 false episodes per 1000 h, against 5.7 for the duty-cycle rule; this result is exploratory. In a blind test on two other units, each with its own calibration, the predictor raised almost no false alarms (0.46 and 0 per 1000 h). It warned none of three air leaks two hours ahead: two were sudden, with normal signals until the pressure collapsed within an hour, and one was invisible in the pressure. A two-hour warning is achievable for leaks that develop over days, not for sudden ruptures. The registered verdict is “partially confirmed”. Finally, a Kalman filter on the balances of the model, fed by the load command and the panel pressure, estimates consumption and delivery at every sample. On synthetic leaks generated by the full model it warned 10–43 h before the 7 bar alarm for leaks growing over 24–72 h, four to seven times earlier than the mass-balance predictor. On the real records the gain is smaller: three of four failures of APU01 warned at 1.8 false episodes per 1000 h, and the sudden leak of MetroPT2 detected 26 min before the alarm. The difference points to what the telemetry lacks: above all a measurement of consumption, which arrives in pulses the model does not describe, then of the delivery and of the exact times of the load switches.