Unbalance identification of high-speed aerostatic conical bearing-rotor system based on Bayesian adaptive optimization CNN-Res time-domain automatic feature extraction algorithm

The aerostatic conical bearing (ACB) is a type of variable clearance bearing that is capable of simultaneously supporting both radial and axial loads. Nevertheless, current unbalance identification methods overlook the variable-clearance characteristic of ACB-rotor system and suffer additional errors from fundamental frequency feature extraction distortion. Therefore, an unbalance identification method based on the Bayesian adaptive optimization convolutional neural network with residual connections (CNN-Res) is proposed, achieving automatic feature extraction in time domain. Through a coupled three-dimensional transient nonlinear dynamics model of ACB-rotor system, we construct the unbalance response dataset with the adaptive gradient Latin hypercube sampling method (AGLHS). The local feature extraction, global feature retention, and hyperparameter adaptive optimization are achieved. The results demonstrate that, the CNN-Res model eliminates the need for Fast Fourier Transform (FFT) and avoids spectral leakage errors. Sub-synchronous, fundamental, and super-synchronous frequency features are retained, thereby circumventing frequency distortion caused by variable clearance in the ACB-rotor system. The Bayesian adaptive optimization can significantly reduce identification errors of outliers. With fewer iterations, global optimal sampling frequency can be determined. Therefore, computational resources are saved. The magnitude and phase of unbalance are accurately identified.