A ship motion trajectory optimization control system based on deep learning

Abstract

The application relates to the technical field of ship swarm intelligence control, and discloses a ship motion trajectory optimization control system based on deep learning, which comprises control devices arranged on ships, receives neighboring ship sailing state information through a ship automatic identification system, generates an asymmetric virtual energy-saving potential field and superimposes the same to form a total potential field, and outputs a sailing track correction instruction based on a negative gradient direction of the potential field. The application enables ships to spontaneously form a whole dynamic sailing manifold with low energy consumption, converts the chain sailing direction correction caused by traditional collision avoidance into cooperative energy-saving motion through the design of an asymmetric potential field of a ship bow repulsive force and a ship tail attractive force, ensures sailing safety by combining a post-position arbitration mechanism, and realizes environment self-adaptive compensation by using a perturbation inversion ocean current.

Description

Technical Field

[0001] This invention relates to a deep learning-based ship motion trajectory optimization control system, belonging to the field of ship swarm intelligent control technology. Background Technology

[0002] Existing technologies mainly rely on autonomous decision-making by individual ships and passive collision avoidance mechanisms. Ships maintain a preset course through automatic steering systems and make discrete turns or decelerations according to international maritime collision avoidance rules when encountering other ships. While this method ensures basic navigation safety, it has systemic limitations: the pursuit of the shortest path by a single ship leads to a chain of course corrections, causing the propagation of micro-disturbances in navigation. For example, a small turn by the preceding ship forces the following ships to adjust their sequence, accumulating into unnecessary energy-consuming ripples. The root cause is that it ignores the energy superposition effect of multi-body motion in hydrodynamics and equates individual optimization with group optimization.

[0003] Existing solutions treat ocean currents as uncontrollable interference and rely on steering gear to continuously correct track deviations. This passive compensation mechanism not only increases energy consumption but also makes it difficult to achieve dynamic coordination between track and hydrodynamics due to the lack of real-time environmental perception capabilities. Coordinated control requires stabilizing the data flow of the Automatic Identification System (AIS). When severe weather or channel congestion causes signal loss, the system is prone to triggering dangerous maneuvers due to sudden changes in the potential field of neighboring ships. Although attempts have been made to introduce linear extrapolation verification, a single decision channel is still insufficient to mitigate safety risks under sudden situations.

[0004] As global shipping density increases and energy efficiency requirements escalate, the aforementioned limitations have evolved from acceptable compromises into bottlenecks restricting industry development: discrete collision avoidance logic exacerbates energy waste, a lack of environmental awareness limits energy efficiency potential, and data link vulnerabilities threaten navigation safety. While the industry has explored path optimization algorithms, it has not broken through the centralized decision-making paradigm of single ships, nor has it resolved the synergistic contradictions between energy consumption, environment, and safety in distributed systems. Therefore, how to construct a distributed mechanism to achieve dynamic optimization of group energy consumption without direct inter-ship communication, while ensuring navigation safety and ocean current adaptability in complex environments, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a deep learning-based ship motion trajectory optimization control system, the main purpose of which is to solve the systemic technical contradiction of the difficulty in balancing global energy consumption optimization, environmental adaptation and navigation safety in distributed cooperative navigation of ship groups.

[0006] To achieve the above objectives, the present invention provides a deep learning-based ship motion trajectory optimization control system, including control devices deployed on each ship, the control devices being configured to perform the following functions:

[0007] The neighboring vessel information receiving module is configured to continuously receive and decode navigation status information broadcast by neighboring vessels through the Automatic Identification System (AIS). The navigation status information includes the position, speed above ground, and heading above ground of the neighboring vessels.

[0008] The virtual potential field generation module is configured to generate an asymmetric virtual energy-saving potential field based on the ship's current navigation status information, acquired through the ship's sensors. This information includes the ship's speed and heading relative to the ground. The asymmetric virtual energy-saving potential field generates a repulsive potential in the region in front of the ship's heading relative to the ground, and the strength of the repulsive potential is positively correlated with the ship's speed relative to the ground. The asymmetric virtual energy-saving potential field generates a gravitational potential in the region behind the ship's heading relative to the ground, and the strength of the gravitational potential is positively correlated with the ship's speed relative to the ground.

[0009] The total potential field superposition module is configured to linearly superimpose the virtual energy-saving potential fields of all neighboring ships received and decoded by the neighboring ship information receiving module to form the total virtual potential field where the ship is currently located.

[0010] The trajectory optimization decision module is configured to determine the negative gradient direction of the potential field that guides the ship's motion direction based on the total virtual potential field. The negative gradient direction of the potential field points to the direction in which the energy consumption of the total virtual potential field decreases the fastest.

[0011] The navigation control output module is configured to generate and output correction control commands to the ship's automatic steering system based on the negative gradient direction of the potential field, so as to drive the ship's trajectory to be adjusted in a controlled manner towards the negative gradient direction of the potential field, enabling the ship group to form a dynamic navigation manifold with reduced overall energy consumption.

