An intelligent ship sea area path planning optimization method

Abstract

The application relates to an intelligent ship sea area path planning optimization method, which comprises the following steps: S1, multi-source heterogeneous data acquisition and space-time fusion, S2, multi-field coupled environment digital twin construction and dynamic correction, S3, DT-DRL double-engine architecture construction and model training, S4, multi-target composite reward path decision output, S5, adaptive rolling re-planning and path smoothing, and S6, optional multi-ship cooperation expansion. The intelligent ship sea area path planning optimization method breaks through the traditional static, single and isolated environment modeling limitation through sea wave-current-tide-sea wind four-field dynamics coupling modeling and the introduction of wave-current interaction correction, significantly improves the ocean environment restoration degree and modeling accuracy, can predict extreme sea conditions in advance through the physical mechanism driven rolling forward deduction and Kalman filter closed loop correction, gets rid of the dependence on historical data, adopts three times B-spline smoothing splicing, the route is smooth and easy to control, and can be extended to multi-ship formation, and is high in universality and practicability.

Description

Technical Field

[0001] This invention relates to the field of ship navigation planning technology, specifically to an intelligent ship navigation planning optimization method. Background Technology

[0002] With the rapid upgrading of the global smart shipping industry, unmanned ships and autonomous navigation vessels are gradually moving from near-shore channel pilot projects to open sea ocean applications. Route planning, as the core underlying technology for autonomous navigation of smart ships, directly determines navigation safety, fuel economy, carbon emission levels, and on-time performance.

[0003] Existing ship routing technologies generally have significant limitations: traditional A Graph search and heuristic methods such as Dijkstra's algorithm are only applicable to static environments and simple hydrological conditions. They often simplify ships into point mass models, ignoring the time-varying characteristics of waves, ocean currents, tides, and sea breezes. They lack the ability to predict future sea conditions and replan online, making them unsuitable for complex open sea areas. Coastal channel-based planning methods only consider static geographical elements, resulting in fixed environmental modeling and a lack of dynamic coupling and extrapolation. This makes them unable to achieve multi-objective optimization and limits their applicability. Statistical methods based on historical AIS big data are retrospective planning methods that lack forward extrapolation driven by physical mechanisms. They are unable to cope with extreme and unknown sea conditions and have large deviations in energy consumption calculations. Coverage-based path planning for inspection and sampling differs significantly from point-to-point navigation tasks in the open ocean and lacks the ability to model large-scale dynamic environments and couple ship hydrodynamics.

[0004] In summary, existing technologies generally suffer from problems such as static and simplistic environmental modeling, oversimplified ship modeling, and poor path executability, which make it difficult to meet the actual needs of autonomous navigation of intelligent ships in open sea areas. Therefore, a method for optimizing intelligent ship sea area path planning is proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent ship marine route planning and optimization method, which has the advantages of accurate multi-field coupled environment modeling, forward-looking prediction of future sea conditions, and adaptive dynamic replanning. It solves the problems of single static environment modeling, oversimplified ship modeling, and poor route executability.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent ship sea route planning and optimization method, comprising the following steps:

