Annoyance Prediction of Civil Aircraft Cabin Noise Based on Generalized Additive Models

Abstract

Noise in civil aircraft cabin is a major contributor to passenger comfort. It is of great significance to predict the noise annoyance by developing a model. However, the low accuracy of linear model and the poor intelligibility of the neural network are the critical problems to be solved. Generalized additive models (GAMs) have become a leading choice for explainable machine learning. This study examines the effectiveness of GAMs in noise annoyance prediction. First, the annoyance of the cabin noise was collected through subjective listening test. Subsequently, three physical acoustic features (sound pressure level, A-weighted sound pressure level, preferred speech interference level) and five psychoacoustic features (loudness, sharpness, roughness, fluctuation strength, tonality) were applied to describe the stimuli. In the third step, three algorithms were used to construct GAMs based on annoyance (dependent variable) and feature parameters (independent variables), and the prediction accuracy was computed by ten-fold cross-validation. The result indicated that GAMs have higher prediction accuracy than linear regression model. The importance of features was evaluated, which is more intelligible than the neural network. This study provides the first assessment of GAMs in the annoyance prediction of civil aircraft cabin noise, and the findings may be of assistance to engineering noise control.