[0012] The navigation micro-disturbance module is configured to periodically apply standardized course micro-disturbance commands to the ship's autopilot system. Based on the vector difference between the ship's actual navigation response caused by the micro-disturbance command and the still water response expected by the ship's maneuverability model, it inversely calculates the local ocean current vector acting on the ship and incorporates the local ocean current vector as a background correction term into the calculation of the total virtual potential field to compensate for the influence of ocean currents.

[0013] Preferably, the control device further includes a ship automatic identification system data preprocessing module, which is configured to: receive raw ship automatic identification system data, and based on the cosine value of the difference in heading angles between ships, dynamically cluster multiple ships whose cosine value of the difference in heading angles is greater than a predetermined heading clustering threshold into a navigation group, and calculate the centroid position, average ground speed and average ground heading for the navigation group, generating a single equivalent navigation state information for the total potential field superposition module.

[0014] Preferably, the virtual potential field generation module is configured to use at least two Gaussian function models with different weights superimposed to form an asymmetric virtual energy-saving potential field, wherein the center of one Gaussian model is located at a predetermined distance ahead of the ship's heading towards the ground, and the center of the other Gaussian model is located at a predetermined distance behind the ship's heading towards the ground.

[0015] Preferably, the control device further includes a post-decision arbitration module, which is configured to: receive the corrected control command output by the navigation control output module, and perform a five-second linear extrapolation based on the original Automatic Identification System (AIS) data received by the neighboring vessel information receiving module to generate a safety verification command; when the minimum encounter distance between the predicted trajectory corresponding to the corrected control command and the linear extrapolation trajectory of any neighboring vessel is not less than a predetermined safety distance threshold, and the initial turn of the corrected control command does not violate the key rules of the International Maritime Collision Avoidance Regulations, the corrected control command is approved for execution; otherwise, a safety verification command is output as the final control command.

[0016] Preferably, the predetermined safe distance threshold for the post-decision arbitration module is 0.1 nautical miles.

[0017] Preferably, the control device is further configured to: when it receives a signal that the ship's trajectory deviates from the negative gradient direction of the potential field due to manual operation, increase the weight of the asymmetric virtual energy-saving potential field generated by the ship, so as to broadcast stronger heading intention information to neighboring ships.

[0018] Preferably, the navigation control output module is configured to determine the correction magnitude of the correction control command based on the magnitude of the potential field gradient indicated by the negative gradient direction, so that the correction magnitude is proportional to the magnitude of the potential field gradient, thereby performing a smaller course adjustment in flat areas of the total virtual potential field to maintain course stability, and performing a larger adjustment in steep areas of the total virtual potential field.

[0019] Preferably, the Automatic Identification System (AIS) data preprocessing module is configured to: when the AIS signal of a vessel in the navigation group is briefly lost, the dead reckoning is performed using the average navigation status information of the navigation group.

[0020] Preferably, the control device further includes a communication status monitoring module, which is configured to monitor the stability of the communication channel by calculating the information entropy of the received Automatic Identification System (AIS) signal timestamp interval; when the information entropy ΔS increases sharply within a preset time period and exceeds a predetermined information entropy threshold S... th When a communication interruption is detected, the system notifies the total potential field overlay module to match the most similar historical pattern to the current navigation state from a pre-stored database containing historical potential field evolution patterns. Based on the matched historical pattern, a predictive virtual potential field P is generated. predict The predictive virtual potential field is compared with the effective potential field P at the last moment before the interruption.lastknown A weighted fusion is performed as part of the total virtual potential field used for superposition in step c, wherein the weighted fusion potential field P fusion P is determined by the following relationship: fusion =w(t)·P predict +(1-w(t))·P lastknown Here, w(t) is a weight function that grows in a controlled manner from zero to one as the communication interruption time t increases.

[0021] Preferably, the navigation control output module is configured to use the International Maritime Collision Avoidance Rules as the highest priority hard constraint when generating corrective control commands, and to adjust the course only without violating the International Maritime Collision Avoidance Rules.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. By constructing an asymmetric virtual energy-saving potential field for each ship and superimposing the potential fields of neighboring ships in real time based on the broadcast information of the Automatic Identification System (AIS), the system guides ships to autonomously adjust their trajectories along the negative gradient direction of the total potential field. This distributed decision-making mechanism enables the ship group to spontaneously form a dynamic navigation manifold, transforming the propagation of micro-disturbances caused by traditional collision avoidance, such as chain-like course corrections, into a cooperative motion with the lowest overall energy consumption. Ships generate a repulsive potential positively correlated with speed in front of them and an attractive potential behind them, thus avoiding the energy consumption of sudden braking from following ships and reducing navigation resistance by utilizing wakes, thereby avoiding the conflict between individual optimality and group energy consumption.

[0024] 2. A post-decision arbitration mechanism is introduced, which generates safety verification instructions through linear extrapolation. When the minimum encounter distance between the energy-saving trajectory generated by the core algorithm and the predicted trajectory of any neighboring vessel is lower than the safety threshold, or when international collision avoidance rules are violated, the system automatically switches to the bottom-line instruction. Combined with the dynamic clustering and dead reckoning of vessels with similar course by the AIS data preprocessing module, this reduces the risk of potential field jumps caused by sudden signal loss and reduces the computational load under dense waterways. The dual mechanism works together to ensure that group energy-saving optimization helps group energy-saving optimization operate within the safety boundary.

