Intelligent ship general collision avoidance decision-making method and system based on ship-crossing knowledge migration

Abstract

The invention provides an intelligent ship universal collision avoidance decision-making method and system based on cross-ship knowledge migration. The method comprises the steps of data acquisition and preprocessing, double-branch feature extraction, candidate strategy retrieval and similarity coefficient calculation, dynamic decision generation, simulation verification and online calibration, instruction execution and circulation and the like. A dual-branch depth feature extraction network is used for decoupling general navigation situation features irrelevant to a ship type and manipulation features strongly relevant to the ship type, a target ship dynamic predictor is used for calculating a dynamic transition probability similarity coefficient, and a candidate strategy is weighted and fused or a safety control instruction is generated based on the coefficient. Decision risk assessment and dynamic predictor online calibration are realized by means of a high-fidelity ship dynamics model and parallel Monte Carlo simulation. According to the method, a single-ship customization mode is broken, cross-ship safe reuse of the collision avoidance strategy is realized, the development and deployment cost of an intelligent navigation system is reduced, and the navigation safety, the operation efficiency and the sustainable evolution capability of a fleet are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent navigation and autonomous ships, and more particularly, relates to an intelligent ship general collision avoidance decision method and system based on cross-ship type knowledge transfer. BACKGROUND

[0002] Currently, the research and development of intelligent navigation systems mostly adopts a "single-ship customization" mode, that is, the perception, decision-making and control modules are tailor-made for specific ships (such as specific tonnage container ships, bulk carriers, etc.). This mode has the following inherent defects: High research and development cost and long cycle: a large amount of simulation, testing and optimization work needs to be done from scratch for each new ship or new model, and existing achievements cannot be reused.

[0003] Low data utilization efficiency: valuable navigation data and experience accumulated by ship A cannot be directly used for the development of intelligent systems for ship B, resulting in knowledge barriers and data islands.

[0004] Poor generalization ability: collision avoidance strategies trained for specific ship types are difficult to adapt to other ship types with different dynamics due to fixed maneuvering characteristics (such as inertia, turning, stopping performance, etc.), which poses a serious safety hazard.

[0005] Strategy library solidification: existing system strategy libraries are usually based on fixed rules or specific scene training and cannot dynamically adapt to the time-varying characteristics of the ship's own state and complex environment.

[0006] In the prior art, although some research attempts to achieve a certain degree of universality through parameter adjustment or adaptive control, most of them only make superficial adjustments and fail to fundamentally solve the problem of decision-making transfer caused by differences in core maneuvering characteristics of ships. Therefore, there is an urgent need for a general intelligent navigation core that can "draw an analogy from one thing to another" to reduce development costs and improve the safety and efficiency of the entire intelligent navigation fleet. SUMMARY

[0007] The purpose of the present application is to provide an intelligent ship general collision avoidance decision method and system based on cross-ship type knowledge transfer to solve the defects of the single-ship customization mode in existing collision avoidance decision methods.

[0008] To achieve the above purpose, in a first aspect, the present application provides an intelligent ship general collision avoidance decision method based on cross-ship type knowledge transfer, comprising: S1, collecting ship real-time state data, ship static parameters and environment perception data, processing the environment perception data into standardized environment grids, and combining the ship real-time state data and the ship static parameters into a state vector; S2, a scene branch of the dual-branch deep feature extraction network extracts a general navigation situation feature vector irrelevant to the ship type, and a maneuver branch of the dual-branch deep feature extraction network extracts a maneuver feature embedding vector strongly related to the ship type; S3, based on the general navigation situation feature vector, a set number of most relevant candidate meta-strategies are retrieved from a meta-strategy network library, and a dynamic transition probability similarity coefficient of an action corresponding to each candidate meta-strategy is calculated by a target ship dynamic predictor; the meta-strategy network library stores pre-trained multiple basic collision avoidance behavior meta-strategies; the target ship dynamic predictor is a prediction model representing the state transition law of the target ship; S4: if the maximum dynamic transition probability similarity coefficient is greater than or equal to a set threshold, a weight is calculated by using a softmax function, and an action control instruction is generated by weighted fusion of the actions of the candidate meta-strategies; if the maximum dynamic transition probability similarity coefficient is less than the set threshold, a safety control instruction is generated; S5: a high-fidelity ship dynamics model is constructed, and parallel Monte Carlo simulation is performed to evaluate the decision risk, and if the simulation result deviates from the prediction by more than a threshold, the target ship dynamic predictor is calibrated online; S6: the control instruction is sent to an executing mechanism, and the next decision cycle is entered, and S1-S6 are executed cyclically.

[0009] Optionally, in S2, the scene branch of the dual-branch deep feature extraction network adopts a convolutional neural network structure, the input is 128x128x6-dimensional environment grid data, and a 64-dimensional general navigation situation feature vector is output by sequentially performing twice convolution-pooling operations and twice fully connected operations; the maneuver branch adopts a fully connected network structure, the input is a 20-dimensional ship real-time state vector and a ship static parameter, and a 32-dimensional maneuver feature embedding vector is output by sequentially performing twice fully connected operations and once embedding layer processing.

[0010] Optionally, in S2, the training loss function of the dual-branch deep feature extraction network is L_total=L_scene+λ'L_maneuver, wherein L_scene=‖F_s-F_s_gt‖² is a scene branch loss function, F_s is a general navigation situation feature vector, F_s_gt is a real scene feature vector labeled by an expert, L_maneuver=-logP(Δs|F_m) is a maneuver branch maximum likelihood loss function, P(Δs|F_m) is a conditional probability that a state change Δs occurs given a maneuver feature embedding vector F_m, and λ' is a maneuver branch loss weight coefficient.

