A documentation-based Modelica model of inverter-fed centrifugal pump units: preregistered blind validation, physics refinement and Kalman filtering on the NLN-EMP data

The open NLN-EMP data of inverter-fed centrifugal pump units have been analysed almost only with data- driven classifiers. We asked two questions: does a physical model built from the documentation predict the electrical power and current of the real units; combined with a Kalman filter, does it solve practical tasks: flow without a flowmeter, impeller damage, speed without a tachometer? Hypotheses, thresholds and predictions were registered before the targets were computed. Object. Two Grundfos NK pumps with MG induction motors (11 kW 4-pole; 22 kW 2-pole), fed by variable-frequency drives at 50–100 % speed. Our model synthesis method produced a Modelica model of the unit from a passport of the data sheets. It passed all 151 behavioural acceptance checks at the first synthesis, with no manual edits. Version 1, blind, documentation only. Power of the first unit: within 4.3 and 5.7 % at full and 75 % speed, but 11.3 % at half speed against a 10 % threshold. Current: within 1.8–5.6 %. Second unit: within 6.7 % (power) and 0.1 % (current). The verdict was “partially confirmed” Version 2. A second registration added laws from the literature and standards with published values: the saturation law of Ranta and Hinkkanen, the loss split of IEC 61800-9-2, seal and bearing friction, and the Reynolds scaling of pump efficiency. One calibration point per unit was added for the pump tolerance (cP = 1.047 and 1.049, within ISO 9906 grade 3B). Power errors became −4.0, −0.2 and 0.0 % at 50, 75 and 100 % speed, and −0.2 % on the second unit; currents are within 1.3 %. All nine registered hypotheses held, also at both ends of the literature ranges and without any calibration. The verdict is “confirmed”. With these laws added to the passport, our method synthesized the refined model anew. It passed all 144 behavioural checks at the first synthesis, with no manual edits, and reproduced the registered predictions for all 617 records of the first unit within 0.005 %. Kalman filters with the refined model: predicted the power and current of each new test within 1–3 % and 0.6–0.9 %; ranked three steps of impeller damage at all speeds; a point-mass filter estimated the flow from power within 1.3–⁠ 1.4 m³/h, below the ±5 m³/h accuracy of the clamp-on flowmeter. Speed from vibration, blind test. A search window set by the slip of the model solved the low-load cases where the published method fails: on throttling, 93 % of records against 55 %. The model predicted the relative change of the slip under throttling within 2.3 %. Over the whole set it did not exceed the published method (85.8 % against 87.5 % for our reimplementation). Learning the window with a Kalman filter from its own estimates made it worse (79.5 %); this registration is not confirmed. Our first reference for this test was wrong: it locked onto the synchronous speed. It was corrected in a second registration and checked against the tachometer. Unbalance from vibration, three registrations. A rotordynamic model of the shaft train built from the documentation did not predict the vibration of the bearing housings. Calibrated on one test, it underestimated weights on the pump side seven- to eight-fold, and a model of the whole rig calibrated on its operational modes did not help: the response is set by the installation, which is not documented. The verdict of both registrations is “not confirmed”. Influence coefficients from the data, with neither a once-per-revolution mark nor a model, estimated weights of 30–90 g on the pump side within 4 % and found the side of every weight from 10 g. The verdict is “confirmed”. A registered threshold on the residual of this fit kept other changes from being taken for a weight: on 41 blind comparisons with misalignment, cavitation and an outdated baseline it raised no false weight, against 24 without the threshold. A digital twin from the documentation. Exported as a Functional Mock-up Unit, the model runs inside the filters at about 40 ms per call, and the unbalance observer adds 0.06 % to the computing time of the loop. Built from the data sheets, refined with published laws and kept in step with the measurements, the model and the filters form a digital twin of the unit; beyond the documentation it needs one calibration point and, for vibration, a baseline record.