Physical informed enhanced marine equipment hydrodynamic performance efficient forecasting method and system
By generating supplementary samples through CFD simulation and potential flow theory, the EPK-XGBoost model was constructed, which solved the problems of generalization and computational efficiency in the prediction of ship hydrostatic resistance in small sample scenarios, and achieved high-precision and low-cost prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-10
AI Technical Summary
Existing methods for predicting ship static resistance have poor generalization ability, high computational cost, and weak physical correlation in small sample scenarios, making it difficult to meet the needs of rapid design.
Ship type sample data is generated through CFD simulation, and supplementary samples are generated by the boundary element method of potential flow theory. An EPK-XGBoost model is constructed, and a density adaptive regularization mechanism and a bias penalty term are embedded to improve the prediction accuracy and generalization ability of the model.
It significantly improves the accuracy and generalization ability of hydrostatic resistance prediction, shortens the modeling and prediction cycle, reduces computational costs, and adapts to rapid design requirements.
Smart Images
Figure CN121835017A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marine engineering technology, and more specifically to a method and system for efficient prediction of the hydrodynamic performance of marine equipment with enhanced physical information. Background Technology
[0002] Ship resistance is a key indicator affecting navigational efficiency and hull design, and its accurate prediction is crucial for ship design optimization. Currently, ship still water resistance prediction mainly relies on three types of methods, as detailed below: (1) CFD simulation methods require the construction of a fine numerical model and a large number of iterative calculations, which have problems such as high computational cost and long time consumption, making it difficult to adapt to the needs of rapid design.
[0003] (2) The empirical formula method is based on statistical laws and lacks a deep connection with hydrodynamic mechanism. The prediction results are easily limited by the range of ship type parameters and have poor applicability.
[0004] (3) Traditional machine learning methods such as SVR, Kriging, and conventional XGBoost have limitations in generalization in small sample scenarios, and the error increases significantly when predicting across the range of parameters of the original sample.
[0005] The core reason for the above problems is that existing methods have not effectively integrated the advantages of hydrodynamic physics and data-driven approaches, the parameter space coverage is insufficient under small sample conditions, and the models lack targeted physical constraints and dynamic adjustment mechanisms, making it difficult to balance prediction accuracy and generalization ability.
[0006] Therefore, how to solve the problems of poor generalization of ship static resistance prediction in small sample scenarios, high calculation cost of traditional methods, or weak physical correlation, so as to provide accurate and efficient prediction technology support for ship preliminary design and energy consumption assessment, is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention
[0007] In view of this, the present invention provides a method and system for efficient prediction of the hydrodynamic performance of marine equipment with enhanced physical knowledge, which solves the problems existing in the background technology.
[0008] To achieve the above objectives, the present invention provides the following technical solution: A physically-informed enhanced method for efficient prediction of the hydrodynamic performance of marine equipment includes the following steps: S1. Based on the CFD method, simulate and generate hydrostatic resistance data for several ship type samples, and compare the hydrostatic resistance data with the experimental values to verify the reliability of the data. S2. Transform the original ship type parameters into dimensionless derived features, and generate the original sample set by combining them with the verified hydrostatic resistance data. S3. Identify the parameter gaps in the original sample set through cluster analysis, generate supplementary samples based on the boundary element method of potential flow theory, and combine them with the original sample set to obtain an extended dataset. S4. Embed density adaptive regularization mechanism and bias penalty term in XGBoost framework to construct EPK-XGBoost model; S5. The EPK-XGBoost model is trained and its performance is evaluated using an extended dataset. The best-performing prediction model is used to predict the hydrostatic drag coefficient.
[0009] Optionally, in S1, the ship type sample is generated by parametric modification using the KCS reference ship type as the parent type and the free deformation method.
[0010] Optionally, in S1, the CFD method is specifically implemented as follows: numerical simulation of hydrostatic resistance of the ship model is carried out using STAR-CCM+ software. The governing equation is the three-dimensional Reynolds-averaged Navier-Stokes equation for incompressible fluids. The finite volume method is used for numerical discretization, and the free surface is captured using the volume fraction method.
[0011] Optionally, in S2, the original hull form parameters include the length between perpendiculars. , width B Design draft T Maximum cross-sectional area of bulbous bow Cross section area The profile coefficient ratio is calculated through feature transformation. Aspect Ratio Wide draft ratio B / T .
