Total longitudinal stress field reconstruction method and device suitable for ships under actual sea conditions

Abstract

The present invention relates to a method and device for reconstructing the total longitudinal stress field of a ship under actual sea conditions. The method comprises: constructing a longitudinal strength member total longitudinal stress prediction model RMMSF for predicting the total longitudinal stress of the longitudinal strength member under a preset sea condition; for any non-measured node of the longitudinal strength member, using the longitudinal strength member total longitudinal stress prediction model RMMSF to predict the total longitudinal stress predicted value of the non-measured node under the preset sea condition; based on the total longitudinal stress measured value of the measured node of the longitudinal strength member under the preset sea condition and the total longitudinal stress predicted value of the non-measured node of the longitudinal strength member under the preset sea condition, the total longitudinal stress field of the longitudinal strength member under the preset sea condition is reconstructed. The present invention can quickly and accurately predict the total longitudinal stress of the longitudinal strength member under actual sea conditions, realize the reconstruction of the total longitudinal stress field of the ship under actual sea conditions, and meet the requirements of the hull total longitudinal strength verification.

Description

Technical Field

[0001] The present invention relates to a total longitudinal stress field reconstruction method and device, in particular to a total longitudinal stress field reconstruction method and device suitable for ships under actual sea conditions. Background Art

[0002] Under the action of wave loads, the deck, outer plate, inner bottom plate and other longitudinal strength components on the hull, which effectively transmit the total longitudinal bending stress, can be effectively combined with the calculated and measured stress results through mathematical and statistical methods to obtain the measured stress field of the longitudinal strength components, which is of great significance for the verification of the total longitudinal strength of the hull.

[0003] The total longitudinal stress of the hull's longitudinal strength members is an important part of the hull strength verification. When conducting the hull strength verification, the ship is generally placed on the waves, the total longitudinal bending moment is calculated, and the calculated total longitudinal bending moment is loaded onto the equivalent beam or the finite element model of the ship. The total longitudinal stress of the hull's longitudinal strength members can be obtained through finite element theory calculation for comparison with the allowable stress. This is the main method for calculating the total longitudinal stress of the hull's longitudinal strength members to date.

[0004] When studying the response of ship structures, full-ship sea trials are the most important research method. They play a decisive role in clarifying the characteristics of wave loads and their statistical distribution laws, verifying and developing theoretical calculation methods, designing reasonable wave load extremes for ship structures, and formulating structural fatigue life analysis and strength standards.

[0005] However, actual ship sea trials are constrained by various objective and subjective factors, and the number of measurement points is limited. Often, only dozens to hundreds of measurement points can be arranged at key locations of the hull structure. The complete structural stress field results of the actual ship in wind and waves are difficult to obtain through testing alone. Therefore, how to quickly and effectively predict the measured stress field of the ship structure is a technical problem that urgently needs to be solved in the current hull strength verification. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and device for reconstructing the total longitudinal stress field of longitudinal strength components of a ship, which can quickly and accurately predict the total longitudinal stress of longitudinal strength components under actual sea conditions, realize the reconstruction of the total longitudinal stress field of the ship under actual sea conditions, and meet the needs of checking the total longitudinal strength of the hull.

[0007] According to the technical solution provided by the present invention, a total longitudinal stress field reconstruction method suitable for ships under actual sea conditions, for a longitudinal strength member, the total longitudinal stress field reconstruction method of the longitudinal strength member includes:

[0008] A longitudinal strength member total longitudinal stress prediction model RMMSF is constructed for predicting the total longitudinal stress of the longitudinal strength member under a preset sea condition, wherein:

[0009] When constructing the longitudinal strength member total longitudinal stress prediction model RMMSF, a neural network-based total longitudinal stress prediction basic model CSFP is first constructed. When constructing the total longitudinal stress prediction basic model CSFP, a stress field standard data set under a preset sea condition in a calculation domain generated based on the longitudinal strength member finite element model is used for training, so as to obtain the total longitudinal stress prediction basic model CSFP after the training reaches the basic model target state;

