Research progress and classification analysis of sound quality prediction models for electric vehicles

Abstract

As interior sound quality has become a critical factor influencing perceived comfort and market competitiveness of electric vehicles (EVs), conventional evaluation approaches relying solely on A-weighted sound pressure levels exhibit limited representativeness under complex, non-stationary in-vehicle acoustic conditions. This article presents a systematic and methodology-oriented review of recent advances in EV interior sound quality modeling, with particular emphasis on data-driven and artificial intelligence–based approaches. Rather than treating modeling techniques in isolation, the review first analyzes commonly adopted acoustic input representations—including psychoacoustic parameters, physical sound pressure descriptors, and time–frequency or structured feature forms—together with the subjective evaluation designs that generate perceptual labels and implicitly shape dataset characteristics. Subsequently, mainstream modeling frameworks, ranging from classical machine learning and shallow neural networks to deep learning architectures and hybrid models incorporating evolutionary optimization strategies, are systematically categorized and comparatively discussed. Through statistical analysis of model usage frequency, dataset scale, preprocessing strategies, and reported performance metrics, this review elucidates the trade-offs among model complexity, training stability, and generalization capability under practical engineering constraints, particularly in small-sample and non-stationary scenarios. Finally, emerging research trends and open challenges—including data-efficient modeling, feature–model matching with interpretability considerations, and multi-source feature fusion—are identified, providing methodological insights to support the robust and intelligent evolution of sound quality evaluation systems for electric vehicles.