An intelligent ship cooperative decision-making and control method

Abstract

The application discloses an intelligent ship cooperative decision-making and control method, which realizes deep fusion and unified expression of navigation, ship body, engine room, cargo, energy efficiency and integrated platform data through unified world space-time topology structure; adopts a recursive cognitive lattice of DeepSeek Ri inference chain, native sparse attention and multi-head potential attention mechanism to realize cross-domain scene understanding and multi-target strategy generation; relies on a multi-domain simulation acceleration platform to realize underwater environment modeling, ship body motion simulation, engine room system simulation, energy efficiency prediction and cargo management; constructs a double-loop control architecture, the inner loop ensures stable control of equipment, and the outer loop provides intelligent adaptive strategy to ensure the verifiability, explainability and closed-loop execution capability of AI decision-making, and realizes digital twin presentation and simulation. The application solves the technical bottlenecks of the existing intelligent ship system in aspects of multi-domain cooperation, environment perception, health management and energy efficiency optimization.

Description

Technical Field

[0001] This invention relates to a ship decision-making and control method, and more particularly to an intelligent ship collaborative decision-making and control method. Background Technology

[0002] With the release of the *Intelligent Ship Specification (2025)*, intelligent navigation (Nx), intelligent hull (Hx), intelligent engine room (Mx), intelligent energy efficiency (Ex), intelligent cargo (Cx), and intelligent integrated platform (Ix) have become core requirements for ship intelligence. However, existing technologies suffer from the following insurmountable shortcomings: 1. Separation of perception and decision-making: Underwater sonar / radar, cameras, cabin sensors, and energy efficiency data are processed independently, lacking a unified data structure and reasoning framework, making cross-domain collaboration impossible.

[0003] 2. The underwater environment is not visible, and the voxel map lacks semantic structure, resulting in insufficient 3D environment modeling: Traditional sonar / radar can only construct low-dimensional or local obstacle information and cannot express the hierarchical relationship between 3D voxels, which makes path planning prone to local optima or over-conservatism.

[0004] 3. The engine room system is a strongly coupled thermal-fluid-dynamic system: The engine room is a strongly coupled thermal-fluid-dynamic system that is difficult to decouple. Traditional condition-based maintenance (CBM) and residual life prediction (RUL) models are centered on single equipment and cannot form joint decisions with navigation parameters such as speed, propulsion load, and energy efficiency.

[0005] 4. Energy efficiency goals and navigation safety goals are inherently conflicting: There are complex constraints among speed, load, draft, and trim, and existing technologies cannot perform unified modeling for global optimization of multiple objectives.

[0006] 5. Lack of a unified data structure to represent the Nx+Mx+Ex three systems: The existing system uses independent modules (navigation / cabin / energy efficiency) and only exchanges values ​​through interfaces, which cannot perform cross-domain reasoning at the structural level.

[0007] 6. Lack of explainable and verifiable AI control: Existing AI decision-making systems are mostly "black boxes" that are difficult to integrate into traditional control loops and cannot meet the requirements of ship safety regulations and engineering reliability.

[0008] Therefore, there is an urgent need for a unified intelligent reasoning and control method that can simultaneously address issues such as underwater voxels, hull motion, engine room health, cargo management, and energy efficiency constraints. This method would solve core problems existing in current intelligent ship systems, such as fragmented multi-domain data, insufficient 3D environment modeling, strong coupling and difficulty in decoupling the engine room, conflict between energy efficiency and safety objectives, lack of unified data structure and interpretable AI control, and achieve full-domain intelligent collaboration among navigation, hull, engine room, cargo, energy efficiency, and the integrated platform. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide an intelligent ship collaborative decision-making and control method, which can realize underwater obstacle avoidance, ship motion perception, engine room health management, cargo status monitoring, energy efficiency optimization and multi-objective collaborative decision-making.

[0010] To address the aforementioned technical problems, this invention provides an intelligent ship collaborative decision-making and control method, comprising the following steps: S1, Multi-source perception data acquisition: collecting navigation environment data and simultaneously collecting engine room equipment data; S2, Unified spatiotemporal data structure construction: adopting a unified world spatiotemporal topology structure, integrating underwater 3D voxels, hull six-degree-of-freedom motion attitude, engine room thermofluid state, cargo loading information, propulsion energy efficiency indicators, and maritime traffic entity data into a hierarchical tree structure, achieving spatiotemporal alignment and semantic association of multi-source heterogeneous data; S3, Multi-domain physical simulation calculation: based on a hardware acceleration platform, parallel execution of sonar signal processing and 3D voxel updating, ship six-degree-of-freedom motion simulation, engine room thermo-fluid-dynamic coupling simulation, propulsion system energy consumption simulation, and cargo stability simulation, constructing a real-time digital twin environment; S4, Deep expert model inference: using DeepSeek MoE... Ri architecture, i=1,2……n, constructs multi-level scene representations through RCL recursive cognitive lattice, filters key features using native sparse attention mechanism, and generates cross-domain strategies through multi-head potential attention experts collaboratively; S5, multi-objective collaborative decision-making: based on digital twin simulation results, comprehensively considers obstacle avoidance safety, energy efficiency optimization, cabin health, cargo management, and regulatory compliance constraints to generate globally optimal navigation strategies and control commands; S6, dual-loop control execution and verification: the inner loop uses traditional control algorithms to execute underlying equipment control, while the outer loop provides adaptive strategy adjustment based on deep expert models. The execution results are fed back to the digital twin environment in real time, forming a complete closed loop of perception-decision-execution-verification.

