Assessing annoyance in automotive seat adjustments; perception and prediction modeling of subjective annoyance responses using advanced approaches; a comparison of regression methods and neural networks

Abstract

In the automotive industry, acoustic comfort is a key aspect of perceived quality, especially in high-end vehicles where even subtle noise such as squeaks, buzzes, or rattles can negatively impact user satisfaction. Sound Quality (SQ) assessments help ensure a premium experience but typically rely on subjective jury testing, which is time-consuming and depends on expert judgment that may not reflect everyday user expectations. This thesis aims to develop a predictive model for annoyance ratings based on objective acoustic parameters. This is a challenging task, as annoyance is inherently subjective and influenced by various perceptual factors. While traditional methods such as linear regression can estimate simple perceptual attributes like loudness, they fall short when modeling more complex, non-linear characteristics. To overcome these limitations, machine learning approaches, including neural networks and random forests, are investigated and compared to linear and polynomial regression models. The study focuses on seat adjustment mechanisms as a use case. These sounds are relatively easy to isolate and analyze, show noticeable variation across vehicle brands, and are less affected by external noise sources. This makes them a suitable candidate for controlled testing and model development. Using those sounds, a large scale listening test is made to assess annoyance and be able to train the models. Results demonstrate that machine learning models can successfully predict perceived annoyance based on objective metrics, offering a promising alternative to traditional jury testing. Such models could significantly improve the efficiency and scalability of SQ evaluations in the automotive industry.