Ship resistance prediction method based on geometric dimension reduction and multi-modal fusion

Abstract

The invention provides a ship resistance prediction method based on geometric dimension reduction and multi-modal fusion, and belongs to the technical field of ship resistance prediction. A public mother type ship is selected, and resistance sample data is obtained through CFD simulation; adjusting local parameters to generate a new ship type, obtaining resistance data through CFD, summarizing original and deformation samples, and forming a training data set containing three-dimensional ship type geometric data, key numerical parameters and corresponding resistance values; preprocessing the three-dimensional ship model point cloud, converting the three-dimensional ship model point cloud into a two-dimensional contour diagram through cross section segmentation and orthogonal projection, and standardizing the two-dimensional contour diagram to obtain structured two-dimensional contour features; constructing an image and numerical value double-branch convolutional neural network, respectively inputting features and parameter extraction features and fusing the features, and outputting a resistance prediction value; training the model through a preset loss function and an optimization strategy; and inputting to-be-predicted ship data to obtain a ship resistance prediction result. According to the invention, a geometric dimensionality reduction method is designed and realized, and resistance prediction is carried out in combination with multi-modal fusion, so that the speed of resistance prediction and the efficiency of ship type design are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of ship resistance prediction, and in particular relates to a ship resistance prediction method based on geometric dimension reduction and multi-modal fusion. BACKGROUND

[0002] With the rapid development of shipbuilding industry, the requirements for ship performance are increasing, among which ship resistance is a key factor affecting the efficiency and energy consumption of ship navigation, and its accurate prediction and optimization are of great significance. Traditional ship resistance prediction methods mainly rely on physical model tests and empirical formulas. Physical model tests measure the resistance of ships at different speeds and fluid environments by making scaled ship models and conducting drag tests in experimental water pools. However, this method is costly, time-consuming, and difficult to fully consider all factors affecting ship resistance. Empirical formulas are based on a large number of experimental data, although they are simple to calculate, but their application range is limited, and it is difficult to accurately predict the resistance of new ship types, and it is difficult to meet the needs of modern ship design rapid iteration. Therefore, developing an efficient and accurate ship resistance prediction method has become an important research topic in the field of shipbuilding industry software design.

[0003] With the rapid development of computer technology, computational fluid dynamics technology has gradually become an important means of ship resistance prediction. CFD solves the Navier-Stokes equation of the flow field around the ship through numerical simulation, thereby obtaining the resistance of the ship. However, the CFD method requires high computational power and is affected by factors such as grid partitioning and turbulence model selection, which affects the accuracy of the simulation results.

[0004] In recent years, with the rapid development of artificial intelligence technology, especially the widespread application of deep learning technology, it has provided a new way of thinking and solution for ship resistance prediction. Convolutional neural networks, as an important model of deep learning, have made remarkable achievements in image processing, speech recognition, and other fields, and their powerful feature extraction and nonlinear mapping capabilities make it possible to accurately predict ship resistance.

[0005] Deep learning models, especially convolutional neural networks, can automatically learn the complex mapping relationship between ship shape features and resistance, achieving fast and accurate ship resistance prediction. However, directly inputting a three-dimensional ship model into a CNN for training will result in a huge amount of computation and low training efficiency. Therefore, how to combine geometric dimension reduction technology and multi-modal data fusion methods to improve the efficiency and accuracy of ship resistance prediction has become a hot research topic. Geometric dimension reduction technology aims to convert high-dimensional geometric data into low-dimensional representations while preserving key geometric features. In ship resistance prediction, geometric dimension reduction technology can convert a three-dimensional ship model into a two-dimensional feature map, reducing computational complexity and improving training efficiency. Common geometric dimension reduction methods include orthogonal projection and cross-sectional segmentation.

[0006] Multi-modal data fusion technology refers to integrating data from different modalities to take full advantage of the complementarity of each modality, improving the performance and accuracy of the model. In ship resistance prediction, multi-modal data fusion technology can combine the geometric characteristics of the ship with the numerical characteristics to improve the accuracy of the prediction.

[0007] The generation of ship resistance is closely related to ship shape, navigation state and fluid environment, etc. The changes of these factors are complex and nonlinear, and traditional physical model test and empirical formula are difficult to fully and accurately describe these complex relationships. Convolutional neural network can realize fast and accurate prediction of ship resistance by automatically learning the complex mapping relationship between ship shape characteristics and resistance. In addition, the convolutional neural network prediction model based on multi-modal fusion has the advantages of strong generalization ability and wide application range, which can adapt to the resistance prediction needs of different ship types and different navigation conditions.

[0008] In the field of ship resistance prediction, the traditional CFD method obtains flow field information by numerically solving the Navier-Stokes equation, relying on structured grid and high-precision difference format, which can completely depict the flow, heat transfer and turbulence details in complex geometry. The single modal machine learning model solves the data exchange problem between heterogeneous grids to some extent, and only trains and reasons on a single data form (such as pure image, pure text or pure numerical field), typical representatives include convolutional neural network for turbulent kinetic energy prediction of flow snapshot, recurrent neural network for time series pressure coefficient regression, etc. However, the above methods still have the following shortcomings: 1. Low grid generation efficiency, poor adaptability to complex geometry, resulting in long modeling cycle and high labor cost.

