A ship type optimization method based on transfer learning

Abstract

The application discloses a ship type optimization method based on transfer learning, and relates to the technical field of ship type optimization.The application improves the precision and generalization performance of the final ship performance model by using a Two Stage TrAdaBoost.R2 regression transfer algorithm.The algorithm uses the knowledge of some known related data sets to assist the target ship type data modeling, so as to improve the precision of the model.The application is divided into two stages for ship type sampling and modeling: a coarse proxy model is constructed based on global scale sampling data, and a fine proxy model is constructed based on the data obtained through the global and local scale sampling stages, and the two models jointly assist the subsequent optimization process.A double proxy assisted optimization method is proposed for the optimization process.The iteration process of the adopted optimization algorithm is improved, and the single proxy model is no longer updated independently, but the fitness values of the two proxy models are considered simultaneously in the optimization process, and the optimal one is selected for iteration.

Description

Technical Field

[0001] This invention belongs to the field of ship type optimization technology, specifically involving a Two StageTrAdaBoost.R2 method based on transfer learning. Background Technology

[0002] In shipbuilding engineering, hull design and optimization are crucial aspects. Optimizing the hull shape can improve a vessel's performance and efficiency, thereby reducing energy consumption and emissions. However, due to the high cost of designing and testing ship models, traditional model testing methods often lead to exorbitant costs, especially for small vessels or non-commercial vessels. Currently, computer simulation-based hull optimization methods have become a popular area in shipbuilding engineering. Computer simulation can significantly reduce the cost of model testing while rapidly obtaining optimized ship solutions. Commonly used CFD simulation software includes Fluent and OpenFORM. Figure 1 As shown. However, in ship hull optimization, obtaining experimental or simulation data also requires significant time and computational resources, primarily in mesh generation and simulation calculations. Therefore, this invention utilizes existing experimental data to assist datasets with only a small number of samples, establishing a high-precision surrogate model to replace CFD simulation calculations for data acquisition, optimizing the ship hull, and providing reasonable optimization evaluation results, thereby improving the ship's performance and efficiency.

[0003] This invention provides a ship hull optimization method based on the Two-Stage TrAdaBoost.R2 regression migration method, characterized by:

[0004] (1) Reduce modeling costs: Obtaining sufficient ship model data requires high costs, but the Two StageTrAdaBoost.R2 algorithm can improve this situation. It can transfer knowledge from other datasets to a sample set with insufficient data for auxiliary modeling.

[0005] (2) Improve model accuracy: The Two Stage TrAdaBoost.R2 algorithm can improve the prediction accuracy of target domain data by using knowledge of source domain data to build target domain model.

[0006] (3) Accelerate convergence speed: The hybrid optimization strategy of dual proxy model is used, which can not only accelerate the convergence speed of the algorithm, but also significantly reduce the probability of the algorithm getting stuck in local optima. Summary of the Invention

[0007] This invention presents a ship hull modeling and optimization process based on the Two-Stage TrAdaBoost.R2 method to address the challenge of establishing a ship proxy model with limited experimental data, followed by optimization of the ship's hull lines. The process includes the following steps:

[0008] A ship hull optimization method based on transfer learning, the specific steps of which are as follows:

[0009] (1) Obtain the publicly available ship model experiment dataset as the source domain dataset, and perform preprocessing and feature processing on the source domain data;

[0010] Furthermore, the preprocessing of the source domain dataset includes: selecting sample data with similar characteristics to the target domain dataset from the source domain dataset, and removing samples from the source domain dataset that are significantly different from the target domain dataset; and feature processing of the source domain dataset includes feature alignment, feature mapping, or feature transformation, so that the data of the source domain and the target domain are represented in the same feature space.

[0011] (2) The target domain dataset is obtained by sampling the initial ship that needs to be optimized. Based on the two aspects of global ship design and local ship design, the sampling and modeling of the target domain data are divided into two corresponding stages: a coarse surrogate model based on global scale sampling data and a fine surrogate model based on local scale sampling data. The two surrogate models are used together in the subsequent optimization process.

[0012] Furthermore, global-scale sampling includes sampling of major dimensions such as ship length, beam, and depth, while local-scale sampling refers to sampling of the dimensions of a specific location on the ship.