[0011] Optionally, in S3, the meta-strategy network library comprises 8 basic collision avoidance behavior meta-strategies, including right giving way in crossing situation, left giving way in overtaking situation, right turning in meeting situation, emergency braking, COLREGs following, collision avoidance in narrow waterway, multi-ship encounter coordination and detour in dangerous area; each meta-strategy network input is a spliced vector of a general navigation situation feature vector and a steering feature embedding vector, and the output is a 2-dimensional action vector; and the set number is 3.

[0012] Optionally, in S3, the target ship dynamic predictor calculates the dynamic transition probability similarity coefficient through the formula κ(s, a) = exp(-‖Δs_T-Δs_S‖² / σ²), wherein σ is a similarity scaling parameter, s is a multi-dimensional vector of the real-time sailing state of the target ship, a is a 2-dimensional control action vector output by the candidate meta-strategy, Δs_T is a state transition prediction value of the target ship after executing the action a output by the target ship dynamic predictor, and Δs_S is a state transition prediction value of the source ship.

[0013] Optionally, in S4, the set threshold is 0.7; and the formula for calculating the weight of the softmax function is: w_i=exp(κ_i / 0.7) / ∑exp(κ_j / 0.7), wherein κ_i is the similarity coefficient of the i-th candidate strategy, w_i is the weight of the i-th candidate strategy, and the final action control instruction is a=∑_{i=1}^3w_i a_i, wherein a_i is the action vector of the i-th candidate strategy.

[0014] Optionally, in S4, the safety control instruction is generated by using a model predictive control framework, and the optimization target is minJ=∑_{k=0}^{19}[‖s_k-s_ref‖_Q²+‖a_k‖_R²]+‖s_H-s_ref‖_P², wherein s_k is a state vector at the k-th step in a prediction time domain, s_ref is a reference state vector, a_k is a control action vector at the k-th step in the prediction time domain, s_H is a state at the end of the prediction time domain, Q is a weight matrix of state tracking error, R is a weight matrix of control quantity, and P is a weight matrix of terminal state. The constraint conditions include state safety constraints and action feasibility constraints, the prediction time domain is 20 steps, and the control time domain is 5 steps.

[0015] Optionally, in S5, the high-fidelity ship dynamics model comprises a longitudinal motion equation M( -rv)=X, a lateral motion equation M( +ur)=N, and a turning motion equation Iz =N, wherein M is the mass of the ship, is the longitudinal acceleration, r is the turning angular velocity, v is the lateral velocity, X is the longitudinal resultant force, for turning angular acceleration, u is longitudinal velocity, N is turning moment of force, Iz is moment of inertia about vertical axis; The parallel Monte Carlo simulation sets 50 parallel simulation trajectories, and evaluates the collision risk through a formula R_collision = ∑_{m=1}^50I(CPA_m < D_safe) / 50, wherein D_safe = max(3L, 1000m), L is the length of the ship, CPA_m is the minimum encounter distance of the mth trajectory, I( ) is an indicator function.

[0016] Optionally, in S5, the triggering condition of the online calibration is ‖Δs_sim-Δs_pred‖²>0.05, wherein Δs_sim is the actual state change of the target ship obtained through simulation, and Δs_pred is the predicted value of the state change output by the target ship dynamic predictor; The loss function of the online calibration is L_calib = ‖Δs_sim-P_T(s,a;θ_T)‖²+λ‖θ_T-θ_T_old‖², wherein P_T(s,a;θ_T) is the state transition prediction value output by the target ship dynamic predictor based on the current state s, action a and parameter θ_T, λ is a regularization coefficient, θ_T is the parameter set of the target ship dynamic predictor, and θ_T_old is the parameter set of the target ship dynamic predictor in the last decision period.

[0017] In a second aspect, the present application provides an intelligent general collision avoidance decision system of a ship based on cross-ship type knowledge transfer, which is used to implement the method of the first aspect, and comprises: a perception input layer, configured to collect real-time state data of the ship, static parameters of the ship and environmental perception data, process the environmental perception data into standardized environmental grids, and combine the real-time state data of the ship and the static parameters of the ship into a state vector; a core processing layer, comprising a double-branch deep feature extraction network, a dynamic transition probability similarity coefficient calculation module and a meta-strategy network library; the double-branch deep feature extraction network is configured to extract a general navigation situation feature vector irrelevant to the ship type by using a scene branch and extract a maneuvering feature embedding vector strongly related to the ship type by using a maneuvering branch; the meta-strategy network library is configured to store a plurality of pre-trained basic collision avoidance behavior meta-strategies, and retrieve a set number of most relevant candidate meta-strategies based on the general navigation situation feature vector; and the dynamic transition probability similarity coefficient calculation module is configured to calculate a dynamic transition probability similarity coefficient of an action corresponding to each candidate meta-strategy by using a target ship dynamic predictor. The decision execution layer includes a strategy migration and dynamic fusion algorithm module and a safety rolling optimization module. The strategy migration and dynamic fusion algorithm module is configured to calculate weights and generate action control instructions by using the softmax function and weighted fusion of candidate meta-policies if the largest dynamic transition probability similarity coefficient is greater than or equal to a set threshold. The safety rolling optimization module is configured to generate safety control instructions if the largest dynamic transition probability similarity coefficient is less than a set threshold. The validation evolution layer includes a digital twin framework, a parallel Monte Carlo simulation module, and an online calibration module. The digital twin framework is configured to construct a high-fidelity ship dynamics model. The parallel Monte Carlo simulation module is configured to perform parallel Monte Carlo simulations to assess decision-making risks. The online calibration module is configured to perform online calibration of the target ship dynamics predictor if the simulation results deviate significantly from the prediction.

