Marine hydrostatic performance prediction large model training, deployment and implementation method based on large model technology

Abstract

The invention discloses a ship hydrostatic performance prediction large model training, deployment and implementation method based on a large model technology. Training a ship hydrostatic performance prediction learning large model according to real sample data of various types, main scales, tonnages and other real ship completion technical data; on the basis of the trained ship hydrostatic performance prediction learning large model, a model structure is adjusted, and a ship hydrostatic performance prediction large model is constructed for deployment; and inputting ship design elements and variable prediction tasks into the ship hydrostatic performance prediction large model to predict ship hydrostatic performance indexes. According to the method, designers can be replaced to complete ship hydrostatic performance estimation work, the working efficiency and level are improved, and basic service is provided for a ship concept designer (intelligent agent) based on artificial intelligence.

Description

Technical Field

[0001] This invention belongs to the field of highly integrated artificial intelligence technology and shipbuilding engineering technology. Specifically, it is a method for training, deploying and implementing a large model for predicting the hydrostatic performance of ships based on large model technology. Background Technology

[0002] The hydrostatic performance of a ship refers to its performance under the influence of gravity and buoyancy in still water, including at least buoyancy, integrity stability, and damage stability. These properties must meet the requirements of the ship's design specifications and ship safety regulations.

[0003] In the process of ship conceptual design, it is necessary to calculate the ship's hydrostatic performance to evaluate the merits of the design scheme. With the development of computer-aided ship design, designers can usually input the design parameters of a ship conceptual design scheme into design software to generate approximate hull lines and estimate the ship's hydrostatic performance values ​​after providing the scheme. However, the generated hull lines are not the optimal hull lines corresponding to the design parameters, and the calculated hydrostatic performance values ​​also contain superposition errors. For a long time, this work has been mainly done manually, which is time-consuming, labor-intensive, and costly, and does not necessarily result in a ship conceptual design scheme with relatively excellent hydrostatic performance.

[0004] Traditional ship design methods are broadly classified into two categories: the parent ship modification method and the successive approximation method. The parent ship modification method uses a prototype ship that is similar in key aspects to the ship to be designed as the parent ship, and then modifies and transforms various design elements according to the design requirements to obtain the corresponding elements of the new ship. The successive approximation method is suitable when design requirements are complex and the designer cannot find a suitable or complete parent ship; in this case, a method of progressively approximating design elements is used to design the new ship.

[0005] For centuries, humanity has designed and built countless ships, accumulating a wealth of ship design, construction, and navigation examples and experience. With the advent of computers, humans have applied computer tools to ship design and calculations of ship performance, such as still water performance. Mature software tools exist for calculating ship still water performance, such as the Finnish software Napa. This work requires inputting hull lines. In computer-aided ship design, the method of modifying a parent ship is to manually adjust the design scheme based on data such as the parent ship's hull lines using design software tools, calculating and evaluating various performance aspects of the designed ship. However, if the desired ship design is significantly different and innovative from the parent ship design, the parent ship data becomes unimportant or nonexistent. In such scenarios, the parent ship's hull lines cannot be used; instead, design elements must be used to generate approximate hull lines and estimate ship performance, leading to significant errors and misleading designers. Artificial intelligence technology provides an excellent tool for learning from human experience. To date, artificial intelligence technology has not been used in the estimation of ship hydrostatic performance. Therefore, the conceptual design of ships still requires a lot of manpower and time, especially for innovative ships. Without suitable or complete reference materials for the parent ship, the design workload is enormous and it limits the designers' innovative capabilities. Summary of the Invention

[0006] To solve the above problems, the technical solution provided by the present invention is as follows: The present invention provides a method for training, deploying, and implementing a large-scale model for predicting the hydrostatic performance of ships based on large-scale model technology, characterized by comprising the following steps: Step 1: Process the general sample set of ship hydrostatic performance. The general sample set includes hydrostatic performance completion data collected from all ship completion technical documents. The collected sample set is then cleaned, organized, and stored. Step 2: Enhance the general sample set of ship hydrostatic performance, including interpolating and filling the sample set based on diffusion model technology to synthesize and generate out-of-distribution samples to enhance and expand the sample set of ship hydrostatic performance. Step 3: Train a general learning model for predicting ship hydrostatic performance, including reading batch training samples, input preprocessing, feature extraction and embedding, sharing the ship hydrostatic knowledge space, guiding the expert module with parallel paths for fixed and variable prediction tasks, traversing variable prediction tasks and guiding the expert module through masks and routing mechanisms to achieve multi-task prediction of ship hydrostatic performance, optimizing model parameters with stochastic gradient descent, and ensuring that the prediction results satisfy Archimedes' law through the loss function of Newton's first law. Step 4: Deploy a general-purpose ship hydrostatic performance prediction model. The general-purpose ship hydrostatic performance prediction model includes two sub-models: a fixed prediction task prediction model and a variable prediction task prediction model. The two sub-models are constructed by adjusting the model structure based on the trained general-purpose ship hydrostatic performance prediction learning model. Step 5: Predict the ship's hydrostatic performance. This involves organizing the information from the variable prediction task and the ship's conceptual design scheme, inputting it into a general ship hydrostatic performance prediction model, and then having the two sub-models work together to predict the ship's hydrostatic performance parameters based on the relationship between the variable and fixed prediction tasks.

[0007] Preferably, the general-purpose ship hydrostatic performance prediction learning model includes an input layer, an input preprocessing module, a feature embedding module, a multi-head self-attention layer, a feedforward neural network layer, and a regression prediction layer; the input layer acquires a training sample set and trains the general-purpose ship hydrostatic performance prediction learning model based on the sample set; the feedforward neural network layer has two parallel data flow paths: a fixed prediction task path and a variable prediction task path. The large-scale learning model for predicting the hydrostatic performance of a ship completes multiple prediction tasks. The parameters of the large-scale learning model for predicting the hydrostatic performance of a ship are optimized based on multiple loss functions, including a prediction task loss function and an Archimedes' law constraint loss function.

