Nonlinear Wave Load Prediction Method and System in Ship Structure Design

Abstract

The present invention relates to a method and system for predicting non - linear wave loads in ship structure design. When constructing the non - linear wave load prediction model STLN, a WLIP network model based on a neural network is constructed by using a numerical calculation data set of the ship to be designed; taking the model test regular wave data set as the target domain, knowledge transfer learning of non - linear wave loads is carried out to obtain a short - term prediction model FTLN for regular waves; for the short - term prediction model FTLN of regular waves, taking the model test irregular wave data set as the target domain, knowledge is transferred again for non - linear wave loads to obtain the non - linear wave load prediction model STLN; when predicting non - linear wave loads, for an irregular wave condition of the ship, the non - linear wave load prediction model STLN is used to obtain the wave load corresponding to the irregular wave condition. The present invention can quickly and accurately predict non - linear wave loads and meet the requirements in the ship structure design stage.

Description

Technical Field

[0001] The present invention relates to a load prediction method and system, in particular to a non-linear wave load prediction method and system in ship structure design. Background Art

[0002] The wave load of a ship is the load on the ship in the marine wind, wave and current environment, including various types of external forces acting on the ship and the overall or local dynamic stress generated inside the ship due to the external forces. The wave load that the ship bears is one of the important factors for evaluating the safety of the ship's hull structure and also affects the performance of the ship such as the sailing speed. Therefore, the calculation of wave load plays an important role in the design stage of the ship and the maintenance stage during the ship operation.

[0003] The wave load is the force and moment on the structure of the ship in the sea wave, which is divided into low-frequency and high-frequency parts. Among them, the low-frequency part is the linear wave load on the hull structure at the regular wave frequency. The high-frequency part is the non-linear wave load such as pounding, wave-induced flutter, and deck wetness on the ship in the wave. After more than 70 years of development, the wave load theory has matured. At present, the prediction of the linear part of the wave load is relatively accurate, while there are many uncertain factors in the non-linear components.

[0004] Currently, the non-linear wave load is usually obtained through model tests and used as the most real value. Therefore, the non-linear wave load prediction method based on model tests is particularly important for the high-precision prediction of the combined bending moment under irregular wave conditions.

[0005] Nowadays, the wave load theory is becoming more and more mature, such as the relatively mature three-dimensional time-domain non-linear wave load theory. However, its calculation is ultimately compared with the model test to verify its accuracy, and the coincidence between the theoretical value and the test value of the non-linear component remains to be discussed.

[0006] In summary, quickly and effectively realizing the prediction of the non-linear wave load of the ship is an urgent technical problem to be solved in the current ship structure design. Summary of the Invention

[0007] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a non-linear wave load prediction method and system in ship structure design, which can quickly and accurately predict the non-linear wave load and meet the requirements of the ship structure design stage.

[0008] According to the technical solution provided by the present invention, a non-linear wave load prediction method in ship structure design, the non-linear wave load prediction method includes:

[0009] Construct a non-linear wave load prediction model STLN for predicting non-linear wave loads, where

[0010] When constructing the non - linear wave load prediction model STLN, a WLIP network model based on neural network is constructed by using the numerical calculation data set of the ship to be designed.

[0011] For the constructed WLIP network model, taking the model test regular wave data set as the target domain, knowledge transfer learning of non - linear wave loads is carried out to obtain a regular wave short - term prediction model FTLN after knowledge transfer learning.

[0012] For the regular wave short - term prediction model FTLN, taking the model test irregular wave data set as the target domain, knowledge of non - linear wave loads is transferred again to obtain a non - linear wave load prediction model STLN after knowledge transfer learning.

[0013] When predicting non - linear wave loads, for an irregular wave condition of the ship, the non - linear wave load prediction model STLN is used to obtain the wave loads corresponding to the irregular wave condition, where

[0014] The irregular wave conditions of the ship include the ship's speed, wave height, period, course angle, and sectional position.

[0015] The wave loads corresponding to the irregular wave conditions include vertical bending moment, horizontal bending moment, and torque.

[0016] When constructing the WLIP network model, it includes

[0017] Providing a neural network model and a numerical calculation data set based on the ship to be designed, where the numerical calculation data set includes several numerical calculation data groups. A numerical calculation data group includes a numerical calculation input subset and a numerical calculation output subset corresponding to the numerical calculation input subset. The numerical calculation input subset includes the ship's speed, wave height, period, course angle, and sectional position. The numerical calculation output subset includes vertical bending moment, horizontal bending moment, and torque.

