Transfer learning on frequency-domain data of vehicle finite-element-simulations in vibroacoustics

In the early stages of vehicle development, surrogate models and machine learning play an increasingly important role. In order to optimize the sound in the interior efficiently, engineers use surrogate models of finite element simulations to minimize high-priced hardware tests. Surrogate models based on neural networks have a high demand of time-consuming simulations to provide a sufficient data basis for training, validation, and testing. In this work, simulations from real development processes are used to investigate the methodology of transfer learning to lower this demand. The transfer learning methodology uses a pre-trained neural network of vehicle A to train a surrogate model for vehicle B with significantly less data. The data basis consists of mechanical point admittances from vibroacoustic simulations in the frequency-domain, which characterize the vibration propensity of a structure and describe a typical evaluation parameter in vibroacoustics. We evaluate this methodology of transfer learning for different vehicles based on the performance of the neural networks. With good performance of the transfer learning step, time-consuming finite element simulations can be saved to train surrogate models for vehicle design. To the best of our knowledge, the investigation of transfer learning on time-consuming, real-world production finite element simulations of different vehicles has not been available in the literature yet. Furthermore, this work is characterized by frequency-domain data.