The development of a deep neural network and its application to evaluating the interior sound quality of pure electric vehicles

Interior noise substantially influences the physiological and psychological sensations of passengers in pure electric vehicles (EVs). Numerous studies have examined the development of acoustic prediction models and acoustic metrics to evaluate EV interior sound quality. However, the existing studies have the following four deficiencies: (1) the interior noise of EVs was studied only on general roads, and few EV samples were tested; (2) the physical acoustical metrics and psychoacoustic metrics did not comprehensively reflect all the characteristics of the interior noise of EVs; (3) features added to the acoustic prediction models were manually extracted and selected and were highly dependent on prior knowledge of acoustic theory and experience; and (4) the most common acoustic prediction models used to evaluate interior noise have shallow architectures. To overcome these deficiencies, we introduce a novel intelligent acoustic model based on deep neural networks (DNNs) called the Laplacian score-deep belief network (LS-DBN). We used the LS-DBN to evaluate the sound quality of EV interior noise. To verify the effectiveness of the proposed method, the interior noises of ten EVs were recorded on eight different road surfaces and corresponding subjective evaluations were conducted. In addition, noise features were extracted adaptively using the LS-DBN, and adaptively extracted features and manually extracted features were compared. The performance of the LS-DBN was validated against a conventional DBN and a back-propagation neural network (BPNN). The results show that the proposed LS-DBN model is superior to the conventional DBN and BPNN in terms of accuracy and stability, and it is highly efficient. Thus, the LS-DBN can achieve good prediction results when evaluating the interior sound quality of EVs.