[0012] Optionally, in S3, the supplementary samples are generated in the following way: Extract ship type and hydrodynamic parameters from the original sample set, and standardize or normalize the hydrodynamic parameters to form a structured parameter matrix; A density-based noise spatial clustering algorithm is used to cluster the parameter matrix to obtain the parameter distribution clusters of the original samples. The parameter gaps between clusters are identified by kernel density estimation. Based on the geometric model of marine equipment, the hydrodynamic calculation domain and boundary conditions under the potential flow theory are determined. The calculation domain is discretized using the boundary element method, and the boundary integral equation of the velocity potential is established and solved. Within the parameter blank interval, Latin hypercube sampling is used to generate several sets of parameter combinations to be supplemented; The sampled parameter combinations are substituted into the boundary element method model to calculate the hydrodynamic performance values under the corresponding parameters, forming a complete supplementary sample containing parameters and hydrodynamic performance values.
[0013] Optionally, in S4, the specific details of the density adaptive regularization mechanism are as follows: Local sample density is calculated using the k-nearest neighbor method. Identify sparse and dense areas of ship operating conditions, as shown below:
[0014] In the formula: For the first i Local density of a sample; The nearest neighbor number controls the statistical range of local density; For the first i one sample Nearest neighbor set, containing neighbors with The closest in features One sample; , The first i , j The input feature vector of each sample; The fitting error is calculated based on the prediction residuals of the current iteration. As shown below:
[0015] In the formula: For the first i The first sample m The true value of each performance target For the first T In the first iteration i The first sample m Predicted values for each performance target T For iteration rounds; M For the number of performance targets, For the first m The weights of each performance objective; Fusion via double exponential function and The dynamic regularization strength is obtained. As shown below:
[0016] In the formula: The basic regularization strength; This is a density-sensitive parameter used to control local density. The degree of influence on the regularization strength >0; This is an error sensitivity parameter that controls the fitting error. The degree of influence on the regularization strength >0; This represents the maximum fitting error during the iteration.
[0017] Optionally, in S4, the deviation penalty term is a penalty term positively correlated with the fluid dynamics law added to the loss function, expressed as:
[0018] In the formula: For multi-objective loss terms, For the true value, For predicted values, n The total number of training samples; For tree structure regularization, K The total number of gradient boosting trees, For the first in the gradient boosting tree k A decision tree; For physical constraint violation degree, For velocity potential, These are the physical constraint weighting coefficients; Among them, the multi-objective loss term adopts a weighted mean square loss, adapted to the engineering priority of multiple performance objectives of the ship, and takes the following form:
[0019] In the formula: M For the number of performance targets, For the first m The weights of each performance objective satisfy... ; The tree structure regularization term uses XGBoost's native tree complexity regularization, in the form of:
[0020] In the formula: For the first k The number of leaf nodes in a tree. For the first k Tree No. j The weight of each leaf node; The penalty coefficient for leaf nodes controls the complexity of the tree; Physical constraint violation The degree to which the physical constraints of velocity potential and wave-making resistance are violated is quantified to ensure that the prediction results conform to the basic laws of ship hydrodynamics. Its expression is:
[0021] In the formula: velocity potential For violations of the Laplace equation, irrotational incompressible fluids satisfy... This ensures the irrotation and incompressibility of the flow field; Wave drag coefficient With Fu Rude The proportional constraint is violated. h It is a proportionality constant to ensure that the wave-making drag changes with speed in accordance with physical expectations.
[0022] Optionally, in S5, the extended dataset is divided into training and test sets in a 7:3 ratio. At the same time, KCS variant ship types that exceed the parameter range of the ship type samples in S1 are selected as cross-interval test sets to verify the model's regular prediction accuracy and cross-parameter interval generalization performance.
[0023] Optionally, in S5, the quantitative indicators for performance evaluation include root mean square error, coefficient of determination, and mean absolute error. The model with the best performance across all indicators is selected as the final forecast model.
[0024] A highly efficient prediction system for the hydrodynamic performance of marine equipment with enhanced physical knowledge, performing any of the methods described above, includes: The data preparation and verification module is used to generate hydrostatic resistance data for several ship type samples through CFD simulation, and compare the hydrostatic resistance data with experimental values to verify the reliability of the data. The feature enhancement module is used to transform the original ship type parameters into dimensionless derived features, and combine them with the verified hydrostatic resistance data to generate the original sample set. The sample expansion module is used to identify parameter gaps in the original sample set through cluster analysis, generate supplementary samples based on the boundary element method of potential flow theory, and combine them with the original sample set to obtain an expanded dataset. The Physically Informed Enhancement Modeling Module is used to embed a density adaptive regularization mechanism and a bias penalty term into the XGBoost framework to build an EPK-XGBoost model. The model training and validation module is used to train and evaluate the performance of the EPK-XGBoost model using an expanded dataset, and to predict the hydrostatic drag coefficient using the best-performing prediction model.
