Engineering Decision-Support Framework with an AI Advisory Layer for Operating-Cost Optimisation Under Railway Disruptions

Effective engineering management of railway disruptions requires a transparent framework that links operating-cost optimisation, detour feasibility and service continuity constraints under uncertain repair durations. This paper presents IAAK (Intelligent Cost Analysis Algorithm), a baseline-gated decision-support framework demonstrated on a selected regional railway line. The deterministic layer computes passenger and freight cost thresholds, while the AI advisory layer is validated on a temporally separated dispatcher decision set and does not replace dispatcher judgement. For railway line No. 30, the framework identifies a direct freight-detour activation threshold of approximately 9.3 h and an upper bound return-plus-detour threshold of approximately 50.6 h. The multi-threshold route inventory is complemented by a 130-case dispatcher decision set for 26 Lubelskie and adjoining lines, with 104 training cases up to 31 December 2025 and 26 test cases from 1 January 2026 to 16 June 2026. On this temporal hold-out, GB and RF reproduce the action classes, while the PNN/RBF comparator reaches macro F1 = 0.927 and balanced accuracy = 0.952. The results improve coverage, temporal validation and operational grounding for dispatcher review. Methodological novelty lies in coupling a hard admissibility gate, deterministic break-even costs, post gate multicriteria ordering and a non-authoritative AI audit in one dispatcher review record rather than in proposing a new cost formula or classifier. On the 26-case temporal hold-out, GB and RF add no action-level disagreement, while PNN/RBF flags 2 of 26 cases; the AI layer is therefore interpreted as an audit channel, not as an accuracy improvement over the deterministic gate. CDRA uses an externally specified scenario probability in the implemented study, whereas probability estimation and online updating remain future extensions. All 26 test lines also occur in the training partition, and the 50.6 h upper bound threshold exceeds the longest recorded closure; these findings are consequently interpreted as within-line scenario evidence rather than cross-network-predictive or empirical-threshold validation.