[0008] Preferably, the specific training method steps of the general-purpose ship hydrostatic performance prediction learning model include: The batch training samples are input into the input layer of the prediction learning model. The input layer adds a physical quantity label to each element in the sample set. The batch training samples do not contain prediction task information. The input preprocessing module normalizes the information output from the input layer and performs linear projection or feature extraction; The feature embedding module performs embedding mapping on the information output by the input preprocessing module to obtain the feature embedding body; The multi-head attention layer performs normalization processing on the output information of the feature embedding module in sequence, calculates multi-head attention and performs residual connection to learn the complex coupling relationship between multiple variables; The feedforward network layer performs both fixed and variable prediction task paths in parallel. Based on the multi-layer stacked structure designed by the large predictive learning model, multi-layer cyclic attention calculation and feedforward network mapping are completed. The regression output layer outputs the ship's hydrostatic performance parameters for the prediction task from the feedforward network layer. Calculate the total loss function and optimize the model parameters using stochastic gradient descent until the training termination condition is met, then save the trained model parameters.

[0009] Preferably, the prediction task includes predicting the ship's hydrostatic performance parameters; the fixed prediction task includes hydrostatic performance prediction tasks that must be completed according to regulations or industry practices, such as ship buoyancy, intact stability, and damaged stability; the variable prediction task includes hydrostatic performance prediction tasks that the designer decides temporarily during the performance evaluation process based on specific design requirements, in addition to the fixed prediction tasks, such as combinations of one or more hydrostatic parameters that have a single-valued functional relationship with other ship design parameters; the fixed prediction task path includes inputting the multi-head attention layer output to multiple directly connected expert modules, including at least: The system comprises three expert modules: a buoyancy expert module, a complete stability expert module, and a damaged stability expert module. Each expert module maps nonlinear relationships and performs its own normalization and residual connection calculations. The variable prediction task path consists of a prediction task traversal module, a router, and expert modules. Normalization is performed sequentially. The prediction task traversal module traverses all variable prediction tasks, generates a prediction task mask matrix based on the input layer information, and multiplies it with the output of the multi-head self-attention layer to generate a data block containing prediction task information. The router guides the masked data block to the corresponding expert module to map nonlinear relationships and perform residual connections according to the prediction task.

[0010] Preferably, the loss function includes Task number , This refers to the number of prediction tasks and sample numbers. , This refers to the number of training samples in a batch. (1) Loss function 1 - Prediction task loss function Mean squared error is chosen as the loss function. ; in, For sample values, This is a predicted value; (2) Loss function 2-Archimedes' law loss function The difference between a ship's weight and its displacement is zero. ; in, This is a predicted value for the ship's weight. , This is the predicted empty ship weight. This is the predicted load capacity. This is a predicted value for ship displacement. , This is the predicted drainage volume. It is the specific gravity of water; (3) Loss function 3 - Loss function for the lateral centering of the ship's center of gravity and center of buoyancy boundary conditions The ship's center of gravity and center of buoyancy are centered laterally, meaning their lateral coordinates are zero. ; in, This is a predicted value for the lateral position of the ship's center of gravity. Predicted lateral position of the ship's center of buoyancy; (4) Loss function 4 - The longitudinal positions of the ship's center of gravity and center of buoyancy are on the same vertical line. Boundary condition loss function The ship's center of gravity and center of buoyancy are longitudinally aligned on the same vertical line. ; in, This is a predicted value for the longitudinal position of the ship's center of gravity. This is the predicted value for the longitudinal position of the ship's center of buoyancy. The longitudinal angle is the ship's trim angle under equilibrium conditions. (5) Monotonicity penalty function Some hydrostatic performance data, such as displacement volume, displacement rate, and moment of inclination per centimeter, increase with increasing draft and decrease with decreasing draft. ; in:" SB represents the increment of the buoyancy parameter vector. T" represents the increase in draft; (6) Total loss function ,in, These are the weights of the loss function.

[0011] Preferably, the general-purpose ship hydrostatic performance prediction model includes two sub-models: a fixed prediction task prediction model and a variable prediction task prediction model. The feedforward neural network layer of the fixed prediction task prediction model consists of multiple concurrent expert modules and two layers of residual connections and normalization. The feedforward neural network layer of the variable prediction task prediction model consists of normalization, a router, expert modules, and residual connections. The router routes to the relevant expert modules according to the prediction task. The ship hydrostatic performance prediction method includes obtaining ship design parameters and variable prediction tasks from the general-purpose ship hydrostatic performance prediction model, determining the collaborative mode of the two sub-models based on the ship design parameters and variable prediction tasks, and completing multiple prediction tasks. The variable prediction task information and ship design parameters are data input manually or passed in by the caller. The collaborative mode of the two sub-models includes parallel and serial modes.

[0012] Preferably, the specific steps of the ship hydrostatic performance prediction method include: (1) Start the prediction task. The start prediction task includes the manual start prediction task and the caller start prediction task. The original input information is input into the "input layer" of the general ship hydrostatic performance prediction model. The manual start prediction task is started by manually inputting or importing information from the human-computer interaction page. The caller start prediction task is started by the information transmitted from the ship concept design system. (2) Organize the input information. The organized information includes adding physical quantity identifiers to the input information by the input layer, and masking the input information and the relevant information of the prediction task according to the prediction task information. It is determined whether the prediction value of the variable prediction task affects the prediction result of the fixed prediction task, and the collaborative mode of the two sub-models, the fixed prediction task prediction model and the variable prediction task prediction model, is determined. (3) Complete the prediction task, which includes predicting the hydrostatic performance values ​​of the ship according to the determined sub-model collaborative mode; (4) Return the predicted value, which includes the prediction result displayed by the initiator or the prediction result returned to the caller.

[0013] Preferably, the collaborative mode of the two sub-models, the fixed prediction task prediction model and the variable prediction task prediction model, includes a series mode and a parallel mode. The serial mode includes inputting the sorted input information into a variable prediction task prediction model, which predicts the ship's hydrostatic performance value and updates the masked information in the input information before inputting it into a fixed prediction task prediction model, which outputs the ship's hydrostatic performance prediction value for the fixed prediction task. The parallel mode includes inputting the sorted input information into a fixed prediction task prediction model and a variable prediction task prediction model, respectively, and the fixed prediction task prediction model and the variable prediction task prediction model output the prediction performance value of the given prediction task, respectively.