[0018] Configuring the training conditions of the neural network model so that after the neural network model reaches the target training state by using the numerical calculation data set, the required WLIP network model is constructed.

[0019] When obtaining the regular wave short - term prediction model FTLN through knowledge transfer learning, it includes

[0020] Adding two layers of neurons in the FTLN model to the constructed WLIP network model, training the two layers of neurons in the FTLN model by using the model regular wave test data set, and when training, adjusting the parameters of the two layers of neurons in the FTLN model by using an automatic parameter - adjusting method until the target training state is reached to obtain the regular wave short - term prediction model FTLN.

[0021] For the obtained regular wave short-term forecast model FTLN, we have:

[0022]

[0023] Among them, R D (X,θ com ,θ exp1 ) is the loss function of the linear wave load correction model FTLN; θ com is the weight and bias between neurons in the WLIP network model; θ exp1 is the weight and bias between neurons in the two-layer FTLN model, f(X (i) θ com θ exp1 ) is the regular wave short-term prediction function obtained using the eighth layer of neurons, n1 is the number of numerical calculation data sets in the numerical calculation data set; is the i-th regular wave output subset in the model regular wave data set, The predicted value of FTLN of the correction model for linear wave loads.

[0024] For the regular wave short-term prediction model FTLN, the nonlinear wave load prediction model STLN is obtained based on knowledge transfer learning, including:

[0025] Two layers of neurons in the STLN model are added to the regular wave short-term prediction model FTLN. The irregular wave test dataset of the model is used to train the neurons in the two layers of the STLN model. During training, the parameters of the neurons in the two layers of the STLN model are adjusted by the automatic parameter adjustment method until the target training state is reached, thereby obtaining the nonlinear wave load prediction model STLN.

[0026] For the nonlinear wave load prediction model STLN, we have

[0027]

[0028] Where: R D (X,θ com ,θ exp1 ,θ exp2 ) is the loss function of the nonlinear wave load prediction model STLN; θ exp2 is the weight and bias between neurons in the STLN model on both sides, f(X (i) θ com θ exp1 θ exp2 ) is the nonlinear wave load prediction function obtained by the tenth layer of neurons, is the i-th model irregular wave test output subset in the model irregular wave test data set, It is the predicted value output by the non - linear wave load prediction function obtained by the tenth - layer neuron.

[0029] For the automatic adjustment of parameter tuning method, we have:

[0030]

[0031] Where: \(x\) is the design variable, \(af\) is the activation function, \(bs\) is the number of batch input samples, \(lr\) is the initial learning rate, \(nh\) is the number of hidden layers of the neural network, \(minf(x)\) is the objective function, which is the root - mean - square error between the predicted value and the experimental value of the test set; \(s.t.\) is the optimization constraint condition;

[0032] When adjusting the parameters, the objective function \(minf(x)\) is calculated in real - time, and the calculated objective function \(minf(x)\) is judged. The judgment conditions are: reaching the maximum number of iterations \(max\_evals\), or the objective function no longer decreases and reaches the early - stop condition \(early\_stop_fn\); when the judgment conditions are met, the automatic parameter adjustment is realized.

[0033] For the objective function \(minf(x)\), \(n\) is the number of test sets in the wave load sample library, is the predicted value of the \(i\) - th working condition in the test set of the sample library, is the data value of the \(i\) - th working condition in the test set of the sample library.

[0034] A non - linear wave load prediction system in ship structure design includes a non - linear wave load predictor for predicting non - linear wave loads. Among them,

[0035] For an irregular wave working condition of the ship, the non - linear wave load predictor uses the above - mentioned prediction method to predict non - linear wave loads.

[0036] Advantages of the present invention:

[0037] The WLIP network model based on deep learning replaces the traditional wave load strip theory. From the perspective of machine learning, this method learns wave load theory knowledge through numerical calculation data sets. Considering the wave load error between model tests and numerical calculations under the same working conditions, a short - term prediction model FTLN for regular waves is constructed based on transfer learning.

[0038] Considering that there is less data for irregular wave working conditions in model tests, a non - linear wave load prediction model STLN is established based on secondary transfer. Thus, based on the fusion of deep learning and transfer learning, wave load theory, short - term prediction knowledge, and historical experience of model tests are integrated, realizing the application of machine learning in the field of ship structure design.