[0025] 3. By periodically adjusting the course micro-disturbance, the system actively stimulates the ship's response and compares it with the expected maneuvering model to deduce the local ocean current vector. This vector is incorporated into the overall potential field calculation as a background correction term, so that the gradient descent direction naturally includes the influence of ocean currents. The ship can perceive environmental interference without additional sensors, avoiding the continuous correction energy consumption caused by track deviation in traditional methods. Combined with the distributed fusion of ocean current data among the group, a self-calibrating environmental perception network is formed, improving energy-saving stability under complex sea conditions. When the AIS communication entropy value changes abruptly, the system matches similar scenarios based on the historical potential field evolution pattern library, generates a predictive virtual potential field, and weights and fuses it with the potential field before the interruption. The weight function is adaptively adjusted with the interruption time, realizing a smooth transition from real-time perception to empirical prediction. The ship can still maintain near-optimal track adjustment under extreme conditions, transforming communication interruption from a system failure risk into a controllable degraded operating state, ensuring the continuous operation of the group's collaboration. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the module structure of the ship motion trajectory optimization control system based on deep learning of the present invention;

[0027] Figure 2 This is a comparison curve of the cumulative energy consumption of the present invention system and the traditional system;

[0028] Figure 3 This is a flowchart of the security verification and decision arbitration process of this invention.

[0029] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0030] To better illustrate the purpose, technical solution, and advantages of this invention, the following detailed description of the invention will be provided in conjunction with specific embodiments. It should be understood that the description herein is merely exemplary and is not intended to limit the scope of protection of this invention.

[0031] This invention discloses a deep learning-based ship trajectory optimization control system, whose functional entity is a control device deployed on each ship. This device logically integrates a neighboring ship information receiving module, a virtual potential field generation module, a total potential field superposition module, a trajectory optimization decision module, and a navigation control output module. To address the complexity and uncertainty of real navigation environments, it further includes a ship automatic identification system (AIS) data preprocessing module, a navigation micro-disturbance module, a post-decision arbitration module, and a communication status monitoring module. These modules operate collaboratively, using navigation status information broadcast by the AIS as input, integrating the ship's own navigation status and environmental inversion data, and outputting trajectory correction commands to the ship's autopilot system through a distributed collaborative decision-making mechanism. This achieves dynamic optimization of the overall energy consumption of the ship group while ensuring navigation safety. Given that single-ship optimal path decisions in densely populated waters can easily trigger cascading collision avoidance, increasing energy consumption, the core operation of this system is designed as follows: the neighboring ship information receiving module continuously captures and decodes the position, speed, and heading information broadcast by neighboring ships through the AIS; simultaneously, a virtual potential field is generated... Based on the ship's real-time ground speed and heading, the module constructs an asymmetric virtual energy-saving potential field. This potential field is specifically composed of the superposition of at least two Gaussian function models with different weights. The center of one Gaussian model is located in front of the ship's ground heading, generating a repulsive potential proportional to the ship's ground speed. The center of the other model is located behind the heading, generating an attractive potential proportional to the speed. Furthermore, when a signal is received indicating that the ship's trajectory deviates from the negative gradient direction of the potential field due to human intervention, the system will increase the weight of the asymmetric virtual energy-saving potential field generated by the ship, thereby broadcasting stronger heading intention information to nearby ships. The virtual energy-saving potential fields of all neighboring ships are linearly superimposed by the total potential field superposition module to form the total virtual potential field in which the ship is currently located. The trajectory optimization decision module then calculates the direction of the negative gradient with the fastest energy consumption reduction based on this total potential field. The navigation control output module generates a correction control command based on this direction. At the same time, the correction magnitude of the command is proportional to the magnitude of the potential field gradient, ensuring that the course is finely adjusted in the flat potential field area to maintain stability, while more significant adjustments are made in the steep potential field area. Through this distributed self-organizing decision, the ship group can spontaneously form a dynamic navigation manifold with the lowest overall energy consumption.

[0032] To address the continuous energy consumption caused by ocean currents in maintaining a course, the system's built-in navigation micro-disturbance module is configured to periodically apply a series of standardized heading micro-disturbance commands to the ship's autopilot system. This module accurately records and compares the ship's actual navigation response caused by the micro-disturbance with the still water response expected by the ship's maneuverability model. The vector difference between the two is then used to inversely calculate the local ocean current vector acting on the ship. This inversely calculated ocean current vector is then seamlessly integrated into the subsequent calculation of the total virtual potential field as a background correction term. Through this active detection and quantification of environmental disturbances, the system naturally incorporates the influence of ocean currents into the course optimization decision, thereby avoiding the continuous rudder correction energy consumption caused by traditional passive compensation mechanisms and achieving efficient environmental adaptation.

