Data-Driven Modeling Approach for Mistuned Cyclic Structures

Abstract

A data-driven approach to predict dynamic responses of mistuned cyclic dynamic structures is presented. Nominally cyclic structures such as turbomachinery bladed disks, or blisks, contain deviations from perfect symmetry, called mistuning, which can significantly increase the forced response of the structure subject to periodic excitations. The ability to accurately model the response is crucial for understanding a structure's dynamics during operation, and thus many reduced-order models have been developed. Data-driven approaches, however, are largely unexplored for predicting physical responses of cyclic structures, and are desirable due to their ability to leverage both computational and experimental data. In this paper, two feed-forward neural networks (FFNNs) are trained using physical responses and parameters from an individual sector of the cyclic structure. The desired physical responses of all sectors are predicted via a cyclic coupling procedure. By only using sector-level data, the number of simulations and/or experiments required to generate training data is substantially reduced, and the approach is not restricted to a specific projection basis unlike previous physics-based methods. A lumped mass model (LMM) of a blisk with stiffness mistuning is used to generate computational surrogate data for training and validation. Both FFNNs achieve high prediction accuracy on both training data and unseen test data. For a test case with an unseen mistuning pattern, this approach shows strong agreement with known actual responses. For this test case, prediction errors in the LMM blade tip amplitudes across all sectors are less than 2%, and the amplification factor error is less than 0.2%.