[0011] Furthermore, the navigation environment data in step S1 includes forward-looking sonar, multibeam echo sounder, side-scan sonar, IMU inertial measurement unit, AIS automatic identification system, radar, gyrocompass, optical camera, satellite positioning system and nautical chart data, and the cabin equipment data includes temperature, pressure, flow rate, vibration, torsional vibration, pump and valve status and electric propulsion parameters.

[0012] Furthermore, the unified world spatiotemporal topology in step S2 adopts a three-layer organization of voxels, fragments, and tiles, specifically: voxel: the smallest spatial unit, with a size configurable from 0.5m×0.5m×0.5m to 2m×2m×2m; fragment: composed of 8×8×8 voxels, serving as the basic unit for parallel computing and storage; tile: composed of 4×4×4 fragments, serving as the spatial carrying unit for spatiotemporal tree nodes; each spatiotemporal tree node contains a timestamp, spatial range, data pointer, and cross-domain association index.

[0013] Furthermore, the RCL recursive cognitive lattice in step S4 includes the following five core layers: voyage layer: defines the entire voyage mission, including the route, timetable, and loading plan; segment layer: divides the voyage into logical segments, each corresponding to a relatively stable environment and operating conditions; scenario layer: describes a specific spatiotemporal region, including environmental characteristics, traffic conditions, and risk distribution; control layer: corresponds to the decision control cycle and generates specific control command sequences; execution layer: provides feedback on the actual execution status and real-time information of the equipment.

[0014] Furthermore, the native sparse attention mechanism in step S4 adopts the following five screening criteria: spatial proximity: distance threshold to the ship's current position or planned path; risk significance: risk score calculated based on occupancy probability, movement speed, and collision time; energy efficiency sensitivity: environmental factors that have a significant impact on fuel consumption; health criticality: monitoring parameters that are strongly correlated with engine room equipment failure prediction; the screening threshold is dynamically adjusted according to the navigation stage, environmental complexity, and system load.

[0015] Furthermore, the multi-head potential attention in step S4 includes six specialized expert modules: an underwater obstacle identification module (processing sonar data to identify reefs, shipwrecks, and fishing net obstacles); a hull motion monitoring module (analyzing the ship's six degrees of freedom motion to assess stability, strength, and comfort); an engine room health diagnosis module (integrating multi-source monitoring data to achieve fault prediction and health management); a cargo status management module (monitoring cargo loading, temperature, humidity, and displacement); a propulsion energy efficiency assessment module (analyzing the energy consumption characteristics of the main engine, propellers, and generators); and a navigation strategy generation module (integrating expert opinions to generate multi-objective optimization strategies). When multiple specialized expert modules issue warnings simultaneously, the RCL recursive cognitive lattice in step S4 performs cross-layer causal analysis to generate a comprehensive diagnostic report including fault location, severity, remaining life, and maintenance recommendations.

[0016] Furthermore, the multi-objective collaborative decision-making in step S5 adopts a three-layer progressive optimization architecture: upper layer: based on the model predictive control framework, a feasible strategy set is generated with a prediction time domain of 10-30 seconds; middle layer: a multi-objective evolutionary algorithm is used to select the Pareto optimal solution from the feasible strategy set; lower layer: according to the current navigation scenario, the optimal execution strategy is selected from the Pareto solution set.

[0017] Furthermore, in step S6, the inner loop of the dual-loop control structure employs a PID controller, model predictive controller, or model reference adaptive controller to ensure the stable operation of underlying equipment such as rudder angle adjustment, speed control, propeller power output, and pump / valve switching. The outer loop generates a global adaptive strategy from the DeepSeek Ri inference chain, including high-level decisions such as dynamic load limiting, intelligent speed adjustment, optimal energy efficiency configuration, hull trim optimization, and dynamic cargo loading adjustment. Specific operating modes and safety protection mechanisms are as follows: Strategy verification mechanism: All control strategies generated by the outer loop are verified through forward-looking simulation in a digital twin environment to ensure compliance with safe navigation requirements; Physical boundary constraints: Equipment safety operating boundaries are set, including maximum rudder angle, minimum draft, maximum main engine speed, and engine room extreme temperature and pressure; Performance monitoring and degradation: Real-time monitoring of the inner loop control performance; Automatic degradation of control mode when tracking error exceeds a threshold; Manual intervention interface: Retaining the highest control priority of the captain and chief engineer, providing physical buttons and software... Dual intervention channels in the interface; normal operation mode: the outer loop DeepSeek provides a strategy update every 2 seconds, and the inner loop controller tracks and executes at a frequency of 100Hz; sensor fault detection mechanism: continuously monitors the confidence level of each sensor data, and when a critical sensor fails: switches to redundant sensors or estimates, reduces the autopilot level, and alarms the shore-based control center; control performance degradation handling mechanism: when the system execution error continues to exceed the preset threshold, the outer loop adjusts the control gain, limits the range of strategy changes, and switches to manual control when necessary; catastrophic failure emergency response: when the system detects the risk of ship loss of control, it automatically executes emergency braking, activates anti-collision maneuvers, and broadcasts an emergency status to surrounding ships.

