Intelligent ship resistance forecasting method and system

Abstract

The invention belongs to the cross technical field of the fusion graph attention network and the ship resistance field, and discloses a ship resistance intelligent forecasting method and system, and the method comprises the steps: constructing nodes and edges in a graph attention network, and carrying out the preprocessing of data; designing and improving a loss function; a dynamic weight multi-attention mechanism is established, and attention weights are automatically distributed according to different ship type features; applying a training strategy of an Adam and SGD combined optimizer to obtain a graph attention network model; and carrying out ship resistance prediction evaluation through the dynamic weight map attention network model. By introducing the ship resistance intelligent forecasting method based on the dynamic weight map attention network and the physical constraint, the ship resistance intelligent forecasting method has the advantages of simple structure and clear algorithm, is specially designed for forecasting resistance under different ship types and solving the matching problem of propellers and power devices, shows remarkable innovativeness and practical value, and is suitable for popularization and application. Powerful support is provided for optimization design, energy conservation and emission reduction and safe navigation of the ship.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of integrating graph attention network and ship resistance field, and particularly relates to a ship resistance intelligent prediction method and system. Background Technique

[0002] Ship resistance is of great significance for aspects such as fuel economy, navigation safety, and environmental impact. Accurately predicting ship resistance is a key link in realizing problems such as efficient ship design and matching main engine power. Traditional ship resistance prediction methods mainly rely on model tests, theoretical calculations, and empirical formulas. Even though they have been widely used in industry, there are still differences from the actual ship resistance.

[0003] Firstly, although empirical formulas are simple and easy to use, their applicable range is limited, and it is often difficult to accurately predict for complex ship types and special working conditions. Theoretical calculation methods need to solve the basic equations of fluid mechanics, the calculation process is complex, and the setting requirements for boundary conditions and initial conditions are relatively high. Secondly, model tests require a large amount of time, manpower, and material resources, and the cost is high. In addition, with the development of computer technology and numerical simulation methods, ship resistance prediction methods based on computational fluid dynamics (CFD) have gradually been applied. However, although the CFD method can relatively accurately simulate the flow field around the ship, the calculation amount is large and the calculation time is long, which is difficult to meet the requirements of rapid prediction in the ship design and optimization process.

[0004] In recent years, to overcome these problems, data - driven intelligent prediction methods have gradually become the research focus. Among them, deep learning technology has emerged in the field of ship resistance prediction due to its excellent ability in pattern recognition and non - linear function fitting. Through learning and training on a large amount of data, the deep learning model can achieve rapid prediction of ship hydrodynamic derivatives and show relatively high prediction accuracy. However, the limitations of traditional neural networks are more obvious when dealing with irregular graph structures or explaining the relationships therein.

[0005] The limitations of traditional neural networks in predicting ship resistance are as follows: 1. When traditional neural networks process data, they often regard data points as independent individuals and it is difficult to make full use of the relationship information between data.

[0006] 2. Traditional neural networks usually assume that data has a regular grid structure. For example, the convolutional attention network (CNN) is suitable for processing data with a two - dimensional grid structure such as images. However, for graph - structured data, the connection relationships between nodes are complex and irregular, and traditional attention networks are difficult to directly process.

[0007] 3. When traditional neural networks process data, the receptive field is usually of a fixed size, making it difficult to consider both local and global features of the data simultaneously. For example, the size of the convolutional kernel in a convolutional neural network (CNN) determines the range of local features it can capture, and it is difficult to obtain global information beyond this range.

[0008] 4. For input data of different sizes or structures, traditional neural networks may need to redesign the network structure or make complex adjustments.

[0009] Through the above analysis, the problems and defects existing in the existing technologies at home and abroad are as follows: (1) When traditional neural networks predict ship resistance, data points are regarded as independent individuals, making it difficult to fully utilize the relational information between data; for graph-structured data, the connection relationships between nodes are complex and irregular, and it is difficult for traditional attention networks to directly process them.

[0010] (2) When traditional neural networks process data, the receptive field is usually of a fixed size, making it difficult to consider both local and global features of the data simultaneously; for input data of different sizes or structures, traditional neural networks may need to redesign the network structure or make complex adjustments.

[0011] (3) In the current ship resistance prediction based on CNN, the relevance of hull geometry, fluid, and working conditions is not considered, resulting in significant prediction errors for complex ship types.

[0012] (4) The application of GAT in flow field prediction only defines edge attributes depending on the Euclidean distance between nodes and does not embed the Navier-Stokes equation, resulting in poor physical interpretability.

[0013] (5) The combined optimization of Adam and SGD adopts a fixed switching strategy and does not dynamically allocate optimizers according to the physical meaning of parameters. Summary of the Invention

[0014] To overcome the problems existing in the related technologies, the disclosed embodiments of the present invention provide a ship resistance intelligent prediction method and system, particularly a ship resistance intelligent prediction method and system based on a dynamic weight graph attention network and physical constraints. The technical solutions are as follows: The present invention is implemented as follows. The ship resistance intelligent prediction method includes the following steps: S1, constructing nodes and edges in the graph attention network and preprocessing the original data; S2, designing and improving a loss function based on physical constraints and node and edge features; S3, establishing a dynamic weight multi-head attention mechanism to automatically allocate attention weights to different ship type features; S4, applying the training strategy of the combined optimizer of Adam and SGD to obtain a graph attention network model; S5. Conduct ship resistance prediction and evaluation through a dynamic weighted graph attention network model; S6. Model verification and continuous optimization.

