Interpretable Passenger Discomfort Prediction in Automated Driving Using Transformers

This study employs a Temporal Fusion Transformer (TFT) model to forecast passenger discomfort in the context of highly automated driving, utilizing time-series data collected from a driving simulator. The objective is to assess whether transformer-based models can provide accurate and interpretable predictions of subjective discomfort levels during vehicle driving. Two variants of the TFT model were developed, referred to as TFT-full and TFT-restricted. Both models were trained on simulator data from 100 participants. Additionally, they were compared to DeepAR, an autoregressive time-series model used as a baseline. Results showed that the TFT-restricted model performed better than both the TFT-full and DeepAR models. The TFT-restricted model achieved the best overall performance, with a Mean Absolute Error (MAE) of 0.042 and a Root Mean Square Error (RMSE) of 0.133. These findings support the suitability of transformer-based models for discomfort detection in automated vehicles.