[0007] S1. Multi-source heterogeneous data acquisition and spatiotemporal fusion: Collect satellite remote sensing ocean data, ocean reanalysis data, global tide forecast data, meteorological wind speed and air pressure data, shipboard AIS data, and shipboard multi-sensor measured data, and uniformly calibrate to the UTC standard time base; use spatiotemporal kriging interpolation algorithm to perform joint interpolation and fusion of spatial and temporal dimensions on multi-source heterogeneous data to construct a standardized marine environment grid field containing multi-dimensional parameters of ocean waves, ocean currents, tides, and sea breezes; S2. Construction and Dynamic Correction of Digital Twin for Multi-Field Coupled Environment: Using the SWAN-ROMS-TPXO coupled model, wave field, ocean current-tidal coupled field, and sea surface wind field are established respectively to achieve dynamic coupled modeling of four fields: wave-ocean current-tidal-sea wind. A wave-current interaction correction mechanism is introduced to dynamically correct the effective wave height based on the angle between ocean current and wave direction. RollingForecast is carried out based on the physical mechanism model to generate a marine environment evolution sequence for a preset duration. A relative error index for environmental extrapolation is defined. When the error exceeds the preset threshold, EnKF ensemble Kalman filter is used to fuse real-time observation data for data assimilation, and the digital twin model parameters are updated online, forming a process from extrapolation to observation and final correction. S3, DT-DRL Dual-Engine Architecture Construction and Model Training: A DT-DRL dual-engine architecture consisting of a digital twin simulation training engine and a ship-side online inference engine was constructed. A multi-field coupled environment digital twin, after closed-loop correction, was used as the deep reinforcement learning simulation training environment to build an improved PPO-Attention network model embedded with a spatiotemporal attention module. A six-degree-of-freedom hydrodynamic response model of the ship was coupled throughout the model training process to establish the ship's six-degree-of-freedom motion equations. The ITTC-1957 friction line formula, Seigel wave drag-increasing empirical formula, and Isherwood wind resistance formula were used, combined with real-time ocean environmental parameters, to calculate static water resistance, wave-added resistance, and wind-added resistance in real time, thereby solving for the ship's total navigation resistance and the real-time power consumption of the main engine. A three-stage hierarchical training strategy was adopted, combined with domain randomization parameter perturbation to achieve Sim-to-Real virtual-real transfer, improving the model's generalization ability in real-world sea areas. S4. Multi-objective composite reward path decision output: Construct a four-objective composite reward function that takes into account navigation safety, fuel consumption, carbon emissions, and navigation timeliness; Deploy the trained and converged PPO-Attention policy network to the ship's edge computing device to achieve millisecond-level low-latency online inference, output the optimal heading increment and speed increment control commands, and generate the dynamic optimal navigation path in open sea areas. S5. Adaptive Rolling Replanning and Path Smoothing: Set time-period trigger conditions and three types of abnormal trigger conditions: marine environment, fuel consumption, and ship encounters. Rolling replanning is initiated immediately when any condition is met. The ship's current real-time latitude and longitude position is used as the starting point for the new planning. The trained PPO-Attention inference model is called to generate local alternative path segments. A cubic B-spline curve interpolation algorithm is used to smoothly splice the new and old route segments to ensure the continuity of the route curvature and the absence of abrupt inflection points. S6. Optional Multi-Ship Collaborative Extension: Based on the Multi-Agent Reinforcement Learning (MARL) framework, a multi-ship formation collaborative reward function is constructed, introducing a formation wake fuel-saving reward term, and utilizing the relative distance and relative speed parameters between ships to achieve multi-ship formation path collaborative optimization and energy-saving navigation.

[0008] Furthermore, step S2, specifically the four-field coupling modeling and wave-current interaction correction, includes: The wave field uses the SWAN wave spectrum model, considering wave energy generation, sea surface whitecap dissipation, and three-wave nonlinear interaction processes to establish a spatiotemporal evolution mapping relationship for waves, outputting the effective wave height, spectral peak period, and wave direction parameters at any spatiotemporal location. The ocean current field uses the ROMS three-dimensional regional ocean model to output a three-dimensional current velocity field, which is then superimposed with the tidal current field output from the TPXO global tidal model to synthesize a spatiotemporally continuous ocean current-tidal coupled current field. The sea breeze field uses ECMWF global meteorological reanalysis wind speed data, combined with sea surface roughness length associated with wave age, and employs a logarithmic wind profile formula to correct the near-surface wind field vertical profile, obtaining the true equivalent wind speed and direction at the sea surface. When the angle between the ocean current direction and the wave propagation direction is less than 30°, the wave-current interaction correction formula is used to correct the effective wave height.

[0009] In the formula, This represents the original, uncorrected significant wave height. The average velocity of the ocean current. The wave phase velocity.

[0010] Furthermore, the Rolling Forecast derivation and EnKF closed-loop correction described in step S2 include: Based on the current reference time Starting from this point, a fixed simulation duration of 24 hours is set. The simulation proceeds forward step by step according to a fixed time interval, generating a continuous time-series sequence of future marine environmental fields. ; Define the formula for calculating the relative error of environmental simulation:

[0011] In the formula, Values ​​extrapolated for a digital twin environment. These are actual observations of the marine environment. Preset error threshold, when When the error exceeds the threshold, the Ensemble Kalman Filter (EnKF) algorithm is activated to assimilate the field-measured observation data into the digital twin model, update the model state variables and coupling parameters online, and achieve continuous self-correction of the digital twin.