[0009] 2. Huge consumption of computing resources, high-fidelity simulation of a single working condition can take several days, which is difficult to meet the rapid iteration demand.

[0010] 3. Only single data form is used, lacking of complementary information from geometry, boundary conditions and other sources, and weak in complex working condition extrapolation ability. SUMMARY

[0011] To solve the problem of ship resistance prediction during navigation, a ship resistance prediction method based on geometric dimension reduction and multi-modal fusion is proposed, a geometric dimension reduction method is designed and implemented, combined with multi-modal fusion for resistance prediction, which provides an effective solution to the above problems to improve the speed of resistance prediction and the efficiency of ship design.

[0012] The present application is implemented by the following technical scheme: a ship resistance prediction method based on geometric dimension reduction and multi-modal fusion: the method specifically includes the following steps: Step 1. Coupling data augmentation with simulation: Select several public ship types as mother ships, obtain resistance sample data through computational fluid dynamics (CFD) simulation; generate different ship type schemes by adjusting local control parameters, perform CFD simulation on the new ship types to obtain resistance data; aggregate the original mother sample and the deformed sample to form a training dataset containing three-dimensional ship geometry data, key numerical parameters, and corresponding resistance values; Step 2. Multiview geometry dimension reduction: preprocess the input three-dimensional ship model as a point cloud, convert the three-dimensional ship model to a two-dimensional contour image through cross-sectional division and orthogonal projection, and then standardize the two-dimensional contour image to obtain structured two-dimensional contour features; Step 3. Multimodal model construction and training: construct a convolutional neural network model with dual branches of image + numerical fusion, input the two-dimensional contour features and ship key numerical parameters obtained in step 2 into the corresponding branches to extract features, concatenate the two types of features in the fusion layer, and output the resistance prediction value through the fully connected regression output layer; train the model using a preset loss function and optimization strategy, and iteratively update the network parameters; Step 4. Output ship resistance prediction value: input the three-dimensional geometry data and key numerical parameters of the ship to be predicted into the trained model after step 2 processing, and obtain the ship resistance prediction result.

[0013] Further, in step 1, comprising: Step 1.1. Select several public ship types as mother ships, and obtain resistance sample data of each mother ship through computational fluid dynamics (CFD) simulation under the same grid strategy and boundary conditions; Step 1.2. Under the premise of keeping the main dimensions of the mother ship unchanged, adjust the local geometry of the ship body through parameterized deformation technology to generate a large number of new ship type samples similar to the mother ship in geometry but with slight differences; Step 1.3. Perform CFD simulation again on each newly generated deformed ship type to calculate resistance, ensuring that the resistance labels of the new samples have the same accuracy and consistency as the mother ship; Step 1.4. Aggregate the original mother sample and the deformed sample to form a complete training dataset containing three-dimensional ship geometry data, key numerical parameters, and corresponding resistance values.

[0014] Further, in step 2, the original three-dimensional model is denoised and simplified through point cloud preprocessing to eliminate noise points in the model and reduce the point cloud density; The point cloud preprocessing includes Poisson disk sampling, voxel downsampling, and statistical filtering; The Poisson disk sampling converts the input STL ship model into a point cloud of naturally and uniformly distributed discrete points through a uniform distribution sampling strategy.

[0015] Further, in step 2, the cross-sectional sectioning and orthogonal projection are specifically: The hull space range is determined by an axis-aligned bounding box (AABB), and a dynamic sectioning strategy is adopted to select several cross-section planes perpendicular to the length direction in the typical area of the hull along the length direction, so as to obtain the corresponding cross-sectional profile by cutting the hull; then, the orthogonal projection is performed on the point cloud at each cross-section, and the three-dimensional point cloud is projected onto a two-dimensional plane, so as to obtain the two-dimensional profile image of the cross-section.

[0016] Further, in step 2, the profile image obtained by projection is standardized, specifically: The convex hull algorithm is used to extract the outer edge of each profile, so as to remove the internal redundant points in the projected point cloud and only keep the peripheral shape of the hull section; the centroid alignment and boundary normalization are performed on each profile image: the coordinate origin of all profile images is translated to the geometric center, and the size of the profile is scaled to a unified range, so as to eliminate the feature inconsistency caused by the position and scale difference of different cross-sections.

[0017] Further, in step 3, the image branch is used to process the two-dimensional profile image obtained by multi-view geometry dimension reduction, and the local detail features and global shape features of the hull profile are extracted through the convolution layer, and a fixed-length feature vector is output; The numerical branch is composed of several layers of fully connected neural networks, and is used to process the key numerical parameters of the ship, and the key numerical parameters processed include length, width, draft, square coefficient, speed, Froude number and wetted area, and a feature vector with a length matching that of the image branch is output.

[0018] Further, in step 3, the preset loss function is a mean square error (MSE) loss function, and the optimization strategy adopts an Adam optimizer, combined with learning rate decay, batch normalization and Dropout regularization to prevent model overfitting.

