Data-Driven Evolutionary Belief Rule Base Modelling for Explainable Arctic Navigation Safety Assessment

Abstract

The Arctic Northeast Passage is of growing economic and strategic importance, while its harsh and highly variable navigation environment poses substantial challenges for vessel operation. Existing assessment approaches often rely on predefined structural assumptions or sacrifice interpretability for numerical performance. This study proposes a data-driven evolutionary belief rule base model that integrates an adaptive reference value field mechanism with differential evolution-based structural optimization for rule pruning and growth. The inputs are observed vessel-operational and environmental variables from the Oden and Tian En datasets, and the output targets are three-grade expert-consensus navigation safety risk distributions aggregated from 100 independent votes per sample. Under the repeated random-split evaluations, ARVF-DEBRB achieved the lowest mean MAE on Oden and errors comparable to the leading learned baselines; on Tian En, its mean errors were comparable to those of the strongest learned baselines while lower than those of several conventional and deep baselines. Complementary chronological rolling-window evaluations over three disjoint future test windows provided future-only evidence as follows: ARVF-DEBRB ranked second in RMSE and third in MAE on Oden, and first in RMSE and second in MAE on Tian En. The rolling factorial ablation obtained the lowest mean errors for the complete configuration on both datasets. These results indicate competitive predictive performance under the reported evaluation protocols, while preserving transparent IF-THEN reasoning semantics and a compact rule structure. The data-adaptive ARVF construction and DE-based rule-structure refinement provide a methodological basis for recalibrating the framework to new operational settings.