[0010] The standard data set of stress field under the preset sea condition of the computational domain includes a plurality of stress field data samples under the preset sea condition of the computational domain, and any stress field data sample under the preset sea condition of the computational domain includes the node coordinates of a node of the longitudinal strength member and the total longitudinal stress value of the node under the preset sea condition;

[0011] Producing a standard dataset of actual ship tests of the longitudinal strength structure under preset sea conditions, configuring the produced standard dataset of actual ship tests as a target domain, and performing knowledge transfer learning on the basic model for total longitudinal stress prediction CSFP based on the configured target domain to generate the longitudinal strength member total longitudinal stress prediction model RMMSF after the knowledge transfer learning;

[0012] For any non-measured node of the longitudinal strength member, a total longitudinal stress prediction model RMMSF of the longitudinal strength member is used to predict the total longitudinal stress predicted value of the non-measured node under the preset sea condition;

[0013] Based on the measured values ​​of the total longitudinal stress of the measured nodes of the longitudinal strength components under the preset sea conditions and the predicted values ​​of the total longitudinal stress of the non-measured nodes of the longitudinal strength components under the preset sea conditions, the total longitudinal stress field of the longitudinal strength components under the preset sea conditions is reconstructed.

[0014] For the constructed total longitudinal stress prediction basic model CSFP, we have:

[0015]

[0016] Among them, R D′ (X (i) ,θ CSFP ) is the loss function of the basic model CSFP for total longitudinal stress prediction; N1 is the number of stress field data samples under the preset sea conditions in the standard data set of stress field under the preset sea conditions in the computational domain; θ CSFP is the weight w trained between neurons in the basic model CSFP for total longitudinal stress prediction ij and bias b j The set of f(X (i) θ CSFP ) is the output function of the output layer of the CSFP basic model for total longitudinal stress prediction, is the total longitudinal stress value of the stress field data sample under the preset sea conditions in the i-th computational domain in the standard dataset of the stress field under the preset sea conditions in the computational domain, X (i) The node coordinates of the i-th data sample in the standard data set of stress field under preset sea conditions for the computational domain, It is the predicted value of the total longitudinal stress of the basic stress prediction model CSFP under the preset sea conditions.

[0017] For the standard stress field data set under the preset sea conditions in the computational domain, we have:

[0018]

[0019] Where D is the basic data set of stress field under the preset sea conditions in the computational domain, is the basic sample of stress field data under preset sea conditions in the i-th computational domain in the basic dataset of stress field under preset sea conditions, X low is the minimum value of the finite element mesh node of the longitudinal strength member, X up is the maximum value of the finite element mesh node of the longitudinal strength member, Mesh is the node coordinate generation function of the longitudinal strength member; Patran is the node coordinate X of the longitudinal strength member (i) Calculation function of total longitudinal stress under preset sea conditions;

[0020] D L ′ is the mean normalized total longitudinal stress of the longitudinal strength member; is the mean value of the total longitudinal stress at all node coordinates of the longitudinal strength member, D Ls is the variance of the total longitudinal stress at all node coordinates of the longitudinal strength member;

[0021] D Z ′ is the mean normalized node coordinate of the longitudinal strength member; is the mean value of the node coordinates of all the node coordinates of the longitudinal strength members, D Zs is the node coordinate variance of all node coordinates of longitudinal strength members;

[0022] Using D Z ′ is the mean normalized node coordinate D of the longitudinal strength member Z ′ and the node coordinates normalized with the mean D Z ′ corresponds to the mean normalized total longitudinal stress D L ’Form a stress field data sample under the preset sea conditions in the calculation domain.

[0023] When the neural network model is trained using the standard stress field data set under the preset sea conditions in the computational domain to obtain the basic model for total longitudinal stress prediction CSFP, we have:

[0024] The mean normalized node coordinates of the longitudinal strength components in the stress field data sample under the preset sea conditions in the computational domain are used as the input of the neural network, and the mean normalized total longitudinal stress values ​​of the corresponding nodes in the stress field data sample under the preset sea conditions in the computational domain are used as the output of the neural network.

[0025] The convergence of the damage function of the neural network is taken as the target state of the basic model; during training, the Adam algorithm is used for gradient descent of the loss function, and the Sigmoid function is used for the activation function.