[0018] Further, step S1 involves data acquisition and preprocessing as follows: the forward-looking sonar outputs multi-beam data at a frequency of 20Hz, with each beam containing 128 sampling points. Matched filtering, beamforming, and range-intensity imaging are processed in parallel using a hardware accelerator. Step S2 constructs a forward-looking sonar-driven 3D underwater digital twin as follows: Spatiotemporal registration: all sonar echo data is precisely registered with IMU, GPS, and ship attitude sensor data in the PETA-X6 data analysis system according to timestamps, and multi-source pose information is fused using extended Kalman filtering; Voxelization: a Bayesian-based dynamic occupancy grid algorithm is used to convert the sonar point cloud into regularized 3D voxels with a resolution of 1m×1m×1m; each voxel contains occupancy probability, material attribute label, confidence factor, and update timestamp; Map update and management: a 200m×200m×50m dynamic update area is constructed centered on the ship's current position, using a voxel-fragment-tile hierarchical storage structure to achieve an efficient balance between local updates and global indexing.

[0019] Furthermore, step S5 implements energy efficiency optimization decision-making under multiple constraints as follows: Objective function construction: Establish an optimization model with minimizing the energy efficiency operating index as the core objective: minJ = ω1 Fuel consumption +ω2 Time cost + ω3 Safety risks; the weighting coefficients ω1, ω2, and ω3 are dynamically adjusted according to the voyage mission; constraint modeling, the main constraints include: (i) safety constraints: the height of the ship above the bottom ≥ the safety margin, and an additional 20% margin in shallow water; (ii) engine room constraints: the long-term operating load rate of the main engine ≤ 85%, and the instantaneous load rate ≤ 95%; (iii) hull constraints: the longitudinal angle ≤ 3°, and the transverse angle ≤ 5°; (iv) cargo constraints: the ballast difference between each compartment ≤ the allowable value, and the stability height ≥ the standard requirement value; (v) regulatory constraints: comply with international collision avoidance rules and lane separation system; multi-objective optimization solution: the NSGA-II multi-objective genetic algorithm is used to generate the Pareto front solution set in parallel on the 3D-EPU multi-domain simulation platform; strategy selection and verification: the DeepSeek outer loop intelligently selects the optimal strategy from the Pareto solution set according to the current navigation scenario, and performs forward simulation verification in the digital twin environment.

[0020] Compared with existing technologies, the present invention has the following advantages: The intelligent ship collaborative decision-making and control method provided by the present invention, through an original unified spatiotemporal topology structure (UNITREE), performs high-precision spatiotemporal alignment and dynamic coupling of navigation (Nx), hull (Hx), engine room (Mx), energy efficiency (Ex), cargo (Cx), and intelligent integration platform (Ix) data; with the help of a multi-domain simulation platform, closed-loop verification under complex sea conditions can be completed in a digital twin environment; the DeepSeek Ri (i=1, 2...n) inference chain, which integrates recursive cognitive lattice (RCL), native sparse attention (NSA), and multi-head latent attention mechanism (MLA), realizes interpretable deduction of the entire process from environmental perception to decision output; the engineering cybernetics dual-loop structure design ensures precise coordination between high-level decision-making strategies and low-level equipment execution, significantly improving the robustness and response speed of the system in complex environments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the intelligent ship collaborative decision-making and control system architecture of the present invention; Figure 2 This is a schematic diagram of the unified world spatiotemporal topology (UNITREE) and voxel-fragment-tile hierarchy of the present invention; Figure 3 This is a diagram illustrating the working mechanism of the DeepSeek Ri (i=1, 2, ..., n) inference chain of this invention; Figure 4 This is a schematic diagram of the engineering cybernetics dual-loop execution structure of the present invention; Figure 5 This is a flowchart of the three-dimensional voxel mesh and A* / D* Lite path search algorithm of the present invention. Detailed Implementation