[0039] After constructing the non - linear wave load prediction model STLN, the research designers can obtain the low - frequency wave loads and synthetic wave loads at any ship speed, wave height, period, heading angle, and sectional position, overcoming the deficiencies of existing extreme value inference methods and realizing the application of machine learning in the field of ship structure design. At the same time, this method can serve both the conceptual design stage of the ship and provide relatively accurate calculations for the detailed design stage. Brief Description of the Drawings

[0040] Figure 1 It is a flowchart of an embodiment for constructing the non - linear wave load prediction model STLN of the present invention.

[0041] Figure 2 It is a model diagram of an embodiment for constructing the non - linear wave load prediction model STLN of the present invention.

[0042] Figure 3 It is a flowchart of an embodiment for the automatic parameter adjustment method of the present invention. Detailed Embodiments

[0043] The present invention will be further described below with reference to specific drawings and embodiments.

[0044] In order to quickly and accurately predict non - linear wave loads and meet the requirements of the ship structure design stage, for the non - linear wave load prediction method in ship structure design, in an embodiment of the present invention, the non - linear wave load prediction method includes:

[0045] Construct a non - linear wave load prediction model STLN for predicting non - linear wave loads, where,

[0046] When constructing the non - linear wave load prediction model STLN, use the numerical calculation data set of the ship to be designed to construct a neural network - based WLIP network model;

[0047] For the constructed WLIP network model, use the model - test regular wave data set as the target domain to perform knowledge transfer learning of non - linear wave loads, so as to obtain a regular wave short - term prediction model FTLN (First Transfer Learning Network, FTLN) after knowledge transfer learning;

[0048] For the regular wave short - term prediction model FTLN, use the model - test irregular wave data set as the target domain to perform knowledge re - transfer of non - linear wave loads, so as to obtain a non - linear wave load prediction model STLN (Secondary Transfer Learning Network, STLN) after knowledge transfer learning;

[0049] When predicting non - linear wave loads, for an irregular wave condition of a ship, the non - linear wave load prediction model STLN is used to obtain the wave loads corresponding to the irregular wave condition, where,

[0050] The irregular wave conditions of the ship include the ship's speed, wave height, period, course angle, and sectional position;

[0051] The wave loads corresponding to the irregular wave condition include vertical bending moment, horizontal bending moment, and torque.

[0052] When constructing the non - linear wave load prediction model STLN, Figure 1 The flowchart of an embodiment is shown, that is, first, a WLIP network model is constructed based on a neural network model, and after knowledge transfer learning based on the WLIP network model, a regular wave short - term prediction model FTLN is constructed. Further, based on the regular wave short - term prediction model FTLN, knowledge transfer learning is carried out to construct the non - linear wave load prediction model STLN.

[0053] After constructing the non - linear wave load prediction model STLN, for any irregular wave condition of the ship, the non - linear wave load prediction model STLN can be used to quickly and accurately predict the corresponding non - linear wave loads. Generally, the irregular wave conditions of the ship include the ship's speed, wave height, period, course angle, and sectional position; the wave loads corresponding to the irregular wave condition include vertical bending moment, horizontal bending moment, and torque.

[0054] Next, according to Figure 1 、 Figure 2 and Figure 3 ,the specific process of constructing the non - linear wave load prediction model STLN will be described in detail.

[0055] In an embodiment of the present invention, when constructing the WLIP network model, it includes

[0056] Providing a neural network model and a numerical calculation data set based on the ship to be designed. The numerical calculation data set includes several numerical calculation data groups. A numerical calculation data group includes a numerical calculation input subset and a numerical calculation output subset corresponding to the numerical calculation input subset. The numerical calculation input subset includes the ship's speed, wave height, period, course angle, and sectional position, and the numerical calculation output subset includes vertical bending moment, horizontal bending moment, and torque;

[0057] Configuring the training conditions of the neural network model to construct the required WLIP network model after the neural network model reaches the target training state using the numerical calculation data set.