[0014] Preferably, the training of the general-purpose ship hydrostatic performance prediction learning model in step three is carried out through a training system, which includes a general-purpose ship hydrostatic performance sample processing module, a general-purpose ship hydrostatic performance training sample set buffer, a general-purpose ship hydrostatic performance sample enhancement model, and a general-purpose ship hydrostatic performance prediction learning model.

[0015] Compared with the prior art, the technical solution provided by this invention has the following advantages: This invention provides a method for training, deploying, and implementing a large-scale model for predicting the hydrostatic performance of ships based on large-scale model technology. By innovatively applying artificial intelligence large-scale model technology, it effectively addresses the core pain point of evaluating the hydrostatic performance of design schemes in the conceptual design of traditional ships, especially innovative ships. Its core advantages and value are reflected in the following aspects: 1. Break through traditional performance evaluation paradigms to achieve data-driven intelligent prediction. Traditional methods heavily rely on empirical formulas and other designer experience or existing parent ship data. To ensure the accuracy of ship hydrostatic performance calculations, the process must use hull lines diagrams, which is time-consuming and hinders innovative design solutions. This invention constructs a multi-task predictive deep learning model trained on massive amounts of real samples and augmented data. Based on this trained model, the structure is adjusted to construct another multi-task predictive model. This model can automatically and quickly predict multiple hydrostatic performance indicators such as buoyancy and stability directly from ship conceptual design parameters. Without compromising the accuracy of the predicted performance indicators, it eliminates the step of calculating ship hydrostatic performance through hull line diagram generation, greatly improving the efficiency and scientific rigor of ship hydrostatic performance prediction and evaluation, and providing ample space for innovative ship design.

[0016] 2. Integrate physical laws to ensure the scientific validity of predictions. The model innovatively adopts a multi-task learning model training method that integrates physical law constraints. It designs a dedicated Archimedes' principle loss function and embeds Archimedes' principle as a constraint into the training process to ensure that the prediction results not only fit the statistical laws of the data, but also conform to the physical laws, thereby making the hydrostatic performance prediction results scientific. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of training, deploying, and implementing the general-purpose large-scale model for predicting the hydrostatic performance of ships according to the present invention. Figure 2 This is a schematic diagram of the structure of the large-scale model training and implementation system for predicting the hydrostatic performance of ships according to the present invention; Figure 3 This is a schematic diagram of the training process of the general ship hydrostatic performance prediction learning model of the present invention; Figure 4 This is a schematic diagram of the general-purpose ship hydrostatic performance prediction learning model of the present invention; Figure 5 This is a schematic diagram of the structure of the fixed prediction task prediction model of the present invention; Figure 6 This is a schematic diagram of the structure of the variable prediction task prediction model of the present invention; Figure 7 The flowchart for predicting the hydrostatic performance of a ship is shown in the example.

[0018] Explanation of the labels in the diagram: 101. General-purpose ship hydrostatic performance sample processing module; 102. General-purpose ship hydrostatic performance training sample set buffer; 103. General-purpose ship hydrostatic performance sample enhancement model; 104. General-purpose ship hydrostatic performance prediction learning large model; 105. Ship hydrostatic performance prediction large model; 105A. Fixed prediction task prediction large model; 105B. Variable prediction task prediction large model; 106. Human-computer interaction interface; 107. Ship concept design system. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0022] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in some cases to indicate a certain dependency or connection relationship. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0023] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linking," and "socketing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] See attached document Figure 1 This embodiment describes a method for training, deploying, and implementing a large-scale model for predicting the hydrostatic performance of ships based on large-scale model technology, comprising the following steps: Step 1: Processing the general sample set of ship hydrostatic performance (1) Collect samples of ship hydrostatic performance: systematically extract sample data related to hydrostatic performance from the completed technical documents of ships of various sizes (including at least the main technical specifications such as ship type and purpose, hull geometric parameters, superstructure geometric parameters, lightship weight and center of gravity details, ship compartment and compartment volume details, buoyancy data, integrity stability and damaged stability data, and other data related to ship hydrostatic performance). (2) Construct a sample set of ship hydrostatic performance: organize the sample data according to the standard format and order; clean the original sample data to remove noise, abnormal data and data that does not satisfy Archimedes' law, and store the sample data to construct a sample set of ship hydrostatic performance.

[0026] Step 2: Enhance the general sample set for the hydrostatic performance of synthetic ships Currently, real-world ship samples are concentrated within local domains of existing ship types, resulting in a sparse distribution of samples and limiting the ability to evaluate the hydrostatic performance of innovative ship conceptual designs. To address this issue, after completing step one, this step prioritizes the use of a diffusion-based sample enhancement and synthesis technique. This technique interpolates and fills existing real-world samples and generates out-of-distribution samples, significantly expanding the diversity and coverage of the samples without violating physical laws, thus constructing a universally applicable sample set of ship hydrostatic performance with a complete distribution.

[0027] Step 3: Training a general-purpose ship hydrostatic performance prediction learning model 104 After completing step two, a general-purpose ship hydrostatic performance prediction learning model 104 is trained. The model trained in this step is a multi-task prediction learning model for ship hydrostatic performance designed and constructed according to the requirements of this invention, referencing the Hybrid Expert (MoE) model concept. The training steps include reading batch training samples, input preprocessing, feature extraction and embedding, sharing the ship hydrostatic knowledge space (using a multi-head self-attention mechanism to capture complex coupling relationships between variables), guiding expert modules through parallel paths for fixed and variable prediction tasks, and a regression prediction layer predicting multi-task ship hydrostatic performance values. The fixed prediction task path step includes all data blocks being directly input into the fixed task expert module (such as buoyancy, integrity, and stability expert modules) in parallel. The variable prediction task path step includes traversing the variable prediction tasks and guiding data blocks to the corresponding variable task expert modules through masking and routing mechanisms. During training, in addition to using a conventional prediction error loss function, a constraint loss function based on physical laws such as Archimedes' principle and the center of gravity-buoyancy relationship is specifically introduced to achieve a dual optimization mechanism of "data-driven + physical guidance". The prediction task involves forecasting the ship's hydrostatic performance parameters, and is divided into fixed prediction tasks and variable prediction tasks. Fixed prediction tasks refer to hydrostatic performance prediction tasks that are required by regulations or industry practice, including at least predictions of buoyancy, intact stability, and damaged stability. Variable prediction tasks refer to hydrostatic performance prediction tasks that, in addition to fixed prediction tasks, are temporarily decided by the designer during the performance evaluation process based on specific design requirements, such as combinations of one or more hydrostatic parameters that have a single-valued functional relationship with other ship design parameters.