[0012] Furthermore, the six-degree-of-freedom hydrodynamic response model and resistance power calculation of the ship described in step S3 include: Establish the equations of motion for a ship's six-degree-of-freedom rigid body:

[0013] In the formula, These represent the six degrees of freedom motion vectors of the ship: pitch, sway, heave, roll, pitch, and bow. For the ship mass matrix, Here is the damping matrix. Here is the stiffness matrix. , , These are the forces of ocean waves, ocean winds, and ocean currents and tidal currents, respectively. The total resistance of a ship consists of three parts: still water resistance, wave resistance, and wind resistance. Still water resistance is calculated using the ITTC-1957 friction resistance line combined with the ship shape factor. Wave resistance is calculated using the Seigel empirical formula with the addition of a second-order wave drift force correction term. Wind resistance is calculated using the Isherwood wind resistance empirical formula. Based on the ship's total sailing resistance, real-time ship speed, and the combined transmission efficiency of the propulsion system, hull, and gearbox, the required output power of the ship's main engine is calculated in real time.

[0014] Furthermore, the spatiotemporal attention module construction method of the improved PPO-Attention network described in step S3 is as follows: The current navigation state vector of the ship is used as the attention query matrix Q, and the key matrix K and value matrix V are constructed by splicing the time sequence of the multi-step marine environment field in the future time window. By adaptively calculating the environmental feature weights at each future moment through an attention mechanism, the system automatically focuses on key dangerous areas such as strong wave zones, strong current zones, and areas of sudden weather changes within the future time window. The weighted and fused spatiotemporal environmental latent features are then input into the PPO policy network and value network to achieve discrete or continuous action decisions for course increments and speed increments.

[0015] Furthermore, the three-stage training and domain randomized transfer of course learning described in step S3 includes: Phase 1 training: Set up a scenario of open sea with low sea state and no surrounding vessels to train the basic obstacle avoidance and stable navigation capabilities of intelligent ships. The second phase of training involves setting up a scenario with moderate sea conditions and regular navigating vessels, with the primary goal of optimizing fuel efficiency, to train energy-saving navigation decision-making capabilities. The third phase of training involves setting up complex scenarios such as high sea states and dense encounters between multiple ships to train the ability to make decisions on multiple objectives such as energy efficiency, carbon emissions, and on-time performance under safety constraints. Domain randomized Sim-to-Real transfer training: During simulation training, the effective wave height of ocean waves is randomly perturbed within a range of ±20%, and the ocean current velocity is randomly perturbed within a range of ±15%. This random perturbation of the ocean environment model parameters enhances the robustness and generalization ability of the training strategy in real and complex marine environments.

[0016] Furthermore, the adaptive rolling replanning triggering conditions in step S5 include: (1) The relative error between the marine environment projection value and the actual observation value exceeds the preset error threshold; (2) The ship's actual cumulative fuel consumption exceeds the planned fuel consumption for the same voyage by 1.1 times; (3) The closest encounter distance between this vessel and the surrounding target vessels is less than 2 nautical miles; If any trigger condition is met, the current planned path is immediately terminated. The ship's current real-time position is used as the new starting point. The PPO-Attention inference model is called again to generate local path segments. Cubic B-spline curves are used to smoothly splice the old and new route segments to ensure a smooth transition of the route without sharp abrupt changes.

[0017] Furthermore, the multi-ship collaborative expansion described in step S6 includes: The Multi-Agent Reinforcement Learning (MARL) framework is adopted to assign an independent agent to each vessel in the formation, construct a fuel-saving reward for the formation, calculate the wake energy-saving gain based on the relative distance and relative speed between the vessels, and achieve the maintenance of the formation spacing, synchronized route planning, and overall fuel-saving and low-carbon navigation of the formation through multi-agent collaborative training.

