Evaluation of Neural Networks as a Technique for Correlating Vehicle Noise with Subjective Response 972016

A number of objective measures are commonly used for an indication of subjective response to vehicle interior noise, including dBA, Loudness, Speech Interference Levels (SIL). Intelligibility etc.. Additional measures have been developed from a combination of specific orthogonal objective parameters with the aim of further improving the correlation with subjective response. For example, the Composite Rating of Preference (CRP) Index has been found to consistently improve the correlation upon dBA, through the inclusion of a spectral balance, and high frequency component However the ability of any objective parameter or index to provide a good correlation will be limited by any non-linearity present in the subject response, typically occurring through saturation of one or more aspect of the vehicle noise. This is further frustrated through subject inconsistencies and disagreements. This paper looks at the extent of such non-linearities and disagreements, and discusses how neural networks can be used to model them.

A robust non-linear model has been developed for the aspects of overall level, spectral balance, and high frequency content within vehicle interior noise. The approach taken here is unique in the field of vehicle noise metrics, since disagreements in subject preferences are accounted for to yield a probability of preference, and an associated level of confidence in that probability. Consideration is also given to the inclusion of data on customer populations (e.g. customer's age) as additional inputs to the training of neural networks. Eventually this will help in the definition of desirable vehicle refinement characteristics for specific market sectors and customers, and consequently will focus design and development efforts for vehicle noise quality.