Perpustakaan Digital ITB

Rotating unbalance may cause large vibrations in machinery, leading to damage or failure. It is addressed using balancing methods such as the Influence Coefficient Method, which rely on vibration measurements. In precision balancing, where noise may dominate the unbalance signal, these measurements must be processed to obtain reliable vibration characteristics. Among the available noise-reduction approaches, the improved Time Synchronous Averaging (TSA) method gives the best balancing performance, since it suppresses non-synchronous noise regardless of frequency rather than only outside a passband. However, it requires many averages to converge, and since each average takes a fixed acquisition time, the measurement becomes increasingly long. This research therefore applies machine learning (ML) to predict the converged TSA characteristics from a reduced number of averages, and verifies the resulting balancing performance. Convergence histories are recorded from the TSA acquisition on a single-stage rotor, storing the current 1X amplitude and phase at every averaging index. Four regression models, namely Random Forest, Gradient Boosting, Gaussian Process, and Support Vector Regression, are trained to predict the residual that carries the current reading to its converged value. The models are evaluated on a test set of increasing generalization difficulty, and the predicted values are applied in single-plane balancing, compared against the no-ML result at equal averages and the full 400-average reference. The models reduce the convergence error across most conditions, with the Gaussian Process giving the best overall prediction and the tree ensembles most robust under severe generalization. In balancing, the reduced-average estimates reproduce the full-average outcome at 0.1 g, where the signal is strong enough to converge at a low count. At 0.05 g, where the weaker signal-to-noise ratio slows convergence, the no-ML method reaches only about 76% and 78% at 50 and 100 averages, while the best model reaches 87.7% at 100 against 87.5% at 400. This corresponds to a fourfold to eightfold reduction in acquisition time, showing that ML prediction of converged TSA values can substantially shorten precision balancing while preserving the balancing outcome.