[0018] Compared with the prior art, the technical solution of this application has the following beneficial effects: 1. This intelligent ship marine path planning optimization method uses a four-field dynamic coupling model of ocean waves, ocean currents, tides, and sea breezes, and introduces a wave-current interaction correction mechanism. It breaks through the limitations of existing single-element, static, and isolated environmental modeling technologies, highly restores the real marine dynamic evolution law of open sea areas, and significantly improves the accuracy of environmental modeling. 2. This intelligent ship sea route planning and optimization method adopts the rolling Forecast future time window driven by physical mechanism, which gets rid of the dependence on traditional big data posterior statistics and can predict extreme sea conditions such as strong waves, strong currents and sudden weather changes 24 hours in advance; combined with EnKF ensemble Kalman filtering to realize the inference-observation-correction closed loop, ensuring the long-term simulation and prediction accuracy of digital twins; 3. This intelligent ship marine path planning optimization method constructs a dual-engine architecture of digital twin simulation training and ship-side online inference. It can simulate various extreme sea conditions and multi-ship dense encounter scenarios in a virtual twin environment at zero cost and zero risk, and complete large-scale parallel training. The ship-side inference latency is low, which meets the engineering implementation requirements of real-time autonomous decision-making for intelligent ships. 4. The intelligent ship sea route planning optimization method embeds a spatiotemporal attention module into the PPO network, enabling the agent to actively identify and focus on key dangerous sea state areas in the future, and has advanced prediction capabilities. Its decision-making foresight is far superior to traditional real-time planning algorithms. 5. This intelligent ship marine route planning and optimization method abandons the traditional simplified ship mass point model, establishes a six-degree-of-freedom hydrodynamic response motion equation, accurately couples the calculation models of hydrostatic resistance, wave resistance, and wind resistance, and solves the total navigation resistance and main engine power in real time, so that each planned route conforms to the real navigation dynamics characteristics of the ship, with strong route feasibility and reliable calculation of energy consumption and carbon emissions. 6. This intelligent ship marine route planning and optimization method constructs a composite reward function with four objectives: safety, fuel consumption, carbon emissions, and on-time performance, achieving a global optimal trade-off among multiple objectives; it is equipped with a multi-condition adaptive rolling replanning mechanism, which can cope with sudden operating conditions such as changes in the marine environment, excessive fuel consumption, and ship encounters, and has strong dynamic adaptive capabilities for the route. 7. The intelligent ship sea route planning optimization method adopts cubic B-spline smooth splicing after route replanning to ensure continuous route curvature and smooth navigation and maneuvering, which meets the ship maneuverability constraints; it can be easily extended to multi-ship MARL formation collaborative scenarios, adapting to the future large-scale formation operation needs of ocean-going intelligent ships, and has excellent versatility and scalability. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Step 1: Multi-source heterogeneous data acquisition and spatiotemporal kriging fusion First, comprehensive, multi-source data collection is conducted. The data sources specifically include: 1) Satellite remote sensing data: data on significant wave height, wave direction, and wind speed and direction at 10m above sea level obtained by satellite inversion; 2) Ocean reanalysis data: Three-dimensional ocean current velocity, temperature, and salinity time-series data output by regional ocean reanalysis products; 3) Global tidal forecast data: tidal harmonic constants and tidal current time series data output by the TPXO global tidal model; 4) Meteorological forecast data: Time series data of sea surface pressure and gradient wind speed from ECMWF meteorological reanalysis and forecast; 5) Vessel navigation data: latitude and longitude, speed, heading, bow direction, and draft data broadcast by the vessel's AIS; 6) Measured data from shipborne sensors: Real-time sampling data from shipborne inertial navigation, anemometer, wave height meter, and current meter.

[0022] All the above multi-source data are unified with a time coordinate and calibrated to the UTC international standard time base to eliminate time misalignment and deviation between different data sources.

[0023] The spatiotemporal kriging interpolation algorithm is adopted, and interpolation fitting is performed simultaneously in the spatial grid dimension and the continuous time dimension to interpolate discrete observation point data into a regular sea area grid field. Each grid node of the grid field contains multi-dimensional environmental parameters such as significant wave height, spectral peak period, wave direction, ocean current northward velocity, eastward velocity, tidal current component, sea surface wind speed, wind direction, and atmospheric pressure, forming a standardized, gridded, and time-seriesd basic field of the sea area environment, providing high-precision input data for subsequent digital twin modeling.

[0024] Step 2: Construction of digital twin of multi-field coupled environment, wave current correction, rolling simulation and EnKF closed-loop correction The four-field coupled modeling implementation of the wave field adopts the SWAN third-generation wave spectrum numerical model. The modeling fully considers the generation of wind and wave energy input, the dissipation of sea surface whitecap breaking, the dissipation of bottom friction, the nonlinear interaction of three waves, and the physical processes of wave refraction and diffraction. The wave spectrum evolution is solved grid by grid and time step by time, and the continuous spatiotemporal wave parameter field of the whole sea area is output.

[0025] The ocean current field was obtained by solving the original ocean equations using the ROMS three-dimensional regional ocean model. The three-dimensional spatiotemporal velocity field was then superimposed with the tidal current fields of each tidal component predicted by the TPXO global tidal model to synthesize the integrated ocean current-tidal coupled current field, which reflects the superposition effect of large-scale circulation and local tides.

