Statistical and connectionist approaches to modeling the assessment of noise annoyance inside passenger vehicles

Abstract

Using neural networks, discriminant analysis and regression, a model for annoyance assessment of vehicle interior noise situations is empirically estimated. The data base for the estimate consists of 12 different vehicle-speed pairs for which 24 subjective assessments are available. Descriptive statistical parameters of the dataset are presented. The estimation of annoyance scores using frequency spectra of the noise signals shows good results, for both the connectionist and the classical statistical approaches. The influence of measuring time and of the frequency range on the classification and on the generalization is evaluated. The background of the used methods is briefly described. Their advantages and disadvantages concerning the application to the data and the interpretation of the estimated classifier concerning generalization of the knowledge extracted by these “black box” tools are discussed and compared.