[0033] Considering that navigation safety is always the highest and non-negotiable principle, the system has specially designed a post-decision arbitration module as the final safety defense. Its working procedure is as follows: When the energy-saving correction control command generated by the navigation control output module is received, the module will first intercept it and perform a five-second linear extrapolation of the trajectories of all nearby vessels based on the received original Automatic Identification System (AIS) data. Subsequently, the module calculates the minimum encounter distance between the predicted trajectory of the ship after executing the energy-saving command and the extrapolated trajectory of any nearby vessel, and checks whether the distance is less than the predetermined safe distance threshold of 0.1 nautical miles. At the same time, it checks whether the initial turning of the command conflicts with the key provisions of the International Maritime Collision Avoidance Regulations (IMC). Only when the safe distance is met and the collision avoidance rules are not violated will the energy-saving correction control command be approved for execution. Otherwise, the module will output a preset safety verification command as the final control command. This mechanism ensures that the system's energy-saving optimization behavior always operates under the rigid constraints of safety and regulations.

[0034] To further enhance the system's robustness in dense waterways or environments with deteriorating communication, the solution integrates a data preprocessing module and a communication status monitoring module for the Automatic Identification System (AIS). The former dynamically clusters multiple ships with course angle differences less than a predetermined course angle clustering threshold into a navigation group based on the cosine of the differences between their course angles. It then calculates the centroid position and average navigation status of each group, using a single equivalent information to participate in the overall potential field superposition. This not only reduces computational load but also allows for dead reckoning when a ship within a group experiences a brief signal loss, using the group's average navigation status to fill the gap. The latter monitors channel stability by calculating the information entropy ΔS of the received AIS signal timestamp interval. When the information entropy ΔS increases sharply within a preset time period and exceeds a predetermined information entropy threshold S... th When the system determines that a communication interruption has occurred, it immediately matches the most similar historical pattern to the current navigation scenario from the pre-stored historical potential field evolution pattern database to generate a predictive virtual potential field P. predict The predicted potential field is the same as the effective potential field P at the last moment before the interruption. lastknownThrough the weighted fusion formula P fusion =w(t)·P predict +(1-w(t))·P lastknown The integration process involves a weighting function w(t) that grows from zero to one in a controlled manner with the communication interruption time t. This smooth transition from real-time perception to empirical prediction transforms the potential risks of communication interruption into a controllable degraded operating state, thereby ensuring the continuity and safety of group collaboration under extreme conditions.

[0035] Example 1: In this example, a cargo ship convoy is sailing densely in a narrow channel with significant transverse ocean currents. If operating with conventional autonomous collision avoidance logic, the course corrections made by the lead ship to counteract the current's influence typically trigger a series of uncoordinated collision avoidance maneuvers by subsequent ships. This propagation of micro-disturbances not only disrupts the entire convoy's formation but also generates additional energy consumption due to frequent turning and speed adjustments by each ship, creating an irreconcilable technical contradiction between navigation safety and convoy energy efficiency. When the control device of this invention is deployed on each cargo ship, the system's synergistic efficiency and problem redefinition characteristics become apparent. The micro-disturbance module on each ship actively detects and independently derives accurate local ocean current vectors. These vectors are then seamlessly integrated into the calculation of the overall virtual potential field, becoming a dynamic... The background correction term of the state; therefore, the negative gradient direction of the potential field calculated by the trajectory optimization decision module inherently contains pre-compensation for ocean currents from the source, which means that the lead ship does not need to make large-scale unilateral corrections, thereby suppressing the initial triggering of chain collision avoidance reactions; furthermore, the establishment of the asymmetric virtual energy-saving potential field transforms the interaction relationship between ships from a simple physical obstacle into an energy-level coupling. The navigation state of any ship affects the whole in the form of a potential field. The trajectory optimization decision of the following ships is no longer to solve for their own shortest path, but to find potential energy depressions in a dynamic energy terrain jointly constructed by all neighboring ships. Therefore, the ship group no longer performs discrete collision avoidance actions, but instead performs smooth, holistic navigation manifold adjustments, spontaneously evolving towards the formation with the lowest overall energy consumption.

[0036] During this voyage, if a high-speed vessel not in the formation attempts to cut laterally, causing the trajectory optimization decision module of a cargo ship in the formation to generate an energy-saving correction control command aimed at rapid avoidance, this command is unconditionally intercepted and verified by the post-decision arbitration module before being output. The arbitration module, through a five-second linear extrapolation of all relevant vessels, determines that executing the energy-saving command would result in the minimum encounter distance between this ship and another adjacent vessel in the formation falling below a predetermined safe distance threshold of 0.1 nautical miles. Therefore, the execution of the energy-saving command is rejected, and the system instead outputs a command that ensures all safety and collision avoidance rules are met. Safety verification instructions; this non-bypass adjudication mechanism ensures that the goal of energy-saving optimization is always under the strict constraints of the safety framework, thereby achieving a balance between efficiency and safety; ultimately, the entire cargo fleet smoothly passed through the narrow channel with a compact and orderly dynamic navigation flow, with each ship's track being smooth, avoiding unnecessary energy consumption caused by chain corrections; under this operating state, the navigation decision of any ship becomes a comprehensive response to the group's energy topography and immediate environmental factors, thus transforming the navigation task from a passive response to a series of discrete conflicts into a continuous and self-organized overall optimization process.