[0015] In step S1, when constructing the graph attention network, the nodes are divided into hull geometric feature nodes, hydrodynamic parameter nodes, and operating condition parameter nodes; The edges are divided into geometric correlation edges, physical action edges, and operating condition influence edges; The geometric correlation edges are used to connect the hull part nodes with geometric adjacent relationships, and the attributes of the edges represent the distance and angle changes between the two parts to reflect the continuity and relevance of the overall hull geometry; The physical action edges represent the edges established between the hydrodynamic parameter nodes and the hull geometric feature nodes, which are used to transmit the interaction information between the fluid and the hull; The attributes of the edges are set as physical quantities describing the action intensity, including the pressure gradient affecting the frictional resistance and the viscous pressure resistance affected by the flow field characteristics; The operating condition influence edges are used to connect the operating condition parameter nodes, and the attributes of the edges represent the influence degree of the operating condition on the parameters; The original data comes from ship tank tests and CFD numerical simulations, and has been normalized and denoised.

[0016] Furthermore, the hull geometric feature nodes include the main dimensions ratio of the ship and the form coefficients. The bow and stern are discretized into nodes, and each node contains the geometric parameters of that part. The geometric parameters include curvature, cross-sectional area, and length-width ratio; In the hydrodynamic parameter nodes, the flow velocity, pressure, density, and viscosity coefficient are set to represent the fluid parameters around the ship; The operating condition of the ship's navigation is used as a node, and the operating conditions include the speed, draft, and trim angle.

[0017] In step S2, the loss function based on physical constraints and node and edge features includes two parts: The first part is the cross-entropy loss function between the estimated value and the actual value of the resistance, which is used to ensure the accurate estimation of the resistance value by the model; The second part is the loss function term for predicting the viscous resistance of the ship based on the Navier-Stokes equation, which is used to ensure that the model prediction results conform to the physical laws of ship resistance prediction.

[0018] In step S2, design and improve the loss function based on physical constraints and node and edge features, including: The physical loss constrains the model output through the integral form of the Navier-Stokes equation, and the expression is: ; In the formula, is the physical constraint term in the loss function, is the square of the difference between two scalar values, that is, the squared error, is the theoretical drag value calculated based on the Navier-Stokes equation, obtained by integrating the pressure gradient on the hull surface; is the drag value predicted by the model; is the total number of samples, is the th item of the sample, is a positive integer representing the numbers from 1 to ; ; In the formula, is the drag value calculated by the theoretical formula, is the area of the outer surface of the ship, is the hydrodynamic viscosity, is the velocity field, is the pressure field, is the normal vector of the hull surface; Data loss is used to measure the difference between the drag value predicted by the model and the actual measured value. The expression is: ; In the formula, is the data loss term in the loss function, is the actual drag value of the th sample, from ship tank experiments or CFD simulations; is the drag value of the th sample predicted by the model; The loss function based on physical constraints and node and edge features, the expression is: ; In the formula, is the total loss function; is a hyperparameter used to balance the weights of data loss and physical loss.

[0019] In step S3, a dynamic weight multi-head attention mechanism is established, including: (1) Dynamic weight generation module; Based on the attention heads of GAT, a weight generation sub-network is added. The input is node features and edge attributes, and the output is the dynamic weight coefficients of each attention head. Specifically, it includes: Feature encoding: Jointly encode the node features and edge attributes , the expression is: ; In the formula, is the feature vector of node after joint encoding, is a multi-layer perceptron for non-linear feature mapping; is node The original characteristics of geometric parameters and hydrodynamic parameters is the side The distance of the geometric correlation side and the pressure gradient attribute of the physical action side is the node The set of neighbor nodes is the vector splicing operation, which splices the node features and edge attributes into an overall input is the expression of the j-th node Weight calculation: Generate the weights of K attention heads through Softmax, and the expression is: ; In the formula, is the th dynamic weight coefficient of the attention head, used to sum and aggregate the outputs of different heads is the learnable parameter vector, representing the physical dimensions of the th head focusing on geometry, fluid, and working conditions is the total number of attention heads is the exponential function, used for Softmax normalization of weight distribution is the node The feature vector after joint encoding (2) Scene-adaptive attention aggregation; In the region of sudden change in bow curvature dominated by local geometry, the attention aggregation specifically includes: Geometry focusing: The weight increases, and the influence of geometric correlation edges is strengthened when calculating attention ; In the formula, is the node 's attention weight to the neighbor node ; is the learnable attention parameter vector, used for weight distribution of geometric features is the learnable weight matrix, used to project the node features into the feature space of the set focusing head is the vector splicing operation, splicing the projection features of the node and ; is the activation function, introducing non-linearity and alleviating the problem of gradient disappearance is the transpose operation to adjust the vector direction, ensuring the dimensional consistency of mathematical operations. Through transpose and dot product, the model automatically learns the influence weight of geometric features on ship resistance is the learnable weight matrix is the node 's geometric parameters is the geometric parameter of node j; Flow field focusing: weight Decrease, and finally the node features are updated to the weighted aggregation of each head; ; In the formula, is the combined parameter after update, is the activation function, enhancing the nonlinear expression ability of the model; is the weight matrix of the th attention head, used for feature transformation; is the attention weight of node to the k-th attention head of neighbor node ; In the formula, is the physical constraint loss term, forcing the attention weight to conform to the laws of hydrodynamics, is a hyperparameter used to control the strength of the physical constraint term; is the ideal weight calculated by equation.

[0020] In step S4, the training strategy of applying the combined optimizer of Adam and SGD includes: Phased combination: Apply the Adam optimizer at the initial stage of training. When the model training approaches the optimal solution, switch to the SGD optimizer; Important parameter combination: Use the SGD optimizer to update the node weights corresponding to the parameters closely related to the ship geometry; for some weights related to the hydrodynamic parameter nodes and operating condition parameter nodes, use the Adam optimizer; Dynamic weight combination: During the training process, dynamically adjust the weights of the Adam optimizer and the SGD optimizer in parameter update, introducing the weight coefficient ; Balance the fast convergence of the Adam optimizer and the stability of the SGD optimizer. The expression is: ; In the formula, is the model parameter of the th iteration of node weights and edge attributes; is the model parameter of the th iteration of node weights and edge attributes; is the global learning rate, controlling the step size of parameter update; is the dynamic weight coefficient, which controls the mixing ratio of the Adam and SGD optimizers; is the momentum term of the Adam optimizer, which combines the first-order moment estimation of the historical gradients; is the current batch gradient of the SGD optimizer, which directly calculates the derivative of the loss function with respect to the parameters.