[0026] The sea wind field is based on the ECMWF10m height wind speed and direction. The sea surface roughness length is calculated according to the real-time wave age. The logarithmic wind profile formula is used to vertically correct the near-sea wind field to obtain the true equivalent wind speed and direction at the ship's waterline height, thus eliminating the wind field calculation error caused by height differences.

[0027] The wave-current interaction correction is implemented in the actual ocean. Waves and ocean currents do not evolve independently. When there is an angle between the direction of the ocean current and the direction of the wave, it will significantly change the effective wave height and the ship's navigation resistance.

[0028] This embodiment sets a judgment threshold: when the angle between the direction of ocean current movement and the direction of wave propagation is less than 30°, the wave-current coupling correction formula is activated.

[0029] In the formula, This represents the original, uncorrected significant wave height. The average velocity of the ocean current. The wave phase velocity.

[0030] The original significant wave height is amplified and corrected to truly reflect the physical law that the downstream wave field increases and the ship's navigation resistance increases; when the included angle is greater than or equal to 30°, no correction is required or a weak correction coefficient is used to conform to the real ocean dynamics mechanism.

[0031] RollingForecast predicts future time windows based on the current UTC time and the current reference time. Using this as the starting point for the extrapolation, a standard extrapolation duration of 24 hours is set. The calculation is iterated forward step by step according to a fixed time step to generate a continuous time series of marine environmental fields within the next 24 hours, realizing a priori physical prospective prediction, which is different from the a posteriori historical statistical method of existing technologies.

[0032] EnKF ensemble Kalman filter online closed-loop correction collects actual marine environmental observation data at a fixed calibration period (e.g., 6 hours) and calculates the relative error between the extrapolated and observed values:

[0033] In the formula, Values ​​extrapolated for a digital twin environment. These are actual observations of the marine environment.

[0034] Preset error threshold, when When the error exceeds the threshold, the Ensemble Kalman Filter (EnKF) algorithm is activated to assimilate the measured observation data into the digital twin model, update the model coupling parameters, boundary conditions and state variables, and realize automatic correction and continuous high-precision evolution of the digital twin.

[0035] Step 3: DT-DRL dual-engine setup, PPO-Attention model training, and six-DOF hydrodynamic coupling. The DT-DRL dual-engine architecture deployment consists of a digital twin simulation training engine and a ship-side online inference engine: The simulation training engine uses a calibrated multi-field coupled digital twin as a virtual navigation environment, which can generate a large number of navigation scenarios in batches and in parallel, including different sea states, different obstacle distributions, and different ship encounter situations, completing agent training with zero risk; The online inference engine deploys the policy network after training convergence to the ship-side edge computing industrial control computer, with inference latency controlled within 100ms, meeting the real-time control requirements of actual ships.

[0036] The spatiotemporal attention-enhanced PPO network is built on the basis of the standard PPO policy network by adding a spatiotemporal attention module. The query matrix Q is composed of the ship's current position, speed, heading, and roll status. The key matrix K and value matrix V are composed of the environmental field features of the next 24 hours. Through adaptive allocation of attention weights, the feature weights of dangerous areas such as strong wave areas, strong current areas, and no-navigation areas are automatically increased, enabling the agent to perceive long-term environmental risks in advance and make predictive path decisions.

[0037] The six-degree-of-freedom hydrodynamic model of the ship is coupled to establish the rigid body motion equations for six degrees of freedom: pitch, sway, heave, roll, pitch, and bow.

[0038] In the formula, These represent the six degrees of freedom motion vectors of the ship: pitch, sway, heave, roll, pitch, and bow. For the ship mass matrix, Here is the damping matrix. Here is the stiffness matrix. , , These are the forces of ocean waves, ocean winds, and ocean currents and tides.

[0039] It comprehensively depicts the motion response of ships under the combined effects of waves, wind, and ocean currents, avoiding the assumptions of traditional point mass models.

[0040] The navigation resistance is calculated precisely in three categories: still water resistance is calculated using the ITTC-1957 friction resistance line formula, combined with the hull shape factor to correct for shape resistance; wave resistance is calculated using the Seigel empirical formula, with the addition of a second-order wave drift force correction term to accurately calculate the additional resistance caused by waves; wind resistance is calculated using the Isherwood wind resistance empirical formula, based on the ship's windward area and wind angle to calculate the additional wind resistance in real time.

[0041] By combining total resistance, ship speed, propulsion efficiency, hull efficiency, and gearbox efficiency, the required output power of the ship's main engine can be calculated in real time, achieving precise quantification of energy consumption and carbon emissions.