[0037] Example 2: In the multi-agent navigation simulation platform test integrating ship hydrodynamic model and marine environmental factors, the calculation update cycle of the key parameter total virtual potential field is set to balance the real-time response of the system and the computational load of the device. This cycle is set to be dynamically related to the ship's speed over land, that is, the higher the ship's speed and the faster the situation changes, the shorter the calculation update cycle will be. At the typical speed of 15 knots in this test, this cycle is set to 1 second. The test includes two parallel simulations that run continuously for 10 hours in the same preset channel and 2 knot constant transverse ocean current environment. The 10 cargo ships in the control group use traditional autopilot and reactive collision avoidance algorithm, while the 10 cargo ships of the same type in the test group use the control device of this invention. Except for the control strategy, all other initial conditions and environmental parameters are kept consistent.

[0038] Table 1: Comparison of key performance indicators between the experimental group and the control group;

[0039]

[0040] Data recorded during the experiment showed that the control group vessels exhibited frequent and large-amplitude reciprocating motions in their rudder angles, with disorderly fluctuations in their spacing. In contrast, the experimental group of vessels displayed a smooth overall transition in their navigation attitude, with significantly reduced amplitude and frequency of rudder angle corrections, and a more stable and spacious spacing (see Table 1). Data also showed that the total navigation energy consumption of the experimental group was approximately 17.9% lower than that of the control group. This was attributed to the combined effect of the asymmetric virtual energy-saving potential field promoting a low-resistance dynamic navigation manifold and the pre-compensation of ocean current effects by the navigation micro-disturbance module. Simultaneously, the average minimum spacing between vessels in the experimental group increased, and the number of high-risk approach events was zero. This phenomenon demonstrates the effectiveness of the post-decision arbitration module as a safety baseline. The experimental results indicate that, under the same external conditions, the overall energy consumption of a vessel group using the control device of this invention is lower than that of a group employing a traditional control strategy, and no high-risk approach events occurred during navigation. Both the safety and economic indicators of the system were improved.

[0041] Example 3: This example combines Figures 1 to 3 This section describes a deep learning-based ship motion trajectory optimization control system, such as... Figure 1 As shown, the system first acquires key data through the input information section, including position, speed to ground, and heading to ground information from the Automatic Identification System (AIS) broadcast data, as well as the ship's speed, heading, and rudder angle information from the ship's own sensor data. The core processing module contains several functional sub-modules: First, the neighboring ship information receiving module decodes and processes the neighboring ship's AIS data stream; simultaneously, the virtual potential field generation module generates an asymmetric energy-saving potential field based on the ship's state; subsequently, the total potential field superposition module linearly superimposes the potential fields of all ships to form a total potential field, providing a basis for energy consumption scenarios for subsequent decisions; then, the trajectory optimization decision module calculates the potential field based on the total potential field. The negative gradient direction is used to determine the direction of the fastest energy consumption reduction. The navigation control output module generates a correction control command based on this direction. To improve the system's environmental perception, the navigation micro-disturbance module periodically applies micro-disturbances. Based on the navigation response results, ocean current vector inversion and environmental adaptive compensation are achieved. The compensation results are returned to the total potential field superposition module as ocean current compensation. In terms of system safety control, the post-decision arbitration module performs safety verification and rule checks on the correction command. It determines that the minimum encounter distance between the trajectory corresponding to the command and the trajectory of the adjacent vessel is ≥0.1 nautical miles, and that it does not violate collision avoidance rules, before it can pass through this module. Finally, the system output will execute the trajectory correction command through the autopilot system.

[0042] like Figure 2As shown in the figure, the horizontal axis represents time in hours, and the vertical axis represents cumulative energy consumption in kilowatt-hours. The solid dotted curve represents the experimental group's system, while the dashed triangular curve represents the control group's traditional system. As can be seen from the figure, as time increases in hours, the cumulative energy consumption in the control group is consistently higher than that in the experimental group, showing a steeper growth trend. In contrast, the cumulative energy consumption in the experimental group's system increases more moderately under the same operating time. This demonstrates that the system reduces the overall energy consumption of the ship group through mechanisms such as asymmetric energy-saving potential field, self-organized trajectory optimization, and environmental adaptive compensation.

[0043] like Figure 3 As shown, firstly, the navigation control output module sends an energy-saving correction control command to the post-decision arbitration module. This module then initiates a safety verification and decision arbitration process for the command, including: intercepting the command for verification, and providing the original navigation data of neighboring vessels from the original AIS data. Based on this, a 5-second linear extrapolation is performed on all neighboring vessels to calculate the predicted trajectory of the ship and the minimum encounter distance with neighboring vessels. If the safety distance detection result is less than 0.1 nautical miles, the system will call the international collision avoidance rule library to provide collision avoidance rule constraints and verify the compliance of the initial turn. If the command meets all safety checks, the energy-saving correction command is approved for execution and the command is output to the autopilot system to execute the optimized trajectory. If the safety check fails, the system outputs a safety verification command, which is received by the autopilot system to execute the safety priority trajectory.

