随着无人技术的发展, 以无人监测船为代表的新型无人平台在渔业活动和非法捕捞行为监测等场景中得到广泛应用。当前, 围绕水中无人平台的智能性的测试和评估工作还处于起步阶段, 同领域可参考的案例样本较少, 急需探索新的方法。该文对无人驾驶汽车智能性等级评定方法进行对比分析, 搭建了适用于合作状态的UP-RAGAs评估框架, 提出了功能映射分级和本体能力评估方法; 在此基础上, 引入天气状况、目标状态等影响因素, 构建了非合作状态下的OODA+E评估方法, 提出了基于Bayes定理的多因素联动评估方法, 以满足不同场景和层级的评估需求。最后通过实例分析验证了相关方法的先进性和可行性。该研究成果拓展了评估结果的维度, 为无人监测船的智能性评估提供了更为科学、系统的解决方案。

Objective: With the rapid development of unmanned technologies, unmanned monitoring ships have become indispensable platforms for marine fishery monitoring and illegal fishing detection, offering advantages such as high flexibility, extended endurance, and remote operability. These platforms significantly expand monitoring scope, improve detection precision, and reduce labor costs. However, the evaluation of intelligence for water-based unmanned platforms remains in its initial stage. Existing evaluation frameworks are predominantly designed for land-based autonomous vehicles and fail to address the unique characteristics of unmanned monitoring ships, including complex marine environments, dynamic non-cooperative targets, and adversarial interactions. Current methods focus primarily on static functional descriptions or virtual environment testing, lacking the capability to reflect real-world dynamic interactions and environmental uncertainties. To address these critical limitations, this study aims to establish a comprehensive, scientific, and targeted intelligence evaluation methodology specifically tailored for unmanned monitoring ships operating in non-cooperative scenarios. Methods: This paper proposes a multi-layered evaluation system comprising three core components. First, the UP-RAGAs (Unmanned Platform-based Retrieval-Augmented Generation Assessment) framework is constructed by reconstructing the traditional RAGAs framework into three stages—task input, platform response, and action output—thereby adapting it to real-scenario operational logic and supporting both algorithm-level software testing and system-level integrated evaluations. Second, a function mapping grading method is developed, classifying intelligence into 0—9 grades across three dimensions: autonomy (independent decision-making capability), cooperativity (multi-sensor data fusion and interoperability), and study ability (adaptive improvement through environmental interaction). Complementing this qualitative approach, an ontology capability evaluation method is proposed, involving 12 key capability indicators distributed across four perception stages: detection (optical imaging, multi-channel fusion, target detection), recognition (target identification, key component detection, behavioral intention analysis), positioning (distance measurement, coordinate transformation, track generation), and tracking (target acquisition, stable tracking, trajectory prediction). Third, to address non-cooperative scenarios characterized by dynamic target behaviors and environmental uncertainties, an innovative OODA+E (Observe-Orient-Decide-Act plus Effect) evaluation method is developed by integrating OODA loop theory with environmental and target state factors. Furthermore, a Bayesian-based multi-factor linkage model is established to dynamically quantify the influence of target movement levels, relative distances, viewing angles, and weather conditions on perception performance, enabling comprehensive assessment under dynamic adversarial conditions. Results: Validation through a typical operational scenario demonstrates the effectiveness and feasibility of the proposed methodologies. The ontology capability evaluation successfully quantifies intelligence using normalized data across the four perception stages, revealing specific performance characteristics. The OODA+E method effectively incorporates dynamic factors, revealing that high target mobility significantly degrades tracking stability, while relative distance inversely affects detection accuracy and positioning precision. Additionally, weather condition variations impact optical imaging performance following expected probabilistic distributions. These quantitative results confirm that the proposed framework can accurately assess intelligence levels under complex, non-cooperative conditions where traditional static evaluation methods would prove inadequate, providing granular insights into factor-specific influences on system intelligence. Conclusions: The proposed evaluation system successfully integrates static grading, quantitative capability assessment, and dynamic multi-factor analysis, effectively addressing the limitations of existing methods that are confined to cooperative scenarios and static testing environments. By realizing multi-dimensional and multi-level intelligence evaluation spanning algorithm performance to system-wide operational effectiveness, this research provides reliable technical support for the standardized application, performance optimization, and mission planning of unmanned monitoring ships in marine-related fields. The innovative incorporation of Bayesian-based multi-factor linkage analysis significantly enhances the scientific rigor and practical applicability of intelligence assessment, offering a systematic and comprehensive solution for advancing the deployment of intelligent unmanned platforms in complex maritime operations characterized by uncertainty and adversarial dynamics.