[0021] In step S5, the trained dynamic weight graph attention network model is used to predict the data in the test set, and the model prediction results are compared and verified with the results obtained by computational fluid dynamics.

[0022] In step S6, according to the results of the prediction evaluation, the model is iteratively optimized, including adjusting the network structure, optimizing the loss function, and selecting the optimization algorithm. Finally, the optimized model is applied to the prediction of the real ship rolling motion.

[0023] Another object of the present invention is to provide a ship resistance intelligent prediction system, which is used to control the ship resistance intelligent prediction method described above. The system includes: A graph attention network construction module, which is used to construct nodes and edges in the graph attention network and preprocess the original data; A data preprocessing module, which is used to design and improve the loss function based on physical constraints and node and edge features; A loss function construction module, which is used to establish a dynamic weight multi-head attention mechanism to automatically assign attention weights to different ship type features; An attention weight automatic assignment module, which is used to establish a dynamic weight multi-head attention mechanism to automatically assign attention weights to different ship type features; A combined training module, which is used to apply the training strategy of the Adam and SGD combined optimizer to obtain the graph attention network model; A ship resistance prediction evaluation module, which is used to perform ship resistance prediction evaluation through the dynamic weight graph attention network model; A model verification and continuous optimization module, which is used for model verification and continuous optimization.

[0024] Combining all the above technical solutions, the beneficial effects of the present invention are: In view of the current limitations of neural networks, the present invention proposes an intelligent ship resistance prediction method based on a dynamic weight graph attention network and physical constraints. By constructing nodes and edges in the graph attention network and introducing a dynamic weight multi-head attention mechanism according to the relationships between different nodes and edges, different weights are dynamically assigned to the neighbor nodes of each node. The specific process is to project the node features into a new space through a linear transformation, calculate the attention scores between nodes using an attention function (such as a shared linear function activated by LeakyReLU), normalize them through the softmax function to obtain the weights of each neighbor node, and add the weighted sum of the neighbor node features according to the weights to update the current node features in the original cross-entropy loss function.

[0025] The present invention constructs the geometric features, hydrodynamic parameters, and navigation conditions of a ship as graph-structured data, and uses the powerful graph data processing ability of the graph attention network to automatically learn the complex non-linear relationships therein. With the cooperation of dynamic weight allocation and physical condition constraints, an accurate ship resistance value can be obtained quickly, overcoming the limitations of traditional ship resistance prediction methods, improving the accuracy and efficiency of ship resistance prediction, and providing strong support for the optimal design, energy conservation and emission reduction, and safe navigation of ships.

[0026] The present invention introduces an intelligent ship resistance prediction method based on a dynamic weight graph attention network and physical constraints, which is designed specifically for resistance prediction under different ship types and for solving the matching problems of propellers and power plants, demonstrating significant innovation and practical value. Compared with the prior art, the present invention has the advantages of a simple structure and a clear algorithm, providing strong support for the optimal design, energy conservation and emission reduction, and safe navigation of ships. In addition, by accurately capturing the characteristics of the ship type, operating conditions, hydrodynamic parameters, etc., this method effectively improves the resistance prediction accuracy under complex ship types, different drafts, and speeds, and solves the limitations of traditional models in dealing with irregular graph structures. Therefore, the present invention not only provides an advanced technical means for resistance prediction under different ship types and for solving the matching problems of propellers and power plants, but also opens up new ways for research and application in related fields, having important theoretical significance and broad application prospects.

[0027] Aiming at the limitations of traditional ship resistance prediction methods, the present invention constructs ship geometric features, hydrodynamic parameters, and navigation conditions into graph-structured data. By constructing a multi-level graph structure including geometric correlation edges, physical action edges, and working condition influence edges, and combining the Navier-Stokes equation to constrain the edge attributes, the deep integration of the graph structure and fluid mechanics mechanism is realized, and the physical interpretability of the model is improved. The present invention introduces a dynamic weight multi-head attention mechanism, which automatically assigns weights according to ship type characteristics and physical scenarios, enhancing the ability to capture complex relationships. A loss function based on physical constraints and node and edge features is designed, and an optimization strategy combining the Adam optimizer and the SGD optimizer is used to train the model. Verified by simulation experiments, compared with traditional neural networks and graph attention networks, the present invention has higher resistance prediction accuracy, stronger generalization ability, and higher computational efficiency under complex ship types and different working conditions, and can effectively support ship optimization design, energy conservation and emission reduction, and safe navigation.

[0028] The intelligent ship resistance prediction method based on the dynamic weight graph attention network and physical constraints of the present invention not only proposes a new construction relationship of the graph attention network theoretically, but also provides an effective ship resistance prediction method in practical applications. By means of optimizing model design, introducing physical constraints, and combining optimizer training strategies, the present invention can accurately predict complex ship type resistance and under different basic working conditions, and the prediction results have physical interpretability, significantly improving the accuracy and reliability of ship resistance prediction. Brief Description of the Drawings

[0029] The drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure; Figure 1 It is a flowchart of the intelligent ship resistance prediction method provided by an embodiment of the present invention; Figure 2 It is a schematic diagram of the principle of the intelligent ship resistance prediction method provided by an embodiment of the present invention; Figure 3 It is a schematic diagram of the multi-head attention mechanism provided by an embodiment of the present invention (k = 3). Detailed Embodiments

[0030] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention is given with reference to the drawings. Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0031] The innovation points of the intelligent ship resistance prediction method and system provided by the embodiments of the present invention are as follows: 1. Deep integration of multi-level graph structure modeling and physical constraints Traditional ship resistance prediction methods (such as CFD) rely on discretized grids and numerical iterations, with high computational complexity and difficulty in capturing multi-physical field coupling effects. The present invention proposes a multi-scale modeling method based on graph structure: Node definition: The hull geometric features (curvature, cross-sectional area), hydrodynamic parameters (flow velocity, pressure gradient), and operating conditions (speed, trim angle) are respectively modeled as independent nodes to form a multi-dimensional feature space.