[0019] A ship resistance prediction system based on geometric dimension reduction and multi-modal fusion, comprising a data enhancement and simulation coupling module, a multi-view geometry dimension reduction module, a multi-modal model construction and training module, and a ship resistance prediction module: The data enhancement and simulation coupling module selects several public ship types as parent ships, obtains resistance sample data through computational fluid dynamics (CFD) simulation; generates different ship type schemes by adjusting local control parameters, and obtains resistance data through CFD simulation of the new ship type; and the original parent sample and the deformed sample are summarized to form a training data set containing three-dimensional ship type geometric data, key numerical parameters and corresponding resistance values; The multi-view geometry dimension reduction module carries out point cloud preprocessing on the input three-dimensional ship model, converts the three-dimensional ship model into a two-dimensional profile image through cross-section division and orthogonal projection, and then carries out standardization processing on the two-dimensional profile image to obtain structured two-dimensional profile features; The multi-modal model construction and training module is used for constructing a convolutional neural network model with double branches of image + numerical value fusion, inputting the two-dimensional profile features obtained by the multi-view geometry dimension reduction module and the key numerical parameters of the ship into corresponding branches to extract features, splicing the two types of features in the fusion layer, and outputting the resistance prediction value through a fully connected regression output layer; the model is trained by using a preset loss function and an optimization strategy, and the network parameters are iteratively updated; The ship resistance prediction module is used for outputting the ship resistance prediction value, inputting the three-dimensional geometric data and key numerical parameters of the ship to be predicted into the trained model after processing by the multi-view geometry dimension reduction module, and obtaining the ship resistance prediction result.

[0020] An electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0021] A computer readable storage medium for storing computer instructions, the computer instructions are executed by a processor to implement the steps of the above method.

[0022] Compared with the traditional ship resistance prediction method, the present application has the following advantages: 1. High prediction accuracy: By introducing the spatial distribution characteristics of the three-dimensional geometry of the ship body and fusing the key physical parameters, the multi-modal model of the present application can more comprehensively depict the mapping relationship between the ship type and the resistance, thereby realizing higher prediction accuracy. Experimental results show that the average absolute percentage error of the model verification set is only about 4.09%, and the maximum error is within ±10%, which is significantly better than the traditional empirical formula or single modal model which only relies on the main dimension parameters. High-precision prediction helps ship designers more accurately evaluate the performance of the scheme and reduce the cost of later modification and testing.

[0023] 2. High prediction efficiency: The trained deep learning model of the present application can complete the resistance prediction of a single ship type in seconds, which is orders of magnitude higher than the calculation period of several days or even several weeks required by traditional physical model tests or high-fidelity CFD simulations. This makes it possible to quickly evaluate and iterate a large number of candidate schemes at the initial stage of ship design, greatly shortening the design cycle. At the same time, the model inference process has relatively low requirements for computing resources and can be run on ordinary workstations or laptops, further improving the engineering practicability.

[0024] 3. Strong Generalization Ability: By preserving the local details of the hull through multi-view geometric dimensionality reduction, the model is more sensitive to local modifications and changes in the hull shape, and can better capture the differences between different hull types. Furthermore, the training data covers multiple hull types and various speed conditions, and parametric deformation expands the sample diversity, giving the model excellent generalization performance. In practical applications, the model is not only applicable to similar hull types in the training set, but also provides reliable resistance estimates for new hull design schemes, demonstrating strong extrapolation capabilities.

[0025] 4. High Robustness and Reliability: This invention employs multiple regularization and optimization strategies during model training, effectively preventing overfitting and improving model stability on unseen data. Simultaneously, the training dataset undergoes rigorous CFD simulation validation, ensuring high and consistent resistance label accuracy. Therefore, the model is insensitive to input noise and minor changes in ship hull shape, resulting in highly reliable predictions. This robustness guarantees stable and reliable model operation in practical engineering, providing strong support for ship design decisions. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a schematic diagram of the geometric dimensionality reduction model construction process of the present invention; Figure 3 This is a flowchart of the voxel downsampling process of the present invention; Figure 4 This is a schematic diagram of the orthogonal projection parameter settings of the present invention; Figure 5 The result of the convex hull algorithm is shown in the image. Figure 6 For DTC bow / stern diagrams; Figure 7 For KCS bow / stern diagram; Figure 8 This is the CNN structure of the present invention; Figure 9 For data comparison. Detailed Implementation

[0027] 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.

[0028] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art.

[0029] This invention presents a ship resistance prediction method based on geometric dimensionality reduction and graph convolutional neural networks. Addressing the problems of high time consumption and high computational resource consumption in traditional CFD simulation for ship resistance calculation, this method proposes a parallel prediction framework that integrates geometric dimensionality reduction processing and graph neural network surrogate modeling. During the ship design phase, due to the high geometric complexity of the hull, directly using high-fidelity CFD simulation leads to a surge in computational costs, while existing surrogate models do not sufficiently extract geometric features, resulting in limited prediction accuracy. To address these issues, this invention proposes a parallel feature extraction method based on geometric dimensionality reduction. By constructing a lightweight geometric representation of a 3D ship model and combining it with the topology learning capability of graph convolutional neural networks, it achieves fast and high-precision prediction of ship resistance. The overall flowchart is as follows. Figure 1 As shown.