[0025] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for efficient prediction of hydrodynamic performance of marine equipment with enhanced physical knowledge, which has the following beneficial effects: (1) Significantly improved prediction accuracy: The mean square error of the EPK-XGBoost model in this invention for predicting hydrostatic resistance coefficient is as low as 0.032, which is far more accurate than traditional methods; (2) Breakthrough in generalization ability: In the interval test beyond the range of the original sample parameters, the model RMSE remains stable at 0.038, while the errors of traditional XGBoost, SVR and Kriging method increase to 0.051, 0.091 and 0.069 respectively, effectively solving the generalization shortcoming of cross-parameter interval prediction. (3) Significantly optimized computational efficiency: This invention can avoid the lengthy computational process of CFD simulation, and does not rely on a large number of experimental samples. The modeling and prediction cycle is significantly shortened compared to CFD simulation, which is suitable for rapid design requirements. (4) Outstanding engineering application value: This invention can directly support the optimization of bulbous bow lines and the energy consumption assessment in the preliminary design stage of ships, reduce redundant design and testing links, reduce ship research and development costs, promote the development of ship design towards precision and efficiency, and conform to the industry trend of energy conservation and consumption reduction. Attached Figure Description
[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0027] Figure 1 A schematic diagram of the computational domain and boundary conditions provided by the present invention; Figure 2 The EPK-XGBoost model framework diagram provided by this invention; Figure 3 The flowchart of the EPK-XGBoost model provided by this invention; Figure 4 A flowchart for predicting hydrostatic resistance provided by this invention. Detailed Implementation
[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0029] This invention discloses a physically-informed enhanced method for efficient prediction of the hydrodynamic performance of marine equipment. This method overcomes the generalization bottleneck of traditional models in small-sample scenarios by using a machine learning framework enhanced by physical mechanisms, combined with feature engineering and sample expansion strategies. Figure 4 As shown, the specific steps include: S1. Based on the CFD method, simulate and generate hydrostatic resistance data for several ship type samples, and compare the hydrostatic resistance data with the experimental values to verify the reliability of the data. S2. Transforming the original ship type parameters into dimensionless derived features and combining them with the verified hydrostatic resistance data to generate the original sample set can enhance the physical correlation of features and reduce dimensional redundancy. S3. By identifying the parameter gaps in the original sample set through cluster analysis, supplementary samples are generated based on the boundary element method (BEM) of potential flow theory. Combined with the original sample set, an extended dataset is obtained, which can expand the coverage of the dataset. S4. Embed density adaptive regularization and bias penalty terms into the XGBoost framework to construct the EPK-XGBoost model. The density adaptive regularization mechanism dynamically adjusts the regularization strength according to the distribution density of samples in the parameter space, and strengthens the constraint in the low-density region. The bias penalty term is a penalty term positively correlated with the laws of fluid mechanics, such as the nonlinear relationship between drag and velocity, added to the loss function to force the model to follow the physical mechanism. S5. The EPK-XGBoost model is trained and its performance is evaluated using an extended dataset. The best-performing prediction model is used to predict the hydrostatic drag coefficient.
[0030] To make the objectives and technical solutions of this invention clearer, the technical solutions of this invention will be clearly described below with reference to the accompanying drawings.
[0031] 1. 3D modeling and CFD simulation of the sample ship type
[0032] In S1 of this embodiment, the ship type sample is generated by parametric modification using the KCS benchmark ship type officially released by KRISO as the parent type and the Free Deformation (FFD) method. In this embodiment, 100 KCS samples with different line types were generated.
[0033] Furthermore, in S1, the specific implementation of the CFD method is as follows: numerical simulation of hydrostatic resistance of the ship model is carried out using STAR-CCM+ software. The governing equation is the three-dimensional Reynolds-averaged Navier-Stokes (RANS) equation for incompressible fluids. Numerical discretization is performed using the finite volume method (FVM), and the free surface is captured using the volume fraction (VOF) method.