[0032] Physical quantification of edge attributes: Through the finite volume method after discretization of the Navier-Stokes equation, the attributes of physical action edges are directly calculated.

[0033] Mechanism-driven modeling: In traditional graph networks (such as GAT), edge attributes mostly rely on Euclidean distance or simple associations, while in this method, edge attributes are directly constrained by physical equations, enabling the graph structure to naturally reflect hydrodynamic laws and significantly improving the interpretability of the model.

[0034] 2. Scenario adaptability of the dynamic weight multi-head attention mechanism Traditional GAT uses static weight allocation and cannot adapt to complex physical scenarios in ship resistance prediction (such as sudden changes in bow curvature, speed changes). The innovation of the present invention lies in: Dynamic weight generation sub-network: Input node features and edge attributes are jointly encoded into feature vectors through a multi-layer perceptron (MLP), and attention head weights are generated.

[0035] Scenario-adaptive attention aggregation: Taking the sudden change in bow curvature as an example, the weights of the focusing heads automatically increase, strengthening the influence of local geometric features on resistance; when the speed changes, the weights of the flow field heads increase, emphasizing the correlation between operating conditions and the flow field. The establishment of this mechanism enables the model to automatically allocate computing resources in complex scenarios, reducing the prediction error compared to traditional GATs.

[0036] 3. Design and optimization of the physical constraint loss function Traditional data-driven models (such as CNN, LSTM) are prone to failure under extreme operating conditions due to the lack of physical law constraints. The present invention proposes a dual-driven loss function: Data loss: The cross-entropy loss measures the difference between the predicted value and the measured value: The cross-entropy loss measures the difference between the predicted value and the measured value.

[0037] Physical loss: The output is constrained by the integral form of the Navier-Stokes equation to ensure that the prediction conforms to hydrodynamic laws.

[0038] Dynamic weight adjustment: The hyperparameter λ balances the data and physical losses. When data is scarce (such as for new ship types), λ is increased to strengthen the physical constraints; when data is sufficient, λ is decreased to focus on fitting accuracy.

[0039] 4. Improvement in the efficiency and stability of the combined optimization strategy Traditional optimizers (such as the Adam optimizer or SGD optimizer) are difficult to balance the convergence speed and accuracy in complex tasks. The present invention proposes a dynamic hybrid optimization strategy: Phased combination: At the initial stage of training, Adam is used for fast convergence. When the model training approaches the optimal solution, it is switched to the SGD optimizer; in the later stage, it is switched to the SGD optimizer for fine-tuning of parameters.

[0040] Important parameter combination: For the key parameters that have a greater impact on the ship resistance prediction result and have clear physical meanings, including the node weights corresponding to the parameters closely related to the ship geometry, the SGD optimizer is used for updating; for some of the weights related to the hydrodynamic parameter nodes and operating condition parameter nodes, the Adam optimizer is used; Dynamic weight combination: During the training process, the weights of the Adam optimizer and the SGD optimizer in parameter update are dynamically adjusted according to the training effect of the model, and a weight coefficient is introduced ; The role of this model is to balance the fast convergence of the Adam optimizer and the stability of the SGD optimizer. At the initial stage of training: mainly use the Adam optimizer to quickly approach the optimal solution. At the later stage of training: enhance the fine-tuning ability of the SGD optimizer to avoid the oscillation caused by the adaptive learning rate of Adam. The expression is: ; In the formula, is the model parameter of the node weight and edge attribute at the th iteration, is the global learning rate, which controls the step size of parameter update; is the dynamic weight coefficient, which controls the mixing ratio of the Adam optimizer and the SGD optimizer; is the momentum term of the Adam optimizer, which combines the first-order moment estimation of the historical gradient; is the current batch gradient of the SGD optimizer, which directly calculates the derivative of the loss function with respect to the parameters; Dynamic hybrid mechanism: Automatically adjust according to the training progress of the loss change and gradient amplitude; ; In the formula, is the attenuation rate, which decreases with the increase of the iteration number and the proportion of the SGD optimizer increases; ​ Sub-parameter optimization: For geometric-related parameters (such as ship form coefficients), force the use of the SGD optimizer to avoid the adaptive noise of the Adam optimizer; for fluid parameters, use the Adam optimizer to automatically adapt to complex relationships.

[0041] Key parameters (such as ship form parameters): Force the use of the SGD optimizer for updating to avoid the noise introduced by the adaptive mechanism of the Adam optimizer.

[0042] General parameters (such as fluid parameters): Use the Adam optimizer to automatically adapt to complex relationships.

[0043] Experimental advantages: In the ship resistance prediction task, the error of this strategy is reduced by 15% compared with a single optimizer, and the convergence speed is increased by 30%.

[0044] Example 1, as Figure 1 shown, the intelligent ship resistance prediction method provided by the embodiment of the present invention includes the following steps: S1, construct nodes and edges in the graph attention network, and preprocess the original data; When constructing the graph attention network, the nodes are divided into hull geometric feature nodes, hydrodynamic parameter nodes, and working condition parameter nodes; input the constructed graph structure data into the graph attention network, and set the input parameters of the ship form whose resistance needs to be predicted as and the input parameters of other similar ship forms are respectively , is the node feature vector, is the number of nodes (the graph structure of the input ship form includes the ship form to be predicted itself), is the node feature dimension, which is used as the initial input and is expressed as: ; The edges are divided into geometric association edges, physical action edges, and working condition influence edges; the geometric association edges are used to connect the hull part nodes with geometric adjacent relationships, and the attributes of the edges represent the distance and angle changes between the two parts to reflect the continuity and relevance of the overall hull geometry; the physical action edges represent the edges established between the hydrodynamic parameter nodes and the hull geometric feature nodes, which are used to transmit the interaction information between the fluid and the hull; the attributes of the edges are set as physical quantities describing the action intensity, including the pressure gradient affecting the frictional resistance and the viscous pressure resistance affected by the flow field characteristics; the working condition influence edges are used to connect the working condition parameter nodes, and the attributes of the edges represent the influence degree of the working condition on the parameters; the original data comes from ship tank tests and CFD numerical simulations, and the data has been normalized and denoised.