[0028] Step 4: Deploy a large-scale model for predicting the hydrostatic performance of general-purpose ships. Based on the general ship hydrostatic performance prediction learning model (104) after completing step three training, retain the trained parameters, adjust the learning model structure to build and deploy the general ship hydrostatic performance prediction model (105).

[0029] Step 5: Predict the ship's hydrostatic performance After completing the deployment of the large-scale prediction model in step four, the general-purpose ship hydrostatic performance prediction model 105 receives information input from the human-computer interaction interface or external design system (including ship design parameters and variable prediction task information), and automatically selects the serial or parallel collaborative mode according to the dependency relationship between the fixed prediction task and the variable prediction task, driving the corresponding sub-model to output the hydrostatic performance parameters of each prediction task.

[0030] Implementation Results: Through the training, deployment, and implementation method of the large-scale ship hydrostatic performance prediction model based on the large-scale model technology, this embodiment realizes the entire process from collecting samples, constructing and enhancing the synthetic training sample set, training and learning the large-scale model, constructing the prediction large-scale model, to completing the ship hydrostatic performance prediction task. It realizes the function of automatically estimating the ship's hydrostatic performance from the ship's conceptual design elements without generating hull line drawings, replacing the work of human ship designers, improving work efficiency, reducing work costs, and providing a highly efficient and high-quality support tool for autonomous ship conceptual design work based on artificial intelligence technology.

[0031] See attached document Figure 2 The training, deployment, and implementation method of the ship hydrostatic performance prediction learning large model based on large model technology described in this embodiment is accomplished through a ship hydrostatic performance prediction large model training and implementation system. The system includes a training part and an implementation part. The training part includes: a general ship hydrostatic performance sample processing module 101, a general ship hydrostatic performance training sample set buffer 102, a general ship hydrostatic performance sample enhancement model 103, and a general ship hydrostatic performance prediction learning large model 104. The implementation part includes: a general ship hydrostatic performance prediction large model 105, a human-computer interaction interface 106, and a ship conceptual design system 107. "General" means that the model and method have universality applicable to all types and sizes of ships. The general-purpose ship hydrostatic performance sample processing module 101: This module collects hydrostatic-related samples from the as-built design documents of ships of all types and sizes, performs cleaning processing on these samples, and stores the processed samples in the "general-purpose ship hydrostatic performance training sample set cache 103".

[0032] General-purpose ship hydrostatic performance training sample set cache 102: responsible for storing ship hydrostatic performance training sample sets, which include actual ship completed samples from the general-purpose ship hydrostatic performance sample processing module 101 and samples generated by the general-purpose ship hydrostatic performance sample enhancement model 103.

[0033] General-purpose ship hydrostatic performance sample enhancement model 103: It collects samples from the general-purpose ship hydrostatic performance training sample set buffer 102, and uses diffusion model technology to expand the diversity and coverage of training samples without violating the physical laws of ship hydrostatics by interpolating and filling existing samples and generating out-of-distribution samples, thus constructing a general-purpose ship hydrostatic performance sample set with complete distribution.

[0034] The general-purpose ship hydrostatic performance prediction model 105 includes an input layer and two sub-models. The input layer represents a variable prediction task through a masking mechanism. The two sub-models include a fixed prediction task prediction model 105A and a variable prediction task prediction model 105B. The fixed prediction task prediction model 105A performs the fixed prediction task, and the variable prediction task prediction model 105B performs the variable prediction task.

[0035] Human-computer interaction interface 106: Provides a graphical operation interface for manual input of ship static performance prediction task information, input or import of data required for predicting ship static performance, and displays prediction results.

[0036] Ship Concept Design System 107: Used for ship concept design, it can be a traditional computer-aided manual design system or an AI-based automated ship concept design system. For each concept design scheme, the general ship hydrostatic performance prediction model 105 of this invention is called to predict the ship's hydrostatic performance to evaluate the merits of the ship concept design scheme.

[0037] See attached document Figure 3 The training method for the general ship hydrostatic performance prediction learning large model in this embodiment includes the following steps: 1. Read batch training samples The "input layer" reads batch training sample sets from the "general-purpose ship hydrostatic performance training sample set buffer 102". The training samples do not contain prediction task information. The input layer adds physical quantity labels to each parameter in the sample set and inputs it into the "input preprocessing module".

[0038] 2. Preprocessing and feature embedding of samples The "Input Preprocessing Module" normalizes the training samples and performs linear projection or extracts features, which are then input into the "Feature Embedding Module" to output embeddings that meet the model requirements.

[0039] 3. Capturing the coupling relationships between multivariate variables The multi-head attention layer performs normalization processing sequentially, captures the coupling relationship between multivariate variables, outputs a representation of the coupling relationship between multivariate variables and residual connections, and inputs them into the "feedforward network layer".

[0040] 4. Perform fixed prediction task paths and variable prediction task paths in parallel. The feedforward network layer performs both fixed and variable prediction task paths in parallel.

[0041] (1) Fixed prediction task expert module mapping nonlinear relationship The input is fed into a fixed prediction task expert module that includes at least buoyancy (SB), integrity stability (IS) and damaged stability (DS) to map the nonlinear relationship and perform their respective normalization and residual connection calculations, and then proceeds to step 5 together with step (3).