[0013] Furthermore, the target domain dataset undergoes feature processing including feature alignment, feature dimensionality reduction, and feature expansion. Preprocessing of the target domain dataset involves extracting the correlations between features based on different ship performance characteristics through correlation analysis, followed by dimensionless and standardized preprocessing to facilitate subsequent model training. After sampling, the required performance data is obtained through CFD simulation, resulting in a labeled target domain dataset.

[0014] (3) The process of constructing the proxy model in step (2): Using the Two Stage TrAdaBoost.R2 method, the source domain data is used to assist in modeling the target domain data. The two stages of modeling yield a coarse proxy model and a fine proxy model, respectively.

[0015] (4) The established coarse and fine surrogate models are applied to the optimization process. The improved QPSO optimization algorithm is used to optimize the ship type data in the target domain to improve the ship's performance and efficiency. The improved QPSO optimization algorithm does not solve the two objectives independently during the optimization process. Instead, it considers both objectives simultaneously during all particle iterations and always selects the better one before proceeding with subsequent iterations.

[0016] Furthermore, under the constraints of the design variables, the ship's shape was optimized under different conditions to meet various practical requirements. Finally, the optimal solution was numerically reconstructed to obtain the ship's shape, and the optimization results were verified through CFD simulation. The improved QPSO optimization algorithm was used to search for the target throughout the optimization process.

[0017] Furthermore, during the optimization process, the function value is selected in the t-th iteration. Where f(X) t ) represents the particle fitness value in the t-th iteration. Represents the i-th particle The fitness value of the coarse surrogate model in the t-th iteration. Represents the i-th particle The fitness value of the fine-matrix model in the t-th iteration. N represents the population size of the particles.

[0018]

[0019] The beneficial effects of this invention are:

[0020] (1) Using the Two Stage TrAdaBoost.R2 method, a proxy model is constructed based on known experimental data to optimize the ship type data of the target domain, which greatly reduces the cost of ship model testing.

[0021] (2) By optimizing the parameters of the proxy model, its prediction accuracy and precision are improved, making the optimization scheme for ships more accurate and reliable.

[0022] (3) This method has good universality and can be used for regression transfer learning in different ship performance and other fields, and has a wide range of application prospects.

[0023] (4) The optimization process is improved so that the two objectives are no longer solved independently, but are considered simultaneously during all particle iterations and the better one is always selected. One advantage of this is that it can help the target escape local optima more quickly when it gets stuck in a local optimum, and accelerate the convergence speed. Attached Figure Description

[0024] Figure 1 The process of creating a mesh for calculating ship performance using CFD simulation software.

[0025] Figure 2 This refers to the relationships between various resistances encountered by a ship during navigation.

[0026] Figure 3 Distribution of cargo vessel data generated by Latin hypercube sampling.

[0027] Figure 4 The image shows the prediction results of ship performance using the AdaboostRegression method.

[0028] Figure 5 This is a diagram showing the prediction results of ship performance using the Decision Tree method.

[0029] Figure 6 The image shows the prediction results of ship performance using the RBF neural network method.

[0030] Figure 7 This is a diagram showing the prediction results of ship performance using the random forest method.

[0031] Figure 8 The image shows the prediction results of ship performance using the Two Stage TrAdaBoost.R2 method.

[0032] Figure 9 This is a plot of the prediction mean square error for large-scale sampled data.

[0033] Figure 10 This is a plot of the prediction mean square error for small-scale sampled data.

[0034] Figure 11 This is a graph showing the fitting effect of large-scale sampled data.

[0035] Figure 12 This is a graph showing the fitting effect for small-scale sampled data.

[0036] Figure 13 This is a schematic diagram of the large-scale sampling points in the first stage.

[0037] Figure 14 This is a schematic diagram of the second stage of fine-tuning the sampling points.

[0038] Figure 15 Flowchart for building an agent model based on transfer learning.

[0039] Figure 16 A comparison chart showing the convergence speed of different optimization strategies.

[0040] Figure 17 A comparison of convergence speeds for different optimization strategies when the ship's width is fixed.

[0041] Figure 18This is a comparison of the convergence speeds of different optimization strategies for fixed depths.

[0042] Figure 19 This is a comparison of the convergence speeds of different optimization strategies when the number of particles is 20.

[0043] Figure 20 This is a comparison of the convergence speeds of different optimization strategies when the number of particles is 50.

