Turbo-generator rotor fault diagnosis via vibration mode decomposition signatures

Frequent load fluctuations from renewable-energy integration increase operational stress on steam turbo-generators and raise the risk of rotor inter-turn short circuits. This paper proposes a measurement-oriented multimodal framework for robust fault-state estimation under complex operating conditions. A fault-sensitive measurand derived from 50 Hz and 100 Hz vibration components at the stator excitation end, is defined and theoretically linked to rotor winding faults. An electromagnetic-mechanical coupled model in ANSYS reveals how rotor inter-turn faults modulate stator vibration spectra; modal characteristics are validated on a 600 MW turbo-generator. A formal measurement model maps the chain from vibration sensing and speed-reference acquisition to feature transformation and fault-state estimation. Variational mode decomposition parameters are optimized via the arithmetic optimization algorithm; band-pass amplitude sequences yield multimodal time-frequency features, processed by a multimodal feature-enhanced temporal neural network. Traceability from transducer to processed feature is ensured, and uncertainty is quantified across signal-to-noise ratio variation, frequency resolution, modal separation, repeatability/reproducibility, and confidence intervals of estimates. Under test conditions, the method achieves robust fault-state estimation and superior noise immunity versus baseline models, demonstrating strong potential for practical rotor winding fault diagnosis in large turbo-generators. • A measurement-oriented framework for rotor winding fault-state estimation. • Fault-sensitive modal signatures are extracted from non-stationary vibrations. • Optimized VMD improves separation of fault-related vibration components. • Multi-physics simulation validates spectra under inter-turn short circuit. • A lightweight temporal network enables robust estimation under noise.