[0030] 1. Coupling of data augmentation and simulation This invention employs a data augmentation and simulation coupling technique in the data preparation phase to obtain high-quality, diverse training data. The specific process is as follows: First, several internationally recognized publicly available ship types are selected as parent ships (e.g., DTMB 5415, DTC, KCS, KVLCC2, and JBC). Computational fluid dynamics (CFD) simulations are used to calculate the resistance of each parent ship under the same mesh strategy and boundary conditions, obtaining high-precision resistance sample data. Then, while ensuring the main dimensions of the parent ships remain unchanged, parametric deformation techniques are used to adjust the local geometry of the hull, such as changing the bow and stern line ratios, generating a large number of new ship type samples that are geometrically similar to the parent ships but with subtle differences. This parametric deformation can be implemented on CAD platforms such as Rhino, generating different ship type schemes by adjusting local control parameters. Next, CFD simulations are performed again on each newly generated deformed ship type to calculate resistance, ensuring that the resistance labels of the new samples have comparable accuracy and consistency to the parent ship. Finally, the original parent model sample and the modified sample are combined to form a complete training dataset containing three-dimensional ship geometry data, key numerical parameters and corresponding resistance values.

[0031] A data augmentation strategy based on parametric deformation and high-precision CFD simulation is proposed to construct a high-quality training dataset. This data augmentation and simulation coupling method effectively expands the design space, ensures the diversity and label accuracy of training samples, and provides a reliable data foundation for model training.

[0032] Through the aforementioned data augmentation and simulation coupling process, this invention effectively expands the training sample space for ship resistance prediction while ensuring the reliability of the newly added sample data. As shown in Table 1, the original five parent ships are parametrically deformed, generating 80 new ship types for each, ultimately forming a massive dataset containing 400 sets of deformation data, providing ample diverse samples for subsequent model training. It is worth mentioning that in simulation verification, this invention compared the CFD calculation results of the original parent ships with experimental values, and the results show that the errors are all within ±5%, proving the accuracy of the simulation method and multi-view geometric dimensionality reduction. The comparison results are shown in the table below: Table 1 Comparison of Ship Type Resistance Calculation Errors

[0033] 2. Construction and Optimization of Multimodal Drag Prediction Model To transform complex 3D ship geometry into a low-dimensional feature representation suitable for machine learning models, this invention proposes a multi-view geometry dimensionality reduction technique. For example... Figure 2 As shown, this technology reduces the 3D geometry into a series of 2D images by projecting and sectioning the 3D ship hull from multiple perspectives, thereby reducing data complexity while preserving key shape features of the hull. The specific implementation steps of multi-view geometry dimensionality reduction are as follows: First, the input 3D ship model (e.g., a hull surface model represented by an STL file) undergoes point cloud preprocessing. This invention employs point cloud processing techniques such as Poisson disk sampling, voxel downsampling, and statistical filtering to denoise and simplify the original 3D model, eliminating noise points and reducing point cloud density, thereby improving the efficiency and accuracy of subsequent processing. After preprocessing, a clean and simplified hull point cloud model is obtained.

[0034] The Poisson disk sampling method uses a uniformly distributed sampling strategy to convert the input STL ship model into a discrete point cloud with naturally uniformly distributed points. Let the original set of triangular facets be T, and the sampling point set P satisfy:

[0035] The voxel downsampling process is as follows Figure 3 As shown: Next, cross-sectional sectioning and orthogonal projection are performed. For example... Figure 4As shown, this invention utilizes axis-aligned bounding boxes (AABBs) to define the spatial extent of the hull and employs a dynamic cutting strategy to automatically generate multiple cross-sections along the ship's length. Specifically, several cutting planes perpendicular to the ship's length are selected in typical areas of the hull (such as the bow and stern) to cut the hull and obtain the corresponding cross-sectional contours. Then, orthogonal projection is performed on the point cloud at each cross-section, projecting the three-dimensional point cloud onto a two-dimensional plane to obtain a two-dimensional contour image of that cross-section. By cutting and projecting at different positions along the ship's length, a series of two-dimensional contour images reflecting different cross-sectional shapes of the hull can be obtained.

[0036] To further improve the consistency and comparability of these two-dimensional contour features, this invention standardizes the contour images obtained from projection. On one hand, a convex hull algorithm is used to extract the outer edge of each profile contour to remove potentially redundant internal points in the projected point cloud, retaining only the outer shape of the hull section. On the other hand, centroid alignment and boundary normalization are performed on each contour image: the origin of all contour images is translated to their geometric center, and the size of the contours is scaled to a uniform range, thereby eliminating feature inconsistencies caused by differences in position and scale between different profiles. After the above processing, profile contours of different positions and sizes are converted into two-dimensional images of uniform size and orientation, facilitating feature extraction by subsequent models. Contour standardization uses a convex hull algorithm to extract the contours. Convex hull algorithms mainly include exhaustive search, divide-and-conquer, Jarvis step method, Graham scan method, and Quickhull algorithm. This paper uses the Quickhull algorithm. The process is as follows: (1) Initial selection: Find the points with the minimum and maximum x-coordinates; (2) Segmentation: Divide the point set into two subsets, upper and lower; (3) Find the farthest point: Find the point farthest from the current line segment in each subset; (4) Recursive processing: Form new line segments and continue recursively processing the points outside each line segment; (5) Eliminate internal points: After finding the farthest point each time, the points inside the triangle are eliminated.

