Application of machine learning on complete vehicle data for interior vibration prediction based on transfer paths in narrowband spectrum

Interior noise and vibration are critical quality criteria for vehicles in the automotive sector. The latter is essential in driving comfort, where many development resources are used for the design and the evaluation. As a result, testing the entire vehicle is very important to assess vibration and ride comfort. The assembly of the entire vehicle is carried out virtually using transfer functions derived either from simulations or from measurements. Since simulations of a complete vehicle are still challenging,transfer functions are usually measured with considerable effort. Therefore, limiting the necessary measurements to specific components is desirable to be efficient and to save costs. In this work, the authors use an artificial neural network to predict the interior vibration profiles of a complete vehicle based on full test drives at different speeds and road conditions. Triaxial acceleration measurements serve as the database. In addition, a criterion is proposed to select essential sensors for the learning process. The results show that if the data is handled carefully, many sensors can be discarded, and the network can predict accurate acceleration spectra for virtual sensors at various test conditions.