[0034] Fluid domain geometry design aims to eliminate boundary disturbances, such as... Figure 1 As shown. Along the ship's length direction (x-axis), take 3 times the length of the ship model, with the front of the ship model extending 1 times the ship's length and the rear extending 2 times the ship's length; along the ship's width direction (y-axis), take 1 time the ship's length to ensure that the wave system on the ship's side develops fully and does not touch the domain wall; along the draft direction (z-axis), take 3 times the ship's length, with the underwater part being 2 times the ship's length and the above-water part being 1 time the ship's length, to accurately simulate the free surface wave pattern.
[0035] The boundary conditions are set as follows: the fluid domain inlet, top, bottom, and sidewalls adopt velocity inlet boundaries, with the incoming flow velocity calculated based on Fr=0.26; the outlet adopts a pressure outlet boundary; the hull surface adopts a no-slip wall boundary with a surface roughness of 0.0015mm; the fluid domain symmetry plane adopts a symmetric boundary. The time step is determined according to the Courant-Friedrichs-Lewy (CFL) condition to ensure stable flow field evolution.
[0036] 2. Calculation of dimensionless characteristics
[0037] In S2 of this embodiment, taking the KCS container ship as an example, the original ship type parameters include the length between perpendiculars. , width B Design draft T Maximum cross-sectional area of bulbous bow Cross section area The profile coefficient ratio is calculated through feature transformation. Aspect Ratio Wide draft ratio B / T The input space constructed from three sets of dimensionless features significantly improves the model's ability to capture changes in the flow field at the bulbous bow.
[0038] 3. Parameter Space Sample Expansion
[0039] In S3 of this embodiment, the supplementary sample is generated in the following way: 1) Preprocessing of parameters in the original sample set: Extract the ship type and hydrodynamic parameters from the original sample set, and standardize or normalize the hydrodynamic parameters to form a structured parameter matrix. 2) Cluster analysis to identify parameter blank intervals: The density-based noise applied spatial clustering (DBSCAN) algorithm is used to cluster the parameter matrix to obtain the parameter distribution clusters of the original samples. Low-density regions between clusters are identified by kernel density estimation. These low-density regions are defined as "parameter blank intervals", that is, the range of parameter combinations not covered by the original samples. 3) Constructing a potential flow theory-boundary element method calculation model: Based on the geometric model of marine equipment, the hydrodynamic calculation domain and boundary conditions under the potential flow theory are determined, the calculation domain is discretized using the boundary element method, and the boundary integral equation of the velocity potential is established and solved. 4) Parameter sampling in blank intervals: Latin hypercube sampling is used in the parameter blank intervals to generate several sets of parameter combinations to be supplemented; 5) Calculation of hydrodynamic performance of supplementary samples: Substitute the sampled parameter combination into the boundary element method model in step 3) to calculate the hydrodynamic performance value under the corresponding parameters, forming a complete supplementary sample containing parameters and hydrodynamic performance values.
[0040] 4. EPK-XGBoost Model Training
[0041] The core principle of the EPK-XGBoost model is to integrate physical constraints and dynamic regularization within the gradient boosting framework, enabling collaborative learning of data features and physical laws under small sample conditions. For example... Figure 2 As shown, its basic architecture is based on the XGBoost gradient boosting tree, supporting multi-objective parallel prediction through a multi-output node design, and also adding a physical awareness module and a dynamic regularization unit. The physical awareness module transforms hydrodynamic constraints into penalty terms of the objective function using the Lagrange multiplier method, while the dynamic regularization unit adjusts the regularization strength in real time according to the sample density and fitting error, strengthening regularization in sparse regions to suppress overfitting and weakening regularization in dense regions to enhance learning ability.
[0042] First, in S4 of this embodiment, the deviation penalty term is a penalty term positively correlated with the fluid dynamics law added to the loss function, and its expression is: (1); In the formula: For multi-objective loss terms, For the true value, For predicted values, n The total number of training samples; For tree structure regularization, K The total number of gradient boosting trees, For the first in the gradient boosting tree k A decision tree; For physical constraint violation degree, It is the velocity potential; The physical constraint weight coefficient is determined through validation set optimization, with an initial value set to 0.2.
[0043] Among them, the multi-objective loss term adopts a weighted mean square loss, adapted to the engineering priority of multiple performance objectives of the ship, and takes the following form: (2); In the formula: M For the number of performance targets, For the first m The weights of each performance objective satisfy... .