[0058] In specific implementation, when constructing the non-linear wave load prediction model STLN, a multi-source wave load data set is required. Among them, the multi-source wave load data set is as follows:

[0059]

[0060] Among them, {D com 、D exp1 、D exp2} is the multi-source wave load data, D com is the numerical calculation data set, D exp1 is the model regular wave test data set; D exp2 is the model irregular wave test data set, n1 is the size of the numerical calculation data set D com ,n2 is the size of the model regular wave test data set D exp1 ,n3 is the size of the model irregular wave test data set D exp2 ,{x1, x2, x3, x4, x5} are the ship speed, period, wave height, course angle and section position respectively, and {y1, y2, y3} are the vertical bending moment, horizontal bending moment and torque respectively.

[0061] In specific implementation, when constructing the WLIP network model, the numerical calculation data set D com is mainly used. To construct the WLIP network model, a neural network model needs to be provided. The provided neural network model can adopt the existing commonly used neural network models. Figure 2 shows the situation of the neural network when forming the WLIP network model. Figure 2 In , the WLIP network model includes six layers of neurons. Figure 2 In , among the six layers of neurons, the number of hidden layers in the second layer of neurons is 256, the number of hidden layers in the third layer of neurons is 128, the number of hidden layers in the fourth layer of neurons is 64, the number of hidden layers in the fifth layer of neurons is 32, and the number of hidden layers in the sixth layer of neurons is 3.

[0062] As can be seen from the above description, for the provided neural network model, the numerical calculation data set needs to be used for training. The specific training method and process can be consistent with the existing training of the neural network model. That is, for the numerical calculation data set, it needs to be divided into a training set, a validation set and a test set. The specific situation of dividing the training set, the validation set and the test set can be selected according to needs, so as to meet the training of the neural network model and construct the WLIP network model.

[0063] Both linear wave loads and non-linear wave loads are used for ship structure design. Wave loads are a prerequisite for ship structure strength calculation. When conducting structural analysis, the loads corresponding to a certain sea condition (speed, period, wave height, course angle) from the bow (section 0) to the stern (section 20) of the ship are usually applied to the ship structure to calculate using existing common technical means. After obtaining the wave load output for a certain sea condition, through multiple sea conditions and the corresponding wave load outputs for each sea condition, the required numerical calculation data set can be obtained. The specific method and process of obtaining the numerical calculation data set can be consistent with the existing technology, specifically based on being able to obtain the required numerical calculation data set.

[0064] The numerical calculation data set includes n1 numerical calculation data groups. For any numerical calculation data group, it includes a numerical calculation input subset and a numerical calculation output subset. Among them, the numerical calculation output subset is the calculation output under the conditions of the numerical calculation input subset within the same numerical calculation data group. When training the neural network model using the numerical calculation data set, the numerical calculation input subset is used as the input of the neural network model, and the numerical calculation output subset is used as the output of the neural network model.

[0065] Specifically in implementation, configure the training conditions of the neural network model, specifically: the loss function gradient descent adopts the Adam algorithm, the number of iterations is 2000 times. Among them, the activation function af adopts the sigmoid function, the batch input sample number bs is 40, and the initial learning rate lr is 0.00157. The target training state is generally the training state where the loss function converges. After iterative training and reaching the target training state, the weights and biases θ of the WLIP network model can be determined. com That is, the WLIP network model can be obtained.

[0066] In an embodiment of the present invention, when obtaining the regular wave short-term prediction model FTLN through knowledge transfer learning, it includes

[0067] Add two layers of neurons in the FTLN model to the constructed WLIP network model, and use the model regular wave test data set to train the two layers of neurons in the FTLN model. And during training, use the automatic parameter adjustment method to adjust the parameters of the two layers of neurons in the FTLN model until the target training state is reached to obtain the regular wave short-term prediction model FTLN.

[0068] As can be seen from the above description, after obtaining the WLIP network model, it is necessary to obtain the short-term prediction model FTLN of regular waves through knowledge transfer learning. The regular wave model test is an experiment to study the relationship between the wave load response and the ratio of ship length to wave length. Among them, the specific value of the regular wave model test is the load response amplitude per unit wave height at different ship speeds, course angles, and wave frequencies. In order to align with the numerical calculation data set and the irregular wave data set of the model test, therefore, it is necessary to make a short-term prediction of the regular wave model test value so that the formed regular wave model data set has the same data form as the numerical calculation data set and the irregular wave data set of the model test, that is, it includes several groups of regular wave inputs and the corresponding regular wave outputs corresponding to the regular wave inputs.