[0044] Figure 21 This is a comparison of the convergence speeds of different optimization strategies when the number of particles is 80.

[0045] Figure 22 This is a comparison of the convergence speeds of different optimization strategies when the number of particles is 100.

[0046] Figure 23 A comparison chart showing the number of times different optimization strategies get stuck in local optima.

[0047] Figure 24 A comparison chart showing the number of times different optimization strategies get stuck in local optima when the ship's width is fixed.

[0048] Figure 25 This is a comparison chart showing the number of times different optimization strategies get trapped in local optima when the model is in a fixed depth.

[0049] Figure 26 A comparison chart showing the number of times the particle count gets trapped in local optima under different optimization strategies for different particle counts. Detailed Implementation

[0050] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0051] 1. Method Overview

[0052] Two-Stage TrAdaBoost.R2

[0053] The Two-Stage TrAdaBoost.R2 method is proposed based on the TrAdaBoost and AdaBoost.R2 methods. TrAdaBoost is used to solve classification problems with a source domain. Its core idea is to transfer knowledge from the source domain data to the target domain data, improving classification performance and generalization ability in the target domain even with insufficient data or different data distributions. AdaBoost.R2's basic idea is to iteratively train a series of weak regression models and then weight them to form a strong regression model. In each iteration, it adjusts the sample weights based on the error between the current model's prediction and the actual result, focusing more attention on samples where previous models predicted poorly in the next iteration, thereby improving the overall model accuracy.

[0054] Combining the principles of AdaBoost.R2 and TrAdaBoost, a new regression algorithm, TrAdaBoost.R2, was created. Pardoe et al. observed that TrAdaBoost.R2 is highly susceptible to overfitting; outside of certain points, prediction accuracy decreases with increasing iteration count N. While AdaBoost.R2 seems to avoid this problem, it tends to gradually reduce the weights of samples similar to the target instance to zero, a phenomenon exacerbated in TrAdaBoost.R2. TrAdaBoost, on the other hand, effectively maintains the final weight distribution of the samples.

[0055] To address the two issues mentioned above, a Two-Stage TrAdaBoost.R2 algorithm is proposed. This algorithm adjusts sample weights in two stages: In the first stage, a lower bound for the weights is determined through F-fold cross-validation, and the weights of the source instances are gradually adjusted downwards until this lower bound is reached during iterations; In the second stage, the weights of the target instances are updated normally in AdaBoost.R2, while the weights of all source instances are no longer updated. Furthermore, only the hypotheses generated in the second stage are stored and used in the output of the final model. Note the weighting factor β. t Instead of selecting based on the assumption error as before, the target instance is selected to impart a certain total weight. In this way, the total weight of the target instance increases uniformly from an initial value m / (n+m) to 1 after S iterations. This is done to impart a certain total weight to the target instance, thereby mitigating the overfitting effect caused by weight shift. The algorithm flow is shown below.

[0056] The specific steps of the Two-Stage TrAdaBoost.R2 algorithm are as follows:

[0057] Step (a), let the source domain dataset be T. source The data volume is n, and the target domain dataset is T. target The dataset has a data size of m, a number of steps of S, a maximum number of augmentation iterations of N, a cross-validation fold number of F, and a learner as the base model. Let the dataset T be T. source and T target The union of the sets has a size of n+m, and the first n instances come from T. source Set the initial weight vector in

[0058] Step (b) involves using the AdaBoost.R2 algorithm on dataset T to create the weight vector w for the t-th iteration. t N and Learner obtain the modelt However, in each iteration, for The weights of the first n instances are not updated. Similarly, F-fold cross-validation is used to obtain the model. t Error estimation t ;

[0059] Step (c) involves using the dataset T and weights w t Calling learner yields a mapping relationship h. t :X→R;

[0060] Where R represents X with respect to mapping h t The image, X represents R with respect to mapping h. t The original image

[0061] Step (d) calculates the error for each sample.

[0062] make

[0063] The relative error is

[0064] Calculate the adjustment error ε t :

[0065]

[0066] Computational model t Weighting factor β t :

[0067] β t =ε t / (1-ε t )

[0068] Among them, E t and These represent the maximum absolute error and the relative error, respectively, x i x j Let y represent the value of the variable in T. j x represents j The true label value, h t (x j ) represents x j The predicted value of the function;

[0069] Step (e), update the weight vector

[0070]

[0071] Among them, Z tFor the normalization coefficients, since the weights of the last m target instances were not updated in the first stage (only simple normalization was performed), the final weights for these m instances will be adjusted as follows:

[0072] Step (f) involves iterating through steps (b) to (e) S times.

