Search

CN-121997462-A - Large model technology-based ship hydrostatic performance prediction large model training, deployment and implementation method

CN121997462ACN 121997462 ACN121997462 ACN 121997462ACN-121997462-A

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 real ship finishing technical data such as various types, main scales, tonnage and the like, adjusting a model structure to construct the ship hydrostatic performance prediction large model for deployment on the basis of the ship hydrostatic performance prediction learning large model for completing training, and inputting ship design elements and variable prediction tasks into the ship hydrostatic performance prediction large model to predict ship hydrostatic performance indexes. The invention can replace designers to complete the ship hydrostatic performance estimation work, improve the working efficiency and the working level, and provide basic service for a ship concept designer (intelligent body) based on artificial intelligence.

Inventors

  • FENG BO

Assignees

  • 上海太科舟智能科技有限公司

Dates

Publication Date
20260508
Application Date
20260213

Claims (9)

  1. 1. A ship hydrostatic performance prediction large model training, deployment and implementation method based on a large model technology is characterized by comprising the following steps: Step one, processing a general sample set of the hydrostatic performance of a ship, wherein the general sample set comprises the steps of collecting the hydrostatic performance finishing data from all ship finishing technical files, and cleaning, finishing and storing the collected sample set; Step two, a general sample set for enhancing the hydrostatic performance of the synthesized ship is obtained, which comprises the steps of interpolating, filling, synthesizing and generating a sample out of distribution to enhance the hydrostatic performance of the expanded ship based on a diffusion model technology; Training a general ship hydrostatic performance prediction learning large model (104), which comprises the steps of reading a batch of training samples, inputting pretreatment, extracting and embedding characteristics, sharing a ship hydrostatic knowledge space, guiding an expert module by a parallel path of a fixed prediction task and a variable prediction task, traversing the variable prediction task, guiding the expert module through a mask and routing mechanism, realizing the multi-task prediction of the ship hydrostatic performance, optimizing model parameters by random gradient descent, and ensuring that a prediction result meets Archimedes 'law through Newton's first law loss function; a general ship hydrostatic performance prediction big model (105) is deployed, wherein the general ship hydrostatic performance prediction big model comprises two sub-models, namely a fixed prediction task prediction big model (105A) and a variable prediction task prediction big model (105B), and the two sub-models are constructed by adjusting a model structure on the basis of a trained general ship hydrostatic performance prediction learning big model (104); Fifthly, predicting the hydrostatic performance of the ship, wherein the variable prediction task and the ship concept design scheme information are arranged and input into a general ship hydrostatic performance prediction large model (105), and the two sub-models work cooperatively to predict the hydrostatic performance parameters of the ship according to the relation between the variable prediction task and the fixed prediction task.
  2. 2. The ship hydrostatic performance prediction large model training and deployment method based on artificial intelligence of claim 1, wherein the general ship hydrostatic performance prediction learning large model (104) comprises 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, wherein the input layer acquires a training sample set, trains the general ship hydrostatic performance prediction learning large model (104) based on a sample set, and the feedforward neural network layer has two parallel data flow paths, namely a fixed prediction task path and a variable prediction task path; And completing a plurality of prediction tasks through the ship hydrostatic performance prediction learning large model, and optimizing parameters of the ship hydrostatic performance prediction learning large model based on a plurality of loss functions, wherein the loss functions comprise prediction task loss functions and Archimedes law constraint loss functions.
  3. 3. The method according to claim 2, wherein the specific training method step of the general ship hydrostatic performance prediction learning large model comprises the following steps of Inputting a batch of training samples to an input layer of the predictive learning large model, wherein the input layer identifies each element in a sample set by physical quantity, and the batch of training samples does not contain predictive task information; the input preprocessing module normalizes the information output by 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 a feature embedded body; The multi-head attention layer sequentially performs normalization processing according to the output information of the feature embedding module, calculates multi-head attention and performs residual connection so as to learn complex coupling relations among multiple variables; The feedforward network layer carries out fixed prediction task paths and variable prediction task paths in parallel; According to the multi-layer stacking structure designed by the prediction learning large model, completing the attention calculation and feedforward network mapping of multi-layer circulation; the regression output layer outputs the feedforward network layer to regress and output the ship hydrostatic performance parameters of the prediction task; and calculating a total loss function, optimizing model parameters based on a random gradient descent method until the training ending condition is met, and storing the trained model parameters.
  4. 4. The method according to claim 3, wherein the prediction tasks comprise prediction ship hydrostatic performance parameters, the fixed prediction tasks comprise ship hydrostatic performance prediction tasks which are required to be completed by regulations or industry practice, such as ship buoyancy, complete stability, cabin breaking stability and the like, the variable prediction tasks comprise ship hydrostatic performance prediction tasks which are temporarily determined in a performance evaluation process according to specific design task requirements except the fixed prediction tasks, such as one or more hydrostatic parameter combinations with a single-value function relation with other ship design parameters, the fixed prediction task path comprises a plurality of expert modules which are directly connected respectively with a multi-head attention layer output, the expert modules at least comprise a buoyancy expert module, a complete stability expert module and a cabin breaking stability expert module, the expert modules are respectively mapped with nonlinear relations and are respectively subjected to normalization processing and residual error connection calculation, the prediction task path comprises a prediction task traversing module, a router and an expert module, all variable prediction tasks are temporarily determined in a performance evaluation process according to specific design task requirements, the prediction task traversing module is used for generating a mask of the prediction task according to input layer information, the mask is generated by the prediction task mapping module and the mask data is generated from the prediction task layer output mask data after the multi-head attention layer output is connected with the expert module.