[0069] As can be seen from the above description, for the regular wave model test of the model, the wave load response amplitude per unit wave height can be measured by setting the ship speed, course angle, and wave frequency. In an embodiment of the present invention, the wave load response amplitude per unit wave height under different conditions can obtain the corresponding short-term prediction value of the wave load through the Jonswap sea wave spectrum, that is, the required short-term prediction value of the wave load can be used to construct a regular wave model data set. The method and process of constructing the regular wave model data set can be selected according to actual needs, which are well known to those skilled in the art and will not be elaborated here.

[0070] For the short-term prediction value of the load, there is:

[0071]

[0072] Among them, Y exp is the short-term prediction value of the load, S ζ (ω, H, T, θ) is the wave spectrum density of the Jonswap sea wave spectrum, G(w, V, β + θ) is the wave load response amplitude under the unit regular wave test, ω is the wave circular frequency, V is the ship speed, θ is the angle between the combined wave and the main wave direction, H is the significant wave height, T is the characteristic period of the wave, and β is the course angle.

[0073] Specifically, when implementing, the wave circular frequency ω, ship speed V, angle θ between the combined wave and the main wave direction, significant wave height H, characteristic period T of the wave, and course angle β constitute the conditions of the regular wave model test. In the ship design stage, the specific values of the conditions can be selected according to experience, etc., so as to meet the requirements of ship structure design.

[0074] As can be seen from the above description, the regular wave model data set includes several regular wave data groups. A regular wave data group includes a regular wave input subset and a regular wave output subset associated with the regular wave input subset. For the specific situation of the regular wave input subset and the regular wave output subset, reference can be made to the description of the numerical calculation input subset and the numerical calculation output subset, which will not be elaborated here.

[0075] Figure 2 In this case, neurons in the two-layer FTLN model are directly added to the WLIP network model, that is, the short-term prediction model for regular waves, FTLN, is obtained. FTLN has eight layers of neurons. When adding two layers of neurons, neurons in one FTLN model are adaptively connected to the sixth-layer neurons of the WLIP network model. Among them, the number of hidden layers of neurons in the FTLN model connected to the sixth-layer neurons is 128, and the number of hidden layers of neurons in the FTLN model that form the eighth layer of neurons is 3. When using the model regular wave test data set as the target domain to train neurons in the two-layer FTLN model, the weights and biases θ of the first six layers of neurons are frozen com , that is, the weights and biases θ of the WLIP network model obtained from the previous training are maintained com .

[0076] When using the model regular wave test data set as the target domain to train neurons in the two-layer FTLN model, the parameters of neurons in the two-layer FTLN model are adjusted using an automatic parameter adjustment method until the target training state is reached. That is, after reaching the target training state, the weights and biases θ between neurons in the two added FTLN models are determined exp1 , realizing the first transfer of knowledge of ship linear wave loads.

[0077] Specifically, the training conditions for neurons in the two-layer FTLN model can be: the loss function gradient descent uses the Adam algorithm and iterates 500 times. Among them, the activation function af, the batch input sample number bs, the initial learning rate lr, and the number of hidden layers nh of the neural network are determined by the automatic parameter adjustment method, that is Figure 1 the adjusted model parameters in. The termination condition for training is: obtaining the minimum and convergent loss function value of the FTLN model.

[0078] In an embodiment of the present invention, for the obtained short-term prediction model for regular waves, FTLN, there is:

[0079]

[0080] Among them, R D (X, θ com , θ exp1 ) is the loss function of the linear wave load correction model FTLN; θ com is the weights and biases between neurons in the WLIP network model; θ exp1 is the weights and biases between neurons in the two-layer FTLN model, f(X (i) ; θ com ; θ exp1 ) is the short-term prediction function for regular waves obtained using the eighth layer of neurons, and n1 is the number of numerical calculation data groups in the numerical calculation data set.

[0081] It is the i-th regular wave output subset in the model regular wave data set. It is the predicted value of the linear wave load correction model FTLN, that is, the predicted value output by the regular wave short-term prediction function f(X (i) ; θ com ; θ exp1 ). The working condition of the predicted value is consistent with the working condition corresponding to the i-th load short-term prediction value in the model regular wave data set. The regular wave output subset and the predicted value both include vertical bending moment, horizontal bending moment and torque.