[0037] After the above steps are completed, the result is as follows: Figure 5 As shown: Depend on Figure 5 As can be seen, the QuickHull algorithm finds the points that form the convex hull from a set of scattered points and connects them in the correct order. The QuickHull algorithm is characterized by constructing the convex hull by recursively finding the point farthest from the current line segment, ultimately obtaining the smallest convex polygon containing all the points.

[0038] Through the aforementioned multi-view geometric dimensionality reduction steps, this invention successfully transforms the complex three-dimensional hull geometry into a set of structured two-dimensional image feature representations. This dimensionality reduction method not only preserves the shape information of key hull parts (such as the line features of the bow and stern), but also compresses the originally high-dimensional three-dimensional data into a low-dimensional space, significantly reducing data complexity and the difficulty of subsequent model training. These two-dimensional contour images obtained from multi-view projection will serve as an important input channel for the multimodal model, used together with other numerical parameters for predicting ship resistance. Some of the dimensionality-reduced results are shown below. Figure 6 , Figure 7 As shown.

[0039] This invention proposes a 3D ship hull geometry dimensionality reduction technique based on orthogonal projection and cross-sectional segmentation. This technique transforms complex 3D ship hull information into a structured 2D contour sequence, which serves as input for a deep learning model. This method effectively preserves the local geometric features of the hull, overcoming the problem of traditional models being insensitive to hull details due to using only principal scale parameters as input. The mesh matching algorithm is optimized: in parallel coupled computation, a method using geometric bounding boxes and spatial hash tables is used to achieve rapid matching and detection of partitioned meshes. This significantly reduces computational load and improves matching efficiency, especially for a large number of parallel processes.

[0040] 3. Construction and optimization of multimodal drag prediction model After obtaining the dimensionality-reduced image features and related numerical parameters, this invention constructs a multimodal resistance prediction model that fuses image and numerical information to achieve accurate prediction of ship resistance. The model employs a convolutional neural network architecture of "image-numerical dual-branch fusion," the core idea of ​​which is to use two channels to process different types of input data separately, and then fuse the two feature paths at a deeper layer of the network for joint regression prediction of resistance values.

[0041] The structure of the model is as follows Figure 8 As shown, the entire network comprises two parallel branches: an image branch and a numerical branch. The image branch processes the 2D contour image obtained through multi-view geometric dimensionality reduction. This branch contains several convolutional-pooling layers (e.g., a 3-layer convolutional neural network is used in this embodiment). Each convolutional layer extracts local texture and shape features from the image, and subsequent pooling layers downsample the feature maps to reduce dimensionality and enhance translation invariance. Through multiple layers of convolutional-pooling operations, the image branch can progressively extract local detail features and global shape features of the hull contour. Finally, the image branch outputs a fixed-length feature vector representing the high-level abstract features learned from the hull contour image.

[0042] The numerical branch handles the ship's key numerical parameters. These parameters include seven key parameters that characterize the ship's main dimensions and operating conditions: length, beam, draft, block coefficient, speed, Froude number, and wetted area. The numerical branch consists of several layers of fully connected neural networks that perform feature transformation and compression on the input numerical parameters. The first fully connected layer maps the original multidimensional parameters to an intermediate-dimensional feature space, extracting the nonlinear relationships between the parameters; the second fully connected layer further compresses the features into a vector matching the length of the image branch output for subsequent fusion.

[0043] After extracting image feature vectors and numerical feature vectors from the image branch and numerical branch respectively, this invention concatenates and fuses these two types of features in the network's fusion layer. The fused feature vector simultaneously contains information about the ship's geometry and key parameters, enabling a more comprehensive characterization of the ship's hydrodynamic characteristics. Following the fusion layer, the network incorporates several fully connected regression output layers to further nonlinearly transform the fused features and ultimately output a scalar value, namely the predicted ship resistance value.

[0044] This invention constructs a convolutional neural network model with dual-modal input of "image + numerical values." The model inputs both the dimensionality-reduced ship profile image and key numerical parameters to achieve accurate prediction of ship resistance. The model employs a "dual-branch fusion" architecture: the image branch extracts local and global features of the ship profile through multiple convolutional-pooling layers, while the numerical branch compresses numerical parameter features through fully connected layers. The two types of features are then concatenated in the fusion layer, and finally, the predicted resistance value is output through another fully connected layer. This multimodal network fully leverages the powerful image feature extraction capabilities of CNNs and combines them with the physical meaning of numerical parameters, significantly improving the model's expressive power and prediction accuracy.

[0045] During the model training phase, this invention employs the mean squared error loss function to measure the difference between the predicted resistance and the actual resistance, and iteratively updates the network parameters through an optimization algorithm. To prevent overfitting and improve its generalization ability, various regularization and optimization strategies are introduced during training, including: learning rate decay strategy, batch normalization, and Dropout random deactivation. Through these optimization measures, the model can learn more stably during training, avoiding overfitting to the training data.

[0046] During model training, mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for parameter optimization. Regularization techniques such as learning rate decay, batch normalization, and Dropout are used to prevent overfitting.