[0022] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0023] Please see Figure 1 and Figure 4 The intelligent ship collaborative decision-making and control method provided by the present invention includes the following steps: S1) Acquisition of multi-source sensing data: Collect navigation environment data (including forward-looking sonar, multibeam echo sounder, side-scan sonar, IMU inertial measurement unit, AIS automatic identification system, radar, gyrocompass, optical camera, satellite positioning system and nautical chart data), and simultaneously collect cabin equipment data (including temperature, pressure, flow rate, vibration, torsional vibration, pump and valve status, electric propulsion parameters); S2) Construction of unified spatiotemporal data structure: The unified world spatiotemporal topology (UNITREE) is adopted to integrate underwater three-dimensional voxels, hull six-degree-of-freedom motion attitude, engine room thermofluid status, cargo loading information, propulsion energy efficiency indicators and maritime traffic entity data into the same hierarchical tree structure, so as to realize spatiotemporal alignment and semantic association of multi-source heterogeneous data. S3) Multi-domain physical simulation computation: Based on the 3D-EPU / ESU (3D stacked electron transfer computing architecture processing unit chip / processing unit) hardware acceleration platform, sonar signal processing and three-dimensional voxel updating, ship six-degree-of-freedom motion simulation, engine room thermal-fluid-dynamic coupling simulation, propulsion system energy consumption simulation and cargo stability simulation are executed in parallel to build a real-time digital twin environment; the above multi-domain computations are executed in parallel on the same hardware platform and maintain strict timing synchronization, effectively solving the bottleneck of inconsistent modes in traditional systems and achieving millisecond-level real-time response of multi-domain simulation; S4) Deep Expert Model Inference: Adopting the DeepSeek MoE Ri (i=1,2……n, i.e., multi-expert model) architecture, it constructs a multi-level scene representation through a recursive cognitive lattice (RCL), uses native sparse attention (NSA) mechanism to filter key features, and generates cross-domain policies through multi-head latent attention (MLA) expert collaboration; RCL represents the ship and its environment as a recursive cognitive lattice (voyage → segment → scene → control cycle); NSA performs native sparse attention on massive sensor / log / radar images, activating only the subset relevant to the current task; MLA combines different "potential heads" corresponding to different control modes (normal economic navigation / wind and wave conditions / ice zone / equipment failure / wartime special situations, etc.) with MoE expert routing; S5) Multi-objective collaborative decision-making: Based on the results of digital twin simulation, the system comprehensively considers obstacle avoidance safety, energy efficiency optimization, cabin health, cargo management and regulatory compliance constraints to generate the globally optimal navigation strategy and control commands. S6) Dual-loop control execution and verification: The inner loop uses traditional control algorithms to execute the control of the underlying equipment, while the outer loop provides adaptive strategy adjustments based on a deep expert model. The execution results are fed back to the digital twin environment in real time, forming a complete closed loop of perception-decision-execution-verification. This uses DeepSeek as a high-level adaptive decision-maker and the traditional controller as a low-level execution controller, unifying the abstraction of "controlled object + controller + feedback + disturbance," which facilitates the implementation of ultra-steady systems, hierarchical control, generalized objects, adaptive and learning control, etc. Please continue reading Figure 2 The intelligent ship collaborative decision-making and control method provided by this invention, wherein the unified world spatiotemporal topology (UNITREE) in step S2) adopts a three-layer organization of voxels-fragments-tiles, specifically as follows: (1) Voxel: The smallest spatial unit, whose size can be configured from 0.5m×0.5m×0.5m to 2m×2m×2m; (2) Fragment: Composed of 8×8×8 voxels, serving as the basic unit for parallel computing and storage; (3) Tile: Composed of 4×4×4 fragments, serving as the spatial carrying unit of the spatiotemporal tree node; (4) Each spatiotemporal tree node contains a timestamp, spatial range, data pointer, and cross-domain association index.

[0024] The intelligent ship collaborative decision-making and control method provided by this invention, in step S4), constructs a multi-level recursive structure of "ship segment → water slice → occupied voxel → engine room equipment → sensor", which solves the problem of structural alignment of data in different domains and realizes full-scale cognition from macro to micro; specifically, it includes the following five core levels: (1) Voyage layer: Defines the entire voyage mission, including route, timetable and loading plan; (2) Segment layer: The flight is divided into logical segments, each segment corresponding to a relatively stable environment and operating conditions; (3) Scene layer: describes a specific spatiotemporal region, including environmental features, traffic conditions and risk distribution; (4) Control layer: corresponding to the decision control cycle, it generates a specific sequence of control instructions; (5) Execution layer: Feedback on the actual execution status and real-time information of the device.

[0025] The intelligent ship collaborative decision-making and control method provided by this invention uses a native sparse attention (NSA) mechanism in step S4) to intelligently filter key data (such as hazardous voxels, sensitive points in the engine room, and energy efficiency gradient factors), reducing the computational complexity from the traditional O(N³) to O(N log N), thus significantly improving real-time processing capabilities. Specifically, the following five filtering criteria are adopted: (1) Spatial proximity: Distance threshold to the ship's current position or planned route; (2) Risk significance: Risk score calculated based on occupancy probability, movement speed, and time to collision (TTC); (3) Energy efficiency sensitivity: Environmental factors that have a significant impact on fuel consumption; (4) Health criticality: Monitoring parameters that are strongly correlated with the prediction of cabin equipment failures; (5) The screening threshold is dynamically adjusted based on the navigation stage, environmental complexity, and system load.

[0026] The intelligent ship collaborative decision-making and control method provided by this invention includes a multi-head potential attention (MLA) step (S4) comprising six specialized expert modules: (1) Underwater obstacle identification expert: Processing sonar data to identify underwater obstacles such as reefs, shipwrecks, and fishing nets; (2) Ship motion monitoring expert: Analyze the six degrees of freedom motion of the ship and evaluate its stability, strength and comfort; (3) Cabin health diagnostic expert: Integrates multi-source monitoring data to achieve fault prediction and health management; (4) Cargo Status Management Expert: Monitors cargo loading, temperature, humidity, displacement and other statuses; (5) Promote energy efficiency assessment experts: analyze the energy consumption characteristics of the main unit, propulsion unit, and generator; (6) Navigation strategy generation expert: Combine the opinions of various experts to generate a multi-objective optimization strategy.

[0027] The intelligent ship collaborative decision-making and control method provided by this invention adopts a three-layer progressive optimization architecture for multi-objective collaborative decision-making in step S5): (1) Upper layer: Based on the model predictive control (MPC) framework, a set of feasible strategies is generated with a prediction time domain of 10-30 seconds; (2) Middle layer: A multi-objective evolutionary algorithm is used to select the Pareto optimal solution from the set of feasible strategies; (3) Lower layer: Select the optimal execution strategy from the Pareto solution set based on the current navigation scenario.

