Pure electric vehicle nonstationary interior sound quality prediction based on deep CNNs with an adaptable learning rate tree

Abstract

Vehicle nonstationary interior noise has nonstationary characteristics that negatively affect the sound annoyance of passengers. Currently, there are some deficiencies in the research of vehicle interior nonstationary noise. (1) Numerous works have studied conventional vehicle interior noise, but limited works have investigated PEV interior noise. (2) Few studies have examined the sound characteristics of vehicle nonstationary interior noise (acceleration and braking conditions). (3) In using intelligent prediction methods such as deep convolutional neural networks (CNNs), reducing the learning rate during training gradually narrows the search range of a solution and becomes trapped in local optima. Consequently, the nonstationary interior noise of PEVs is studied in this paper. A method for quantitative sound quality prediction of the PEV nonstationary interior noise based on tacho-tracking psychoacoustic metrics and deep CNNs with adaptable learning rate trees (ALRT-CNNs) is presented to solve the aforementioned problems. There are two original contributions of this paper. First, ALRT-CNNs can adaptively reduce and increase the learning rate based on the training loss, and an appropriate search range for a better solution can be obtained. Second, the proposed prediction method can comprehensively reflect the nonstationary sound characteristics of PEV nonstationary interior noise as well as their influence on human subjective annoyance.