[0018] The beneficial effects of this invention are that it provides a general intelligent ship collision avoidance decision-making method based on cross-ship type knowledge transfer. This method decouples the ship collision avoidance decision-making process into two independent dimensions: situational awareness and maneuvering decision-making. A dual-branch deep feature extraction network extracts general navigation situational features independent of ship type and maneuvering features strongly correlated with ship type, respectively. A target ship dynamic predictor is introduced to quantify the transfer safety of the source ship's strategy in the target ship's current scenario (dynamic transfer probability similarity coefficient). Based on this coefficient, a strategy fusion or safety fallback scheme is dynamically selected. A high-fidelity ship dynamics model and parallel Monte Carlo simulation are used to achieve decision risk assessment and online calibration of the dynamic predictor. This ultimately forms a closed-loop decision-making process of data acquisition, feature extraction, strategy transfer, decision generation, simulation verification, and online optimization. Compared with existing technologies, this method has the following significant advantages: Breaking away from the customized paradigm of "one ship, one solution," the new ship's intelligent system can quickly acquire core collision avoidance capabilities through transfer learning, saving more than the cost of algorithm debugging and verification, and significantly reducing development costs and time.

[0019] The dynamic transition probability similarity coefficient can effectively warn of unsafe policy migrations. Combined with online simulation verification, it enables the system to have a clear understanding of its own decision boundaries, avoids forced decisions in unsuitable scenarios, and improves decision security and environmental adaptability.

[0020] A novel and efficient collision avoidance strategy learned by one ship can be safely transferred to other ship types in the fleet through this system, achieving "experience" download in the air, improving the intelligence level of the entire fleet, and realizing fleet knowledge sharing.

[0021] By combining digital twins and online learning mechanisms, the system can continuously optimize itself during use, adapt to slow changes in ship performance (such as hull fouling) and new navigation environments, and achieve continuous evolution.

[0022] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0023] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the invention.

[0024] Figure 1 The flowchart of the intelligent ship general collision avoidance decision method based on cross-ship type knowledge transfer according to Embodiment 1 of the present invention is shown.

[0025] Figure 2 The diagram shows the architecture of the intelligent ship general collision avoidance decision system based on cross-ship type knowledge transfer according to Embodiment 2 of the present invention.

[0026] Figure 3 A schematic diagram of the structure of the dual-branch deep feature extraction network of Embodiment 2 of the present invention is shown.

[0027] Figure 4 The flowchart of strategy migration and dynamic fusion based on dynamic transition probability similarity coefficient is shown in Embodiment 2 of the present invention.

[0028] Figure 5 A closed-loop calibration principle diagram of the digital twin framework of Embodiment 2 of the present invention is shown. Detailed Implementation

[0029] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0030] Example 1

[0031] See Figure 1 This embodiment provides a general collision avoidance decision-making method for intelligent ships based on cross-ship type knowledge transfer, with the target ship (current ship) as the executing entity, including: S1 Data Acquisition and Preprocessing: Acquire real-time ship status data, ship static parameters, and environmental perception data; process the environmental perception data into standardized environmental grids; combine the real-time ship status data and ship static parameters into a state vector. In this embodiment, the environmental perception data includes radar point cloud data, AIS message data, and visual sensor data. The real-time ship status data includes the ship's real-time load condition, draft, main engine speed, and rudder angle. The ship's static parameters include length L, beam B, design speed V_d, and block coefficient C_b, where the block coefficient C_b is a dimensionless parameter characterizing the fullness of the hull, calculated as displacement volume / (L × B × draft). The environmental perception data is processed into a 128 × 128 × 6-dimensional standardized environmental raster (a multi-dimensional matrix, where each pixel contains encoded data such as obstacle position, velocity vector, heading information, and target type). The real-time ship status data and the ship's static parameters are combined into a 20-dimensional state vector, providing standardized input for subsequent feature extraction.

[0032] S2. Use the scene branch of the dual-branch deep feature extraction network to extract a general navigation situation feature vector that is independent of the ship type, and use the maneuvering branch of the dual-branch deep feature extraction network to extract a maneuvering feature embedding vector that is strongly related to the ship type. In this step, the scene branch of the dual-branch deep feature extraction network adopts a convolutional neural network structure, with input being 128×128×6 dimensional environmental grid data. It then undergoes two convolution-pooling operations and two fully connected operations to output a 64-dimensional general navigation situation feature vector. The manipulation branch adopts a fully connected network structure, with input being a 20-dimensional real-time ship state vector and ship static parameters. It then undergoes two fully connected operations and one embedding layer to output a 32-dimensional manipulation feature embedding vector. The training loss function of the dual-branch deep feature extraction network is L_total = L_scene + λ'L_maneuver, where L_scene = ||F_s - F_s_gt||² is the scene branch loss function, F_s is the general navigation situation feature vector, and F_s_gt is the expert-annotated real scene feature vector, used to constrain the extraction accuracy of the general navigation situation features. L_maneuver=-logP(Δs|F_m) is the maximum likelihood loss function for the maneuver branch. P(Δs|F_m) is the conditional probability of the state change Δs occurring given the maneuver feature embedding vector F_m. Δs includes six components: x-direction displacement, y-direction displacement, heading angle ψ, longitudinal velocity u, lateral velocity v, and bow angular velocity r, used to ensure a strong correlation between maneuver features and ship state changes. λ' is the loss weight coefficient for the maneuver branch, with a value range of [0.5, 2.0], used to balance the training priority of the scenario branch and the maneuver branch.