[0028] Please refer to the above. Figure 3 and Figure 4 The intelligent ship collaborative decision-making and control method provided by this invention, in step S6), the inner loop adopts mature traditional controllers such as PID, MPC (Model Predictive Control), and MRAC (Model Reference Adaptive Control) to ensure the stable operation of underlying equipment such as rudder angle adjustment, speed control, propeller power output, and pump valve switching; the outer loop generates a global adaptive strategy from the DeepSeek Ri (i=1, 2...n) inference chain, including high-level decisions such as dynamic load limiting, intelligent speed adjustment, optimal energy efficiency configuration, hull trim optimization, and dynamic cargo loading adjustment; this dual-loop control structure achieves for the first time the verifiable and closed-loop execution of AI decision-making strategies within a preset safety boundary, significantly improving the system's operational reliability and decision interpretability; it has the following safety protection mechanisms: (1) Strategy verification mechanism: All control strategies generated in the outer loop must be verified by forward simulation in a digital twin environment to ensure that they meet the requirements for safe navigation; (2) Physical boundary constraints: Set the safe operating boundaries of the equipment, including the maximum rudder angle (±35°), minimum draft (according to the ship load line specification), maximum main engine speed (105% of the rated speed), and engine room limit temperature and pressure (according to the equipment manual). (3) Performance monitoring and degradation: Real-time monitoring of inner loop control performance. When the tracking error exceeds the threshold (e.g., track deviation > 5m), the control mode is automatically degraded (e.g., switching from fully automatic to assisted driving). (4) Manual intervention interface: retain the highest control priority of operators such as the captain and chief engineer, and provide dual intervention channels of physical buttons and software interface.

[0029] The intelligent ship collaborative decision-making and control method provided by this invention supports the following intelligent functions defined in the "Intelligent Ship Specification (2025)": (1) Nx Intelligent Navigation: Enables autonomous obstacle avoidance, dynamic route optimization and precise speed control; (2) Hx intelligent hull: Real-time monitoring of structural stress, trim, corrosion status and vibration noise level; (3) Mx Intelligent Cabin: Implement equipment status monitoring, early fault prediction and condition-based maintenance decision-making; (4) Ex Smart Energy Efficiency: Optimizes fuel consumption, reduces carbon emissions and improves operational efficiency; (5) Cx Smart Cargo: Monitor cargo status, optimize loading schemes and ensure transportation safety; (6) Ix Intelligent Platform: Integrates various subsystems and provides a unified human-computer interaction and remote control interface.

[0030] The intelligent ship collaborative decision-making and control method provided by this invention also includes fault injection and emergency drill functions, which can simulate the following abnormal scenarios in a digital twin environment: (1) Sensor fault simulation: Simulate scenarios such as GPS signal failure, sonar interference, and communication link interruption; (2) Equipment failure simulation: Simulate failures such as main engine shutdown, servo motor jamming, and power system interruption; (3) Environmental mutation simulation: Simulate severe sea conditions (such as Force 9 wind waves), sudden drop in visibility (<1 nautical mile), and sudden underwater obstacles; (4) Human error simulation: Simulates situations such as incorrect operation instructions, conflicting instructions from multiple devices, and delayed emergency response; The above simulation training enhances the system's robustness and the crew's emergency response capabilities.

[0031] The intelligent ship collaborative decision-making and control method provided by this invention has the following advantages: 1. Achieving unified reasoning for six systems Nx+Hx+Mx+Ex+Cx+Ix for the first time: By leveraging UNITREE spatiotemporal topology and RCL recursive cognitive lattice, underwater obstacle risks, cabin load limitations, cargo status monitoring, and energy efficiency optimization objectives are integrated into the same reasoning framework for the first time to achieve collaborative optimization.

[0032] 2. Voxel map complexity reduced from O(N³) to O(N log N): By using the NSA’s native sparse attention mechanism, only voxels in dangerous areas are calculated, resulting in an exponential improvement in path planning and simulation efficiency. This is a breakthrough that traditional voxel navigation technology cannot achieve.

[0033] 3. A unified solution for multi-objective optimization of energy efficiency and safety is achieved through collaborative optimization: The planning in this invention not only enables precise obstacle avoidance and path planning, but also simultaneously generates and outputs multi-dimensional cross-domain strategies such as optimal speed control, dynamic power allocation, intelligent trim optimization, refined cargo management, pump and valve coordinated scheduling, and propulsion system efficiency improvement, ensuring that the dual objectives of navigation safety and energy efficiency improvement are achieved simultaneously. This cross-domain collaborative decision-making capability is difficult to achieve with traditional methods.

[0034] 4. Verifiable AI outer-loop control improves reliability: The dual-loop control structure ensures that the AI ​​strategy generated by DeepSeek Ri (i=1, 2...n) is strictly executed within physical constraints such as rudder angle range, thruster load limit, nacelle temperature / pressure threshold, draft and UKC safety boundary, avoiding the risk of exceeding the limits and realizing "transparent and explainable AI decision-making".

[0035] 5. Millisecond-level global inference and simulation synchronization: Relying on the 3D-EPU / ESU multi-domain simulation accelerator, real-time synchronization of multi-domain simulation and DeepSeek inference is achieved, realizing millisecond-level response of global dynamic optimization, and solving the technical problem that traditional systems cannot synchronously process 6DOF motion, three-dimensional voxel and thermal fluid simulation.

[0036] Example 1: Construction of a 3D Underwater Digital Twin Driven by Forward-Looking Sonar 1. Data Acquisition and Preprocessing: The forward-looking sonar outputs multi-beam data at a frequency of 20Hz, with each beam containing 128 sampling points (depending on equipment performance). Matched filtering, beamforming, and range-intensity imaging are processed in parallel using a 3D-EPU hardware accelerator. Figure 3 and Figure 4 As shown.