[0042] The course learning involves three stages of training and domain randomized transfer, including: Phase 1 training: Low sea state scenario with no other vessels, training basic obstacle avoidance and stable navigation; Phase Two Training: Conventional navigation scenarios in moderate sea states, focusing on fuel-saving and energy efficiency optimization; The third phase of training involves scenarios involving multiple ships meeting in high sea states, emphasizing the trade-offs between safety, energy efficiency, and carbon emissions.

[0043] Meanwhile, during training, random perturbations of wave height ±20% and current velocity ±15% are applied to achieve Sim-to-Real virtual-real transfer, thereby improving the model's robustness in real sea areas.

[0044] Step 4: Four-objective composite reward function and path decision output Construct a composite reward function:

[0045] In the formula: For the overall reward value, For safety rewards, For fuel consumption incentives, For carbon emission incentives, As a reward item for on-time performance, , , , They are respectively , , , The weighting coefficients, and , , , all ,satisfy .

[0046] Safety bonuses include penalties for obstacle distance and penalties for exceeding ship roll limits; The fuel consumption bonus is calculated using real-time relative fuel consumption normalization. The carbon emission incentive is obtained by multiplying fuel consumption by the carbon emission factor. The on-time performance bonus applies a second penalty for flight delays.

[0047] The trained policy network performs real-time inference, outputting heading and speed increments, and generating the globally optimal navigation path segment by segment.

[0048] Step 5: Adaptive Rolling Replanning and Cubic B-Spline Path Smoothing Three types of replanning trigger conditions are set: environmental simulation error exceeds the standard, fuel consumption exceeds the planned value by 1.1 times, and the CPA of other vessels is less than 2 nautical miles. Any trigger will result in immediate replanning.

[0049] The local path segment is regenerated from the current ship position. Cubic B-spline curve interpolation is used to smoothly connect the old and new routes, eliminating planning inflection points and ensuring smooth ship navigation and maneuvering that conforms to the ship's turning maneuvering characteristics.

[0050] Step Six: Multi-ship Formation Collaborative Implementation The system employs MARL multi-agent reinforcement learning, with each ship acting as an independent agent. It introduces a formation-following fuel-saving reward, calculates energy-saving gains based on the relative distance and relative speed between ships, and achieves synchronized routes, maintained spacing, and low-carbon and energy-saving navigation of the entire formation.

[0051] The working principle of the above embodiments is as follows: This invention employs a four-field dynamic coupling modeling approach, combining ocean waves, ocean currents, tides, and sea breezes, and introduces wave-current interaction correction. This overcomes the limitations of traditional static, singular, and isolated environmental modeling, significantly improving the accuracy and realism of marine environment simulation. Through physical mechanism-driven rolling forward inference and Kalman filter closed-loop correction, extreme sea conditions can be predicted in advance, eliminating reliance on historical data. It utilizes a dual-engine approach of digital twin simulation training and ship-side online inference, combined with a PPO network embedded with spatiotemporal attention, resulting in more forward-looking and real-time decision-making. It can safely and efficiently complete training and deployment on real ships in complex scenarios. By abandoning the point mass model and coupling the ship's six-degree-of-freedom hydrodynamic response, it accurately calculates navigation resistance and energy consumption, significantly improving path feasibility and the realism of energy efficiency calculations. Through four-objective composite rewards and multi-condition adaptive replanning, it balances navigation safety, fuel economy, low carbon emissions, and punctuality. It uses cubic B-spline smooth splicing, resulting in smooth and easy-to-control routes, and can be expanded to multi-ship formations, demonstrating strong versatility and practicality.

