Sustainable Marine Structures

Efficient predictive maintenance is vital for maintaining operational reliability in marine logistics infrastructure, especially within ports and offshore hubs where equipment failure can result in costly downtime and disrupted supply chains. This study introduces an AI-driven predictive maintenance and cost-risk optimization framework that integrates advanced machine learning techniques with Mixed-Integer Linear Programming (MILP) to enable dynamic and data-driven maintenance scheduling. The proposed framework utilizes real-time asset data, including sensor readings, environmental variables, and operational logs collected from 124 marine logistics assets over a six-month monitoring period. Predictive models were developed using the Random Forest algorithm and rigorously validated through time-blocked and grouped cross-validation to prevent data leakage and ensure temporal consistency. The model achieved strong predictive performance, with an AUC of 0.86 and a PR-AUC of 0.71, while calibration reliability was verified using the Brier score and decision curve analysis. The MILP-based optimization component incorporated operational constraints, such as maintenance crew availability, failure probabilities, and environmental stressors, to generate cost-effective maintenance schedules. Implementation of the proposed system resulted in a 21.4% decrease in unplanned downtime, a 16.2% improvement in Mean Time Between Failures (MTBF), and a 13.8% reduction in overall maintenance costs compared with historical benchmarks. This research offers a scalable, interpretable, and data-driven framework for predictive maintenance in complex marine environments, contributing to the advancement of smart port operations, sustainable asset management, and AI-enhanced infrastructure resilience.