[0042] (2) Traverse the variable prediction task and embed it into the input of the feedforward network layer. In sync with step (1), first normalization is performed, and then the "prediction task traversal module" traverses all variable prediction tasks and masks the vectors related to the prediction task from the output of the "multi-head self-attention layer".

[0043] (3) Based on the prediction task, guide the corresponding expert module to map the nonlinear relationship. After completing step (2), the "router" guides the masked data blocks to the corresponding expert modules to map nonlinear relationships according to the prediction task, and then calculates the residual connections.

[0044] 5. Determine if all stacking layers have been calculated. After completing the fixed and variable prediction task paths, determine whether all stacking layers have been calculated. If not, proceed to the "next layer" and return to step 3.

[0045] 6. The regression prediction layer outputs the regression values ​​for the ship's hydrostatic prediction given the prediction task. 7. Calculate all loss functions 8. Determine whether the loss function meets the requirements. If the loss function does not meet the requirements, then "stochastic gradient descent optimizes the model parameters" and proceeds to step 2.

[0046] 9. Determine whether the training meets the termination criteria. If the termination condition is not met, "reorder the samples" and proceed to step 1.

[0047] 10. Save the trained model parameters. See attached document Figure 4 The general ship hydrostatic performance prediction learning model 104 in this embodiment is designed and constructed based on Archimedes' principle and with reference to the Hybrid Expert (MoE) architecture concept, according to the requirements of this invention.

[0048] 1. Learn the principles of large-scale models When a ship floats in still water and is in equilibrium, it obeys Archimedes' principle. A ship's hydrostatic performance refers to its performance under the influence of gravity and buoyancy in still water, including at least buoyancy, integrity stability, and damage stability. These properties are closely related to the ship's hull geometry, superstructure geometry, center of gravity (lightweight), deadweight, and draft. The center of gravity (lightweight) is closely related to design specifications such as class, regulations, navigation area, structural materials, and machinery. The deadweight parameters are also related to design specifications and determine cargo capacity. These complex relationships can be expressed by the following formula: (1) Archimedes' principle (a) Force equilibrium equations When a ship floats in still water and is in equilibrium, the condition that buoyancy and gravity are equal according to Archimedes' principle is: ; in, For the ship's weight, including both lightship weight and deadweight, i.e., a single-valued function. ; in, The empty weight of the vessel. ; The design task indicators are derived from the "Design Technical Task Book" and are implemented upon ship delivery. These include indicators that affect hydrostatic performance, such as ship type, load capacity, class, specifications, navigation area, structural materials, power plant and mechanical equipment, speed, endurance, and self-sufficiency. Variables representing hull geometry features, including principal dimensions, hull form factors, and hull line feature parameters; : Geometric characteristic parameters of the superstructure; : Draft of the midships of a ship.

[0049] The load capacity, derived from the "Design Technical Specification," includes cargo / passenger capacity, crew and baggage, fuel oil, lubricating oil and boiler water, fresh water and food, spare parts and supplies, etc. A single-valued function exists. ; Ship capacity (Capacity) characteristic parameters include volume (cargo hold volume, oil and water tank volume, dedicated ballast water tank volume, engine room volume, and other tank volume), deck area, and other variables.

[0050] Variables are physical properties of cargo, ballast, and passengers, including cargo name, cargo density, cargo stowage factor, ballast water density, crew, passenger and baggage weight, fuel and lubricating oil density, and other physical properties related to load capacity.

[0051] The gravity acting on a ship can be expressed as ; This refers to the ship's displacement.

[0052] It is the density of water.

[0053] Let be the total displacement volume, and be the vector of ship buoyancy characteristic parameters. One of the components can be represented as a single-valued function. ; The force equilibrium equation can also be expressed as: ; (b) Torque balance equation When a ship is upright in still water, the ship's center of gravity must also be satisfied. and floating heart On the same plumb line ; in, The center of gravity lever arm vector, It is the vector of the buoyancy center lever arm.

[0054] According to the traditional ship coordinate system (the origin is the intersection of the midship section and the baseline, and the direction is towards the bow), Forward, pointing to the starboard side (positive), the equation can be decomposed into: ; ; in, Longitudinal position of the center of buoyancy The trim angle is the angle of inclination of the ship in equilibrium. The central longitudinal position, The center of gravity is the lateral position. Variables related to the center of gravity position are categorized into those related to the empty ship's weight and those related to the center of gravity position. and load capacity, weight, center of gravity, position, and other variables All of these can be represented as characteristic parameters of the ship set. A single-valued function.

[0055] Due to the weight of the ship Center of gravity position, drainage volume Since both are single-valued functions of the independent variable, Archimedes' principle can be expressed by the following implicit function: ;

[0056] (2) Floatability ; in, This is a vector of ship buoyancy characteristic parameters, such as hydrostatic curves, draft at the bow and stern, and freeboard.

[0057] (3) Integrity and stability ; in, This is a vector of characteristic parameters of the ship's intrinsic stability, such as initial metacentric height and large tilt angle stability eigenvalues.

[0058] (4) Damaged stability ; in, This is a vector of characteristic parameters for ship damage stability.

[0059] (5) Other hydrostatic properties ; in, For the first A vector of hydrostatic performance characteristic parameters, .

[0060] 2. Learning the structure of large models See attached document Figure 4 The general-purpose ship hydrostatic performance prediction learning model 104 in this embodiment includes an input layer, an input preprocessing module, a feature embedding module, a multi-head self-attention layer, a feedforward neural network layer, and a regression prediction layer. The multi-head self-attention layer consists of normalization processing, multi-head attention calculation, and residual connection calculation. The feedforward neural network layer has two paths: a fixed task path and a variable task path. The fixed task path does not go through a router and is directly connected to the expert module, where normalization and residual connection calculation are performed. The variable task path consists of normalization processing, a prediction task traversal module, a router, an expert module, and residual connection calculation. The multi-head self-attention layer and the feedforward neural network layer are the same as those in the Transformer model, consisting of multiple layers stacked together.

[0061] (1) Input layer: Input training sample information, add physical quantity identification information to the training sample information, and do not include prediction task information.