[0026] The knowledge transfer learning of the basic total longitudinal stress prediction model CSFP includes:

[0027] Two layers of neurons are added to the output layer of the basic model for total longitudinal stress prediction CSFP in sequence. The standard data set of actual ship test is used as the target domain, and the weight and bias set θ between neurons in the basic model for total longitudinal stress prediction CSFP is frozen. CSFP Under the above conditions, the two added layers of neurons are trained until the target training state of the prediction model is reached to generate the total longitudinal stress prediction model RMMSF of the longitudinal strength member.

[0028] For the total longitudinal stress prediction model RMMSF of longitudinal strength members, we have:

[0029]

[0030] Among them, R De′xp (X exp ,θ CSFP θ exp ) is the loss function of the total longitudinal stress prediction model RMMSF of longitudinal strength members; θ exp To increase the weights and biases between two layers of neurons; is the output function of the output layer of the RMMSF model for predicting the total longitudinal stress of longitudinal strength components, is the total longitudinal stress value of the ith actual ship test in the actual ship test standard data set, is the predicted value of the total longitudinal stress of the longitudinal strength member total longitudinal stress prediction model RMMSF.

[0031] For the production of standard data sets for actual ship tests under preset sea conditions, there are:

[0032]

[0033] Among them, D exp It is the basic data set for the actual ship test; is the boundary node set of the longitudinal strength member, is the set of calculated values ​​of total longitudinal stress at the boundary nodes of longitudinal strength members, and N2 is the number of boundary nodes of longitudinal strength members; is the measured node set of the longitudinal strength member, is the set of total longitudinal stress measured values ​​corresponding to the measured nodes of the longitudinal strength member, N3 is the number of measured nodes of the longitudinal strength member; {y exp} is the measured value of the total longitudinal stress at the measured node;

[0034] Basic data set for real ship test

[0035] D e ' xp D is the basic data set for the actual ship test exp The mean normalization process is performed to form a standard data set for actual ship test.

[0036] The longitudinal strength members include decks, outer plates or inner bottom plates;

[0037] For any non-measured node of the longitudinal strength member, the coordinates of the non-measured node are determined, and the coordinates are processed in sequence through the total longitudinal stress prediction basic model CSFP and the two layers of neurons added to the output layer of the total longitudinal stress prediction basic model CSFP, so as to output the total longitudinal stress prediction value of the non-measured node under the preset sea conditions through the output layer of the two added layers of neurons.

[0038] The predicted value of the total longitudinal stress of the non-measured node under the preset sea condition includes the tensile and compressive stresses of the non-measured node in the x-direction.

[0039] A total longitudinal stress field reconstruction device suitable for ships under actual sea conditions includes a total longitudinal stress field reconstruction processor, wherein:

[0040] For any longitudinal strength member, the total longitudinal stress field of the longitudinal strength member under the preset sea condition is reconstructed using the above-described method.

[0041] The advantages of the present invention are as follows: the basic model for total longitudinal stress prediction CSFP based on neural network replaces the traditional finite element theory. From the perspective of machine learning, the finite element theoretical knowledge is learned through the standard data set of stress field under the preset sea conditions in the calculation domain, and after verification, the basic model for total longitudinal stress prediction CSFP that satisfies the smaller loss function is obtained. Therefore, the total longitudinal stress value of any node in the longitudinal strength component space can be obtained.

[0042] For the same longitudinal strength member, considering the stress error between the actual ship test and the finite element calculation, a total longitudinal stress prediction model RMMSFF of longitudinal strength members is established based on transfer learning. That is, two layers are added on the basis of the CSFP network to fine-tune the model, and finally the total longitudinal stress prediction model RMMSFF of longitudinal strength members is obtained. The total longitudinal stress prediction model RMMSFF of longitudinal strength members can be used to obtain the predicted value of the total longitudinal stress of non-measured nodes, which provides support for the reconstruction of the total longitudinal stress field, so that the total longitudinal stress of longitudinal strength members under actual sea conditions can be predicted quickly and accurately, and the reconstruction of the total longitudinal stress field of the ship under actual sea conditions can be realized, meeting the needs of the hull total longitudinal strength verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention provides a flowchart of an embodiment for determining the predicted value of total longitudinal stress in total longitudinal stress field reconstruction.

