Neural Network Models For The Subjective And Objective Assessment Of A Propeller Aircraft Interior Sound Quality

摘要

This paper reports on the use of neural networks for modelling the relation between the salient objective and subjective psychoacoustic attributes of a propeller aircraft interior sound. The developed model grounds on a modular approach consisting in a series of two stages. The first stage is devoted to the data-driven estimation of sound quality features (loudness, sharpness, etc.) in time domain. In the second stage the estimated sound quality attributes are adopted to classify the input sounds in terms of passenger annoyance. This second module consists in an Artificial Neural Network model, trained on the basis of a subjective evaluation test. The paper describes the approach followed for the neural networks definition and for the collection of the subjective and objective propeller aircraft in-cabin psychoacoustic attributes. The adopted model has been compared with alternative machine learning instruments. We finally assessed the accuracy of the model in predicting the passenger response by validating it on experimental propeller aircraft in-cabin noise recordings whose annoyance was evaluated by a pool of jurors in a subjective evaluation test. Such a tool, integrated in a virtual prototyping framework, paves the way for the inclusion of the human perception in the aircraft design optimization process.