Transfer Learning-Based Neural Network for Natural Frequency Prediction of Linear Dynamic Systems

Abstract

The prediction of natural frequencies is a crucial aspect of engineering design and analysis. Traditional methods involve finite element analysis (FEA) which is a standard method for calculating natural frequencies of dynamic systems. For each design variant, FEA calculation can be time-consuming and computationally expensive. In this study, we propose a novel method for predicting the natural frequencies of design variants using transfer learning and artificial neural networks (ANN).

The proposed method involves the use of FEA to generate the stiffness and mass matrices of the brake disc, which are then used as inputs to the neural network. However, the prediction can become tedious when there is a change in the design. To address this, we employ transfer learning followed by linear regression using a design variant of the previous structure as test data. The neural network learns through transfer learning and fine-tunes its outputs using regression for final frequency prediction.

The proposed approach can predict the natural frequencies of new structures efficiently without compromising the quality of the outcome, even when the degree of freedom changes due to design alterations. The effectiveness of this method is demonstrated by calculating frequencies of brake disc with different material property, and the results are compared with FEA to measure its accuracy. The results indicate that this method can accurately predict the natural frequencies of new design variants with high prediction accuracy and computational efficiency. This method has potential applications in engineering design and analysis, especially for structures that require iterations to finalize design and where there is a need to calculate the dynamic characteristics of the system.