[0033] In this embodiment, the scene branch adopts a convolutional neural network (CNN) structure, and the input is the 128×128×6 dimensional environmental raster data processed in step 1. The network sequentially passes through a convolutional layer 1 with 64 3×3 convolutional kernels (stride 1, ReLU activation function), a 2×2 max pooling layer 1 (stride 2), a convolutional layer 2 with 128 3×3 convolutional kernels (stride 1, ReLU activation function), a 2×2 max pooling layer 2 (stride 2), a fully connected layer 1 with 256 neurons (ReLU activation function), and a fully connected layer 2 with 128 neurons (ReLU activation function), finally outputting a 64-dimensional general navigation situation feature vector F_s (∈R^64). This vector encodes the scene's geometric relationships and dynamic core information (such as the relative positions of other ships, CPA / TCPA, channel boundaries, obstacle distribution, etc.), and is independent of ship type. F_s extracted from different ship types in the same scene represents the same situation. The maneuvering branch employs a fully connected network (FCN) structure, with inputs consisting of a 20-dimensional real-time ship state vector (combined from step 1) and ship static parameters. The network sequentially passes through a fully connected layer 1 with 128 neurons (ReLU activation function), a fully connected layer 2 with 64 neurons (ReLU activation function), and an embedding layer with 32 neurons (Tanh activation function), ultimately outputting a 32-dimensional maneuvering feature embedding vector F_m (∈R^32). This vector serves as the ship's "maneuvering fingerprint," implicitly encoding core ship-related characteristics such as turning performance, inertial properties, and maneuvering response speed.

[0034] S3. Based on the general navigation situation feature vector, retrieve a set number of the most relevant candidate meta-policies from the meta-policy network library, and calculate the dynamic transition probability similarity coefficient of the action corresponding to each candidate meta-policy through the target ship dynamic predictor; the meta-policy network library stores a variety of pre-trained basic collision avoidance behavior meta-policies; the target ship dynamic predictor is a prediction model that characterizes the state transition law of the target ship. In this step, the meta-policy network library contains eight basic collision avoidance behavior meta-policies, including right yield in crossover situations, left yield in overtaking situations, right turn in encounter situations, emergency braking, COLREGs following, collision avoidance in narrow waterways, multi-ship encounter coordination, and danger zone detour. Each meta-policy network takes as input a concatenated vector of a general navigation situation feature vector and a manipulation feature embedding vector, and outputs a 2D action vector; the number of such vectors is set to three. The target ship dynamic predictor calculates the dynamic transition probability similarity coefficient using the formula κ(s,a)=exp(-‖Δs_T-Δs_S‖² / σ²), where σ is the similarity scaling parameter, s is the multi-dimensional vector of the target ship's real-time navigation state, a is the 2D control action vector output by the candidate meta-policy, Δs_T is the predicted state transition value of the target ship after executing action a, and Δs_S is the predicted state transition value of the source ship.

[0035] In this embodiment, during candidate policy retrieval, based on the general navigation situation feature vector F_s, the cosine similarity d_i = cos_sim(F_s, F_s^i) is calculated, where F_s^i is the scene feature vector corresponding to each pre-trained policy in the meta-policy network library. Then, three candidate meta-policies most relevant to the current scene are retrieved from the meta-policy network library. , , The meta-policy network library stores eight pre-trained basic collision avoidance meta-policies: right yield in crossing situations, left yield in overtaking situations, right turn in encounter situations, emergency braking, COLREGs following, collision avoidance in narrow waterways, multi-ship encounter coordination, and danger zone detour, covering common navigation situations and emergency scenarios. The input to each meta-policy network is a 96-dimensional vector (∈R^(64+32)) formed by concatenating the general navigation situation feature vector F_s and the manipulation feature embedding vector F_m, and the output is a 2-dimensional control action vector a (∈R^2), including the rudder angle δ (range ±35°) and thrust T (range 0~100%). In the calculation of the dynamic transfer probability similarity coefficient, for each candidate meta-policy π_i, its output action vector a_i=π_i(F_s,F_m) is obtained. The migration safety of this action on the target ship (current ship) is calculated through the target ship dynamic predictor, i.e., the dynamic transfer probability similarity coefficient κ_i=κ(s,a_i). The target ship dynamic predictor input includes a 10-dimensional real-time navigation state vector s (∈R^10, specifically composed of: x position, y position, heading angle ψ, longitudinal velocity u, lateral velocity v, turning angular velocity r, wind speed, wind direction, current speed, and current direction), a 2-dimensional action vector a_i of the candidate meta-strategy, and a target ship maneuvering feature embedding vector F_m. The output is a 6-dimensional state transition prediction value Δs_T (∈R^6, including Δx longitudinal displacement, Δy lateral displacement, Δψ heading angle change, Δu longitudinal velocity change, Δv lateral velocity change, and Δr turning angular velocity change) after the target ship executes action a_i. The dynamic transition probability similarity coefficient is calculated using the following formula: κ(s,a)=exp(-‖Δs_T-Δs_S‖² / σ²), where Δs_S is the predicted state transition value after the source ship performs the same action a (pre-calculated based on the source ship's dynamic predictor and stored in the meta-policy network library); σ=0.1 is the similarity scaling parameter, which controls the sensitivity of the coefficient to state transition differences; κ(s,a) ranges from 0 to 1, the closer to 1, the higher the safety of the source ship's policy migration to the target ship, and the closer to 0, the greater the migration risk.