[0073] Step (g), output model k , where k = argmin v error v v = 1, 2, ..., S.

[0074] Finally, it's important to note that if the source and target domains have different feature dimensions, operations such as feature alignment, feature mapping, or feature transformation may be necessary to ensure that the data in both domains are represented in the same feature space. This is essential to provide the appropriate input for the Two-Stage TrAdaBoost.R2 algorithm, enabling effective knowledge transfer between the source and target domains.

[0075] Two-stage sampling

[0076] The sampling process combines sampling methods with object deformation methods to expand the target ship data sample. An initial ship requiring optimization is selected as the mother ship, and hull deformation and reconstruction are achieved through both large-scale and local small-scale deformations. This is conducted in two stages: the first stage primarily alters the ship's main dimensions (length, width, and height) to change its overall shape; the second stage modifies variables affecting the ship's local features. Through these two stages, sub-ship types with different length-to-beam ratios, beam-to-draft ratios, and different local features can be obtained.

[0077] Hybrid optimization strategy

[0078] For the constructed dual-proxy model, the final optimal value cannot be absolutely guaranteed to appear in either proxy model. In this case, independent optimization is a feasible method, but it suffers from two problems: First, the optimization time issue. When iterating independently, the total number of iterations will be the sum of the two parts, requiring significant iteration time and computational resources, which increase rapidly with the dimensionality of the problem or the size of the search space. Second, the convergence speed issue. Solving independently is prone to getting stuck in local optima, or requiring many iterations to escape a local solution, and the probability of successfully escaping a local solution is still relatively small. Often, all particles will converge to the same local optimum. This slows down the convergence to the final solution, and the final result may only converge to a single optimal solution.

[0079] Based on the two issues mentioned above, this patent modifies the iterative process of the optimization algorithm. Instead of solving the two objectives independently, both objectives are considered simultaneously during all particle iterations, and the superior objective is always selected. The following equation illustrates the function value selection in the t-th iteration. This ensures a faster escape from local optima when one objective is trapped, thus accelerating convergence. Furthermore, the convergence performance will differ with different particle numbers. The following equation illustrates the function value selection in the t-th iteration. Where f(X) t ) represents the fitness value in the t-th iteration. This represents the fitness value of the coarse surrogate model in the t-th iteration. This represents the fitness value of the fine-grained proxy model in the t-th iteration.

[0080]

[0081] 2. Experimental Dataset

[0082] To evaluate the proposed method, we conducted experiments on the public dataset Series 60 ship model experimental dataset and our target datasets: the cargo_vessel dataset and the KCS sampled dataset.

[0083] The series 60 ship model experimental dataset consists of 60 ship models with a block coefficient C. b The series of ship types were formed by varying the length-to-breadth ratio (L / B) from 0.6 to 0.8, the length-to-breadth ratio (L / B) from 5.5 to 8.5, and the breadth-to-draft ratio (B / H) from 2.5, 3, and 3.5. The resistance data accumulated from ship model tests within the range of speed-to-length ratio (V / √Lwl) from 0.9 to 1.6 was used in the 60 series. A total of 45 ship models were used, resulting in 838 sets of data.

[0084] The cargo_vessel dataset consists of data on civilian cargo ships with a length-to-beam ratio (L / B) of 6.0–7.5 and a block coefficient of 0.45–0.6. Resistance data was obtained by importing ship models created using MAXSURF software into the CFD software Fluent for calculation. This data can be used to verify the effectiveness of the method used in this invention, as well as the error between the model's predictions and the CFD simulation results.

[0085] The KCS sampling dataset is a dataset obtained by sampling a KCS ship model using a sampling method. KCS is a modern container ship model with a bulbous bow, developed by the Korea Research Institute of Shipbuilding and Ocean Engineering (KRISO), and is an internationally recognized standard ship type. Two datasets were obtained by sampling the ship at both a large scale and locally, and the aforementioned transfer learning method was used for modeling. Finally, the ship's hull form was optimized.