[0047] After multiple rounds of iterative training, the multimodal drag prediction model of this invention achieved excellent performance on the validation set, with its prediction accuracy far exceeding that of traditional methods. Specifically, for example…Figure 9 As shown.

[0048] Experimental results show that the model's mean absolute percentage error (MAPE) on the validation set is reduced to 4.09%, and the maximum relative error is controlled within ±10%, significantly outperforming traditional empirical formulas and neural network models using only a single modality (such as image or numerical data). This result verifies the effectiveness and superiority of the multimodal fusion model proposed in this invention for ship resistance prediction tasks.

[0049] In summary, this invention acquires high-quality training data through data augmentation and simulation coupling, transforms complex three-dimensional ship geometry into easily processed two-dimensional features through multi-view geometric dimensionality reduction, and achieves efficient and accurate prediction of ship resistance by fusing image and numerical information through a multimodal resistance prediction model. The combination of these technical solutions effectively solves the problems of long cycles and high costs in traditional ship resistance prediction methods, significantly improving prediction efficiency and accuracy, and providing strong technical support for resistance performance evaluation during the ship design phase.

[0050] Accordingly, in addition to the cross-sectional contour extraction method based on orthogonal projection, the geometric dimensionality reduction of this invention can also employ other geometric representations and dimensionality reduction techniques. For example, the three-dimensional hull model can be projected into two-dimensional images from different perspectives to replace the cross-sectional sequence, thereby obtaining the image features of the overall hull shape; parametric curve and surface representation can also be used to describe the hull geometry, and the shape parameters of key curves can be extracted as model input; furthermore, direct point cloud dimensionality reduction is also an alternative approach, using methods such as principal component analysis or manifold learning to reduce the dimensionality of the hull point cloud and extract a few main patterns to characterize hull shape changes. These alternative geometric dimensionality reduction methods can also transform three-dimensional hull shape information into a low-dimensional feature input model, enabling resistance prediction.

[0051] The multimodal model designed in this invention, while maintaining the "dual-modal input + feature fusion" approach, can employ different deep learning model architectures to achieve multimodal fusion. For example, in addition to using convolutional neural networks, the image branch can also use novel neural networks such as the Visual Transformer to extract image features; in addition to fully connected networks, the numerical branch can consider introducing graph neural networks or other networks suitable for structured data; in terms of fusion methods, besides simple feature concatenation, gating fusion, attention fusion, and other mechanisms can be used to dynamically adjust the fusion weights according to feature importance to improve model performance. Furthermore, a multi-task learning architecture can be adopted, treating drag prediction and other related hydrodynamic performance predictions as a joint task, sharing some network layers to improve feature utilization efficiency. These different model architectures and fusion strategies all fall within the scope of multimodal fusion prediction and can achieve similar functions to the solution of this invention.

[0052] In this embodiment, to expand the training data, in addition to the parametric deformation method based on the Rhino platform, other geometric deformation and simulation strategies can be employed. For example, random perturbation methods can be used to randomly offset and twist the hull lines of the parent ship to generate diverse deformed hull shapes; or techniques such as shape grammars and generative adversarial networks can be used to automatically generate new hull shapes, creating training samples while ensuring the geometric rationality of the hull shape. Regarding simulation calculations, in addition to using high-fidelity CFD solvers, experimental design methods can be combined to select several key design points near the parent ship for model experiments to obtain resistance data, or a surrogate model-assisted sampling strategy can be used, first using simplified model predictions to select schemes before CFD verification, thereby reducing the amount of simulation computation. Furthermore, the concept of transfer learning can be considered, using resistance data from existing hull shape databases or literature to pre-train the model, and then fine-tuning it with a small amount of simulation data from the target hull shape, thereby reducing the dependence on large-scale simulation data. These alternative data augmentation and simulation schemes all aim to build high-quality training sets, which can also achieve the goal of enriching model training samples and improving prediction accuracy.

[0053] During model training, in addition to MSE loss and the Adam optimizer, other loss functions and optimization algorithms can be used. For example, mean absolute error or Huber loss can be used instead of MSE to improve the model's robustness to outliers. Regarding optimizers, besides Adam, SGD, RMSprop, and other optimization algorithms can also be used for model parameter updates; the appropriate optimization strategy should be selected based on the actual training situation. For regularization methods, in addition to Dropout and batch normalization, weight decay, early stopping, or data augmentation can be used to prevent overfitting. Furthermore, ensemble learning can be considered, training multiple sub-models with different initializations or structures, and then weighted averaging their outputs to improve prediction stability. For inference deployment, model compression techniques can be used to obtain lightweight models for use in embedded devices or real-time simulations. These alternative training optimization methods can also be used to improve model performance and practicality, and are equivalent extensions of the technical solution of this invention.