[0036] S4: If the maximum dynamic transition probability similarity coefficient is greater than or equal to the set threshold, the softmax function is used to calculate the weights and weighted fuse the candidate meta strategies to generate action control instructions; if the maximum dynamic transition probability similarity coefficient is less than the set threshold, safety control instructions are generated. In this step, the threshold is set to 0.7; the formula for calculating the weights using the softmax function is: w_i=exp(κ_i / 0.7) / ∑exp(κ_j / 0.7), where κ_i is the similarity coefficient of the i-th candidate policy, w_i is the weight of the i-th candidate policy, and the final action control command is a=∑_{i=1}^3w_i a_i, where a_i is the action vector of the i-th candidate policy. A model predictive control framework is used to generate safe control commands. The optimization objective is minJ=∑_{k=0}^{19}[‖s_k-s_ref‖_Q²+‖a_k‖_R²]+‖s_H-s_ref‖_P², where s_k is the state vector at the k-th step in the prediction time domain, s_ref is the reference state vector, a_k is the control action vector at the k-th step in the prediction time domain, s_H is the terminal state in the prediction time domain, Q is the weight matrix of the state tracking error, R is the weight matrix of the control quantity, and P is the weight matrix of the terminal state. Constraints include state safety constraints and action feasibility constraints. The prediction time domain has 20 steps, and the control time domain has 5 steps.

[0037] In this embodiment, the similarity coefficients corresponding to the three candidate strategies calculated based on S3 are { , , The maximum value max(κ_i) is extracted and compared with a set threshold of 0.7 to dynamically select the decision generation method. When there are safe and transferable candidate policies, policy weighted fusion is performed (max(κ_i)≥0.7). The softmax function is used to calculate the weight of each candidate policy, with the formula: w_i=exp(κ_i / 0.7) / ∑exp(κ_j / 0.7) (j=1,2,3). The action vectors of each candidate policy are fused by weighted fusion to generate the final action control command: a=∑_{i=1}^3w_i a_i, where a_i is the action vector of the i-th candidate policy and w_i is the weight of the i-th candidate policy, ensures that the decision takes into account both safety and adaptability. When the migration risk of all candidate policies exceeds the limit, a safety control instruction is generated (max(κ_i) < 0.7), triggering the safety rolling optimization module to generate a safety control instruction based on the model predictive control (MPC) framework. The optimization objective of the model predictive control framework is: min J = ∑_{k=0}^{19}[‖s_k-s_ref‖_Q²+‖a_k‖_R²]+‖s_20-s_ref‖_P², where s_k is the state vector at the k-th step in the prediction time domain, s_ref is the reference state vector output by the global path planner (i.e., the desired navigation state), s_20 is the state vector at the 20th step at the end of the prediction time domain, a_k is the control action vector at the k-th step in the prediction time domain (consistent with the action vector structure output by the meta-policy, including rudder angle δ and thrust T), Q is the state tracking error weight matrix, marking the tracking priority coefficients of each state parameter, R is the control quantity weight matrix, used to suppress abrupt changes in control actions, and P is the terminal state weight matrix, used to enhance the consistency between the terminal state and the reference state. The prediction time domain is 20 steps, covering the navigation state for the next 20 seconds, and the control time domain is 5 steps, optimizing only the first 5 control actions, keeping the 5th step action unchanged for the subsequent 15 steps, balancing optimization accuracy and computational efficiency. The constraints include state prediction constraints, state safety constraints, and action feasibility constraints. The state prediction constraint is s_{k+1}=f_T(s_k,a_k), where f_T is the target ship's dynamics model, used to predict the state at the next moment based on the current state and actions; the state safety constraint is s_k∈S_safe, ensuring that the ship does not collide, run aground, or cross boundaries during navigation; the action feasibility constraint is a_k∈A_feasible, which meets the physical limits of the ship's actuators (such as rudder angle ±35°, thrust 0~100%).

[0038] S5: Construct a high-fidelity ship dynamics model and perform parallel Monte Carlo simulation to assess decision-making risks. If the simulation results deviate from the predictions, the target ship dynamics predictor is calibrated online. In this step, a high-fidelity ship dynamics model is used, including the longitudinal motion equation M( -rv)=X, Equation of lateral motion M( +ur)=N and the equation of motion for head turning Iz =N, where M is the ship's mass. Let r be the longitudinal acceleration, r be the angular velocity of the turn, v be the lateral velocity, and X be the resultant longitudinal force. is the yaw angular acceleration, u is the longitudinal velocity, N is the yaw moment, and Iz is the moment of inertia about the vertical axis; 50 parallel simulation trajectories are set in the parallel Monte Carlo simulation. The collision risk is evaluated by the formula R_collision = ∑_{m = 1}^50I(CPA_m < D_safe) / 50, where D_safe = max(3L, 1000m), L is the ship length, CPA_m is the minimum distance of approach of the m-th trajectory, and I( ) is the indicator function. The trigger condition for online calibration is ‖Δs_sim - Δs_pred‖² > 0.05, where Δs_sim is the actual state change of the target ship obtained from simulation, and Δs_pred is the predicted value of the state change output by the target ship dynamic predictor; the loss function for online calibration is L_calib = ‖Δs_sim - P_T(s, a; θ_T)‖² + λ‖θ_T - θ_T_old‖², where P_T(s, a; θ_T) is the predicted value of the state transition output by the target ship dynamic predictor based on the current state s, action a, and parameters θ_T, λ is the regularization coefficient, θ_T is the set of parameters of the target ship dynamic predictor, and θ_T_old is the set of parameters of the target ship dynamic predictor in the previous decision period.