[0062] (2) Input preprocessing module: Normalizes the elements of the training samples and performs linear projection or feature extraction.

[0063] (3) Feature embedding module: Projects the output of the "input preprocessing module" into an embedding that meets the model requirements.

[0064] (4) Multi-head self-attention layer: It consists of normalization, multi-head self-attention and residual connection in sequence, and performs its own calculations to learn and express the complex coupling relationship between multiple variables shared by multiple tasks.

[0065] (5) Feedforward neural network layer: divided into fixed prediction task path and variable prediction task path.

[0066] (a) The fixed prediction task path consists of two layers: expert modules and residual connections + normalization. The expert modules are feedforward neural networks (FFN), including at least fixed prediction task expert modules such as buoyancy expert module FFN-SB, completion stability expert module FFN-IS, and damage stability expert module FFN-DS. Each expert module learns the multivariate nonlinear mapping relationship of its own task.

[0067] (b) The variable prediction task path consists of normalization, prediction task traversal module, router, expert module, and residual connection. The prediction task traversal module traverses all prediction tasks, generates a prediction task mask matrix based on the input layer information, and multiplies it with the output of the multi-head self-attention layer to generate an output containing prediction task information. The router routes to the expert module of the relevant task according to the prediction task. The expert module is a feedforward neural network (FFN), and each expert module learns the multivariate nonlinear mapping relationship of its own task.

[0068] (6) Regression prediction layer: Output the predicted hydrostatic performance parameters of the ship corresponding to the prediction task.

[0069] 3. Loss function for training the model The loss function is calculated for each batch of training samples under the same prediction task, where the task number is... , This refers to the number of prediction tasks and sample numbers. , This refers to the number of training samples in a batch. (1) Loss function 1 - Prediction task loss function Mean squared error is chosen as the loss function. ; in, For sample values, This is a predicted value; (2) Loss function 2-Archimedes' law loss function The difference between a ship's weight and its displacement is zero. ; in, This is a predicted value for the ship's weight. , This is the predicted empty ship weight. This is the predicted load capacity. This is a predicted value for ship displacement. , This is the predicted drainage volume. It is the specific gravity of water; (3) Loss function 3 - Loss function for the lateral centering of the ship's center of gravity and center of buoyancy boundary conditions The ship's center of gravity and center of buoyancy are centered laterally, meaning their lateral coordinates are zero. ; in, This is a predicted value for the lateral position of the ship's center of gravity. Predicted lateral position of the ship's center of buoyancy; (4) Loss function 4 - The longitudinal positions of the ship's center of gravity and center of buoyancy are on the same vertical line. Boundary condition loss function The ship's center of gravity and center of buoyancy are longitudinally aligned on the same vertical line. ; in, This is a predicted value for the longitudinal position of the ship's center of gravity. This is the predicted value for the longitudinal position of the ship's center of buoyancy. The longitudinal angle is the ship's trim angle under equilibrium conditions. (5) Monotonicity penalty function Some hydrostatic performance data, such as displacement volume, displacement rate, and moment of inclination per centimeter, increase with increasing draft and decrease with decreasing draft. ; in:" SB represents the increment of the buoyancy parameter vector. T" represents the increase in draft; (6) Total loss function ,in, These are the weights of the loss function.

[0070] See attached document Figure 5 The fixed multi-task prediction large model 105A structure includes an input preprocessing module, a feature embedding module, a multi-head self-attention layer, a feedforward neural network layer, and a regression prediction layer.

[0071] (1) Input preprocessing module: Normalizes the output information of the input layer and performs linear projection or feature extraction.

[0072] (2) Feature embedding module: Projects the output of the "input preprocessing module" into an embedding that meets the model requirements.

[0073] (3) Multi-head self-attention layer: It consists of normalization, multi-head self-attention and residual connection in sequence, and performs its own calculations to learn and express the complex coupling relationship between multiple variables shared by multiple tasks.

[0074] (4) Feedforward Neural Network Layer: It consists of two layers: expert modules and residual connection + normalization. The expert modules are feedforward neural networks (FFN), including at least fixed prediction task expert modules such as buoyancy expert module FFN-SB, completion stability expert module FFN-IS and damage stability expert module FFN-DS. Each expert module learns the multivariate nonlinear mapping relationship of its own task.

[0075] (5) Regression prediction layer: Outputs the regression values ​​of the ship hydrostatic performance parameters corresponding to the prediction task.

[0076] See attached document Figure 6 The variable prediction task prediction model 105B includes an input preprocessing module, a feature embedding module, a multi-head self-attention layer, a feedforward neural network layer, and a regression prediction layer.

[0077] (1) Input preprocessing module: Normalizes the output information of the input layer and performs linear projection or feature extraction.

[0078] (2) Feature embedding module: Projects the output of the "input preprocessing module" into an embedding that meets the model requirements.

[0079] (3) Multi-head self-attention layer: It consists of normalization, multi-head self-attention and residual connection in sequence, and performs its own calculations to learn and express the complex coupling relationship between multiple variables shared by multiple tasks.

[0080] (4) Feedforward neural network layer: It consists of normalization, router, expert module and residual connection. The router routes to the expert module of the relevant task according to the variable prediction task. The expert module is a feedforward neural network (FFN). Each expert module learns the multivariate nonlinear mapping relationship of its own task. (5) Regression prediction layer: Outputs the regression values ​​of the ship hydrostatic performance parameters corresponding to the prediction task.

