A neural network model with filtering-based time-domain data augmentation method for predicting vibration discomfort

This study investigated the vibration-induced discomfort experienced at the second-row seat in multi-purpose vehicles (MPVs) and proposed a deep learning-based prediction model for discomfort. Experiments were conducted on a four-poster test rig to collect time-domain accelerations in three directions at the backrest, seat pan and armrests, with different seat absorber designs and vehicle operating conditions. A Long Short-Term Memory (LSTM) neural network was developed to model the nonlinear relationship between the objective vibration features and subjective discomfort ratings. To increase the sample size and enhance generalisation, a data augmentation strategy was implemented by filtering white noise signals to match the spectral characteristics of the measured accelerations. Results demonstrated that using the three-directional backrest acceleration as input yielded high prediction accuracy; however, expanding input features to include vibrations from other seat components led to degraded performance. This study established a framework for modelling the relationship between seat vibration and occupant discomfort.