Driving operation state transition prediction under moving bottleneck scenario via Graph-Attention LSTM with Bayesian survival features

Moving bottlenecks create localized traffic disturbances that trigger sequential transitions in driving operation states, from stable cruising to a series of braking, steering, and recovery maneuvers. However, how these states evolve under different driving automation levels and whether their transitions can be predicted remain insufficiently understood. This study develops an integrated framework for identifying, interpreting, and predicting driving operation state transitions in a moving-bottleneck scenario. A driving-simulator experiment was conducted under manual driving (Manual), advanced driver-assistance system (ADAS), and automated driving (AUTO) modes at three traffic-density levels. Vehicle-control signals, eye-tracking measures, and surrounding-vehicle interactions were collected. First, a hierarchical Dirichlet process hidden semi-Markov model (HDP-HSMM) was used to identify latent driving operation states. Eight driving operation states were identified: Mixed residual state (S0), Hard braking (S1), Brief braking adjustment (S2), Brake release (S3), Rightward evasive braking (S4), Stable cruising (S5), Leftward steering correction (S6), and Sustained braking (S7). These states provide a scenario-specific empirical representation, with weaker stability for S0 and, to a lesser extent, S3 and S6. A Bayesian multi-state survival model was then applied to estimate transition hazards and their duration-dependent variations across driving automation levels. Compared with Manual driving, ADAS and AUTO were associated with higher S5-to-S2 and lower early S7-to-S1 transition hazards. Manual driving was associated with an earlier S4-to-S3 transition. Finally, a graph-attention long short-term memory (GAT-LSTM) model incorporating eye-tracking and Bayesian survival features was developed to predict transition occurrence, the destination driving operation state, and time to transition. Across driver-level out-of-fold predictions for all 41 drivers (19,284 windows), the full model achieved accuracy 0.816, F1-score 0.868, AUC 0.868, direction accuracy 0.598, and a time-to-transition mean absolute error of 0.493 s. Driver-cluster bootstrap comparisons showed target-specific trade-offs: the graph-related configuration improved transition F1, eye features reduced F1 and increased timing error in the evaluated configuration, and survival-derived features improved timing but reduced destination accuracy. The target-specific predictive results inform future evaluation of warning timing and human-automation coordination around moving maintenance operations.