[0044] Example 4: In this example, before being deployed on a specific vessel, the control device of the present invention can execute a set of offline calibration and online adaptive procedures to match its internal model and operating parameters with the hydrodynamic characteristics and expected navigation environment of the vessel, thereby avoiding a decrease in control performance due to parameter mismatch. During the offline calibration phase, to set the parameters of the Gaussian function model in the virtual potential field generation module, the system uses the vessel's maneuverability model for simulation. By changing the vessel's ground speed, the system records the intensity and range of the repulsive and gravitational potentials required to achieve optimal energy-saving effects at different speeds, and establishes a navigation... A mapping table between velocity and Gaussian function weights and variance parameters is embedded in the control device for online operation. Simultaneously, to construct the historical potential field evolution pattern database required for the communication status monitoring module, the system runs large-scale Monte Carlo simulations covering various ship encounter scenarios. Each scenario generates a scenario feature vector consisting of three dimensions: number of ships, average speed, and formation density, as well as time-series data of the total virtual potential field evolution generated by the system under that scenario. These key-value pairs, composed of the scenario feature vector and the potential field evolution time-series data, together form the pattern database.

[0045] Once the vessel enters a stable navigation state, the system initiates an online adaptive procedure. The communication status monitoring module statistically analyzes the timestamp intervals of the received Automatic Identification System (AIS) signals and calculates the information entropy of one thousand consecutive samples to obtain the baseline information entropy mean μ. S With standard deviation σ S Based on this, the information entropy threshold S is set. th Set as S th =μ S +4σ S When channel conditions deteriorate, causing the information entropy ΔS to exceed the threshold S... th When the system determines that it has entered a communication interruption state, it initiates the fusion prediction mechanism, and its fusion weight function w(t) is defined as a saturated linear function w(t) = min(t / T). predict ,1), where the time constant T predict The setting of 60 seconds ensures that the system relies entirely on the predictive virtual potential field P after a 60-second communication interruption. predict The P predict The generation of the data is achieved by extracting the scene feature vector before the interruption, retrieving the three most similar historical scenes from the historical potential field evolution pattern database using the k-nearest neighbor algorithm, and then taking a weighted average of the corresponding potential field evolution time series data based on the inverse distance in the feature space. This combination of offline calibration and online adaptive procedures ensures that the internal model and key parameters of the control device have been initialized and contextualized before it is put into use. When faced with a sudden communication interruption, the system can deterministically transition from a control closed loop based on real-time data to a predictive control closed loop based on historical statistical patterns, thereby maintaining the stability of the navigation trajectory in the absence of external information input.

[0046] Example 5: This example aims to demonstrate the internal decision-making priority and control switching logic of the control device of the present invention when facing complex encounter scenarios defined by international maritime collision avoidance rules and when manual commands are involved. When a ship A equipped with this system forms a head-on encounter with ship B, and another ship C forms a cross encounter from the starboard side of ship A, the system faces a complex decision-making scenario constrained by multiple collision avoidance rules. In this case, the energy-saving trajectory correction command calculated by the trajectory optimization decision module of ship A based on the total virtual potential field may be a slight left turn. Before the command is output, its built-in rule compliance review logic immediately identifies and determines that the head-on encounter is the highest priority collision avoidance event, and that the left turn command is contrary to international maritime collision avoidance rules. Accordingly, the post-decision arbitration module immediately rejects the energy-saving command and bypasses the trajectory optimization decision module, directly outputting a clear right turn safety verification command to the autopilot system. This mechanism ensures that the ship's actions comply with maritime traffic regulations, even if the action is not the optimal solution in terms of energy consumption.

[0047] If, during the aforementioned process, the crew of vessel C performs manual operations, and their steering behavior deviates continuously from the system-generated instructions, vessel C's control device, upon detecting this state, will not only increase the broadcast weight of its own asymmetric virtual energy-saving potential field but also broadcast a status code indicating priority for manual control through the Automatic Identification System (AIS). Upon receiving this status code, the control devices of neighboring vessels A and B will adjust the operating logic of their total potential field superposition modules, no longer treating vessel C as a potential field source that can be collaboratively optimized, but reclassifying it as a dynamic physical obstacle with the highest collision avoidance priority, and generating a more reinforced virtual repulsive force region around its linear extrapolated trajectory. Therefore, the subsequent trajectory optimization of vessels A and B will be carried out in a potential field environment that has taken vessel C's manual navigation intentions as a hard constraint, thereby achieving a high-priority response to human driving decisions at the system level.