[0086] Example

[0087] Two-Stage TrAdaBoost.R2

[0088] To verify that the known series of 60 ship model experimental data significantly contributes to the surrogate model established for the target domain, 100 data points of different ship sizes were generated using Latin hypercube sampling from a typical cargo ship (cargo_vessel). The distribution of these 100 randomly perturbed data points is shown below. Figure 3 As shown, from Figure 3 The total drag coefficient C of the cargo_vessel data can be seen. T The distribution of this data differs from that of the data from the 60 series of ship models. 90% of this data is distributed between 0.742 and 0.941, with a minimum value of 0.742 and a maximum value of 1.151.

[0089] Twenty rows from the cargo_vessel dataset were used as the training set for the target domain, and the remaining 85 rows were used as the test set. To verify the significance of the transfer learning prediction effect, the fitting effects of common machine learning algorithms were compared. Figures 4 to 8 This paper showcases the performance of Adaboost regression, decision trees, RBF neural networks, random forests, and the Two-Stage TrAdaBoost.R2 method on the test set. Table 1 presents their best error (MSEbest), worst error (MSEbest), and mean error (MSEmean) on the test set. The results in the table show that the Two-Stage TrAdaBoost.R2 method significantly outperforms the machine learning algorithms on the test set, demonstrating the effectiveness and accuracy of transfer learning in predicting ship performance in small-sample scenarios within the maritime field.

[0090] Table 1. Ship forecasting performance of different algorithms on the cargovessel dataset.

[0091]

[0092] We collected 20 and 40 simulation data points in the first and second phases, respectively, and performed regression transfer modeling using a series of 60 ship model data points. The total drag coefficient C was determined based on the simulation results from both phases. T Proxy models for the total resistance coefficient are established separately. A Decision Tree is used as the base learner in this method, and the basic parameter settings are shown in Table 2. The table shows the number of weak learners (NoE), the number of algorithm iterations (Steps), the number of cross-validation rounds (Fold), and the maximum depth of the decision tree (MaxD).

[0093] First, we will use some commonly used machine learning algorithms to perform fitting and prediction on our target domain. Figure 9 and Figure 10The mean squared error (MSE) in the target domain is shown separately for each of the two stages. The x-axis represents the number of training iterations, and the total number of training iterations plus the number of predictions equals the total number of target domains. It can be seen that as the number of training iterations increases, the Two StageTrAdaBoost.R2 algorithm exhibits good robustness and prediction performance, maintaining the lowest prediction error in most cases.

[0094] Finally, the target domain was trained using this regression-transfer algorithm, and the trained MSEs were 0.0038 and 0.0086, respectively. Figure 11 and Figure 12 The regression fit results for all target domain data are shown.

[0095] Table 2. Basic parameter settings for Two Stage TrAdaBoost.R2

[0096]

[0097] Ship type optimization

[0098] Two-stage sampling

[0099] KCS was selected as the ship type to be optimized, and its sample was expanded by Latin hypercube sampling.

[0100] Because the KCS is bilaterally symmetrical, only half of the ship's shape needs to be changed; therefore, the design variables only need to be selected for half of the ship. Then, a cube mesh of appropriate size is created, and the area to be deformed is placed within the mesh. This invention uses data sampled from the Latin hypercube to control the design variables to achieve deformation of the object within the mesh.

[0101] The hull form sampling process is divided into two stages. The first stage mainly involves changing the ship's principal dimension ratios, including variations in beam and depth, to create different length-to-beam ratios and beam-to-draft ratios. The first stage sampling process follows ship dimension ratio specifications. The second stage includes local fine-tuning of the bow and stern to obtain sub-hull forms with different shapes. The second stage sampling process satisfies the displacement constraint: |Δ′ / Δ-1|≤0.01, and the hull wetted surface area constraint: |S′ / S-1|≤0.01.

[0102] We now define the design variables as: Where i = 1, 2, ..., 11 represents 11 deformation point groups composed of different vertex sets. j = 1, 2, 3 represent FFD stretching transformations along the x-axis, y-axis, and z-axis, respectively. Figure 13 and Figure 14The points corresponding to specific variables are shown. The direction of deformation j of different points i follows different constraints. For example, in order to ensure the rationality of the changes in the ship, when the ship's width and depth change, the changes are only made along the y-axis of the width direction and the z-axis of the depth direction, respectively. When making local adjustments to the ship, non-ship boundary points are selected to ensure that the ship's length, width, and draft remain unchanged.