[0037] 2. Spatiotemporal Registration: All sonar echo data are precisely registered with IMU (100Hz), GPS (10Hz), and ship attitude sensor (50Hz) data in the PETA-X6 (Pipeline Electron Transfer Architecture in 6G or THz) data analysis system according to timestamps. Extended Kalman filtering is used to fuse multi-source pose information, such as... Figure 3 and Figure 4 As shown.

[0038] 3. Voxelization: A Bayesian-based dynamic occupancy mesh algorithm is used to convert the sonar point cloud into regularized 3D voxels with a resolution of 1m×1m×1m. Each voxel contains four core parameters: occupancy probability (0-1), material attribute label, confidence factor, and update timestamp, such as... Figure 5 As shown.

[0039] 4. Map Update and Management: Centered on the ship's current location, a dynamic update area of ​​200m×200m×50m is constructed. A hierarchical storage structure of "voxels → fragments (8×8×8 voxels) → tiles (4×4×4 fragments)" is adopted to achieve an efficient balance between local updates and global indexing. Figure 5 As shown.

[0040] Example 2: Implementation of cabin health diagnosis using DeepSeek Ri (i=1, 2, ..., n) inference chain (see...) Figure 3 , Figure 4 ) 1. RCL Structure Construction: A five-layer recursive cognitive structure is established for the cabin system, with the following hierarchical relationship: (1) Classification: Overall ship health status (2) System level: power system, electrical system, auxiliary system (3) Equipment level: main unit, auxiliary unit, pump set, valve, etc. (4) Component level: cylinder liners, bearings, seals, etc. (5) Sensor level: temperature, pressure, vibration, flow rate and other measurement points 2. NSA Feature Screening: For host vibration signals, NSA automatically identifies three key features: (1) Key frequency band: 0-200Hz (fundamental frequency and low harmonics) (2) Fault-sensitive frequency band: 2kHz-5kHz (early bearing fault characteristics) (3) Abnormal time window: Abnormal amplitude lasting more than 3 seconds 3. MLA Expert Collaborative Diagnosis: Diagnosis is completed through the collaboration of four specialized expert heads. (1) Combustion expert: Analyzes cylinder pressure curves, exhaust temperature, and fuel consumption rate. (2) Lubricating oil expert: Monitors oil viscosity, acid value, and metal particle content. (3) Cooling experts: Evaluate cooling efficiency, temperature gradient, and flow stability. (4) Vibration expert: Identifies shaft misalignment, bearing wear, and blade damage. Joint reasoning output: When multiple experts issue warnings simultaneously, RCL performs cross-layer causal analysis to generate a comprehensive diagnostic report that includes fault location, severity, remaining life, and maintenance recommendations.

[0041] Example 3: Energy efficiency optimization decision under multiple constraints (see Figure 3 , Figure 4 ) 1. Objective Function Construction: Establish an optimization model with minimizing EEOI (Energy Efficiency Operating Index) as the core objective. minJ =ω1 Fuel consumption +ω2 Time cost + ω3 Security risks The weighting coefficients ω1, ω2, and ω3 are dynamically adjusted according to the voyage mission.

[0042] 2. Constraint Modeling: The main constraints include: (1) Safety constraints: The height of the vessel above the bottom (UKC) ≥ the safety margin (an additional 20% margin is added in shallow water). (2) Cabin constraints: Main engine load rate ≤ 85% (long-term operation), instantaneous load rate ≤ 95%. (3) Hull constraints: longitudinal angle ≤ 3°, transverse angle ≤ 5° (4) Cargo constraints: Ballast difference between compartments ≤ permissible value, stability height ≥ standard requirement value. (5) Regulatory constraints: Strictly comply with international collision avoidance regulations (COLREGs) and lane separation systems. 3. Multi-objective optimization solution: The NSGA-II (Nondominated Sorting Genetic Algorithm II) multi-objective genetic algorithm is used to generate the Pareto front solution set in parallel on the 3D-EPU multi-domain simulation platform.

[0043] 4. Strategy Selection and Verification: The DeepSeek outer loop intelligently selects the optimal strategy from the Pareto solution set based on the current navigation scenario and performs forward simulation verification in a digital twin environment to ensure the strategy is safe and feasible.

[0044] Example 4: Adaptive path replanning in dynamic environments. Please refer to [link to example]. Figure 5 .

[0045] 1. Initial path generation: An improved A* algorithm is used, and the heuristic function incorporates the following three key factors: h(n) = α d(n,goal)+β Risk(n)+γ EnergyCost(n) Where α, β, and γ are adaptive weight parameters, n represents the current node or current position / state, and goal represents the goal node or goal position / state.

[0046] 2. Dynamic Obstacle Detection: The sonar system updates the voxel map every 5 seconds. When a new obstacle is detected, the following steps are performed: (1) Calculate the obstacle's movement trend (speed, direction) (2) Predict possible location within the next 10 seconds (3) Update the risk weights of affected voxels 3. Incremental path optimization: The D* Lite algorithm is used for local path replanning, specifically including: (1) Only update search nodes affected by obstacles. (2) Keep the unaffected path segments unchanged. (3) Calculate the minimum cost of incremental modification. 4. Smoothing and control interface conversion: (1) Smooth the discrete path points using a cubic B-spline curve. (2) Calculate the continuous curvature change to ensure the servo motor can perform the operation. (3) Generate time-parameterized trajectories: position sequence $p(t)$, velocity sequence $v(t)$, and heading sequence $\psi(t)$ (4) Convert to inner loop controller commands: rudder angle $\delta(t)$, main engine speed $n(t)$, thrust distribution $T(t)$.