[0054] A ship resistance prediction system based on geometric dimensionality reduction and multimodal fusion includes a data augmentation and simulation coupling module, a multi-view geometric dimensionality reduction module, a multimodal model construction and training module, and a ship resistance prediction module. The data augmentation and simulation coupling module selects several publicly available ship types as parent ships and obtains resistance sample data through computational fluid dynamics (CFD) simulation; it generates different ship type schemes by adjusting local control parameters, and obtains resistance data by performing CFD simulation on the new ship types; it summarizes the original parent sample and the deformed sample to form a training dataset containing three-dimensional ship type geometric data, key numerical parameters and corresponding resistance values. The multi-view geometry dimensionality reduction module preprocesses the point cloud of the input 3D ship model, transforms the 3D ship model into a 2D contour image through cross-sectional segmentation and orthogonal projection, and then performs standardization processing on the 2D contour image to obtain structured 2D contour features. The multimodal model construction and training module is used to construct a two-branch fusion convolutional neural network model of "image + numerical value". The two-dimensional contour features obtained by the multi-view geometric dimensionality reduction module and the key numerical parameters of the ship are respectively input into the corresponding branches to extract features. The two types of features are concatenated in the fusion layer, and the resistance prediction value is output through the fully connected regression output layer. The model is trained using a preset loss function and optimization strategy, and the network parameters are iteratively updated. The ship resistance prediction module is used to output ship resistance prediction values. The three-dimensional geometric data and key numerical parameters of the ship to be predicted are processed by the multi-view geometric dimensionality reduction module and then input into the trained model to obtain the ship resistance prediction results.

[0055] Through the above technical solution, the present invention 1. This invention addresses the issues of feature preservation and accuracy improvement during the geometric dimensionality reduction process of 3D ship models: In existing technologies, when reducing a 3D ship model to a 2D feature map for rapid resistance prediction, the limitations of the dimensionality reduction algorithm often lead to geometric feature distortion, thus affecting prediction accuracy. This invention aims to solve the problem of effectively preserving the key geometric features of the ship model during geometric dimensionality reduction while improving the accuracy of the reduced 2D feature map, thereby enhancing the accuracy of ship resistance prediction.

[0056] 2. This invention addresses the effectiveness and optimization issues of multimodal data fusion methods: Currently, the application of multimodal data fusion technology in ship resistance prediction is still immature. In particular, how to efficiently fuse image features and numerical features to fully utilize the complementarity of each modality and improve the performance and accuracy of the prediction model is a key problem this invention aims to solve. This invention proposes an optimized multimodal data fusion method, aiming to improve the representational power of the fused data and the generalization ability of the prediction model.

[0057] 3. This invention addresses the problem of accurate geometric feature extraction under complex deformations of ship surfaces: Local deformations of ship surfaces, especially the sharp deformations of the bow and stern, can lead to difficulties in geometric feature extraction, affecting dimensionality reduction and resistance prediction accuracy. To address this issue, this invention proposes a multi-region alignment method based on a divide-and-conquer strategy, effectively solving the problem of accurate geometric feature extraction under complex deformations of ship surfaces.

[0058] 4. This invention addresses the problem of optimizing the structure and parameters of convolutional neural network (CNN) models: In ship resistance prediction, the structure and parameter settings of CNN models significantly impact prediction performance. How to optimize the structure and parameters of CNN models based on the characteristics of ship resistance prediction to improve prediction accuracy and computational efficiency is the technical challenge this invention aims to solve. Through experiments and algorithm optimization, this invention has determined an efficient CNN model structure and parameter configuration scheme.

[0059] 5. This invention addresses the problem of constructing and expanding ship resistance prediction datasets: High-quality datasets are fundamental for training efficient ship resistance prediction models. However, due to the diversity and complexity of real-world ship models, constructing a dataset that is both sufficient and representative is a challenge. This invention proposes a dataset construction and expansion method based on ship model deformation and simulation, effectively solving the problems of data scarcity and insufficient representativeness in the dataset construction process.

[0060] 6. Solved the problem of efficient training and real-time prediction of ship resistance prediction models: In practical engineering applications, ship resistance prediction models need to have the ability to train efficiently and predict in real time. How to optimize the model training process and improve prediction speed while ensuring prediction accuracy to meet the needs of practical engineering applications? This invention achieves efficient training and real-time prediction of ship resistance prediction models through optimized algorithms and parallel computing technology.

[0061] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0062] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0063] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0064] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means such as coaxial cable, optical fiber, digital subscriber line, DSL, or wireless means such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium such as a floppy disk, hard disk, magnetic tape; an optical medium such as a high-density digital video disc, DVD; or a semiconductor medium such as a solid-state disk, SSD, etc.

[0065] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0066] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0067] The above provides a detailed description of the ship resistance prediction method based on geometric dimensionality reduction and multimodal fusion proposed in this invention, and elucidates the principle and implementation of this invention. The above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for predicting ship resistance based on geometric dimensionality reduction and multimodal fusion, characterized in that: The method specifically includes the following steps: Step 1. Data Augmentation and Simulation Coupling: Select several publicly available ship types as parent ships and obtain resistance sample data through computational fluid dynamics (CFD) simulation; generate different ship type schemes by adjusting local control parameters, and obtain resistance data by performing CFD simulation on the new ship types; summarize the original parent sample and the modified sample to form a training dataset containing three-dimensional ship type geometric data, key numerical parameters and corresponding resistance values. Step 2. Multi-view geometric dimensionality reduction: The input 3D ship model is preprocessed with point cloud. The 3D ship model is converted into a 2D contour image by cross sectioning and orthogonal projection. The 2D contour image is then standardized to obtain structured 2D contour features. Step 3. Multimodal Model Construction and Training: Construct a two-branch convolutional neural network model that fuses "image + numerical values". Input the two-dimensional contour features and key numerical parameters of the ship obtained in Step 2 into the corresponding branches to extract features. Concatenate the two types of features in the fusion layer and output the predicted resistance value through a fully connected regression output layer. Train the model using a preset loss function and optimization strategy, and iteratively update the network parameters. Step 4. Output ship resistance prediction: Input the three-dimensional geometric data and key numerical parameters of the ship to be predicted into the trained model after processing in step 2 to obtain the ship resistance prediction result.