[0052] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing intelligent ship navigation routes in marine areas, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data acquisition and spatiotemporal fusion: Collect satellite remote sensing ocean data, ocean reanalysis data, global tide forecast data, meteorological wind speed and air pressure data, shipboard AIS data, and shipboard multi-sensor measured data, and uniformly calibrate to the UTC standard time base; use spatiotemporal kriging interpolation algorithm to perform joint interpolation and fusion of spatial and temporal dimensions on multi-source heterogeneous data to construct a standardized marine environment grid field containing multi-dimensional parameters of ocean waves, ocean currents, tides, and sea breezes; S2. Construction and Dynamic Correction of Digital Twin for Multi-Field Coupled Environment: Using the SWAN-ROMS-TPXO coupled model, wave field, ocean current-tidal coupled field, and sea surface wind field are established respectively to achieve dynamic coupled modeling of four fields: wave-ocean current-tidal-sea wind. A wave-current interaction correction mechanism is introduced to dynamically correct the effective wave height based on the angle between ocean current and wave direction. RollingForecast is carried out based on the physical mechanism model to generate a marine environment evolution sequence for a preset duration. A relative error index for environmental extrapolation is defined. When the error exceeds the preset threshold, EnKF ensemble Kalman filter is used to fuse real-time observation data for data assimilation, and the digital twin model parameters are updated online, forming a process from extrapolation to observation and final correction. S3, DT-DRL Dual-Engine Architecture Construction and Model Training: Construct a DT-DRL dual-engine architecture consisting of a digital twin simulation training engine and a ship-side online inference engine; use a digital twin of a multi-field coupled environment after closed-loop correction as a deep reinforcement learning simulation training environment to build an improved PPO-Attention network model with embedded spatiotemporal attention modules; couple the ship's six-degree-of-freedom hydrodynamic response model throughout the model training process to establish the ship's six-degree-of-freedom motion equations; The ITTC-1957 friction line formula, Seigel's empirical formula for wave drag, and Isherwood's wind resistance formula were used respectively, combined with real-time ocean environmental parameters, to calculate still water resistance, wave-added resistance, and wind-added resistance in real time, and then to solve for the total ship navigation resistance and the real-time power consumption of the main engine. A three-stage hierarchical training strategy was adopted for course learning, and Sim-to-Real virtual-real transfer was achieved by combining domain randomization parameter perturbation to improve the model's generalization ability in real sea areas. S4. Multi-objective composite reward path decision output: Construct a four-objective composite reward function that takes into account navigation safety, fuel consumption, carbon emissions, and navigation timeliness; Deploy the trained and converged PPO-Attention policy network to the ship's edge computing device to achieve millisecond-level low-latency online inference, output the optimal heading increment and speed increment control commands, and generate the dynamic optimal navigation path in open sea areas. S5. Adaptive Rolling Replanning and Path Smoothing: Set time period trigger conditions and three types of abnormal trigger conditions: marine environment, fuel consumption, and ship encounters. Rolling replanning will be started immediately when any condition is met. Using the ship's current real-time latitude and longitude position as the starting point for a new route planning, the trained PPO-Attention inference model is invoked to generate local alternative route segments; a cubic B-spline curve interpolation algorithm is used to smoothly splice the new and old route segments to ensure continuous curvature of the route and no abrupt inflection points; S6. Optional Multi-Ship Collaborative Extension: Based on the Multi-Agent Reinforcement Learning (MARL) framework, a multi-ship formation collaborative reward function is constructed, introducing a formation wake fuel-saving reward term, and utilizing the relative distance and relative speed parameters between ships to achieve multi-ship formation path collaborative optimization and energy-saving navigation.

2. The intelligent ship navigation optimization method according to claim 1, characterized in that, Step S2, specifically the four-field coupling modeling and wave-current interaction correction, includes: The wave field uses the SWAN wave spectrum model, considering wave energy generation, sea surface whitecap dissipation, and three-wave nonlinear interaction processes to establish a spatiotemporal evolution mapping relationship for waves, outputting the effective wave height, spectral peak period, and wave direction parameters at any spatiotemporal location. The ocean current field uses the ROMS three-dimensional regional ocean model to output a three-dimensional current velocity field, which is then superimposed with the tidal current field output from the TPXO global tidal model to synthesize a spatiotemporally continuous ocean current-tidal coupled current field. The sea breeze field uses ECMWF global meteorological reanalysis wind speed data, combined with sea surface roughness length associated with wave age, and employs a logarithmic wind profile formula to correct the near-surface wind field vertical profile, obtaining the true equivalent wind speed and direction at the sea surface. When the angle between the ocean current direction and the wave propagation direction is less than 30°, the wave-current interaction correction formula is used to correct the effective wave height. In the formula, This represents the original, uncorrected significant wave height. The average velocity of the ocean current. The wave phase velocity.