[0047] Example 5: Fail-safe modes of a dual-loop control system, see [link / reference] Figure 4 1. Normal operating mode: The outer loop DeepSeek provides a policy update every 2 seconds, and the inner loop controller tracks and executes at a frequency of 100Hz.

[0048] 2. Sensor Fault Detection Mechanism: The system continuously monitors the confidence level of data from each sensor. When a critical sensor (such as IMU or GPS) fails: (1) Switch to redundant sensors or estimates (2) Reduce the level of autonomous driving (3) Alarm to shore-based control center 3. Control performance degradation handling mechanism: When the system execution error continues to exceed the preset threshold: (1) Adjusting the control gain in the outer loop (2) Limit the range of policy changes (3) Switch to manual control when necessary. 4. Catastrophic Failure Emergency Response: When the system detects a risk of loss of ship control: (1) Automatically perform emergency braking (2) Activate collision avoidance maneuver (3) Broadcast an emergency to surrounding vessels.

[0049] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be defined by the claims.

Claims

1. A method for collaborative decision-making and control of intelligent ships, characterized in that, Includes the following steps: S1. Multi-source sensing data acquisition: Collect flight environment data and simultaneously collect cabin equipment data; S2. Construction of a unified spatiotemporal data structure: A unified world spatiotemporal topology is adopted to integrate underwater three-dimensional voxels, hull six-degree-of-freedom motion attitude, engine room thermofluid status, cargo loading information, propulsion energy efficiency indicators, and maritime traffic entity data into the same hierarchical tree structure, so as to realize the spatiotemporal alignment and semantic association of multi-source heterogeneous data. S3, Multi-domain physical simulation calculation: Based on a hardware acceleration platform, it performs sonar signal processing and three-dimensional voxel updating, ship six-degree-of-freedom motion simulation, engine room thermal-fluid-dynamic coupling simulation, propulsion system energy consumption simulation and cargo stability simulation in parallel, and builds a real-time digital twin environment; S4. Deep Expert Model Inference: The DeepSeek MoE Ri architecture is adopted, i=1,2……n. Multi-level scene representation is constructed through RCL recursive cognitive lattice, key features are screened using native sparse attention mechanism, and cross-domain strategies are generated through multi-head potential attention experts in collaboration. S5. Multi-objective collaborative decision-making: Based on the results of digital twin simulation, it comprehensively considers obstacle avoidance safety, energy efficiency optimization, cabin health, cargo management and regulatory compliance constraints to generate the globally optimal navigation strategy and control commands. S6. Dual-loop control execution and verification: The inner loop uses traditional control algorithms to execute the control of the underlying equipment, while the outer loop provides adaptive strategy adjustment based on a deep expert model. The execution results are fed back to the digital twin environment in real time, forming a complete closed loop of perception-decision-execution-verification.

2. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, The navigation environment data in step S1 includes forward-looking sonar, multibeam echo sounder, side-scan sonar, IMU inertial measurement unit, AIS automatic identification system, radar, gyrocompass, optical camera, satellite positioning system and nautical chart data. The cabin equipment data includes temperature, pressure, flow rate, vibration, torsional vibration, pump and valve status and electric propulsion parameters.

3. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, The unified world spacetime topology in step S2 adopts a three-layer organization of voxels, fragments, and tiles, specifically as follows: Voxel: The smallest spatial unit, with dimensions configurable from 0.5m×0.5m×0.5m to 2m×2m×2m; Fragments: Composed of 8×8×8 voxels, serving as the basic unit for parallel computing and storage; Tile: Composed of 4×4×4 fragments, serving as the spatial carrier unit of the spacetime tree node; Each spatiotemporal tree node contains a timestamp, spatial range, data pointer, and cross-domain association index.

4. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, The RCL recursive cognitive lattice in step S4 comprises the following five core levels: Voyage layer: Defines the entire voyage mission, including the route, timetable, and loading plan; Segment layer: Divides the flight path into logical segments, each segment corresponding to a relatively stable environment and operating conditions; Scene layer: describes a specific spatiotemporal region, including environmental characteristics, traffic conditions, and risk distribution; Control layer: corresponding to the decision-making and control cycle, it generates specific control command sequences; Execution layer: Provides feedback on the actual execution status and real-time information of the device.

5. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, The native sparse attention mechanism in step S4 adopts the following five selection criteria: Spatial proximity: Distance threshold to the ship's current position or planned route; Risk significance: Risk score calculated based on occupancy probability, movement speed, and collision time; Energy efficiency sensitivity: Environmental factors that have a significant impact on fuel consumption; Health criticality: Monitoring parameters strongly correlated with cabin equipment failure prediction; The screening threshold is dynamically adjusted based on the navigation stage, environmental complexity, and system load.

6. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, The multi-head potential attention in step S4 includes six specialized expert modules: Underwater obstacle recognition module: processes sonar data to identify obstacles such as reefs, shipwrecks, and fishing nets; Hull motion monitoring module: Analyzes the six degrees of freedom motion of the ship to assess stability, strength and comfort; Cabin health diagnostic module: Integrates multi-source monitoring data to achieve fault prediction and health management; Cargo status management module: Monitors cargo loading, temperature, humidity, and displacement status; The propulsion energy efficiency assessment module analyzes the energy consumption characteristics of the main unit, propeller, and generator. Navigation strategy generation module: Integrates expert opinions to generate multi-objective optimization strategies; When multiple specialized expert modules issue warnings simultaneously, the RCL recursive cognitive lattice in step S4 performs cross-layer causal analysis to generate a comprehensive diagnostic report that includes fault location, severity, remaining lifespan, and maintenance recommendations.

7. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, The multi-objective collaborative decision-making in step S5 adopts a three-layer progressive optimization architecture: Upper layer: Based on the model prediction control framework, a set of feasible strategies is generated with a prediction time domain of 10-30 seconds; Middle layer: Employs a multi-objective evolutionary algorithm to select the Pareto optimal solution from the set of feasible strategies; Lower layer: Select the optimal execution strategy from the Pareto solution set based on the current navigation scenario.

8. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, The inner loop of the dual-loop control structure in step S6 employs a PID controller, model predictive controller, or model reference adaptive controller to ensure the stable operation of underlying equipment such as rudder angle adjustment, speed control, propeller power output, and pump / valve switching. The outer loop generates a global adaptive strategy using the DeepSeek Ri inference chain, including high-level decisions such as dynamic load limiting, intelligent speed adjustment, optimal energy efficiency configuration, hull trim optimization, and dynamic cargo loading adjustment. The specific operating modes and safety protection mechanisms are as follows: Strategy verification mechanism: All control strategies generated in the outer loop are verified by forward simulation in a digital twin environment to ensure that they meet the requirements for safe navigation; Physical boundary constraints: Set the safe operating boundaries of the equipment, including maximum rudder angle, minimum draft, maximum main engine speed, and extreme temperature and pressure of the engine room; Performance monitoring and degradation: Real-time monitoring of inner loop control performance; automatic degradation of control mode when tracking error exceeds the threshold. Manual intervention interface: Retains the highest control priority of the captain and chief engineer, and provides dual intervention channels through physical buttons and software interface; Normal operating mode: The outer loop DeepSeek provides a policy update every 2 seconds, and the inner loop controller tracks and executes at a frequency of 100Hz; Sensor fault detection mechanism: Continuously monitor the confidence level of each sensor data. When a critical sensor fails: switch to a redundant sensor or estimated value, reduce the autonomous driving level, and alarm the shore-based control center. Control performance degradation handling mechanism: When the system execution error continues to exceed the preset threshold, the outer loop adjusts the control gain, limits the magnitude of strategy changes, and switches to manual control when necessary; Catastrophic failure emergency response: When the system detects a risk of loss of control of the vessel, it automatically performs emergency braking, activates anti-collision maneuvers, and broadcasts an emergency status to surrounding vessels.

9. The intelligent ship collaborative decision-making and control method as described in claim 3, characterized in that, The step S1 involves data acquisition and preprocessing as follows: the forward-looking sonar outputs multi-beam data at a frequency of 20Hz, with each beam containing 128 sampling points, and the data is processed in parallel by a hardware accelerator for matched filtering, beamforming, and range-intensity imaging. Step S2 constructs a forward-looking sonar-driven three-dimensional underwater digital twin as follows: Spatiotemporal registration: All sonar echo data are precisely registered with IMU, GPS, and ship attitude sensor data in the PETA-X6 data analysis system according to timestamps, and multi-source pose information is fused using extended Kalman filtering; Voxelization: A dynamic occupancy grid algorithm based on Bayesian update is used to convert the sonar point cloud into regularized 3D voxels with a resolution of 1m×1m×1m; each voxel contains occupancy probability, material attribute label, confidence factor and update timestamp; Map update and management: A dynamic update area of ​​200m×200m×50m is constructed with the ship's current location as the center. A hierarchical storage structure of voxels-fragments-tiles is adopted to achieve an efficient balance between local updates and global indexing.

10. The intelligent ship collaborative decision-making and control method as described in claim 1, characterized in that, Step S5 achieves energy efficiency optimization decision-making under multiple constraints in the following manner: Objective function construction: Establish an optimization model with minimizing the energy efficiency operating index as the core objective: minJ =ω1 Fuel consumption +ω2 Time cost + ω3 Safety risks; the weighting coefficients ω1, ω2, and ω3 are dynamically adjusted according to the voyage mission. Constraint modeling, the main constraints include: (i) Safety constraints: The height of the vessel above the bottom must be greater than or equal to the safety margin, with an additional 20% margin in shallow waters; (ii) Cabin constraints: The main engine's long-term operating load rate is ≤85%, and the instantaneous load rate is ≤95%; (iii) Hull constraints: longitudinal angle ≤ 3°, transverse angle ≤ 5°; (iv) Cargo constraints: Ballast difference between compartments ≤ permissible value, stability height ≥ standard requirement value; (v) Regulatory constraints: Comply with international collision avoidance rules and lane separation schemes; Multi-objective optimization solution: The NSGA-II multi-objective genetic algorithm is used to generate the Pareto front solution set in parallel on the 3D-EPU multi-domain simulation platform; Strategy Selection and Verification: The DeepSeek outer loop intelligently selects the optimal strategy from the Pareto solution set based on the current navigation scenario and performs forward-looking simulation verification in a digital twin environment.

Images