[0082] In an embodiment of the present invention, when obtaining the non-linear wave load prediction model STLN based on knowledge transfer learning for the regular wave short-term prediction model FTLN, it includes:

[0083] Add two layers of neurons in the STLN model to the regular wave short-term prediction model FTLN, and use the model irregular wave test data set to train the neurons in the added two layers of the STLN model. When training, use the automatic adjustment parameter method to adjust the parameters of the neurons in the two layers of the STLN model until the target training state is reached to obtain the non-linear wave load prediction model STLN.

[0084] Specifically, the model irregular wave test mainly obtains the load response value of the ship in a relatively real sea wave. Therefore, the model irregular wave test data set can be obtained by using the model irregular wave test. The model irregular wave test data set generally also includes several model irregular wave test data groups. A model irregular wave test data group includes a model irregular wave test input subset and a model irregular wave test output subset associated with the model irregular wave test input subset. The associated corresponding situation can refer to the above description and will not be elaborated here.

[0085] As can be seen from the above description, the obtained regular wave short-term prediction model FTLN includes eight layers of neurons. In order to form the non-linear wave load prediction model STLN, two more layers of neurons are added on the basis of the regular wave short-term prediction model FTLN. The two added layers of neurons are the neurons in the two layers of the STLN model. The number of hidden layers of the neurons in the STLN model as the ninth layer is 64. The neurons in the STLN model as the tenth layer are used as the output layer of the non-linear wave load prediction model STLN.

[0086] As can be seen from the above description, after adding two layers of neurons in the STLN model, it is necessary to train the added two layers of neurons in the STLN model. When training, keep the weights and biases of the neurons in the regular wave short-term prediction model FTLN, that is, freeze the weights and biases θ com, and freeze the weights and biases θ exp1 .

[0087] When training the neurons in the two-layer STLN model, use the model irregular wave test data set for training. Among them, the model irregular wave test input subset is used as the input for training, and the model irregular wave test output subset is used as the output for training.

[0088] When training the neurons in the two-layer STLN model, the training conditions can be configured as follows: the number of training iterations can be 500, the loss function gradient descent uses the Adam algorithm, and the neural network activation function, learning rate, batch size, and number of hidden layers are adjusted by the automatic tuning method, that is Figure 1 the adjusted model parameters in

[0089] In an embodiment of the present invention, for the nonlinear wave load prediction model STLN, there is

[0090]

[0091] Where: R D (X, θ com , θ exp1 , θ exp2 ) is the loss function of the nonlinear wave load prediction model STLN; θ exp2 is the weight and bias between the neurons in the two-sided STLN model, and f(X (i) ; θ com ; θ exp1 ; θ exp2 ) is the nonlinear wave load prediction function obtained by the tenth-layer neurons, is the i-th model irregular wave test output subset in the model irregular wave test data set, is the predicted value output by the nonlinear wave load prediction function obtained by the tenth-layer neurons.

[0092] As can be seen from the above description, the i-th model irregular wave test output subset in the model irregular wave test data set is the vertical bending moment, horizontal bending moment, and torque; therefore, the predicted value output by the nonlinear wave load prediction function obtained by the tenth-layer neurons also correspondingly includes the vertical bending moment, horizontal bending moment, and torque. In addition, the model irregular wave test output subset and the predicted value have the same model irregular wave conditions, that is, under the input conditions of the same model irregular wave test input subset.

[0093] In an embodiment of the present invention, for the automatic tuning method, there is:

[0094]

[0095] Where: x is the design variable, af is the activation function, bs is the number of batch input samples, lr is the initial learning rate, nh is the number of hidden layers of the neural network, minf(x) is the objective function, which is the root mean square error between the predicted value and the experimental value of the test set; s.t. are the optimization constraints.

[0096] Figure 3 An embodiment of automatically adjusting parameters during training using the automatic tuning method is shown. Among them, the wave load sample library is the data required during training. For example, when training neurons in a two-layer FTLN model, the wave load sample library uses the determined short-term load prediction value as the target domain; when training neurons in a two-layer STLN model, the wave load sample library is the irregular wave test data set of the model.

[0097] For the wave load sample library, it is necessary to define the number of batch input samples, the initial learning rate, the activation function, the number of hidden layers of the neural network, and the optimization number constraints. After that, call the Hyperopt library, set the objective function minf(x), the maximum number of iterations max_evals, and the early stop condition early_stop_fn. Specifically, when defining the number of batch input samples, the initial learning rate, the activation function, and the number of hidden layers of the neural network, they are taken within the above value ranges. The Hyperopt library is a python library for optimization calculations. The method and process of using the Hyperopt library for optimization calculations are the same as those in the prior art and will not be elaborated here.