[0103] Specifically, these include:

[0104]

[0105] A total of 27 design variables control all adjustments to the ship. For example, among them... This represents the adjustment value of the ship's beam along the y-axis for the first vertex group. and The values ​​represent the adjustments made to the ship's depth along the z-axis by the second and third vertex groups, respectively, through the upper and lower surfaces of the hexahedron. The remaining points represent local adjustments made to the ship. Therefore, the design variables for the first stage sampling are j=2, 3, corresponding to... and The second phase of sampling covers all 27 design variables.

[0106] The constraints for the design variables are shown in Table 3, where L represents the ship's length. The initial position of the ship is indicated when all variables are set to 0.

[0107] Table 3 Design Variable Constraint Range

[0108]

[0109] Dual-surrogate model for total drag coefficient

[0110] Transfer learning can be used to transfer knowledge from the existing series of 60 ship model test datasets to our target ship type set, enabling the establishment of a surrogate model with good performance in terms of the total resistance coefficient of ships on a small dataset. The series of 60 ship model data is 4-dimensional, and includes... C B L / B and B / H, where V represents the boat speed, Lwl represents the waterline length, and C B Let L and B represent the ship's length and beam, respectively, and H represent the draft. There are 27 design variables controlling the ship's deformation, resulting in a 31-dimensional target domain data. To enable better knowledge transfer in the Two-Stage TrAdaBoost.R2 algorithm, feature processing is needed to fill in the missing dimensions of the source domain data. For convenience, this invention uses the mean of the corresponding dimensions in the target domain. Then, based on the iteration... Figure 15 The steps shown are for building the proxy model.

[0111] Two surrogate models were built for the data from the two sampling stages using transfer learning. These models represent a coarse surrogate model f1(X1) that only accounts for the main scale variation, and a fine surrogate model f2(X2) that includes local fine-tuning variations.

[0112] Hybrid Optimization of Dual Agent Model

[0113] Optimization definition

[0114] In this invention, the data obtained from KCS sampling is divided into two parts, and the two parts need to satisfy different constraints. Therefore, when optimizing the surrogate models f1(X1) and f2(X2), different constraints will also be satisfied. The optimization process of f1(X1) needs to satisfy the main dimensions of the ship design, while f2(X2) needs to keep the main dimensions of the ship unchanged and satisfy the constraints of displacement and wetted surface area changes.

[0115] The optimization goals are as follows:

[0116] Large-scale ship hull optimization Opt1:

[0117] Min C T =f1(X1), F r =0.26

[0118] st. 6.0 ≤ L / B ≤ 7.5

[0119] 2.7 ≤ B / d ≤ 3.8

[0120] 1.5≤D / d≤2.0

[0121]

[0122] X1∈D1

[0123] Ship hull fine-scale optimization Opt2:

[0124] Min C T =f2(X2), F r =0.26

[0125] st.l = L - L0 = 0

[0126] b = B - B0 = 0

[0127] t = d - d0 = 0

[0128] v = |Δ′ / Δ-1| ≤ 0.01

[0129] s=|S′ / S-1|≤0.01

[0130] X2∈D2

[0131] Where C TB represents the total drag coefficient. KCS The original mother ship KCS represents the width, B represents the beam, L represents the length, D represents the depth, d represents the draft, and F represents the length. r Let Froude number be an expression. Equality constraints l, b, and t represent the changes in the ship's length, width, and draft, respectively, while these values ​​must be kept to 0 during the fine-tuning phase. L0, B0, and d0 represent the ship's length, width, and draft before deformation, respectively. Inequality constraints v and s indicate that the changes in the ship's displacement and wetted surface area during the fine-tuning phase must not exceed 1% of the original ship's dimensions. Δ and Δ′ represent the ship's displacement before and after deformation, respectively, and S and S′ represent the ship's wetted surface area before and after deformation, respectively. D1 and D2 are the feasible regions of the variables shown in Table 1.

[0132] Hybrid optimization strategy

[0133] The hybrid optimization of the constructed f1(X1) and f2(X2) does not guarantee that the optimal value will appear in any particular surrogate model. This invention modifies the iterative process of the QPSO algorithm. During the iteration process, all particles need to simultaneously calculate the fitness values ​​of f1(X1) and f2(X2) and select the optimal one before proceeding to subsequent iterations.