[0045] The nodes include the main dimensions ratio of the ship and the ship form coefficients. The bow and stern are discretized into nodes, and each node contains the geometric parameters of that part. The geometric parameters include curvature, cross-sectional area, and length-width ratio. In the nodes of hydrodynamic parameters, the flow velocity, pressure, density, and viscosity coefficient are set to represent the fluid parameters around the ship. The operating conditions of the ship's navigation are used as nodes, and the operating conditions include the speed, draft, and trim angle.

[0046] S2. Design and improve the loss function based on physical constraints and node and edge features; The loss function based on physical constraints and node and edge features consists of two parts: The first part is the cross-entropy loss function between the estimated value and the actual value of the resistance, which is used to ensure the accurate estimation of the resistance value by the model; The second part is the loss function term for predicting the viscous resistance of the ship based on the Navier-Stokes equation, which is used to ensure that the model prediction results conform to the physical laws of ship resistance prediction.

[0047] To calculate the features of the edges between nodes in the graph attention network, the relevance function needs to consider the influence of two nodes simultaneously, as shown in the following formula: ; In the formula, is a -dimensional learnable parameter matrix. Therefore, is to map a -dimensional vector to a -dimensional space. The calculation dimension of is . For the features of all vertices, mapping is required, so the calculation dimension is . Among them, is to map a -dimensional vector to a real number, and its calculation dimension . When calculating the attention coefficient, each edge in the graph needs to be calculated, so its calculation dimension is . That is, when outputting the ship form resistance, the influence of its neighboring nodes will be considered.

[0048] ; The above formula concatenates node and , and then maps it to a scalar. At this time, , indicating that it is not symmetric at this time. In addition to concatenation, it is also necessary to aggregate neighbor information, so the attention of all neighbors of each node needs to be normalized. The attention coefficient obtained after normalization can be used as the real aggregation coefficient.

[0049] Calculate its weight through the dynamic weight multi - head attention mechanism.

[0050] The weight of head 1 (geometric focus) Increases, strengthening the influence of geometric correlation edges when calculating attention; ; In the formula, is the attention weight of node to neighbor node ; is a learnable attention parameter vector for weight allocation of geometric features; is a learnable weight matrix for projecting node features into the feature space of the set focus head; is the vector concatenation operation, concatenating the projected features of node and ; is an activation function, introducing non - linearity and alleviating the vanishing gradient problem; is the transpose operation to adjust the vector direction, ensuring the dimensional consistency of mathematical operations. Through transpose and dot product, the model automatically learns the influence weight of geometric features on ship resistance; is a learnable weight matrix, is the geometric parameter of node , is the geometric parameter of node j; The weight of head 2 (flow field focus) Decreases, and finally the node features are updated as the weighted aggregation of each head; ; In the formula, is the combined parameter of node after update, is an activation function to enhance the non - linear expression ability of the model; is the weight matrix of the th attention head for feature transformation; is the attention weight of node to neighbor node for the k - th attention head; Since the multi - head attention has multiple heads working in parallel, each head can focus on different aspects of information, thus being able to capture richer and more diverse feature relationships. As shown in Figure 3 , its value is taken as 3. For example, in resistance prediction, different heads can respectively focus on different dimensional information such as the main scale ratios of different ship types and various ship type parameters, making the model's understanding of the graph structure and the relationships between nodes more comprehensive.

[0051] Design and improve the loss function based on physical constraints and node and edge features, including: The physical loss constrains the model output through the integral form of the Navier-Stokes equation, and the expression is: ; In the formula, is the physical constraint term in the loss function, is the square of the difference between two scalar values, that is, the squared error, is the theoretical drag value calculated based on the Navier-Stokes equation, obtained by integrating the pressure gradient on the hull surface; is the drag value predicted by the model; is the total number of samples, is the th item of the sample, is a positive integer representing the numbers from 1 to ; ; In the formula, is the drag value calculated by the theoretical formula, is the area of the outer surface of the ship, is the hydrodynamic viscosity, is the velocity field, is the pressure field, is the normal vector of the hull surface; The data loss is used to measure the difference between the drag value predicted by the model and the actual measured value, and the expression is: ; In the formula, is the data loss term in the loss function, is the actual drag value of the th sample, from ship tank experiments or CFD simulations; is the drag value of the th sample predicted by the model; The loss function based on physical constraints and node and edge features, the expression is: ; In the formula, is the total loss function; is a hyperparameter used to balance the weights of data loss and physical loss.