[0048] Example 6: Before deploying the control device on a ship of a specific class, a standardized offline parameter calibration and logic verification procedure must be performed to ensure that the system's operating parameters are precisely matched with the hydrodynamic characteristics and expected operating environment of the ship of that class. This procedure is performed in a certified simulation environment that can reproduce the handling characteristics of the ship of that class. To determine the preset distance between the center of the Gaussian function model in the virtual potential field generation module and the amplitude and frequency of the standardized heading micro-disturbance command applied by the navigation micro-disturbance module, the simulation system runs a multi-objective optimization program. This program aims to minimize a weighted cost function J = w1·E + w2·Δv + w3·δ, which includes total energy consumption, track deviation, and maximum rudder angle. max With the goal of [missing information], iterative operations are performed within a preset parameter space to ultimately output a Pareto optimal parameter set that allows the cost function to achieve convergence under various simulated sea conditions. This parameter set is then fixed as a dedicated configuration for ships of this class. For the heading clustering threshold in the data preprocessing module of the Automatic Identification System (AIS), its calibration is achieved by analyzing historical navigation data and setting the threshold to a specific cosine value. Under this cosine value, the expected value of the lateral separation of the tracks of clustered ships based on 15-minute linear extrapolation does not exceed one standard beam of the ship of this class.

[0049] To verify the effectiveness of the rule compliance review logic in the post-decision arbitration module, the regulations pre-compile key clauses of the International Maritime Collision Avoidance Code (IMC) into a deterministic rule state machine and load it into the simulation environment. This state machine determines the encounter situation in real time based on changes in the relative bearing and distance between the vessel and other vessels, such as head-on, crossing, or overtaking, and pre-sets a set of compliant permissible maneuvers for each situation. During the simulation, any corrective control command generated by the trajectory optimization decision module or that does not belong to the set of permissible maneuvers under the current situation is recorded as a successful logical interception. The logic verification of this module is completed by traversing all typical encounter scenarios and confirming the accuracy of the interceptions. Furthermore, to enhance the robustness of the local ocean current vectors retrieved by the navigation micro-disturbance module, the system... An additional verification step was added: before the vector is incorporated into the calculation of the total virtual potential field, it must pass through a low-pass filter to remove high-frequency noise, and its vector magnitude must not exceed a maximum reasonable flow velocity limit set based on historical hydrological data of the ship's current navigation area. Any calculation result exceeding this limit is considered an anomaly of the sensor or algorithm and is discarded to avoid the system overreacting to false extreme environmental disturbances. Through this standardized procedure, the internal parameters and logical judgment benchmarks of the control device are supported and verified by engineering data based on specific ship types, thereby transforming a general control scheme into a fully calibrated and hardened deployable system for a specific ship. Meanwhile, in this embodiment, the cost function J = w1·E + w2·Δv + w3·δ max The determination of the weighting coefficients and the parameterization procedure for micro-perturbation commands, specifically the offline calibration phase, involves the following technical steps: The weighting coefficients (w1, w2, w3) are assigned values ​​using the Analytic Hierarchy Process (AHP), a mathematical process for structured decision-making. During this process, a parameterization algorithm is constructed by shipping experts regarding the target E, Δv, and δ... max For each pairwise comparison judgment matrix, the calculated consistency ratio CR value must be less than 0.1. The normalized eigenvector obtained after passing this test is the weight coefficient of that group. Simultaneously, the standardized heading micro-disturbance command's signal shape is determined to be a pseudo-random binary sequence, i.e., PRBS, a deterministic signal with wide-spectrum characteristics, whose amplitude A... p The simulation was set to generate a minimum signal-to-noise ratio exceeding 10 dB in the ship's response, with the amplitude causing a lateral displacement of the ship not exceeding the width of a standard beam, and its symbol switching period Δt. p It is then set within the range of 0.5 to 2.0 times the ship's maneuverability time constant T, i.e., 0.5T ≤ Δt. p ≤2T, where T is the inherent maneuverability index calibrated offline for this ship type.

[0050] Furthermore, the execution logic of trajectory prediction in the post-decision arbitration module, and the initial construction of the historical potential field evolution pattern database, are implemented in the following details: The trajectory prediction adopts an adaptive model switching mechanism. By default, it uses a linear extrapolation model with low computational resource consumption. However, when it detects that any neighboring vessel's turning rate (ROT), a physical quantity representing the vessel's angular velocity, broadcast by the Automatic Identification System (AIS), exceeds a predetermined threshold of 3 degrees per minute, the prediction model automatically switches to a constant turning rate model for that vessel. This model solves the trajectory based on the current ground speed and the ROT value. It is considered as an arc; to support the predictive function during communication interruption, the historical database is filled in the early stage of device deployment through a cloud-ground collaborative guidance mode. That is, in the first 100 voyages of the device in the new operating waters, all eligible non-conflict encounter scenario data are uploaded to the local database, and an anonymized basic pattern library matching the ship type and the current navigation area is subscribed from a cloud service platform for the fleet. This method ensures that the device has the predictive ability based on extensive statistical laws before accumulating sufficient local operating data. All of these are extended implementation methods known to those skilled in the art.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based ship motion trajectory optimization control system, characterized in that, This includes control devices deployed on each vessel, configured to perform the following functions: The neighboring vessel information receiving module is configured to continuously receive and decode navigation status information broadcast by neighboring vessels through the Automatic Identification System (AIS). The navigation status information includes the position, speed above ground, and heading above ground of the neighboring vessels. The virtual potential field generation module is configured to generate an asymmetric virtual energy-saving potential field based on the ship's current navigation status information, acquired through the ship's sensors. This information includes the ship's speed and heading relative to the ground. The asymmetric virtual energy-saving potential field generates a repulsive potential in the region in front of the ship's heading relative to the ground, and the strength of the repulsive potential is positively correlated with the ship's speed relative to the ground. The asymmetric virtual energy-saving potential field generates a gravitational potential in the region behind the ship's heading relative to the ground, and the strength of the gravitational potential is positively correlated with the ship's speed relative to the ground. The total potential field superposition module is configured to linearly superimpose the virtual energy-saving potential fields of all neighboring ships received and decoded by the neighboring ship information receiving module to form the total virtual potential field where the ship is currently located. The trajectory optimization decision module is configured to determine the negative gradient direction of the potential field that guides the ship's motion direction based on the total virtual potential field. The negative gradient direction of the potential field points to the direction in which the energy consumption of the total virtual potential field decreases the fastest. The navigation control output module is configured to generate and output correction control commands to the ship's automatic steering system based on the negative gradient direction of the potential field, so as to drive the ship's trajectory to be adjusted in a controlled manner towards the negative gradient direction of the potential field, enabling the ship group to form a dynamic navigation manifold with reduced overall energy consumption. The navigation micro-disturbance module is configured to periodically apply standardized course micro-disturbance commands to the ship's autopilot system. Based on the vector difference between the ship's actual navigation response caused by the micro-disturbance command and the still water response expected by the ship's maneuverability model, it inversely calculates the local ocean current vector acting on the ship and incorporates the local ocean current vector as a background correction term into the calculation of the total virtual potential field to compensate for the influence of ocean currents.

