How a Four‑Layer Pyramid Powers AI‑Driven Autonomous Ship Navigation

In April 2025 Dalian Maritime University and industry partners launched "Linghang", China’s first port‑to‑port autonomous navigation large model, demonstrating a shift from human‑centric monitoring to AI‑pre‑control in shipping.

1. AI Era in Maritime

By 2024 the maritime AI market reached $4.1 B, nearly tripling year‑over‑year. Notable deployments include the "VTS Sprite" agent at Shanghai Yangshan Port, the PortGPT model at Tianjin Port, and the Linghang model. The core challenge is enabling a large model to truly read electronic charts (S‑57/ENC), understand navigation rules, and make safe decisions.

2. Four‑Layer Pyramid Architecture

The solution builds a "cognitive ladder" of four layers:

┌─────────────────────────────────────┐
│  第4层：决策指令层（自然语言）      │ ← "建议左转15°，避开前方沉船"
├─────────────────────────────────────┤
│  第3层：场景推理层（语义网络）      │ ← 船舶状态+环境约束+规则匹配
├─────────────────────────────────────┤
│  第2层：要素关系层（知识图谱）      │ ← 障碍物‑航道‑助航设施的空间拓扑
├─────────────────────────────────────┤
│  第1层：结构化数据层（GeoJSON）    │ ← 几何+属性+时态信息
└─────────────────────────────────────┘

Layer 1 – Structured Data (GeoJSON)

Raw chart data is converted into a "maritime element model". Example JSON shows a vessel with draft 10.5 m, position, and nearby hazards with depth and distance:

{
  "vessel": {
    "name": "海丰测试轮",
    "draft": 10.5, // 吃水深度
    "position": {"lat": 38.9333, "lon": 121.6500}
  },
  "hazards": [
    {"type": "wreck", "depth": -8.5, "distance": 1.2} // 沉船水深‑8.5 m，距本船1.2 海里
  ]
}

Key technique: By juxtaposing vessel draft (10.5 m) with obstacle depth (‑8.5 m), the AI can infer a shallow‑water risk: with tide +2.1 m the remaining under‑keel clearance is only 0.1 m, triggering a grounding warning.

Layer 2 – Knowledge Graph

Using Neo4j‑style relationships, chart elements are linked:

(本船:集装箱船)-[:吃水10.5m]->(需要水深>11.5m)
(前方沉船)-[:水深-8.5m]->(危险障碍物)
(沉船)-[:标记有]->(浮标DL1#)
(当前航道)-[:属于]->(大连港主航道)

Inference rules are expressed as concise statements:

Shallow‑water risk: if vessel draft + safety margin > available depth → grounding alert.

Obstacle avoidance: if obstacle depth < vessel draft and angle < 30° → suggest turn.

Lane compliance: if heading deviation > 15° → remind TSS rule.

Layer 3 – Scene Understanding

Structured data is rendered into natural‑language scenario descriptions. Example prompt: "海丰测试轮（集装箱船，长180 m，吃水10.5 m）目的港大连港D3泊位。当前涨潮中，潮高+2.1 m，能见度8 海里。若保持当前航向，CPA约0.8 海里，安全。但前方2 海里将进入VTS报告线，需呼叫大连VTS。" Prompt engineering tips include using maritime terminology with explanations, quantifying risk (distance, time, margin), and explicitly marking uncertainty. Layer 4 – Executable Navigation Advice The final output mimics a seasoned officer’s recommendation, formatted as JSON: <code>{ "alert_level": "caution", "recommendations": [ { "action": "建议向左调整航向至242°", "reason": "当前偏右100m，中线水深更充裕（-14m vs -11m）", "execution": "小幅度左转3°，观察1分钟后回正" }, { "action": "注意右侧沉船浮标DL1#", "reason": "沉船有效水深仅-6.4m，远低于船舶吃水", "execution": "雷达设置0.8海里guard zone" } ] }</code> 3. Route Planning: Hybrid Intelligence Pure LLM planning is slow and unreliable; pure algorithmic planning lacks business context. The optimal practice is "algorithmic coarse screening + LLM fine‑tuning". Stage 1 – Algorithmic Candidate Generation A* or RRT* searches the chart grid, respecting depth constraints (depth > draft + margin), exclusion zones, and lane rules, outputting 2‑3 candidate routes with key waypoints. Stage 2 – LLM Evaluation Each candidate is evaluated on four dimensions: Safety: under‑keel clearance, obstacle distance, traffic density. Economics: distance, time, fuel cost. Compliance: adherence to fixed‑line and reporting requirements. Comfort: exposure to wind‑wave, turning frequency. Example assessment: Route A is shortest (240 nm) but traverses a shallow area (‑9.8 m) with UKC 1.2 m, yielding a medium‑high risk and a recommendation to verify tidal windows or switch to Route B. Stage 3 – Full Navigation Plan The final plan lists segment distances, headings, ETA, key nodes (tidal windows, VTS report points, emergency alternatives), and risk alerts (depth constraints, traffic hotspots, weather impacts). 4. Technical Stack Recommendations Suggested stacks evolve with project maturity: MVP validation: Claude 3.5 Sonnet / Moonshot Kimi API (≈¥2K‑5K/月) for rapid product logic proof. V1.0 production: Cloud‑hosted LLM + fine‑tuned Qwen2.5‑7B (≈¥1K‑2K/月) for routine routes and real‑time risk detection. Private deployment: Qwen2.5‑72B fine‑tuned on edge hardware for offline ship‑board operation and data security. Key components: Spatial database: PostGIS stores chart features; Redis Geo provides real‑time indexing. Knowledge graph: Neo4j captures navigation rules and element relationships. Path algorithms: A*/RRT* combined with OR‑Tools constraint solving. Multimodal perception: Qwen‑VL identifies chart symbols and weather maps. 5. Industry Rollout (2025) Linghang model (DeepSeek‑R1 fine‑tuned) – port‑to‑port autonomous routing, online testing across scenarios. Yangshan Port "VTS Sprite" – traffic pre‑control, average transit time reduced by 17 minutes. Tianjin Port PortGPT – video‑based port‑operation recognition, potential 24‑hour safety monitoring. COSCO "ShipVision" – 200+ intelligent agents supporting natural‑language vessel queries. 6. Implementation Roadmap for Digital Shipping Leaders Three‑phase plan: Phase 1 (1‑3 months): MVP for single‑vessel static route planning – ingest ENC data, generate textual route description, validate LLM rule comprehension. Phase 2 (3‑6 months): Real‑time monitoring – integrate AIS and tide APIs, produce risk alerts and avoidance suggestions, build knowledge graph of port regulations. Phase 3 (6‑12 months): Multi‑vessel coordination – construct scenario library, generate COLREG‑compliant avoidance maneuvers, deploy edge‑computing on ships. Conclusion AI will not replace captains but will act as a "never‑tiring watch‑assistant", scanning the entire sea in 0.1 seconds, tracking dozens of vessels, and remembering every navigation rule. The four‑layer pyramid bridges human judgment with super‑computing, enabling the next generation of intelligent shipping.