[0044] The tree structure regularization term uses XGBoost's native tree complexity regularization, in the form of: (3); In the formula: For the first k The number of leaf nodes in a tree. For the first k Tree No.j The weight of each leaf node; The penalty coefficient for leaf nodes controls the complexity of the tree.
[0045] Physical constraint violation The degree to which the physical constraints of velocity potential and wave-making resistance are violated is quantified to ensure that the prediction results conform to the basic laws of ship hydrodynamics. Its expression is: (4); In the formula: velocity potential For violations of the Laplace equation, irrotational incompressible fluids satisfy... This ensures the irrotation and incompressibility of the flow field; Wave drag coefficient With Fu Rude The proportional constraint is violated. h It is a proportionality constant to ensure that the wave-making drag changes with speed in accordance with physical expectations.
[0046] The execution flow of the EPK-XGBoost model can be divided into six core stages: data preparation, model initialization, objective function construction, dynamic regularization calculation, gradient boosting iterative optimization, prediction, and evaluation. Figure 3 As shown, the main steps are as follows: (1) Data preparation and preprocessing Ship performance data is collected, covering input features and output performance indicators. The data is then processed, missing and outlier values are handled, and the dataset is divided into training, validation, and test sets. Simultaneously, features are standardized or normalized to eliminate the impact of dimensional differences on the model.
[0047] (2) Model initialization
[0048] Set the basic parameters of XGBoost, including the number of trees K, the maximum depth of a single tree max_depth, and the learning rate. Etc. Initialize physical constraint weights. The initial value is set to 0.2, and subsequent optimization is performed using a validation set; dynamic regularization parameters are initialized, etc.
[0049] (3) Construction of the objective function
[0050] Construct a comprehensive objective function as shown in formula (1), integrating multi-objective loss terms, tree structure regularization terms, and physical constraint violation terms.
[0051] (4) Dynamic regularization calculation
[0052] In S4 of this embodiment, the specific content of the density adaptive regularization mechanism is as follows: Local sample density is calculated using the k-nearest neighbor method. Identify sparse and dense areas of ship operating conditions, as shown below: (5); In the formula: For the first i Local density of a sample; The nearest neighbor number controls the statistical range of local density; for ship scenarios, it is usually set to 5~10. For the first i one sample Nearest neighbor set, containing neighbors with The closest in features One sample; , The first i , j The input feature vector of each sample.
[0053] The fitting error is calculated based on the prediction residuals of the current iteration. As shown below: (6); In the formula: For the first i The first sample m The true value of each performance target For the first T In the first iteration i The first sample m Predicted values for each performance target T For iteration rounds; M For the number of performance targets, For the first m The weights of each performance objective.
[0054] Fusion via double exponential function and The dynamic regularization strength is obtained. As shown below: (7); In the formula: The basic regularization strength is typically set to 0.1 to 0.3. This is a density-sensitive parameter used to control local density. The degree of influence on the regularization strength >0; This is an error sensitivity parameter that controls the fitting error. The degree of influence on the regularization strength >0; This represents the maximum fitting error during the iteration.
[0055] (5) Gradient boosting iterative optimization
[0056] For each iteration Perform the following operations: (I) Calculate the first gradient and second gradient Based on the current model prediction The changing trend of quantified loss.
[0057] (II) Constructing a regression tree: combining dynamic regularization strength The leaf node weights are optimized using equation (8); the optimal splitting feature and threshold are selected based on the gain maximization criterion (equation (9)) to generate the first splitting node. T Tree of return ; (8); (9); In the formula: For the first t Tree No. j The weight vector of leaf nodes, where each element corresponds to a weight of a performance objective. ; For the first t Tree No. j The sample set contained in each leaf node; For the first T In the first iteration, the... i The second gradient of each sample; For the first T In the first iteration, the... i The first-order gradient of each sample; , For the first t The total number of leaf nodes in the tree; The sample set of the left child nodes after the split. The set of samples for the right child nodes after the split. This is the sample set of the parent node before the split.
[0058] (III) Update model predictions .
[0059] (IV) Determine if the iteration has converged: If the number of iterations reaches the preset number of trees. K or loss function If the convergence condition is met, stop the iteration; otherwise, return to (Ⅰ) to continue the next round.
[0060] (6) Model prediction and performance evaluation
[0061] The trained model is used to predict the performance of the test set samples to obtain the predicted value of the target. The quantitative indicators for performance evaluation include root mean square error, coefficient of determination, and mean absolute error. The model with the best performance across all indicators is selected as the final forecast model.