[0039] In this embodiment, a high-fidelity ship dynamics model consistent with the dynamic characteristics of the target ship (the current ship) is constructed as the core of the digital twin simulation. The model includes longitudinal motion equations, lateral motion equations, and yaw motion equations. The longitudinal motion equation is M( - rv) = X, where M is the ship mass, is the longitudinal acceleration, r is the yaw angular velocity, v is the lateral velocity, and X is the longitudinal resultant force (X = X_H + X_R + X_P, X_H is the hull hydrodynamic force, X_R is the rudder force, and X_P is the propeller thrust); the lateral motion equation is M( + ur) = N, where is the yaw angular acceleration, u is the longitudinal velocity, and N is the yaw moment; the yaw motion equation is Iz = N, where Iz is the moment of inertia of the ship about the vertical axis, and N = N_H + N_R + N_P (N_H is the hydrodynamic moment of the hull, N_R is the rudder moment, and N_P is the propeller thrust moment). The parallel Monte Carlo simulation and risk assessment are based on the above high-fidelity ship dynamics model. 50 parallel simulation trajectories are set (considering different values of random factors such as environmental disturbances and sensor noises), and the ship navigation process after the execution of the control instructions generated in step 4 is simulated. The collision risk is evaluated by the following formula: R_collision = ∑_{m = 1}^50I(CPA_m < D_safe) / 50, where CPA_m is the minimum distance of approach of the m-th simulation trajectory; D_safe = max(3L, 1000m) is the safety distance threshold (L is the length of the target ship); I( ) is the indicator function, which takes the value of 1 (there is a collision risk) when CPA_m < D_safe, and 0 otherwise. During the online calibration process, the deviation between the actual state change Δs_sim of the target ship obtained by simulation and the predicted state change value Δs_pred output by the target ship dynamic predictor is calculated. When the deviation satisfies ‖Δs_sim - Δs_pred‖² > 0.05 (0.05 is the calibration trigger threshold), online calibration is performed on the target ship dynamic predictor. The loss function for online calibration is L_calib = ‖Δs_sim - P_T(s, a; θ_T)‖² + λ‖θ_T - θ_T_old‖², where P_T(s, a; θ_T) is the predicted state transition value output by the target ship dynamic predictor based on the current state s, action a, and parameters θ_T; θ_T is the parameter set of the target ship dynamic predictor, including the weights of the fully connected layer, bias parameters, and the correlation coefficients of the dynamic characteristics; θ_T_old is the parameter set of the target ship dynamic predictor in the previous decision cycle, which is used for regularization constraints to avoid parameter mutations; λ is the regularization coefficient, and its value range is [0.01, 0.1], which balances the prediction accuracy and parameter stability; the stochastic gradient descent algorithm with momentum is used to update θ_T, and the update formula is: θ_T ← θ_T - η L_calib + α(θ_T - θ_T_prev), where η is the learning rate, α is the momentum coefficient, and θ_T_prev is the value before the previous parameter update.

[0040] S6: Send the control instruction to the actuator, enter the next decision cycle, and loop through S1 - S6.

[0041] In this embodiment, the action control instruction or safety control instruction generated in S4 is sent to the ship actuators (rudder and propeller) to control the ship to adjust the rudder angle and thrust according to the instruction; after waiting for 1.0 seconds, return to step 1 to enter the next decision cycle, realizing continuous and real-time collision avoidance decision-making.

[0042] The method in this embodiment ensures its technical advantages through the following quantitative indicators: Safety indicators: Collision avoidance rate ≥99.5%, COLREGs compliance rate ≥98%, minimum encounter distance ≥3L; Efficiency metrics: Collision avoidance decisions result in a flight time increase of ≤15% and a fuel consumption increase of ≤20%; Migration performance indicators: Adaptation time to new ship type ≤ 24 hours, strategy migration success rate ≥ 90%.

[0043] This embodiment achieves high collision avoidance capability and rule compliance rate in terms of safety, while controlling the reasonable increase in voyage time and fuel consumption in terms of efficiency. In terms of migration performance, it significantly shortens the adaptation time for new ship types and improves the success rate of strategy migration. System verification is conducted through a professional maritime simulation platform, covering typical navigation areas and major merchant ship types. The real-ship verification plan is implemented in phases, ultimately achieving commercial deployment and online update capabilities. Simulation verification is based on the Real-time Marine Simulator (RMS) professional maritime simulation platform to build the verification environment. Test scenarios cover typical navigation areas such as Tokyo Bay, the Dover Strait, and the Singapore Strait, and test ship types include mainstream merchant ship types such as container ships, bulk carriers, tankers, and LNG carriers. Real-ship verification is implemented in three phases. In the first phase, three different types of test ships are selected to complete a total of 1000 hours of real-ship navigation verification; in the second phase, this is expanded to 10 ships, covering major merchant ship types and complex navigation scenarios; in the third phase, commercial deployment is achieved, supporting OTA (Over-The-Air) strategy updates and system upgrades.

[0044] Example 2

[0045] See Figures 2-5 This embodiment provides an intelligent ship general collision avoidance decision-making system based on cross-ship type knowledge transfer, used to implement the method in Embodiment 1. It includes a perception input layer, a core processing layer, a decision execution layer, and a verification evolution layer connected in sequence. Each layer module works collaboratively to complete the decision-making closed loop. The perception input layer is configured to collect real-time ship status data, ship static parameters, and environmental perception data; process the environmental perception data into a standardized environmental grid; and combine the real-time ship status data and ship static parameters into a state vector. In this embodiment, the specific functions of the perception input layer include: collecting environmental perception data, real-time ship status data, and ship static parameters through devices such as radar, AIS receiver, visual sensors, and ship status sensors; fusing and processing the environmental perception data into a 128×128×6 dimensional standardized environmental grid; and combining the real-time ship status data and static parameters into a 20-dimensional state vector to provide standardized input data for the core processing layer.

