Application of VMD-enhanced POD-deep learning model in internal field prediction of aeroengines

Compared with traditional CFD methods and experimental methods, the POD method combined with deep learning can achieve prediction of engine internal flow field data while significantly reducing time and resource consumption. However, the linear computational characteristics of POD itself limit its description of nonlinear features. The primary objective of this study is to investigate the effectiveness of incorporating variational mode decomposition (VMD) into the "POD decomposition-deep learning" architecture. We aim to analyze the effect of VMD decomposition and denoising on organizing nonlinear features in POD decomposition time coefficients and improving the prediction accuracy of deep learning models. The results show that the POD-VMD process has a better effect on improving the prediction accuracy of deep learning models; truncated VMD can better improve the prediction of detailed parts of the engine internal flow field by the deep learning model when K is low, but when K is high, signal loss causes errors that negatively affect the prediction; in terms of neural network adjustment, the dual-channel path cleaning model improves the prediction accuracy for physical fields with lower nonlinearity, but has limited benefits for improving the prediction accuracy of physical fields with higher nonlinearity.