Adaptive collision-avoidance timing recognition for autonomous ship encounters using multi-evidence Bayesian backward evidence detection

Avoidance timing is a critical interface between collision-risk assessment and collision-avoidance planning for maritime autonomous surface ships (MASSs). Existing approaches usually trigger avoidance using fixed collision-risk thresholds, closest point of approach (CPA) criteria, ship-domain boundaries, or single maneuvering indicators, which cannot fully capture the combined effects of maneuvering behavior, relative-motion evolution, and encounter-type-dependent risk evolution. This study proposes a multi-evidence Bayesian backward evidence-detection framework to identify avoidance timing and learn encounter-specific adaptive trigger thresholds from historical automatic identification system (AIS) trajectories. Maneuvering, relative-motion, and risk-evolution evidence are integrated into a unified feature representation. Normal navigation and active avoidance are modeled as two latent states through a regularized Gaussian likelihood-ratio formulation, and avoidance onset is identified by accumulating backward evidence from the collision-risk peak. The collision-risk values at the detected avoidance-start moments are then reconstructed using weighted kernel density estimation (KDE) for overtaking, head-on, and crossing encounters, and adaptive trigger thresholds are derived from density modes with safety-advance corrections. Experiments using 6,186 paired AIS encounter files from the Yangtze River Estuary produced 4,361 valid avoidance-timing samples and 4,036 effective threshold-learning samples. The detected avoidance onsets occurred 8.42, 8.60, and 6.50 min before the collision-risk peak for crossing, head-on, and overtaking encounters, respectively. The resulting adaptive thresholds were 0.245, 0.267, and 0.368. Additional train-test validation, sensitivity analysis, and multi-baseline comparison further demonstrate the stability and interpretability of the proposed framework.