[0052] S3. Establish a dynamic weight multi-head attention mechanism to automatically assign attention weights to different ship type features; Establish a dynamic weight multi-head attention mechanism, including: (1) Dynamic weight generation module; based on the attention heads of GAT, a weight generation sub-network is added. The input is node features and edge attributes, and the output is the dynamic weight coefficients of each attention head, specifically including: Feature encoding: Jointly encode the node features and edge attributes The expression is: ; In the formula, is the feature vector of node after joint encoding, is a multi-layer perceptron for non-linear feature mapping; is the original feature of the geometric parameters and hydrodynamic parameters of node , is the distance of the geometrically associated edge and the pressure gradient attribute of the physically acting edge of edge , is the set of neighbor nodes of node ; is the vector concatenation operation, which concatenates the node features and edge attributes into an overall input; is the expression of the j-th node; Weight calculation: Generate the weights of K attention heads through Softmax. The expression is: ; In the formula, is the dynamic weight coefficient of the -th attention head, which is used to sum and aggregate the outputs of different heads; is a learnable parameter vector, representing the physical dimensions of geometry, fluid, and working conditions that the -th head focuses on; is the total number of attention heads; is the exponential function, which is used for Softmax normalization of weight distribution; is the feature vector of node after joint encoding; (2) Scene-adaptive attention aggregation; in the region of sudden change in bow curvature dominated by local geometry, the attention aggregation specifically includes: Geometric focusing: The weight increases, and the influence of geometrically associated edges is strengthened when calculating attention; ; In the formula, is the attention weight of node to neighbor node ; is a learnable attention parameter vector for weight distribution of geometric features; is a learnable weight matrix used to project node features into the feature space of the set focusing head; is a vector concatenation operation that concatenates the projected features of nodes and ; is an activation function that introduces non-linearity and alleviates the vanishing gradient problem; is a transpose operation to adjust the vector direction and ensure dimensional consistency in mathematical operations. Through transpose and dot product, the model automatically learns the influence weights of geometric features on ship resistance; is a learnable weight matrix, is the geometric parameter of node ; is the geometric parameter of node j; Flow field focusing: The weight decreases, and finally the node features are updated as the weighted aggregation of each head; ; In the formula, is the combined parameter after the update of node ; is an activation function that enhances the non-linear expression ability of the model; is the weight matrix of the th attention head for feature transformation; is the attention weight of node to the kth attention head of neighbor node ; (3)Physics-constrained weight optimization; Add a physical regularization term for attention weights in the loss function to ensure that the weight distribution conforms to the laws of hydrodynamics; ; In the formula, is the physical constraint loss term, forcing the attention weights to conform to the laws of hydrodynamics, is a hyperparameter used to control the strength of the physical constraint term; is the ideal weight calculated through the equation.

[0053] S4, Apply the training strategy of the Adam and SGD combined optimizer to obtain the graph attention network model; Phase combination: Apply the Adam optimizer at the initial stage of training. When the model training approaches the optimal solution, switch to the SGD optimizer; Important parameter combination: For the node weights corresponding to the parameters closely related to the ship geometry, use the SGD optimizer for update; For some weights related to the hydrodynamic parameter nodes and operating condition parameter nodes, use the Adam optimizer; Dynamic weight combination: During the training process, the weights of the Adam optimizer and the SGD optimizer in parameter update are dynamically adjusted according to the training effect of the model, and a weight coefficient is introduced. ; Balance the fast convergence of the Adam optimizer and the stability of the SGD optimizer. The expression is: ; In the formula, is the model parameter of the node weight and edge attribute at the th iteration; is the model parameter of the node weight and edge attribute at the th iteration; is the global learning rate, which controls the step size of parameter update; is the dynamic weight coefficient, which controls the mixing ratio of the Adam optimizer and the SGD optimizer; is the momentum term of the Adam optimizer, which combines the first-order moment estimation of the historical gradient; is the current batch gradient of the SGD optimizer, which directly calculates the derivative of the loss function with respect to the parameter.

[0054] S5. Through the dynamic weight graph attention network model, conduct ship resistance prediction and evaluation; Use the trained dynamic weight graph attention network model to predict the data in the test set, and compare the model prediction results with the results obtained by computational fluid dynamics for verification. Use the trained graph attention network model to predict the test set data, focusing on its performance in terms of resistance prediction speed and accuracy. By comparing the model prediction results with the results obtained by computational fluid dynamics (CFD), verify the effectiveness and superiority of the model in constructing an intelligent ship resistance prediction. According to the prediction and evaluation results, iteratively optimize the model, such as adjusting the network structure, optimizing the loss function, and selecting an optimization algorithm, etc., to further improve the prediction accuracy and generalization ability of the model. Finally, apply the optimized model to the prediction of the real ship's sway motion to provide decision support for ship performance evaluation and navigation state prediction. The schematic diagram of the intelligent ship resistance prediction method provided by the embodiment of the present invention is as Figure 2 shown.

[0055] S6. Verification and continuous optimization of the model.

[0056] According to the results of the prediction and evaluation, iteratively optimize the model, including adjusting the network structure, optimizing the loss function, and selecting an optimization algorithm, and finally apply the optimized model to the prediction of the real ship's sway motion.

[0057] Example 2. The intelligent ship resistance prediction system provided by the embodiment of the present invention includes: The graph attention network construction module is used to construct nodes and edges in the graph attention network. Among them, the nodes are divided into hull geometric feature nodes, hydrodynamic parameter nodes, and operating condition parameter nodes, and the edges are divided into geometric association edges, physical action edges, and operating condition influence edges; The data preprocessing module is used to preprocess the raw data from ship tank tests and CFD numerical simulations by normalizing and denoising; The loss function construction module is used to design and improve the loss function based on physical constraints and node and edge features, including the physical constraint loss based on the resistance prediction model and the prediction error loss based on data; The attention weight automatic allocation module is used to establish a dynamic weight multi-head attention mechanism to automatically allocate attention weights to different ship type features; The combined training module is used to adopt the training strategy of the Adam and SGD combined optimizer in ship resistance prediction to obtain the dynamic weight graph attention network model; The ship resistance prediction and evaluation module is used to predict the data in the test set through the dynamic weight graph attention network model and compare the prediction results with the results obtained by computational fluid dynamics; The model verification and continuous optimization module is used to iteratively optimize the model according to the results of the prediction evaluation, and finally apply the optimized model to the prediction of the real ship's sway motion.

[0058] To further prove the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions. The comparison between the present invention (GATs) and traditional neural networks is shown in Table 1.

[0059] Table 1 Comparison with traditional neural networks

[0060] The comparison between the present invention and the loss functions of existing models is shown in Table 2. In the simulation experiments carried out on the model, after cleaning the original data, the node features obtained are further optimized by constructing the relationship of the corresponding edges. After adjusting the hyperparameters, the combined optimization algorithm of the Adam optimizer and the SGD optimizer is adopted to ensure the stability and convergence of the model training. In the experiments on existing ship types such as KCS and KVLCC2 and relatively complex ship types, comparisons are made with traditional finite element models and other neural network models (PINN, LSTM, MLP, etc.) to verify the response accuracy and real-time performance of the present invention under complex ship types and different basic operating conditions.

