面向集运MaaS多式联运的多模态大模型研究与应用

Research and Application of MLLMs for Container Shipping MaaS Multimodal Transportation

摘要: 多模态大模型作为人工智能的前沿方向，正逐步成为推动智慧物流升级的核心动力。在集运MaaS多式联运这一复杂场景中，传统单模态方法难以融合船舶AIS轨迹、港区监控视频、电子单证与气象文本等多源异构数据，导致跨模态语义理解与决策支持不足。针对上述问题，本文提出一种基于多模态大模型的智能系统架构，并构建四大核心应用场景：智能单证识别、货物异常识别、智能规划决策与业务数据问答。系统利用Qwen2.5-VL-32B实现单证自动解析与结构化输出，通过图文一致性比对提升货物核验效率，整合实时多源数据为客户提供多式联运路径优化方案，并以Qwen3-235B-A22B支撑自然语言查询与多轮交互。研究表明上述应用在提升物流效率、优化运营决策与降低作业成本方面具有显著潜力。

Abstract: Multimodal large models, as a frontier of artificial intelligence, are increasingly becoming a principal driving force for advancing intelligent logistics. In the complex context of container consolidation under Mobility-as-a-Service (MaaS) multimodal transport, traditional unimodal approaches struggle to integrate heterogeneous data sources—such as vessel AIS trajectories, port surveillance video, electronic documents, and meteorological texts—resulting in insufficient cross-modal semantic understanding and decision support. To address these challenges, this paper proposes an intelligent system architecture based on multimodal large models and develops four core application scenarios: intelligent document recognition, cargo anomaly detection, intelligent planning and decision-making, and business-data question answering. The system employs Qwen2.5-VL-32B to automate document parsing and structured output, leverages image-text consistency checks to enhance cargo verification efficiency, integrates real-time multisource data to provide multimodal route-optimization solutions, and utilizes Qwen3-235B-A22B to enable natural-language queries and multi-turn interaction. Empirical analysis demonstrates that the proposed applications have significant potential to improve logistics efficiency, optimize operational decision-making, and reduce operational costs.