[0098] When automatically adjusting parameters, the objective function minf(x) is calculated in real time, and the calculated objective function minf(x) is judged. The judgment conditions are: reaching the maximum number of iterations max_evals, or the objective function no longer decreases and reaches the early stop condition early_stop_fn; when the judgment conditions are met, the training target is reached and the parameter adjustment is realized.

[0099] For the objective function minf(x), n is the number of test sets in the wave load sample library. is the predicted value of the i-th working condition in the test set of the sample library. Here, θ is the above-mentioned weight and bias. is the data value of the i-th working condition in the test set of the sample library. As can be seen from the above description, the data value is a calculated value or an experimental value. RMSE is the objective function minf(x), max_evals is the maximum number of optimization iterations, which can be set to 3000 times specifically in implementation, and early_stop_fn is the early stop condition, which can be set to 500 times specifically in implementation, that is, when the objective function minf(x) no longer decreases after 500 iterations, the optimization is terminated early.

[0100] For the test set in the wave load sample library, generally, it is necessary to divide the wave load sample library. After division, a training set, a test set, a validation set, etc. are formed. Therefore, the specific number n of the test set can be determined according to actual needs to meet the requirements of training and automatic adjustment, which will not be elaborated here.

[0101] In summary, for a non - linear wave load prediction system, in an embodiment of the present invention, it includes a non - linear wave load predictor for predicting non - linear wave loads, where,

[0102] For an irregular wave condition of a ship, the non - linear wave load predictor uses the above - mentioned prediction method to predict non - linear wave loads.

[0103] The non - linear wave load predictor can generally be a computer device, that is, a non - linear wave load prediction model STLN is configured in the computer device, and the configured non - linear wave load prediction model STLN is used to predict non - linear wave loads.

[0104] When predicting wave loads, for an irregular wave condition, the WLIP network model, the neurons in the two - layer FTLN model, and the neurons in the two - layer STLN model are processed in sequence. Finally, the wave load output by the entire non - linear wave load prediction model STLN is the non - linear wave load.

[0105] In summary, the WLIP network model based on deep learning in the present invention replaces the traditional wave load strip theory. From the perspective of machine learning, this method learns wave load theory knowledge through a numerical calculation data set. Considering the wave load error between the model test and numerical calculation under the same working condition, a regular wave short - term prediction model FTLN is constructed based on transfer learning.

[0106] Considering that there is less data for the irregular wave condition in the model test, a non - linear wave load prediction model STLN is established based on secondary transfer. Thus, based on deep learning and transfer learning, wave load theory, short - term prediction knowledge, and historical experience of model tests are integrated, realizing the application of machine learning in the field of ship structure design.

[0107] After constructing the non - linear wave load prediction model STLN, researchers and designers can obtain low - frequency wave loads and synthetic wave loads at any ship speed, wave height, period, course angle, and sectional position, overcoming the deficiencies of existing extreme value inference methods, realizing the application of machine learning in the field of ship structure design. At the same time, this method can serve both the conceptual design stage of ships and provide relatively accurate calculations for the detailed design stage.

Claims

1. A method for predicting non - linear wave loads in ship structure design, characterized in that, The non - linear wave load prediction method includes: Constructing a non - linear wave load prediction model STLN for predicting non - linear wave loads, where When constructing the non - linear wave load prediction model STLN, a neural - network - based WLIP network model is constructed using a numerical calculation data set of the ship to be designed; For the constructed WLIP network model, using the model test regular wave data set as the target domain, knowledge transfer learning of non - linear wave loads is carried out to obtain a regular wave short - term prediction model FTLN after knowledge transfer learning; For the regular wave short - term prediction model FTLN, using the model test irregular wave data set as the target domain, knowledge transfer of non - linear wave loads is carried out again to obtain a non - linear wave load prediction model STLN after knowledge transfer learning; During non - linear wave load prediction, for an irregular wave condition of the ship, the non - linear wave load prediction model STLN is used to obtain the wave load corresponding to the irregular wave condition, where The irregular wave condition of the ship includes the ship's speed, wave height, period, course angle, and section position; The wave load corresponding to the irregular wave condition includes vertical bending moment, horizontal bending moment, and torque.