  5. 5. The method of claim 2, wherein the loss function comprises The loss function is calculated according to the batch training samples under the same prediction task, wherein the task numbers , Is the predicted task number, sample number , Is the number of batch training samples, (1) loss function 1-predictive task loss function Selecting a mean square error as a loss function ; Wherein, the In order to be a sample value, Is a predicted value; (2) Loss function 2-Archimedes' law loss function The difference between the weight of the ship and the displacement is zero ; Wherein, the As the predicted value of the weight of the ship, , Is a predicted value of the weight of the empty ship, Is a load capacity predicted value; Is a predicted value of the water displacement of the ship, , As a predicted value of the volume of water to be discharged, Is water severe; (3) Loss function 3-center of gravity and center of buoyancy lateral position centered boundary condition loss function The centre of gravity and the transverse position of the floating center of the ship are centered, i.e. the transverse coordinates are zero ; Wherein, the Is a predicted value of the transverse position of the center of gravity of the ship, Predicting a ship floating center transverse position value; (4) Loss function 4-loss function of hull center of gravity and longitudinal position of center of buoyancy at same plumb line boundary condition Center of gravity of ship and floating center of ship body longitudinal position at the same plumb line ; Wherein, the Is a predicted value of the longitudinal position of the center of gravity of the ship, Is a predicted value of the longitudinal position of the floating center of the ship body, Is the trim angle in the balanced state of the ship; (5) Monotonicity penalty function Some hydrostatic performance data, such as displacement volume, displacement, pitch moment per cm, increases with increasing draft, whereas decreases, ; Wherein, is' SB is buoyancy parameter vector increment " T' is draft increment; (6) Total loss function , wherein, Weights are the penalty functions.
  6. 6. The method of claim 1, wherein the general purpose marine hydrostatic performance prediction big model (105) comprises two sub-models of a fixed prediction task prediction big model (105A) and a variable prediction task prediction big model (105B), a feedforward neural network layer of the fixed prediction task prediction big model (105A) consists of a plurality of concurrent expert modules and residual connections and normalization two layers, a feedforward neural network layer of the variable prediction task prediction big model (105B) consists of normalization, a router, an expert module and residual connections, wherein the router routes the prediction task to the expert module of the relevant task, the marine hydrostatic performance prediction method comprises the general purpose marine hydrostatic performance prediction big model (105) obtaining a marine design parameter and a variable prediction task, determining a cooperative mode of the two sub-models of the general purpose marine hydrostatic performance prediction big model (105) and completing a plurality of prediction tasks based on the marine design parameter and the variable prediction task, the variable prediction task information and the marine design parameter manual input or call input data, and the two sub-models comprise a cooperative mode and a parallel mode.
  7. 7. The method of claim 6, wherein the ship hydrostatic performance prediction method comprises the following specific steps: (1) The method comprises the steps that a starting prediction task is started, wherein the starting prediction task comprises a manual starting prediction task and a calling party starting prediction task, original input information is input into an input layer of a general ship hydrostatic performance prediction large model (105), the manual starting prediction task is started by manually inputting or importing information from a man-machine interaction page (106), and the calling party starting prediction task is started by information imported by a ship concept design system (107); (2) Sorting input information, wherein the sorting information comprises an input layer adding physical quantity identification to the input information, masking the input information and prediction task related information according to prediction task information, judging whether a variable prediction task prediction value influences a fixed prediction task prediction result, and determining a cooperative mode of two sub-models of a fixed prediction task prediction big model (105A) and a variable prediction task prediction big model (105B); (3) Completing a prediction task, wherein the completion of the prediction task comprises predicting a ship hydrostatic performance value according to the determined sub-model cooperative mode; (4) And returning the predicted value, wherein the returning the predicted value comprises the predicted result displayed by the initiator or the predicted result returned to the calling party.
  8. 8. The method of claim 7, wherein the collaborative modes of the two sub-models of the fixed predictive task prediction big model (105A) and the variable predictive task prediction big model (105B) include a series mode and a parallel mode; the serial mode comprises the steps that the tidied input information is input into a variable prediction task prediction big model (105B), the variable prediction task prediction big model (105B) predicts a ship hydrostatic performance value, information masked in the input information is updated and then is input into a fixed prediction task prediction big model (105A), and the fixed prediction task prediction big model (105A) outputs a ship hydrostatic performance predicted value of a fixed prediction task; The parallel mode comprises the steps of respectively inputting the well-arranged input information into a fixed prediction task prediction big model (105A) and a variable prediction task prediction big model (105B), and respectively outputting the prediction performance values of given prediction tasks by the fixed prediction task prediction big model (105A) and the variable prediction task prediction big model (105B).
  9. 9. The method according to claim 1, wherein the step three training of the general purpose ship hydrostatic performance prediction learning large model (104) is performed by a training system, and the training system comprises a general purpose ship hydrostatic performance sample processing module (101), a general purpose ship hydrostatic performance training sample set buffer (102), a general purpose ship hydrostatic performance sample enhancement model (103) and a general purpose ship hydrostatic performance prediction learning large model (104).

