Condition monitoring and predictive maintenance run on data, and the data that matters most — a machine with a fault, at a known severity, in a known location — is the data nobody has. Real faults are rare, expensive to wait for, and almost never labelled. RotorDynamics is a Modelica library built by Model Based Innovation for one purpose:

to create credible synthetic data for fault detection in rotating machinery and gearboxes.

The library models shaft lines, rotors, rolling-element, plain and fluid-film bearings, couplings, housings and foundations, and the gearing that connects them — spur, helical, double-helical, internal, bevel, spiral-bevel, hypoid, crossed-helical, worm, rack-and-pinion and complete planetary stages. On top of that physics sit the fault models: bearing race defects, spalled teeth, eccentric wheels, misaligned couplings, planet-bearing faults and rotor instabilities. Run the same machine healthy and faulted, and you get vibration, displacement, speed and torque signals whose fault content is known exactly — ready for training and benchmarking diagnostic algorithms, or as the physics core of a digital twin.

What sets RotorDynamics apart is not the component list. It is that every claim the library makes about its own physics is checked, recorded and traceable — see Verification & Validation below.

The bearing-fault model implements the ball-impact formulation of Ishibashi, Han and Kawai (Modelica Conference 2017) and extends it considerably. Impacts are triggered by shaft angle, not by time, so the impulse train stays phase-locked through a run-up — the ratio of impact spacing to ball-pass period is 1 at every impact to better than 10⁻⁶. The contact pulse follows the Hertzian profile and transfers exactly the momentum an elastic collision must.

The result is the signature an envelope-based diagnostic keys on: lines at the ball-pass frequencies (BPFO, BPFI) and their harmonics, shaft-speed sidebands for an inner-race defect — and nothing at those lines in the healthy control run.

Other fault and excitation mechanisms in the library:

• Spalled teeth, eccentric wheels and runout on any gear mesh
• Planet-bearing faults on a carrier that moves
• Coupling offset and angular misalignment
• Internal-damping (Newkirk–Kimball) whirl instability
• Angle-synchronous motor torque ripple
• Belt pull and spring loads for test-rig reconstructions

IMS Bearing Run-to-Failure Rig

A reconstruction of the IMS / University of Cincinnati endurance rig (NASA PCoE dataset): four bearings on one shaft, a 13.3 kN radial preload, one accelerometer per housing. The pedestal, shaft mass and fault impulse were calibrated against the real accelerometer recordings, bringing the simulated-to-measured RMS ratio from 96×–737× down to 0.84×–1.03× across all four channels. The calibration also found that one bearing needed an extra degree of freedom — an outer-ring resonance near the dataset’s own BPFO — rather than a retuned damper.

Laboratory Rotor Kit

The rotor kit from the original paper, rebuilt from its published parameter table. With no parameter tuned, the model predicts the first critical speed at 1669 rpm against the paper’s 1600 rpm (+4.3 %) at the paper’s own run-up rate, and within about 1 % once the sweep-rate bias is removed. The one parameter the paper never reports — bearing radial stiffness — moves the answer by less than 0.1 % over two decades.

Planetary Gearbox with a Flexible Ring

Why does a fixed accelerometer on a planetary ring see sidebands at all? On a rigid ring, equally spaced planet forces cancel and there is nothing to measure. RotorDynamics models the ring rim as a flexible structure, so the sidebands are an output of the physics rather than of an assumed weighting window. The model reproduces the published sideband pattern (Inalpolat 2009, case i) from first principles — and the control run with a stiff rim correctly loses them. A digital twin of the public NLR Gearbox Reliability Collaborative 750 kW gearbox is under way on the same components.

Right-Angle Spiral-Bevel Drive

A complete right-angle drive with its thrust reacted where a real gearbox reacts it, in the locating bearings. This example also carries a machine-readable context of use: the quantities of interest for a specific engineering task, declared on the model itself, so the tooling can check before a single simulation runs whether the components used can deliver them.

Most simulation libraries ship with examples. RotorDynamics ships with a credibility report: a Jupyter notebook that runs every check live against a Modelon Impact workspace and regenerates every number, table and figure from scratch. Nothing in it is transcribed. It follows the Credible Modeling Process (the model-level tier of the prostep ivip SmartSE Simulation Credibility Assessment framework) and uses ASME VVUQ 1-2022 terminology throughout.

The report’s credibility factor assessment follows SmartSE:

The overall level is set by the weakest factor. The report states plainly what holds it there: a single developer, no independent reviewer yet, and only a subset of features has been validated with real validation data. However, many values in gearboxes are hard or impossible to measure, and the code verification is extensive.

The credibility evidence is not a separate document that drifts out of date. It lives in the library itself: every component carries an evidence annotation, every test model declares which claims it establishes, every connector declares its interface, and task-level examples declare their context of use. From these, tooling generates machine-readable credibility records in the SSP Traceability format — one per component family — signed by the model owner and checked by a release gate before every release.

For commercial, encrypted distributions the same evidence is carried as a side-car in the library’s Resources folder: same content, different serialization. We are working to bring this library-evidence process into the next version of the SSP Traceability standard.

• Modelica Standard Library 4.1.0
• Developed and verified in Modelon Impact (OCT compiler)
• Credibility report: Python, modelon-impact-client, numpy / scipy / matplotlib
• Regression testing against stored reference trajectories
• Export to FMI for use in digital-twin platforms

Reproducible in ten years: the report pins model file hashes, the git commit, the compiler and library versions and every experiment definition as a literal. Re-running the notebook end to end regenerates every result.