2. The method for predicting non-linear wave loads in ship structure design according to claim 1, characterized in that, When constructing the WLIP network model, it includes Providing a neural network model and a numerical calculation data set of the ship to be designed, where the numerical calculation data set includes several numerical calculation data groups. A numerical calculation data group includes a numerical calculation input subset and a numerical calculation output subset corresponding to the numerical calculation input subset. The numerical calculation input subset includes the ship's speed, wave height, period, course angle, and section position, and the numerical calculation output subset includes vertical bending moment, horizontal bending moment, and torque; Configuring the training conditions of the neural network model so that after the neural network model reaches the target training state using the numerical calculation data set, the required WLIP network model is constructed.

3. The method for predicting non-linear wave loads in ship structure design according to claim 2, characterized in that, When obtaining the regular wave short - term prediction model FTLN through knowledge transfer learning, it includes Adding two layers of neurons in the FTLN model to the constructed WLIP network model, training the two layers of neurons in the FTLN model using the model regular wave test data set, and during training, using an automatic parameter adjustment method to adjust the parameters of the two layers of neurons in the FTLN model until the target training state is reached to obtain the regular wave short - term prediction model FTLN.

4. The method for predicting non-linear wave loads in ship structure design according to claim 3, characterized in that, For the obtained regular wave short - term prediction model FTLN, there is: Among them, R D (X, θ com , θ exp1 ) is the loss function of the linear wave load correction model FTLN; θ com is the weight and bias between neurons in the WLIP network model; θ exp1 is the weight and bias between neurons in the two-layer FTLN model, f(X (i) ; θ com ; θ exp1 ) is the short-term prediction function of regular waves obtained by using the neurons in the eighth layer, and n2 is the number of numerical calculation data groups in the numerical calculation data set; is the i-th regular wave output subset in the model regular wave data set, is the predicted value of the linear wave load correction model FTLN.

5. The method for predicting non-linear wave loads in ship structure design according to claim 3, characterized in that, When obtaining the non - linear wave load prediction model STLN based on knowledge transfer learning for the regular wave short - term prediction model FTLN, it includes: Adding two layers of neurons in the STLN model to the regular wave short - term prediction model FTLN, training the two layers of neurons added in the STLN model using the model irregular wave test data set, and during training, using an automatic parameter adjustment method to adjust the parameters of the two layers of neurons in the STLN model until the target training state is reached to obtain the non - linear wave load prediction model STLN.

6. The method for predicting non-linear wave loads in ship structure design according to any one of claims 1 to 5, characterized in that, For the non - linear wave load prediction model STLN, there is Where: R D (X, θ com , θ exp1 , θ exp2 ) is the loss function of the non - linear wave load prediction model STLN; θ exp2 is the weight and bias between neurons in the STLN models on both sides, f(X (i) ; θ com ; θ exp1 ; θ exp2 ) is the non - linear wave load prediction function obtained by the tenth - layer neurons, is the i - th model irregular wave test output subset in the model irregular wave test data set, is the predicted value output by the non - linear wave load prediction function obtained by the tenth - layer neurons.

7. The method for predicting non - linear wave loads in ship structure design according to claim 4 or 5, characterized in that, For the automatic parameter adjustment method, there is: Where: x is the design variable, af is the activation function, bs is the number of batch input samples, lr is the initial learning rate, nh is the number of hidden layers of the neural network, minf(x) is the objective function, and the root mean square error between the predicted value and the experimental value of the test set; s.t. is the optimization constraint condition; When adjusting the parameters, the objective function minf(x) is calculated in real time, and the calculated objective function minf(x) is judged. The judgment conditions are: reaching the maximum number of iterations max_evals, or the objective function no longer decreases and reaching the early stop condition early_stop_fn; when the judgment conditions are met, the automatic parameter adjustment is realized; For the objective function min f(x), n is the number of test sets in the wave load sample library, is the predicted value of the i-th working condition in the test set of the sample library, is the data value of the i-th working condition in the test set of the sample library.

8. A non-linear wave load prediction system in ship structure design, characterized by: including for For the non-linear wave load predictor for non-linear wave load prediction, where For an irregular wave condition of a ship, the non-linear wave load predictor uses the prediction method of any one of claims 1 to 7 above to perform non-linear wave load prediction.

Images