Description

Large model technology-based ship hydrostatic performance prediction large model training, deployment and implementation method Technical Field The invention belongs to the field of high integration of artificial intelligence technology and ship engineering technology, and particularly relates to a ship hydrostatic performance prediction large model training, deployment and implementation method based on a large model technology. Background The ship hydrostatic performance is the performance of the ship under the influence of gravity and buoyancy in a hydrostatic environment, and at least comprises buoyancy, integrity stability and cabin breaking stability. These properties are required to meet the requirements of the vessel design task book and the vessel safety regulations. The hydrostatic performance of the ship needs to be calculated in the ship conceptual design process so as to evaluate the advantages and disadvantages of the design scheme. With the development of computer-aided ship design, after a design scheme of a ship concept is given, a designer may input design parameters of the scheme into design software to generate an approximate profile and estimate a ship hydrostatic performance value. However, the generated profile is not the optimal profile corresponding to the design parameters, and the calculated ship hydrostatic performance value also has a superposition error. For a long time, the manual workload in the work accounts for a main part, and the design method is time-consuming, labor-consuming and high in cost, and a ship concept design scheme with relatively good hydrostatic performance is not necessarily designed. The traditional ship design method is divided into two major categories, namely a mother ship modification method and a gradual approximation method. The method for modifying the mother ship uses a real ship similar to the main aspect of the ship to be designed as the mother ship, and transforms each design element by a proper method according to the design task requirement to obtain the corresponding element of the new ship. The progressive approximation method is suitable for the situation that the design requirement is complex, and a designer cannot find a proper or complete master ship, and the adopted design element progressive approximation method is used for new ship design. For hundreds of years, mankind designed and built a large number of vessels, accumulating a large number of vessel design, construction and navigation examples and experiences. After the computer is developed, the human beings use the computer tools in the performance calculation of ship design, ship hydrostatic force and the like. There are already well established software tools for calculating the hydrostatic performance of ships with computers, such as those in finland Napa. This work requires the input of a hull pattern. The method for reforming the mother ship in the computer-aided ship design work is to manually adjust the design scheme to calculate and evaluate each performance of the designed ship by using a design software tool on the basis of data such as a mother ship pattern and the like. However, if a large variance and innovation of the designed vessel pattern relative to the parent vessel design is desired, the parent vessel data becomes unimportant or no parent vessel data at all, and in such a scenario the parent vessel pattern cannot be used, but only the design elements can be used to generate an approximation pattern and estimate the vessel properties, resulting in large errors misleading the designer. Artificial intelligence techniques provide an excellent tool for machine learning of human experiences. Heretofore, artificial intelligence technology has not been used in the estimation of hydrostatic performance of ships, and thus the design of ship concepts still requires a lot of manpower and time, especially for innovative ships, without proper or complete data of female ship references, the design effort is enormous and limits the innovation ability of designers. Disclosure of Invention In order to solve the problems, the technical scheme provided by the invention is as follows: the invention discloses a ship hydrostatic performance prediction large model training, deployment and implementation method based on a large model technology, which is characterized by comprising the following steps: Step one, processing a general sample set of the hydrostatic performance of a ship, wherein the general sample set comprises the steps of collecting the hydrostatic performance finishing data from all ship finishing technical files, and cleaning, finishing and storing the collected sample set; Step two, a general sample set for enhancing the hydrostatic performance of the synthesized ship is obtained, which comprises the steps of interpolating, filling, synthesizing and generating a sample out of distribution to enhance the hydrostatic performance of the expanded ship based on a diffus