2. The deep learning-based ship motion trajectory optimization control system according to claim 1, characterized in that, The control device also includes a ship automatic identification system data preprocessing module, which is configured to: receive raw ship automatic identification system data, and based on the cosine value of the difference in heading angles between ships, dynamically cluster multiple ships whose cosine value of the difference in heading angles is greater than a predetermined heading clustering threshold into a navigation group, and calculate the centroid position, average ground speed and average ground heading for the navigation group, generating a single equivalent navigation state information for the total potential field superposition module.

3. The deep learning-based ship motion trajectory optimization control system according to claim 1, characterized in that, The virtual potential field generation module is configured to use at least two Gaussian function models with different weights to form an asymmetric virtual energy-saving potential field. The center of one Gaussian model is located at a predetermined distance ahead of the ship's heading towards the ground, and the center of the other Gaussian model is located at a predetermined distance behind the ship's heading towards the ground.

4. The deep learning-based ship motion trajectory optimization control system according to claim 1, characterized in that, The control device also includes a post-decision arbitration module, which is configured to: receive the corrected control command output by the navigation control output module, and perform a five-second linear extrapolation based on the original Automatic Identification System (AIS) data received by the neighboring vessel information receiving module to generate a safety verification command; when the minimum encounter distance between the predicted trajectory corresponding to the corrected control command and the linear extrapolation trajectory of any neighboring vessel is not less than a predetermined safety distance threshold, and the initial turn of the corrected control command does not violate the key rules of the International Maritime Collision Avoidance Regulations, the corrected control command is approved for execution; otherwise, a safety verification command is output as the final control command.

5. A deep learning-based ship motion trajectory optimization control system according to claim 4, characterized in that, The predetermined safe distance threshold for the post-decision arbitration module is 0.1 nautical miles.

6. The deep learning-based ship motion trajectory optimization control system according to claim 1, characterized in that, The control device is also configured to increase the weight of the asymmetric virtual energy-saving potential field generated by the ship when it receives a signal that the ship's track deviates from the negative gradient direction of the potential field due to manual operation, so as to broadcast stronger heading intention information to neighboring ships.

7. The deep learning-based ship motion trajectory optimization control system according to claim 1, characterized in that, The navigation control output module is configured to determine the correction magnitude of the correction control command based on the magnitude of the potential field gradient indicated by the negative gradient direction, so that the correction magnitude is proportional to the magnitude of the potential field gradient. This allows for smaller heading adjustments in flat areas of the total virtual potential field to maintain heading stability, and larger adjustments in steep areas of the total virtual potential field.

8. The deep learning-based ship motion trajectory optimization control system according to claim 1, characterized in that, The control device also includes a communication status monitoring module, which is configured to monitor the stability of the communication channel by calculating the information entropy of the timestamp interval of the received Automatic Identification System (AIS) signals. When the information entropy ΔS increases sharply within a preset time period and exceeds the predetermined information entropy threshold S th When a communication interruption is detected, the system notifies the total potential field overlay module to match the most similar historical pattern to the current navigation state from a pre-stored database containing historical potential field evolution patterns. Based on the matched historical pattern, a predictive virtual potential field P is generated. predict The predictive virtual potential field is compared with the effective potential field P at the last moment before the interruption. lastknown A weighted fusion is performed as part of the total virtual potential field used for superposition in step c, wherein the weighted fusion potential field P fusion P is determined by the following relationship: fusion =w(t)·P predict +(1-w(t))·P lastknown , where w(t) is a weight function that grows in a controlled manner from zero to one as the communication interruption time t increases.

Images