An approach for predicting vibration discomfort via machine learning: integrating objective measurements of sound and vibration with passengers' individual characteristics

Artikel i vetenskaplig tidskrift, 2026

Evaluating ride comfort, particularly vibration discomfort, is crucial in automotive studies. Traditional field studies face challenges such as high costs and inefficiency, prompting the integration of machine learning. However, ensuring data quality, addressing subjective comfort perception, extracting relevant features, standardising evaluations, and improving model interpretability remain key challenges. This study proposes a machine learning framework to tackle these issues, focusing on data limitations, model interpretability, and cross-study comparisons. A random permutation method expanded the dataset from 160 to 1.5 × 10^8 data points, enhancing gradient boosting machine and random forest performance. F1 score was used for evaluation. Focus group discussions highlighted the framework's strengths in improving predictive reliability and cross-study comparisons but noted challenges in computational resources and model generalisability. Overall, the proposed framework demonstrates potential in advancing vibration discomfort research through machine learning.