[0046] The core processing layer includes a dual-branch deep feature extraction network, a dynamic transition probability similarity coefficient calculation module, and a meta-policy network library. The dual-branch deep feature extraction network is configured to extract a general navigation situation feature vector independent of ship type using the scene branch, and extract a manipulation feature embedding vector strongly related to ship type using the manipulation branch. The meta-policy network library is configured to store a variety of pre-trained basic collision avoidance behavior meta-policies, and retrieve a set number of the most relevant candidate meta-policies based on the general navigation situation feature vector. The dynamic transition probability similarity coefficient calculation module is configured to calculate the dynamic transition probability similarity coefficient of the action corresponding to each candidate meta-policy through the target ship dynamic predictor. In this embodiment, the dual-branch deep feature extraction network extracts a 64-dimensional general navigation situation feature vector using the scene branch and extracts a 32-dimensional manipulation feature embedding vector using the manipulation branch, thereby decoupling scene cognition and manipulation characteristics. The meta-policy network library stores 8 pre-trained basic collision avoidance behavior meta-policies and retrieves 3 of the most relevant candidate meta-policies based on the general navigation situation feature vector. The decision execution layer includes a strategy transfer and dynamic fusion algorithm module and a safety rolling optimization module. The strategy transfer and dynamic fusion algorithm module is configured to use a softmax function to calculate weights and weightedly fuse the actions of candidate meta-policies to generate action control commands if the maximum dynamic transfer probability similarity coefficient is greater than or equal to a set threshold. The safety rolling optimization module is configured to generate safety control commands if the maximum dynamic transfer probability similarity coefficient is less than a set threshold. In this embodiment, when the maximum dynamic transfer probability similarity coefficient is greater than or equal to 0.7, the strategy transfer and dynamic fusion algorithm module uses a softmax function to calculate weights and weightedly fuse the actions of candidate policies to generate action control commands. When the maximum dynamic transfer probability similarity coefficient is less than 0.7, the safety rolling optimization module generates safety control commands based on a model predictive control framework to ensure navigation safety in extreme scenarios.

[0047] The verification evolution layer includes a digital twin framework, a parallel Monte Carlo simulation module, and an online calibration module. The digital twin framework is configured to construct a high-fidelity ship dynamics model. The parallel Monte Carlo simulation module is configured to perform parallel Monte Carlo simulations to assess decision-making risks. The online calibration module is configured to perform online calibration of the target ship's dynamic predictor if the simulation results deviate significantly from the predictions. In this embodiment, the digital twin framework synchronously simulates the ship's navigation state, providing a core platform for simulation verification. The parallel Monte Carlo simulation module sets 50 parallel simulation trajectories to simulate the execution effect of control commands and assesses decision-making safety through a collision risk formula. When the simulation results deviate significantly from the predictions, the online calibration module adjusts the parameter set θ_T of the target ship's dynamic predictor based on a calibration loss function, enabling continuous system evolution.

[0048] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A general collision avoidance decision-making method for intelligent ships based on cross-ship type knowledge transfer, characterized in that, include: S1. Collect real-time ship status data, ship static parameters and environmental perception data, process the environmental perception data into standardized environmental grids, and combine the real-time ship status data and ship static parameters into a state vector. S2. Use the scene branch of the dual-branch deep feature extraction network to extract a general navigation situation feature vector that is independent of the ship type, and use the maneuvering branch of the dual-branch deep feature extraction network to extract a maneuvering feature embedding vector that is strongly related to the ship type. S3. Based on the general navigation situation feature vector, retrieve a set number of the most relevant candidate meta-policies from the meta-policy network library, and calculate the dynamic transition probability similarity coefficient of the action corresponding to each candidate meta-policy through the target ship dynamic predictor; the meta-policy network library stores a variety of pre-trained basic collision avoidance behavior meta-policies; the target ship dynamic predictor is a prediction model that characterizes the state transition law of the target ship. S4. If the largest dynamic transition probability similarity coefficient is greater than or equal to the set threshold, the softmax function is used to calculate the weights and weighted fuse the candidate meta strategies to generate action control instructions; if the largest dynamic transition probability similarity coefficient is less than the set threshold, safety control instructions are generated. S5. Construct a high-fidelity ship dynamics model and perform parallel Monte Carlo simulation to assess decision-making risks. If the simulation results deviate from the predictions, the target ship dynamics predictor is calibrated online. S6. Send control commands to the actuator and enter the next decision cycle, repeating S1-S6.

2. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 1, characterized in that, In S2, The scene branch of the dual-branch deep feature extraction network adopts a convolutional neural network structure. The input is 128×128×6 dimensional environmental grid data, which is sequentially processed by two convolution-pooling operations and two fully connected operations to output a 64-dimensional general navigation situation feature vector. manipulate The branch adopts a fully connected network structure. The input is a 20-dimensional real-time ship state vector and ship static parameters. After passing through two fully connected operations and one embedding layer, the output is a 32-dimensional maneuver feature embedding vector.

3. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 2, characterized in that, In S2, The training loss function of the dual-branch deep feature extraction network is L_total = L_scene + λ'L_maneuver, where L_scene = ||F_s - F_s_gt||² is the scene branch loss function, F_s is the general navigation situation feature vector, F_s_gt is the expert-annotated real scene feature vector, L_maneuver = -logP(Δs|F_m) is the maximum likelihood loss function of the manipulation branch, P(Δs|F_m) is the conditional probability of the state change Δs occurring given the manipulation feature embedding vector F_m, and λ' is the weight coefficient of the manipulation branch loss.

4. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 1, characterized in that, In S3, The meta-policy network library contains eight basic collision avoidance behavior meta-policies, including right yield in crossover situations, left yield in overtaking situations, right turn in encounter situations, emergency braking, COLREGs following, collision avoidance in narrow waterways, multi-ship encounter coordination, and danger zone detour; the input of each meta-policy network is a concatenated vector of a general navigation situation feature vector and a maneuver feature embedding vector, and the output is a 2D action vector; the set number is 3.

5. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 4, characterized in that, In S3, The target ship dynamic predictor calculates the dynamic transition probability similarity coefficient using the formula κ(s,a)=exp(-‖Δs_T-Δs_S‖² / σ²), where σ is the similarity scaling parameter, s is the multi-dimensional vector of the target ship's real-time navigation state, a is the 2-dimensional control action vector output by the candidate meta-policy, Δs_T is the state transition prediction value of the target ship after executing action a, and Δs_S is the state transition prediction value of the source ship.

6. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 1, characterized in that, In S4, The set threshold is 0.7; The formula for calculating the weights using the softmax function is: w_i = exp(κ_i / 0.7) / ∑exp(κ_j / 0.7), where κ_i is the similarity coefficient of the i-th candidate policy, w_i is the weight of the i-th candidate policy, and the final action control command is a = ∑_{i=1}^3w_i a_i, where a_i is the action vector of the i-th candidate policy.

7. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 1, characterized in that, In S4, The safety control command is generated using a model predictive control framework. The optimization objective is minJ=∑_{k=0}^{19}[‖s_k-s_ref‖_Q²+‖a_k‖_R²]+‖s_H-s_ref‖_P², where s_k is the state vector at the k-th step in the prediction time domain, s_ref is the reference state vector, a_k is the control action vector at the k-th step in the prediction time domain, s_H is the end state in the prediction time domain, Q is the weight matrix of the state tracking error, R is the weight matrix of the control quantity, and P is the weight matrix of the end state. The constraints include state safety constraints and action feasibility constraints, with a prediction time domain of 20 steps and a control time domain of 5 steps.

8. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 1, characterized in that, In S5, The high-fidelity ship dynamics model includes the longitudinal motion equation M( -rv)=X, Equation of lateral motion M( +ur)=N and the equation of motion for head turning Iz =N, where M is the ship's mass. Let r be the longitudinal acceleration, r be the angular velocity of the turn, v be the lateral velocity, and X be the resultant longitudinal force. Let θ be the angular acceleration, u be the longitudinal velocity, N be the angular torque, and Iz be the moment of inertia about the vertical axis. The parallel Monte Carlo simulation sets 50 parallel simulation trajectories, and evaluates the collision risk through the formula R_collision = ∑_{m = 1}^50I(CPA_m < D_safe) / 50, where D_safe = max(3L, 1000m), L is the ship's length, CPA_m is the minimum distance of approach of the m-th trajectory, and I( ) is the indicator function.

9. The intelligent ship general collision avoidance decision-making method based on cross-ship type knowledge transfer according to claim 1, characterized in that, In S5, The trigger condition for the online calibration is ||Δs_sim-Δs_pred||²>0.05, where Δs_sim is the actual state change of the target ship obtained from the simulation, and Δs_pred is the predicted state change value output by the target ship dynamic predictor. The loss function for online calibration is L_calib=‖Δs_sim-P_T(s,a;θ_T)‖²+λ‖θ_T-θ_T_old‖², where P_T(s,a;θ_T) is the state transition prediction value output by the target ship dynamic predictor based on the current state s, action a and parameter θ_T, λ is the regularization coefficient, θ_T is the parameter set of the target ship dynamic predictor, and θ_T_old is the parameter set of the target ship dynamic predictor in the previous decision cycle.

10. A general intelligent ship collision avoidance decision-making system based on cross-ship type knowledge transfer, used to implement the method described in any one of claims 1-9, characterized in that, include: The perception input layer is configured to collect real-time ship status data, ship static parameters, and environmental perception data, process the environmental perception data into standardized environmental grids, and combine the real-time ship status data and ship static parameters into a state vector. The core processing layer includes a dual-branch deep feature extraction network, a dynamic transition probability similarity coefficient calculation module, and a meta-policy network library. The dual-branch deep feature extraction network is configured to extract a general navigation situation feature vector independent of ship type using the scene branch, and extract a maneuver feature embedding vector strongly related to ship type using the maneuver branch. The meta-policy network library is configured to store various pre-trained basic collision avoidance behavior meta-policies and retrieve a set number of most relevant candidate meta-policies based on the general navigation situation feature vector. The dynamic transition probability similarity coefficient calculation module is configured to calculate the dynamic transition probability similarity coefficient of the action corresponding to each candidate meta-policy through a target ship dynamic predictor. The decision execution layer includes a strategy migration and dynamic fusion algorithm module and a security rolling optimization module; The strategy migration and dynamic fusion algorithm module is configured such that if the maximum dynamic transition probability similarity coefficient is greater than or equal to a set threshold, the softmax function is used to calculate the weights and weighted fuse the candidate meta-strategies to generate action control instructions; the safe rolling optimization module is configured such that if the maximum dynamic transition probability similarity coefficient is less than a set threshold, a safe control instruction is generated. The validation evolution layer includes a digital twin framework, a parallel Monte Carlo simulation module, and an online calibration module. The digital twin framework is configured to construct a high-fidelity ship dynamics model. The parallel Monte Carlo simulation module is configured to perform parallel Monte Carlo simulations to assess decision-making risks. The online calibration module is configured to perform online calibration of the target ship dynamics predictor if the simulation results deviate significantly from the prediction.

Images