[0134] Ship type optimization results

[0135] Now we set the parameters for the optimization phase. Let the maximum number of iterations of the QPSO algorithm be 1000*D, where D is the problem dimension and the population size is 100. All optimizations are taken as the optimal value after 31 independent runs.

[0136] First, we optimized the two surrogate models using a hybrid optimization approach. Table 4 shows the changes in the final optimization results on the principal scale and the final total drag coefficient.

[0137] C T The table shows the changes. The final optimized total drag coefficient (Opt) decreased by 17.93%, and the CFD simulation yielded a total drag coefficient of 3.0013. The error between the optimized result and the actual simulation value is 1.67%. Furthermore, since the ship length remains unchanged, the optimized beam is easily obtained as 30.69m and the depth as 16.88m. The wetted surface area decreased by 126.87m². 2 (1.34%), drainage volume decreased by 2568.62 m³. 3 (4.93%).

[0138] Table 4 Final Ship Type Optimization Results

[0139]

[0140] Comparison of convergence results

[0141] To compare the effectiveness of our proposed hybrid optimization strategy, we compare the convergence results of the single search strategy and the hybrid search strategy based on the above optimization results of the total drag coefficient. Figure 16 The convergence speed of the optimal particle in each generation is shown. The figure reveals that the method employing the hybrid optimization strategy has the fastest convergence speed. Furthermore, thanks to the efficiency of the QPSO algorithm and the limitation of the search space size, the proposed method exhibits rapid convergence performance from the outset. It can be seen that when only the fine surrogate model is used for optimization (Opt2), the number of iterations required for convergence is more than twice that of the hybrid optimization strategy. For the coarse surrogate model optimization (Opt1), i.e., optimization only at large scales, as shown in Section 4.4.1, the final result is not as good as the result with local adjustments. Therefore, it only converged to 3.0931.

[0142] Finally, to analyze the differences in the optimization results between the two stages, we conducted optimization experiments under different constraints. Table 5 shows our optimization results for fixed beam, depth, and fine-tuning, and the actual errors obtained through CFD simulation calculations. The first four columns in the table only show the values ​​corresponding to the better of the two optimization results. In this experiment, the Opt2 results were consistently better than the Opt1 results.

[0143] Table 5 shows that the optimization results of the two optimization stages differ under different constraints. Generally, the results obtained by performing large-scale optimization only in the first stage are generally worse than those obtained in the second stage. The second stage includes not only changes in the main dimensions but also further detailed changes to the bow and stern. Therefore, the superior results of the second stage are consistent with reality, but not absolute. Specifically, when changes to the main dimensions are not permitted, i.e., only adjustments to the bow and stern shapes are made based on the initial ship, the optimization result is 3.2574 × 10⁻⁶. -3 The optimization margin was 9.41%, with an error of 1.12% compared to the CFD simulation results. Miao et al., in optimizing the bow and stern deformations of the KCS, selected the optimal hull form from three different alternative schemes. r When the coefficient of resistance is 0.26, the total drag coefficient decreases by 1.52%, 4.24%, and 9.24%, respectively.

[0144] Table 5 Optimization results under different constraints

[0145]

[0146] Similarly, we conducted the same convergence experiment with a fixed ship under the constraint of the same height or width. The results are as follows: Figure 17 and Figure 18The results from the graph show that the method employing the hybrid optimization strategy consistently achieved the fastest convergence speed.

[0147] The above convergence results are all for the case where the number of particles is 100. In order to analyze the impact of the number of particles on the convergence of the optimization process, we will control the number of particles at different values ​​and analyze the convergence of the optimal value with and without the hybrid optimization strategy.

[0148] We controlled the particle number at 20, 50, 80, and 100 respectively. Since the optimization effect of Opt1 was consistently lower than that of Opt2 in the experiments of this invention, and because the problem dimension was very low, it quickly converged to its optimal value, making its inclusion in the convergence results comparison meaningless. Therefore, the following experimental comparison only focuses on the simple Opt2 process and the case using a hybrid optimization strategy. Figures 19 to 22 As can be seen, the convergence speed of the hybrid optimization strategy is higher than that of the simple Opt2 process in all cases.

