DETERMINING THE IMPACT OF THE HUMAN OPERATOR FACTOR USING A NAVIGATION SIMULATOR

https://doi.org/10.33815/2313-4763.2025.1.30.158-170

Abstract

As maritime operations become increasingly complex, there is a growing need for objective, scalable methods to evaluate seafarer competence beyond traditional instructor-based assessments. This study presents a comprehensive, data-driven framework for analyzing cadet performance in maritime simulator exercises, utilizing state-of-the-art unsupervised learning and explainable AI. The research draws on multivariate time-series data from navigational simulations, capturing vessel dynamics, control actions, and environmental parameters across dozens of features. A rigorous preprocessing pipeline was applied, combining statistical feature aggregation and redundancy reduction through Pearson correlation and Mutual Information. This yielded a compact, yet informative feature set encompassing control inputs, navigational states, and vessel motions. To offset limited number of sessions small in the dataset, each simulation was split into meaningful intervals by applying rolling window statistics. Each window was then encoded as a summary vector, reflecting both central tendencies and temporal variability. The analysis employed the HDBSCAN clustering algorithm, which excels in detecting groups of variable density and naturally identifies outlier behaviors—critical in the context of training evaluation. The resulting clusters were projected into lower-dimensional space via t-SNE, providing interpretable visualizations of cadet performance patterns. To further elicit the distinguishing characteristics of each group, a linear Support Vector Machine was trained to predict cluster membership, with SHapley Additive exPlanations (SHAP) attributing each decision to underlying features. Key findings reveal that clusters align with distinct navigational strategies: stable, conservative approaches are differentiated from more dynamic or risk-prone styles by features such as roll velocity, yaw rate, and engine RPM. Sessions flagged as outliers typically exhibited abrupt maneuvers or inconsistent control usage, highlighting potential skill gaps. The SHAP-based interpretability layer transforms complex model outputs into actionable instructional feedback, enabling targeted interventions and tailored training. Overall, this automated approach has potential to become a transparent, scalable alternative to subjective grading in maritime education, with significant implications for enhancing safety and developing individualized learning pathways. The proposed system demonstrates strong potential for integration into real-world training environments and continuous improvement as more operational data becomes available.