2. The method according to claim 1, characterized in that: Step 1 includes: Step 1.

1. Select several publicly available ship types as parent ships, and obtain resistance sample data for each parent ship through computational fluid dynamics (CFD) simulation under the same mesh strategy and boundary conditions; Step 1.

2. While ensuring that the main dimensions of the parent ship remain unchanged, the local geometry of the hull is adjusted through parametric deformation technology to generate a large number of new ship type samples that are similar to the parent ship in geometry but have subtle differences. Step 1.

3. Perform CFD simulation again to calculate the resistance for each newly generated deformed ship type, ensuring that the resistance label of the new sample has the same accuracy and consistency as the parent type; Step 1.

4. Combine the original parent model sample and the modified sample to form a complete training dataset containing three-dimensional ship geometry data, key numerical parameters and corresponding resistance values.

3. The method according to claim 2, characterized in that: In step 2, the original 3D model is denoised and simplified through point cloud preprocessing to eliminate noise points in the model and reduce the point cloud density. The point cloud preprocessing includes Poisson disk sampling, voxel downsampling, and statistical filtering; The Poisson disk sampling uses a uniformly distributed sampling strategy to convert the input STL ship model into a naturally uniformly distributed discrete point cloud.

4. The method according to claim 3, characterized in that: In step 2, the cross-sectional division and orthogonal projection are specifically as follows: The hull space is defined by axis-aligned bounding boxes (AABB). A dynamic cutting strategy is used to select several cutting planes perpendicular to the hull length in typical areas of the hull along the hull length direction to cut the hull and obtain the corresponding cross-sectional profile. Then, orthogonal projection is performed on the point cloud at each cross-section to project the three-dimensional point cloud onto a two-dimensional plane, thereby obtaining the two-dimensional profile image of the cross-section.

5. The method according to claim 4, characterized in that: In step 2, the contour image obtained by projection is standardized, specifically as follows: The convex hull algorithm is used to extract the outer edge of each profile contour to remove any redundant internal points in the projected point cloud and retain only the outer shape of the hull section. Centroid alignment and boundary normalization are performed on each contour image: the origin of all contour images is translated to its geometric center and the size of the contour is scaled to a uniform range, thereby eliminating feature inconsistencies caused by differences in position and scale between different profiles.

6. The method according to claim 5, characterized in that: In step 3, the image branch is used to process the two-dimensional contour image obtained by multi-view geometric dimensionality reduction. It extracts local detail features and global shape features of the hull contour through convolutional layers and outputs a fixed-length feature vector. The numerical branch consists of several layers of fully connected neural networks, used to process key numerical parameters of the ship, including length, beam, draft, block coefficient, speed, Froude number, and wetted area, and outputs a feature vector that matches the length of the image branch.

7. The method according to claim 6, characterized in that: In step 3, the preset loss function is the mean squared error (MSE) loss function, and the optimization strategy adopts the Adam optimizer, combined with learning rate decay, batch normalization and Dropout regularization to prevent model overfitting.

8. A ship resistance prediction system based on the ship resistance prediction method of geometric dimensionality reduction and multimodal fusion according to any one of claims 1 to 7, characterized in that: The ship resistance prediction system includes a data augmentation and simulation coupling module, a multi-view geometric dimensionality reduction module, a multi-modal model construction and training module, and a ship resistance prediction module. The data augmentation and simulation coupling module selects several publicly available ship types as parent ships and obtains resistance sample data through computational fluid dynamics (CFD) simulation; it generates different ship type schemes by adjusting local control parameters, and obtains resistance data by performing CFD simulation on the new ship types; it summarizes the original parent sample and the deformed sample to form a training dataset containing three-dimensional ship type geometric data, key numerical parameters and corresponding resistance values. The multi-view geometry dimensionality reduction module preprocesses the point cloud of the input 3D ship model, transforms the 3D ship model into a 2D contour image through cross-sectional segmentation and orthogonal projection, and then performs standardization processing on the 2D contour image to obtain structured 2D contour features. The multimodal model construction and training module is used to construct a two-branch fusion convolutional neural network model of "image + numerical value". The two-dimensional contour features obtained by the multi-view geometric dimensionality reduction module and the key numerical parameters of the ship are respectively input into the corresponding branches to extract features. The two types of features are concatenated in the fusion layer, and the resistance prediction value is output through the fully connected regression output layer. The model is trained using a preset loss function and optimization strategy, and the network parameters are iteratively updated. The ship resistance prediction module is used to output ship resistance prediction values. The three-dimensional geometric data and key numerical parameters of the ship to be predicted are processed by the multi-view geometric dimensionality reduction module and then input into the trained model to obtain the ship resistance prediction results.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images