[0062] 5. Model Validation and Comparison
[0063] In S5 of this embodiment, the extended dataset (100 samples) is divided into a training set (70 samples) and a test set (30 samples) in a 7:3 ratio. At the same time, 10 KCS variant ship types that exceed the parameter range of the ship type samples in S1 are selected as cross-interval test sets to verify the model's conventional prediction accuracy and cross-parameter interval generalization performance.
[0064] Four models were trained separately. The traditional XGBoost model had the same base parameters as EPK-XGBoost, the SVR model used a radial basis function kernel, and the Kriging model used a Gaussian covariance function. Test results showed: EPK-XGBoost RMSE = 0.032 on the test set and RMSE = 0.038 on the cross-interval test set; traditional XGBoost RMSE = 0.041 on the test set and RMSE = 0.051 on the cross-interval test set; SVR RMSE = 0.063 on the test set and RMSE = 0.091 on the cross-interval test set; and Kriging RMSE = 0.069 on the test set and RMSE = 0.102 on the cross-interval test set, verifying the superiority of the method of this invention.
[0065] and Figure 4 Corresponding to the method described above, this embodiment of the invention also provides a physically-informed enhanced high-efficiency prediction system for the hydrodynamic performance of marine equipment, used for... Figure 4 The specific implementation of the method, the physical knowledge-enhanced high-efficiency prediction system for hydrodynamic performance of marine equipment provided in this embodiment of the invention, can be applied to computer terminals or various mobile devices, specifically including: The data preparation and verification module is used to generate hydrostatic resistance data for several ship type samples through CFD simulation, and compare the hydrostatic resistance data with experimental values to verify the reliability of the data. The feature enhancement module is used to transform the original ship type parameters into dimensionless derived features, and combine them with the verified hydrostatic resistance data to generate the original sample set. The sample expansion module is used to identify parameter gaps in the original sample set through cluster analysis, generate supplementary samples based on the boundary element method of potential flow theory, and combine them with the original sample set to obtain an expanded dataset. The Physically Informed Enhancement Modeling Module is used to embed a density adaptive regularization mechanism and a bias penalty term into the XGBoost framework to build an EPK-XGBoost model. The model training and validation module is used to train and evaluate the performance of the EPK-XGBoost model using an expanded dataset, and to predict the hydrostatic drag coefficient using the best-performing prediction model.
[0066] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0067] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment, characterized in that, The method comprises the following steps: S1, based on the CFD method simulation to generate a number of ship type sample static water resistance data, the static water resistance data and experimental value comparison, verify the reliability of the data; S2, the original ship type parameters into dimensionless derived features, combined with the verified static water resistance data to generate the original sample set; S3, through cluster analysis to identify the parameter blank interval of the original sample set, based on the boundary element method of potential flow theory to generate supplementary samples, combined with the original sample set to obtain the expanded data set; S4, embedding density adaptive regularization mechanism and bias penalty term in XGBoost framework, constructing EPK-XGBoost model; S5, using the expanded data set to train and evaluate the performance of the EPK-XGBoost model, and using the performance optimal prediction model to predict the static water resistance coefficient.
2. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S1, the ship type sample is based on KCS benchmark ship type, and the free deformation method is used for parameterization modification.
3. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S1, the specific implementation of the CFD method is: through STAR-CCM+ software to carry out numerical simulation of ship model static water resistance, control equation selects three-dimensional Reynolds average Navier-Stokes equation of incompressible fluid, adopts finite volume method for numerical discretization, and captures free surface through volume fraction method.
4. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S2, the original ship type parameters include the length between perpendiculars , the breadth B , the designed draft T , the maximum sectional area of the bulbous bow , the midship sectional area , and the sectional coefficient ratio , the length-breadth ratio , and the breadth-draft ratio B / T are calculated by feature conversion.
5. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S3, the generation method of supplementary samples is as follows: Extract the ship type and hydrodynamic parameters in the original sample set, standardize or normalize the hydrodynamic parameters to form a structured parameter matrix; Using the density-based noise application spatial clustering algorithm to cluster the parameter matrix, obtaining the parameter distribution cluster of the original sample, and identifying the parameter blank interval between the clusters through kernel density estimation; Based on the geometric model of marine equipment, determine the hydrodynamic calculation domain and boundary conditions under the potential flow theory, discretize the calculation domain by using the boundary element method, establish the boundary integral equation of velocity potential and solve it; In the parameter blank interval, use Latin hypercube sampling to generate several groups of parameter combinations to be supplemented; Put the parameter combinations obtained by sampling into the boundary element method model to calculate the hydrodynamic performance value under the corresponding parameters, and form a complete supplementary sample containing parameters and hydrodynamic performance value.
6. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S4, the specific content of the density adaptive regularization mechanism is: Local sample density is computed by k-neighbor method Sparse and dense areas of ship working conditions are identified as follows: where: is the local density of the i th sample; is the number of neighbors, controlling the statistical range of the local density; is the local density of the i th sample; is the set of neighbors of the th sample, containing the samples closest to the th sample; are the input feature vectors of the i th and j th samples, respectively; calculating a fitting error based on the prediction residual of the current iteration as follows: In the formula: For the first i The first sample m The true value of each performance target For the first T In the first iteration i The first sample m Predicted values for each performance target T For iteration rounds; M For the number of performance targets, For the first m The weights of each performance objective; By a double exponential function fusion and , obtaining a dynamic regularized intensity as follows: wherein: is the base regularization strength; is a density sensitivity parameter, controlling the local density to which the regularization strength is affected, > 0; is an error sensitivity parameter, controlling the fitting error to which the regularization strength is affected, > 0; is the maximum fitting error for iteration.
7. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S4, the bias penalty term is to add a penalty term that is positively related to the law of fluid mechanics in the loss function, and the expression is: In the formula: For multi-objective loss terms, For the true value, For predicted values, n The total number of training samples; For tree structure regularization, K The total number of gradient boosting trees, For the first in the gradient boosting tree k A decision tree; For physical constraint violation degree, For velocity potential, These are the physical constraint weighting coefficients; Wherein, the multi-objective loss term adopts weighted mean square loss, which adapts to the engineering priority of ship multi-performance target, and the form is: In the formula: M is the number of performance targets, is the weight of the m th performance target, satisfying ; The tree structure regularization term adopts the original tree complexity regularization of XGBoost, and the form is: wherein: is the number of leaf nodes of the tree k is the number of leaf nodes of the tree is the weight of the leaf node of the tree k is the weight of the leaf node of the tree j is the weight of the leaf node of the tree is the weight of the leaf node of the tree Physical constraint violation degree The physical constraint violation degree of the quantized velocity potential and the wave-making resistance is calculated to ensure that the prediction results conform to the basic law of ship hydrodynamics, and the expression is as follows: where is the velocity potential is the amount of violation to Laplace equation , which ensures the irrotationality and incompressibility of the flow field is the wave-making resistance coefficient is the Froude number is the proportional constraint violation amount h is the proportional constant, which ensures that the wave-making resistance varies with the speed in line with physical expectations.
8. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S5, the expanded data set is divided into training set and test set according to the ratio of 7:3, and the KCS variation ship type beyond the ship type sample parameter range in S1 is selected as the cross-interval test set, to verify the general prediction accuracy and cross-parameter interval generalization performance of the model.
9. A physically informed enhanced method for efficient prediction of hydrodynamic performance of marine equipment according to claim 1, characterized in that, In S5, the quantitative indicators of performance evaluation include root mean square error, determination coefficient and mean absolute error, and the model with the optimal indicators is selected as the final prediction model.
10. A physically informed enhanced marine equipment hydrodynamic performance high- efficiency prediction system performing the method according to any one of claims 1-9, characterized by, It comprises: The data preparation and verification module is used to generate hydrostatic resistance data for several ship type samples through CFD simulation, and compare the hydrostatic resistance data with experimental values to verify the reliability of the data. The feature enhancement module is used to transform the original ship type parameters into dimensionless derived features, and combine them with the verified hydrostatic resistance data to generate the original sample set. The sample expansion module is used to identify parameter gaps in the original sample set through cluster analysis, generate supplementary samples based on the boundary element method of potential flow theory, and combine them with the original sample set to obtain an expanded dataset. The Physically Informed Enhancement Modeling Module is used to embed a density adaptive regularization mechanism and a bias penalty term into the XGBoost framework to build an EPK-XGBoost model. The model training and validation module is used to train and evaluate the performance of the EPK-XGBoost model using an expanded dataset, and to predict the hydrostatic drag coefficient using the best-performing prediction model.