[0081] See attached document Figure 7 The method for predicting the general ship hydrostatic performance using the large-scale prediction model 105 in this embodiment includes the following steps: 1. Start the prediction task. The start prediction task includes manual start prediction task and caller start prediction task. Input the prediction task information and related information into the "input layer" of the general ship hydrostatic performance prediction model 105. The manual start prediction task is started by manually inputting or importing information from the human-computer interaction page 106. The caller start prediction task is started by information transmitted from the ship concept design system 107. 2. Organize the input information. This includes adding physical quantity identifiers to the input information at the input layer, masking the input information and related information based on the prediction task information, determining whether the predicted value of the variable prediction task affects the prediction result of the fixed prediction task, and determining the collaborative mode of the two sub-models, the fixed prediction task prediction model 105A and the variable prediction task prediction model 105B. The collaborative mode of the two sub-models includes parallel and series modes. If the predicted value of the variable prediction task affects the prediction result of the fixed prediction task, proceed to step "3. Branch Flow 1 - Variable Prediction Task and Fixed Prediction Task in Series"; if the predicted value of the variable prediction task does not affect the prediction result of the fixed prediction task, proceed to step "4. Branch Flow 2 - Variable Prediction Task and Fixed Prediction Task in Parallel". 3. Branch Flow 1 - Cascaded Variable Prediction Task and Fixed Prediction Task (1) Input the input information into the variable task model Enter the compiled input information into "Variable Prediction Task Prediction Large Model 105B".

[0082] (2) Variable task model output prediction performance value The "Variable Prediction Task Prediction Model 105B" outputs the predicted values ​​of the ship's hydrostatic performance for the given prediction task.

[0083] (3) Update the input information of the fixed task model The masked information in the input information is updated by updating the ship hydrostatic performance prediction value output by "Variable Prediction Task Prediction Model 105B" and then input into "Fixed Prediction Task Prediction Model 105A".

[0084] (4) Fixed task model output prediction performance value The "Fixed Prediction Task Prediction Model 105A" outputs predicted values ​​of ship hydrostatic performance.

[0085] 4. Branch Flow 2 - Variable prediction tasks and fixed prediction tasks in parallel (1) Input the input information into the fixed task and variable task models respectively. Enter the compiled input information into "Fixed Prediction Task Prediction Model 105A" and "Variable Prediction Task Prediction Model 105B" respectively.

[0086] (2) The fixed task model and the variable task model output prediction performance values ​​respectively. The "Fixed Prediction Task Prediction Model 105A" and "Variable Prediction Task Prediction Model 105B" output the predicted values ​​of the ship's hydrostatic performance for the given prediction task, respectively.

[0087] 5. Return the predicted value, which may include the prediction result displayed by the initiator or the prediction result returned to the caller. If the prediction task is initiated manually, the output predicted value of the ship's hydrostatic performance is displayed on the "Human-Computer Interaction Page 107". If the prediction task is initiated by the caller, the output predicted value of the ship's hydrostatic performance is returned.

[0088] The above-described embodiments are merely illustrative of certain implementations of the present invention, and are described in a relatively specific and detailed manner. However, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for training, deploying, and implementing a large-scale model for predicting the hydrostatic performance of ships based on large-scale model technology, characterized in that: Includes the following steps: Step 1: Process the general sample set of ship hydrostatic performance. The general sample set includes hydrostatic performance completion data collected from all ship completion technical documents. The collected sample set is then cleaned, organized, and stored. Step 2: Enhance the general sample set of ship hydrostatic performance, including interpolating and filling the sample set based on diffusion model technology to synthesize and generate out-of-distribution samples to enhance and expand the sample set of ship hydrostatic performance. Step 3: Train a general learning model for predicting the hydrostatic performance of ships (104), including reading batch training samples, input preprocessing, feature extraction and embedding, sharing the knowledge space of hydrostatic performance of ships, guiding the expert module with parallel paths for fixed prediction tasks and variable prediction tasks, traversing the variable prediction tasks and guiding the expert module through masking and routing mechanisms to realize multi-task prediction of hydrostatic performance of ships, optimizing model parameters with stochastic gradient descent, and ensuring that the prediction results satisfy Archimedes' law through the loss function of Newton's first law; Step 4: Deploy the general ship hydrostatic performance prediction model (105). The general ship hydrostatic performance prediction model includes two sub-models: a fixed prediction task prediction model (105A) and a variable prediction task prediction model (105B). The two sub-models are constructed by adjusting the model structure based on the trained general ship hydrostatic performance prediction learning model (104). Step 5: Predict the hydrostatic performance of the ship. This includes organizing the information on the variable prediction task and the ship's conceptual design scheme and inputting it into the general ship hydrostatic performance prediction model (105). Based on the relationship between the variable prediction task and the fixed prediction task, the two sub-models work together to predict the ship's hydrostatic performance parameters.

2. The method for training and deploying a large-scale model for predicting the hydrostatic performance of ships based on artificial intelligence, as described in claim 1, is characterized in that: The general ship hydrostatic performance prediction learning model (104) includes an input layer, an input preprocessing module, a feature embedding module, a multi-head self-attention layer, a feedforward neural network layer, and a regression prediction layer. The input layer obtains a training sample set and trains the general ship hydrostatic performance prediction learning model (104) based on the sample set. The feedforward neural network layer has two parallel data flow paths: a fixed prediction task path and a variable prediction task path. The large-scale learning model for predicting the hydrostatic performance of a ship completes multiple prediction tasks. The parameters of the large-scale learning model for predicting the hydrostatic performance of a ship are optimized based on multiple loss functions, including a prediction task loss function and an Archimedes' law constraint loss function.

3. The method according to claim 2, characterized in that: The specific training method steps for the general-purpose ship hydrostatic performance prediction learning model include: The batch training samples are input into the input layer of the prediction learning model. The input layer adds a physical quantity label to each element in the sample set. The batch training samples do not contain prediction task information. The input preprocessing module normalizes the information output from the input layer and performs linear projection or feature extraction; The feature embedding module performs embedding mapping on the information output by the input preprocessing module to obtain the feature embedding body; The multi-head attention layer performs normalization processing on the output information of the feature embedding module in sequence, calculates multi-head attention and performs residual connections to learn the complex coupling relationship between multiple variables; The feedforward network layer performs both fixed and variable prediction task paths in parallel. Based on the multi-layer stacked structure designed by the large predictive learning model, multi-layer cyclic attention calculation and feedforward network mapping are completed. The regression output layer outputs the ship's hydrostatic performance parameters for the prediction task from the feedforward network layer. Calculate the total loss function and optimize the model parameters using stochastic gradient descent until the training termination condition is met, then save the trained model parameters.