3. The intelligent ship navigation optimization method according to claim 1, characterized in that, The Rolling Forecast derivation and EnKF closed-loop correction described in step S2 include: Based on the current reference time Starting from this point, a fixed simulation duration of 24 hours is set. The simulation proceeds forward step by step according to a fixed time interval, generating a continuous time-series sequence of future marine environmental fields. ; Define the formula for calculating the relative error of environmental simulation: In the formula, Values ​​extrapolated for a digital twin environment. These are actual observations of the marine environment. Preset error threshold, when When the error exceeds the threshold, the Ensemble Kalman Filter (EnKF) algorithm is activated to assimilate the field-measured observation data into the digital twin model, update the model state variables and coupling parameters online, and achieve continuous self-correction of the digital twin.

4. The intelligent ship navigation optimization method according to claim 1, characterized in that, The six-degree-of-freedom hydrodynamic response model and resistance power calculation of the ship mentioned in step S3 include: Establish the equations of motion for a ship's six-degree-of-freedom rigid body: In the formula, These represent the six degrees of freedom motion vectors of the ship: pitch, sway, heave, roll, pitch, and bow. For the ship mass matrix, Here is the damping matrix. Here is the stiffness matrix. , , These are the forces of ocean waves, ocean winds, and ocean currents and tidal currents, respectively. The total resistance of a ship consists of three parts: still water resistance, wave resistance, and wind resistance. Still water resistance is calculated using the ITTC-1957 friction resistance line combined with the ship shape factor. Wave resistance is calculated using the Seigel empirical formula with the addition of a second-order wave drift force correction term. Wind resistance is calculated using the Isherwood wind resistance empirical formula. Based on the ship's total sailing resistance, real-time ship speed, and the combined transmission efficiency of the propulsion system, hull, and gearbox, the required output power of the ship's main engine is calculated in real time.

5. The intelligent ship navigation optimization method according to claim 1, characterized in that, The spatiotemporal attention module construction method of the improved PPO-Attention network described in step S3 is as follows: The current navigation state vector of the ship is used as the attention query matrix Q, and the key matrix K and value matrix V are constructed by splicing the time sequence of the multi-step marine environment field in the future time window. By adaptively calculating the environmental feature weights at each future moment through an attention mechanism, the system automatically focuses on key dangerous areas such as strong wave zones, strong current zones, and areas of sudden weather changes within the future time window. The weighted and fused spatiotemporal environmental latent features are then input into the PPO policy network and value network to achieve discrete or continuous action decisions for course increments and speed increments.

6. The intelligent ship marine route planning and optimization method according to claim 1, characterized in that, The three-stage training and domain randomization transfer of the course learning described in step S3 include: Phase 1 training: Set up a scenario of open sea with low sea state and no surrounding vessels to train the basic obstacle avoidance and stable navigation capabilities of intelligent ships. The second phase of training involves setting up a scenario with moderate sea conditions and regular navigating vessels, with the primary goal of optimizing fuel efficiency, to train energy-saving navigation decision-making capabilities. The third phase of training involves setting up complex scenarios such as high sea states and dense encounters between multiple ships to train the ability to make decisions on multiple objectives such as energy efficiency, carbon emissions, and on-time performance under safety constraints. Domain randomized Sim-to-Real transfer training: During simulation training, the effective wave height of ocean waves is randomly perturbed within a range of ±20%, and the ocean current velocity is randomly perturbed within a range of ±15%. This random perturbation of the ocean environment model parameters enhances the robustness and generalization ability of the training strategy in real and complex marine environments.

7. The intelligent ship navigation optimization method according to claim 1, characterized in that, The adaptive rolling replanning triggering conditions mentioned in step S5 include: (1) The relative error between the marine environment projection value and the actual observation value exceeds the preset error threshold; (2) The ship's actual cumulative fuel consumption exceeds the planned fuel consumption for the same voyage by 1.1 times; (3) The closest encounter distance between this vessel and the surrounding target vessels is less than 2 nautical miles; If any trigger condition is met, the current planned path is immediately terminated. The ship's current real-time position is used as the new starting point. The PPO-Attention inference model is called again to generate local path segments. Cubic B-spline curves are used to smoothly splice the old and new route segments to ensure a smooth transition of the route without sharp abrupt changes.

8. The intelligent ship navigation optimization method according to claim 1, characterized in that, The multi-ship collaborative expansion mentioned in step S6 includes: The Multi-Agent Reinforcement Learning (MARL) framework is adopted to assign an independent agent to each vessel in the formation, construct a fuel-saving reward for the formation, calculate the wake energy-saving gain based on the relative distance and relative speed between the vessels, and achieve the maintenance of the formation spacing, synchronized route planning, and overall fuel-saving and low-carbon navigation of the formation through multi-agent collaborative training.

Images