LAUSR

Abstract The perception of vehicle interior sound quality is important for passengers. In this paper, a feature fusion process for extracting the characteristics of vehicle interior noise is studied, and an improved deep belief network (DBN) that uses continuous restricted Boltzmann machines (CRBMs) to model continuous data is proposed. Six types of vehicles are used for recording interior noise under different working conditions, and a corresponding subjective evaluation is implemented. Psychoacoustic metrics and energy-based criteria using the wavelet transform (WT), wavelet packet transform (WPT), empirical mode decomposition (EMD), critical-band-based pass filter, and Mel-scale-based triangular filer approaches have been applied to extract interior noise features and then develop a fusing feature set combining psychoacoustic metrics and critical band energy based on comparisons. Using the obtained fusion feature set, a CRBM--based DBN (CRBM-DBN) model is developed through experiments. The newly developed model is verified by comparing its performance relative to multiple linear regression (MLR), backpropagation neural network (BPNN), and support vector machine (SVM) models. The results show that the proposed CRBM-DBN model has a lower prediction error and higher correlation coefficient with human perception compared to the other considered methods. In addition, CRBM-DBN outperforms BPNN and SVM in terms of stability and reliability. The presented approach may be regarded as a promising method for evaluating vehicle noise.