[0149] Based on the experimental results of this invention, since Opt2 consistently outperforms Opt1, a comparison is made below of the local optima traps encountered when using only Opt2 and when employing a hybrid optimization strategy. To examine the local optima traps during the optimization process, we performed the optimization experiment a total of different runs, among which... Figures 23 to 25 The bar chart shows the number of times different optimization processes get trapped in local optima. It can be seen that the solution process using a hybrid optimization strategy significantly reduces the probability of getting trapped in local optima. The effect is most pronounced, especially when the scale constraint is not increased. Figure 23 The reason is that at this point, all variables in the optimization process of Opt1 and Opt2 can be searched within the feasible region, which greatly increases the probability of finding a better solution.

[0150] Figure 26 The results show the number of times the Opt2 and hybrid optimization strategies got trapped in local optima after 50 iterations for particle numbers of 20, 50, 80, 100, and 200, respectively, without increasing the scale constraint. It can be seen that the number of times the hybrid optimization strategy gets trapped in local optima is consistently higher when using only Opt2 for iteration than when using the hybrid optimization strategy.

Claims

1. A ship morphology optimization method based on transfer learning, characterized in that, The specific steps are as follows: (1) Obtain the publicly available ship model experiment dataset as the source domain dataset, and perform preprocessing and feature processing on the source domain data; (2) The target domain dataset is obtained by sampling the initial ship that needs to be optimized. Based on the two aspects of global ship design and local ship design, the sampling and modeling of the target domain data are divided into two corresponding stages: a coarse surrogate model built based on global scale sampling data and a fine surrogate model built based on local scale sampling data. The two surrogate models are used together in the subsequent optimization process. (3) The process of constructing the proxy model in step (2): Using the Two Stage TrAdaBoost.R2 method, the source domain data is used to assist in modeling the target domain data. The two stages of modeling yield a coarse proxy model and a fine proxy model, respectively. (4) The established coarse surrogate model and fine surrogate model are applied to the optimization process. The improved QPSO optimization algorithm is used to optimize the ship type data in the target domain to improve the performance and efficiency of the ship. The improved QPSO optimization algorithm does not solve the two objectives independently in the optimization process. Instead, it considers the two objectives at the same time and always selects the better one when all particles are iterated and updated, and then proceeds with the subsequent iterations.

2. The ship morphology optimization method based on transfer learning according to claim 1, characterized in that, In step (1), the preprocessing of the source domain dataset includes: selecting sample data from the source domain dataset that have similar characteristics to the target domain dataset, and removing samples from the source domain dataset that are significantly different from the target domain dataset. The feature processing of the source domain dataset is as follows: This includes feature alignment, feature mapping, or feature transformation, so that the data in the source and target domains are represented in the same feature space.

3. The ship morphology optimization method based on transfer learning according to claim 1, characterized in that, In step (2), global scale sampling includes sampling of main dimensions such as ship length, ship width and molded depth, while local scale sampling refers to sampling of the dimensions of a certain local position of the ship.

4. The ship morphology optimization method based on transfer learning according to claim 1, characterized in that, In step (2), the feature processing of the target domain dataset includes feature alignment, feature dimensionality reduction, and feature expansion; the preprocessing of the target domain dataset includes: for different performances of ships, it is necessary to extract the correlation between features through correlation analysis, and perform dimensionless and standardized preprocessing operations to facilitate subsequent model training; after sampling, the required performance data is obtained through CFD simulation, thereby obtaining the labeled target domain dataset.

5. The ship morphology optimization method based on transfer learning according to claim 1, characterized in that, In step (4), the ship's shape is optimized under different conditions to meet different practical requirements while satisfying the design variable constraints. Finally, the optimal solution is numerically reconstructed to obtain the ship's shape, and the optimization results are verified by CFD simulation. The improved QPSO optimization algorithm is used to search for the target in the optimization process.

6. The ship morphology optimization method based on transfer learning according to claim 1, characterized in that, In step (4), during the optimization process, the function value is selected in the t-th iteration; where f(X) t ) represents the particle fitness value in the t-th iteration. Represents the i-th particle The fitness value of the coarse surrogate model in the t-th iteration. Represents the i-th particle The fitness value of the fine-matrix model in the t-th iteration; N represents the population size of the particles;

Images