4. The method according to claim 3, characterized in that, The prediction tasks include predicting the ship's hydrostatic performance parameters; the fixed prediction tasks include hydrostatic performance prediction tasks that must be completed according to regulations or industry practices, such as ship buoyancy, integrity stability and damaged stability; the variable prediction tasks include hydrostatic performance prediction tasks that the designer decides temporarily during the performance evaluation process according to the specific design task requirements, in addition to the fixed prediction tasks, such as combinations of one or more hydrostatic parameters that have a single-valued functional relationship with other ship design parameters. The fixed prediction task path includes inputting the multi-head attention layer output to multiple directly connected expert modules, including at least: a buoyancy expert module, a complete stability expert module, and a damaged stability expert module. Each expert module maps nonlinear relationships and performs its own normalization and residual connection calculation. The variable prediction task path consists of a prediction task traversal module, a router, and expert modules. Normalization is performed sequentially. The prediction task traversal module traverses all variable prediction tasks, generates a prediction task mask matrix based on the input layer information, multiplies it with the multi-head self-attention layer output to generate a data block containing prediction task information, and the router guides the masked data block to the corresponding expert module to map nonlinear relationships and perform residual connections according to the prediction task.

5. The method according to claim 2, characterized in that: The loss function includes The loss function is calculated for each batch of training samples under the same prediction task, where the task number is... , This refers to the number of prediction tasks and sample numbers. , The number of training samples in a batch; (1) Loss function 1 - prediction task loss function Mean squared error is chosen as the loss function. ； in, For sample values, This is a predicted value; (2) Loss function 2-Archimedes' law loss function The difference between a ship's weight and its displacement is zero. ； in, This is a predicted value for the ship's weight. , This is the predicted empty ship weight. This is the predicted load capacity. This is a predicted value for ship displacement. , This is the predicted drainage volume. It is the specific gravity of water; (3) Loss function 3 - Loss function for the lateral centering of the ship's center of gravity and center of buoyancy boundary conditions The ship's center of gravity and center of buoyancy are centered laterally, meaning their lateral coordinates are zero. ； in, This is a predicted value for the lateral position of the ship's center of gravity. Predicted lateral position of the ship's center of buoyancy; (4) Loss function 4 - The longitudinal positions of the ship's center of gravity and center of buoyancy are on the same vertical line. Boundary condition loss function The ship's center of gravity and center of buoyancy are longitudinally aligned on the same vertical line. ； in, This is a predicted value for the longitudinal position of the ship's center of gravity. This is the predicted value for the longitudinal position of the ship's center of buoyancy. The longitudinal angle is the ship's trim angle under equilibrium conditions. (5) Monotonicity penalty function Some hydrostatic performance data, such as displacement volume, displacement rate, and moment of inclination per centimeter, increase with increasing draft and decrease with decreasing draft. ； in:" SB represents the increment of the buoyancy parameter vector. T" represents the increase in draft; (6) Total loss function ,in, These are the weights of the loss function.

6. The method according to claim 1, characterized in that: The general-purpose ship hydrostatic performance prediction model (105) includes two sub-models: a fixed prediction task prediction model (105A) and a variable prediction task prediction model (105B). The feedforward neural network layer of the fixed prediction task prediction model (105A) consists of multiple concurrent expert modules and two layers of residual connections and normalization. The feedforward neural network layer of the variable prediction task prediction model (105B) consists of normalization, router, expert modules, and residual connections. The router routes to the expert modules of the relevant tasks according to the prediction tasks. The ship hydrostatic performance prediction method includes the general-purpose ship hydrostatic performance prediction model (105) acquiring ship design parameters and variable prediction tasks. Based on the ship design parameters and variable prediction tasks, the collaborative mode of the two sub-models of the general-purpose ship hydrostatic performance prediction model (105) is determined and multiple prediction tasks are completed. The variable prediction task information and ship design parameters are data that is manually input or passed in by the caller. The collaborative mode of the two sub-models includes parallel and serial modes.

7. The method according to claim 6, characterized in that: The specific steps of the method for predicting the hydrostatic performance of a ship include: (1) Start the prediction task. The start prediction task includes the manual start prediction task and the caller start prediction task. The original input information is input into the "input layer" of the general ship hydrostatic performance prediction model (105). The manual start prediction task is started by manually inputting or importing information from the human-computer interaction page (106). The caller start prediction task is started by the information passed in by the ship concept design system (107). (2) Organize the input information. The organized information includes adding physical quantity identifiers to the input information by the input layer, and masking the input information and the relevant information of the prediction task according to the prediction task information. It is determined whether the prediction value of the variable prediction task affects the prediction result of the fixed prediction task, and the collaborative mode of the two sub-models, the fixed prediction task prediction model (105A) and the variable prediction task prediction model (105B), is determined. (3) Complete the prediction task, which includes predicting the hydrostatic performance values ​​of the ship according to the determined sub-model collaborative mode; (4) Return the predicted value, which includes the prediction result displayed by the initiator or the prediction result returned to the caller.

8. The method according to claim 7, characterized in that, The collaborative modes of the two sub-models, the fixed prediction task prediction model (105A) and the variable prediction task prediction model (105B), include serial mode and parallel mode. The serial mode includes inputting the sorted input information into the variable prediction task prediction model (105B), the variable prediction task prediction model (105B) predicts the ship's hydrostatic performance value and updates the masked information in the input information before inputting it into the fixed prediction task prediction model (105A), the fixed prediction task prediction model (105A) outputs the ship's hydrostatic performance prediction value for the fixed prediction task. The parallel mode includes inputting the sorted input information into a fixed prediction task prediction model (105A) and a variable prediction task prediction model (105B), respectively. The fixed prediction task prediction model (105A) and the variable prediction task prediction model (105B) respectively output the prediction performance value of the given prediction task.

9. The method according to claim 1, characterized in that: The training of the general ship hydrostatic performance prediction learning model (104) in step three is carried out through a training system, which includes a general ship hydrostatic performance sample processing module (101), a general ship hydrostatic performance training sample set buffer (102), a general ship hydrostatic performance sample enhancement model (103), and a general ship hydrostatic performance prediction learning model (104).

Images