[0061] Table 2 Prediction accuracy tests of four models in different ship types

[0062] The comparison between the dynamic weight GATs of the present invention and the traditional GATs is shown in Table 3.

[0063] Table 3 Comparison between the traditional GAT and the model of the present invention

[0064] 2. Experimental verification (1) Scenario of sudden change in bow curvature Test ship type: bulbous bow container ship (the proportion of the sudden change area of curvature is 15%).

[0065] Result comparison: Dynamic weight GAT: Geometric head weight , accurately capturing the contribution of curvature to friction (prediction error of 2.1%).

[0066] Traditional GAT: The weights of each head are evenly distributed , with an error of up to 8.7%.

[0067] (2) Scenario of change in ship speed Working condition: The ship speed increases from 15 knots to 25 knots (the Reynolds number changes significantly).

[0068] Result comparison: Dynamic weight GAT: Flow field head weight.

[0069] Traditional GAT: Due to the static weight being unable to adapt to the change of the flow field, the error reaches 12.3%.

[0070] In summary, the intelligent ship resistance prediction method based on the dynamic weight graph attention network and physical constraints proposed by the present invention not only proposes a new construction relationship of the graph attention network theoretically, but also provides an effective ship resistance prediction method in practical applications. By means of optimizing the model design, introducing physical constraints, combining optimizer training strategies, etc., the present invention can accurately predict the complex ship type resistance and the prediction results with physical interpretability under different basic working conditions, significantly improving the accuracy and reliability of ship resistance prediction.

[0071] The technical solution of the present invention has significant commercial value in the fields of ship design, manufacturing, and operation; Economic benefits: By optimizing the ship hull design through high-precision resistance prediction, the ship design cycle can be shortened by approximately 40%, the CFD simulation cost can be reduced by 60%, and the annual fuel cost per ship can be saved by approximately 12% (corresponding to a CO2 emission reduction of 500 tons). Market competitiveness: Provide shipyards with a rapid resistance optimization tool to help seize the high-energy efficiency ship market. For example, in the design of ultra-large container ships (24,000 TEU), the resistance error of this method can be controlled within 3%, which is better than traditional CFD methods (error > 5%). Environmental compliance: Meet IMO (International Maritime Organization) carbon emission regulations, and optimize the navigation strategy through accurate resistance prediction to help shipping companies reduce carbon tax expenditures.

[0072] The present invention realizes the deep coupling of graph neural network and fluid mechanics mechanism for the first time: By directly constraining the edge attributes and attention weights through the Navier-Stokes equation, the problem of poor physical interpretability of traditional GATs in ship engineering is solved. Dynamic weight multi-head attention mechanism: Breaks through the static weight limitation of traditional GATs and realizes the automatic focusing on geometry, fluid, and working conditions under complex ship hulls for the first time. Combined optimization strategy: Proposes a dynamic hybrid optimization method with parameter grouping (Adam optimizer + SGD optimizer), which improves the convergence speed by 30% and reduces the error by 15% compared with the existing stage switching strategy.

[0073] The disconnection between data and physical models: In traditional methods, data-driven models (such as CNN, LSTM) lack physical constraints, while CFD relies on numerical simulation with a large amount of calculation. This technology realizes data-mechanism dual drive for the first time through physical loss functions and NS equation embedding, and the prediction error is reduced by 60% compared with pure data models.

[0074] Adaptive modeling for complex ship hulls: Traditional neural networks need to redesign the network for different ship hulls, while this technology automatically adapts through a dynamic weight mechanism (such as when the ship speed changes, the weight of the flow field head is increased to 72%), and the generalization ability is improved by 50%.

[0075] Stability under extreme working conditions: Under non-design working conditions with a longitudinal inclination angle > 5°, the error of this technology is only 2.8%, while the error of traditional CFD reaches 4.5% due to grid distortion.

[0076] The above is only a relatively optimal specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. An intelligent prediction method for ship resistance, characterized in that The method includes the following steps: S1. Construct nodes and edges in the graph attention network, and preprocess the original data; S2. Design and improve the loss function based on physical constraints and node and edge features; S3. Establish a dynamic weight multi-head attention mechanism to automatically assign attention weights to different ship type features; S4. Apply the training strategy of the combined Adam and SGD optimizers to obtain the graph attention network model; S5. Through the dynamic weight graph attention network model, conduct ship resistance prediction and evaluation; S6. Verification and continuous optimization of the model.

2. The intelligent prediction method for ship resistance according to claim 1, wherein In step S1, when constructing the graph attention network, the nodes are divided into hull geometric feature nodes, hydrodynamic parameter nodes, and operating condition parameter nodes; The edges are divided into geometric association edges, physical action edges, and operating condition influence edges; the geometric association edges are used to connect hull part nodes with geometric adjacent relationships, and the physical action edges represent the edges established between the hydrodynamic parameter nodes and the hull geometric feature nodes, and the attribute of the edge is set as the physical quantity describing the action intensity; The original data comes from ship tank tests and CFD numerical simulations, and is data that has been normalized and denoised.

3. The intelligent prediction method for ship resistance according to claim 2, wherein The hull geometric feature nodes include the main dimensions ratio and form coefficients of the ship. The bow and stern are discretized into nodes, and each node contains the geometric parameters of this part. The geometric parameters include curvature, cross-sectional area, and length-width ratio; in the hydrodynamic parameter nodes, the flow velocity, pressure, density, and viscosity coefficient are set to represent the fluid parameters around the ship.

4. The intelligent prediction method for ship resistance according to claim 1, wherein In step S2, the loss function based on physical constraints and node and edge features includes two parts: The first part is the cross-entropy loss function between the estimated value and the actual value of the resistance, which is used to ensure the accurate estimation of the resistance value by the model; The second part is the loss function term for predicting the viscous resistance of the ship based on the Navier-Stokes equation, which is used to ensure that the model prediction results conform to the physical laws of ship resistance prediction.

