A novel multi-coupled neural network for nonlinear dynamic prediction of mistuned bladed disk

In order to overcome the limitations of low computational efficiency in existing solution methods for analyzing the nonlinear dynamic characteristics of gas turbine, and the lack of physical interpretability of traditional surrogate models, a multi-coupled neural network (MCNN) has been proposed for predicting three-level pivotal dynamic parameters, including node-level temporal dynamics at contact surfaces, component-level spatial responses of mistuned blades, and system-level damping performance of bladed disk. A high-fidelity finite element reduced-order model of a full-circle turbine-bladed disk is constructed, and the nonlinear solution method is employed to generate the training dataset for the proposed neural network. The results show that the MCNN framework achieves superior accuracy and computational efficiency in predicting dynamic characteristics of bladed disks compared to traditional solution methods. The prediction errors in both time-domain and frequency-domain responses are within 0.1%, and the prediction errors in the amplitude amplification coefficient and modal damping ratio are less than 2% and 2.5%, respectively, outperforming traditional machine learning methods. In terms of computational time, the calculation speed of MCNN for a single case is improved by five orders of magnitude over traditional methods. The proposed MCNN framework can comprehensively and efficiently present the nonlinear dynamic characteristics of gas turbine bladed disks under multiple operating conditions, providing a rapid and accurate analysis tool for preliminary design and further optimization of gas turbine bladed disks in engineering scenarios.