Empirical identification of squeeze-film damper bearings using neural networks

To date empirically obtained SFD models have been based upon the determination of linearised force coefficients; such models are severely limited in their range of applicability since they are only valid for small perturbations from a mean position. The present research provides the introduction and validation of a nonlinear SFD identification technique that uses neural networks, trained from experimental data, to reproduce the input-output function over the full range of the SFD clearance. Details of the commissioning of a specially designed identification test rig and its associated data acquisition system are presented. The neural network's construction and training process is described and relevant testing is detailed. The empirically identified neural network is progressively validated, culminating in remarkably accurate nonlinear vibration response prediction of an SFD test rig subjected to external dual-frequency orthogonal excitation, as present in twin-spool engines (where the nonlinear vibrations are driven by the unbalance on the two rotors turning at different speeds). When used within the dynamic analysis of the test rig, the trained neural network is shown to be capable of predicting complex nonlinear phenomena with excellent accuracy. By comparison to an advanced theoretical model, the results show that the neural networks are able to capture the effects of features that are difficult to include in a hydrodynamic model or are particular to a given SFD. (C) 2012 Elsevier Ltd. All rights reserved.