5. The intelligent prediction method for ship resistance according to claim 4, wherein, In step S2, designing and improving the loss function based on physical constraints and node and edge features includes: The physical loss constrains the model output through the integral form of the Navier-Stokes equation, and the expression is: ； In the formula, is the physical constraint term in the loss function, is the square of the difference between two scalar values, is the theoretical drag value calculated based on the Navier-Stokes equation, is the drag value predicted by the model; is the total number of samples, is the th item of the sample, is a positive integer representing the numbers from 1 to ; ； wherein, is the resistance value calculated by the theoretical formula, is the area of the outer surface of the hull, is the hydrodynamic viscosity, is the velocity field, is the pressure field, is the normal vector of the hull surface; The data loss is used to measure the difference between the model predicted resistance value and the actual measured value, and the expression is: ； In the formula, is the data loss term in the loss function, is the actual resistance value of the th sample, and is the resistance value of the th sample predicted by the model; The loss function based on physical constraints and node and edge features, the expression is: ； In the formula, is the total loss function; is the hyperparameter.

6. The intelligent prediction method for ship resistance according to claim 1, wherein In step S3, establishing a dynamic weight multi-head attention mechanism includes: (1) Dynamic weight generation module; based on the attention heads of GAT, add a weight generation sub-network, with the input being node features and edge attributes, and the output being the dynamic weight coefficients of each attention head, specifically including: Feature encoding: encoding the node features and edge attributes jointly, with the expression: ； In the formula, is the jointly encoded feature vector of the node , is a multi-layer perceptron, is the original features of the geometric parameters and hydrodynamic parameters of the node , is the distance of the geometric correlation edge and the pressure gradient attribute of the physical action edge of the edge , is the set of neighbor nodes of the node ; is a vector concatenation operation that concatenates the node features and edge attributes into a single overall input; is the expression of the th node. Weight calculation: Generated by Softmax weights of attention heads, and the expression is: ； Wherein, is the dynamic weight coefficient of the -th attention head, is the learnable parameter vector, is the total number of attention heads; is the exponential function, is the feature vector after joint encoding of node ; (2) Scene-adaptive attention aggregation; in the region of sudden change of bow curvature dominated by local geometry, the attention aggregation includes: Geometric Focus: Weight Increase, and strengthen the influence of geometric correlation edges when calculating attention; ； Wherein, is the node of the attention weight for neighbor nodes ; is the learnable attention parameter vector, is the learnable weight matrix, is the vector concatenation operation, is the activation function, is the transpose operation to adjust the vector direction, is the learnable weight matrix, is the node of the geometric parameter, is the node of the geometric parameter; Flow field focusing: Weight Reduce, and update the node features to the weighted aggregation of each head; ； In the formula, is the combined parameter after node update, is the activation function, is the weight matrix of the th attention head, is the attention weight of node to neighbor node for the th attention head; (3) Physical constraint-guided weight optimization; add a physical regularization term for the attention weights in the loss function to ensure that the weight distribution conforms to the hydrodynamic laws; ； In the formula, is the physical constraint loss term, which forces the attention weights to conform to the laws of hydrodynamics, is a hyperparameter, is the ideal weight calculated through the equation.

7. The intelligent prediction method for ship resistance according to claim 1, characterized in that In step S4, applying the training strategy of the combined Adam and SGD optimizers includes: Stage combination: Apply the Adam optimizer at the initial stage of training. When the model training approaches the optimal solution, switch to the SGD optimizer; Important parameter combination: For the node weights corresponding to the parameters closely related to the ship geometry, use the SGD optimizer for update; for some of the weights related to the hydrodynamic parameter nodes and the operating condition parameter nodes, use the Adam optimizer. Dynamic weight combination: During the training process, dynamically adjust the weights of the Adam optimizer and the SGD optimizer in parameter updates according to the training effect of the model, and introduce a weight coefficient ; Balance the fast convergence of the Adam optimizer and the stability of the SGD optimizer. The expression is: ； In the formula, is the model parameter of the node weight and edge attribute for the -th iteration; is the model parameter of the node weight and edge attribute for the -th iteration; is the global learning rate, controlling the step size of parameter update; is the dynamic weight coefficient, controlling the mixing ratio of the Adam optimizer and the SGD optimizer; is the momentum term of the Adam optimizer, combining the first-order moment estimation of the historical gradient; is the current batch gradient of the SGD optimizer, directly calculating the derivative of the loss function with respect to the parameter.

8. The intelligent prediction method for ship resistance according to claim 1, wherein In step S5, use the trained dynamic weight graph attention network model to predict the data in the test set, and compare and verify the model prediction results with the results obtained by computational fluid dynamics.

9. The intelligent prediction method for ship resistance according to claim 1, wherein In step S6, according to the results of the prediction evaluation, iteratively optimize the model, including adjusting the network structure, optimizing the loss function, and selecting the optimization algorithm. Finally, apply the optimized model to the prediction of the real ship's sway motion.

10. An intelligent ship resistance prediction system, characterized in that, This system is used to control the ship resistance intelligent prediction method described in any one of claims 1-9. The system includes: A graph attention network construction module, used to construct nodes and edges in the graph attention network and preprocess the original data; A data preprocessing module, used to design and improve the loss function based on physical constraints and node and edge features; A loss function construction module, used to establish a dynamic weight multi-head attention mechanism to automatically assign attention weights to different ship type features; An attention weight automatic assignment module, used to establish a dynamic weight multi-head attention mechanism to automatically assign attention weights to different ship type features; A combined training module, used to apply the training strategy of the Adam and SGD combined optimizer to obtain the graph attention network model; A ship resistance prediction evaluation module, used to perform ship resistance prediction evaluation through the dynamic weight graph attention network model; A model verification and continuous optimization module, used for model verification and continuous optimization.

Images