Full vehicle NVH end-of-line testing and data-driven fault detection of a high-performance hybrid vehicle front e-axle

This paper presents an AI-based approach to detect Noise, Vibration and Harshness (NVH) anomalies during end-of-line testing, leveraging a database of experimental recordings from one healthy vehicle and seven other vehicles with anomalies. This study focuses on e-drive maneuvers to detect faults on the front e-axle. To obtain a sufficiently large database of signals to train data-driven models, an NVH simulator is used to reproduce, in a sound quality equivalent manner, the in-cabin sound and, after sound quality validation, to synthesize 20000 realistic vehicles' acoustic performances in healthy and faulty conditions. The whole dataset is labeled to distinguish between healthy and faulty samples. Key psychoacoustic metrics extracted from the sounds are used to train various autoencoder architectures for fault detection, which are evaluated on unseen synthesized sound samples and experimental recordings, demonstrating the system's efficacy.