[0044] Figure 2 This is a schematic diagram of an embodiment of constructing a total longitudinal stress prediction model RMMSF for longitudinal strength members according to the present invention.

[0045] Figure 3 This is a schematic diagram of the features of the standard data set for actual ship tests using the deck as an example of longitudinal strength member selection in the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described below with reference to specific drawings and embodiments.

[0047] In order to quickly and accurately predict the total longitudinal stress of longitudinal strength components under actual sea conditions, realize the reconstruction of the total longitudinal stress field of the ship under actual sea conditions, and meet the requirements of the hull total longitudinal strength verification, a total longitudinal stress field reconstruction method suitable for ships under actual sea conditions is provided. In one embodiment of the present invention, for a longitudinal strength component, the total longitudinal stress field reconstruction method of the longitudinal strength component includes:

[0048] A longitudinal strength member total longitudinal stress prediction model RMMSF is constructed for predicting the total longitudinal stress of the longitudinal strength member under a preset sea condition, wherein:

[0049] When constructing the longitudinal strength member total longitudinal stress prediction model RMMSF, a neural network-based total longitudinal stress prediction basic model CSFP is first constructed. When constructing the total longitudinal stress prediction basic model CSFP, a stress field standard data set under a preset sea condition in a calculation domain generated based on the longitudinal strength member finite element model is used for training, so as to obtain the total longitudinal stress prediction basic model CSFP after the training reaches the basic model target state;

[0050] The standard data set of stress field under the preset sea condition of the computational domain includes a plurality of stress field data samples under the preset sea condition of the computational domain, and any stress field data sample under the preset sea condition of the computational domain includes the node coordinates of a node of the longitudinal strength member and the total longitudinal stress value of the node under the preset sea condition;

[0051] Producing a standard dataset of actual ship tests of the longitudinal strength structure under preset sea conditions, configuring the produced standard dataset of actual ship tests as a target domain, and performing knowledge transfer learning on the basic model for total longitudinal stress prediction CSFP based on the configured target domain to generate the longitudinal strength member total longitudinal stress prediction model RMMSF after the knowledge transfer learning;

[0052] For any non-measured node of the longitudinal strength member, a total longitudinal stress prediction model RMMSF of the longitudinal strength member is used to predict the total longitudinal stress predicted value of the non-measured node under the preset sea condition;

[0053] Based on the measured values ​​of the total longitudinal stress of the measured nodes of the longitudinal strength components under the preset sea conditions and the predicted values ​​of the total longitudinal stress of the non-measured nodes of the longitudinal strength components under the preset sea conditions, the total longitudinal stress field of the longitudinal strength components under the preset sea conditions is reconstructed.

[0054] It can be seen from the above description that the longitudinal strength member can be one of the deck, outer plate or inner bottom plate mentioned above. Of course, it can also be other longitudinal strength members in the ship. The type of longitudinal strength member can be selected according to needs to meet the reconstruction of the total longitudinal stress field of the longitudinal strength structure.

[0055] After selecting the type of longitudinal strength member, it is necessary to construct a longitudinal strength member total longitudinal stress prediction model RMMSF (Reconstruction Method of Measured Stress Field) for the selected longitudinal strength member. Figure 1 The flowchart of an embodiment of constructing the total longitudinal stress prediction model RMMSF of longitudinal strength members is shown in FIG. Figure 1 The process shown in the figure details the process of constructing the total longitudinal stress prediction model RMMSF for longitudinal strength members.

[0056] Figure 1 In this process, a neural network-based basic model for total longitudinal stress prediction (CSFP) is first constructed. The neural network is then trained using a standard dataset of stress fields under preset sea conditions in the computational domain. Once the training reaches the target state for the basic model, the basic model for total longitudinal stress prediction (CSFP) is obtained. The neural network underlying or employed in the basic model for total longitudinal stress prediction (CSFP) can be a commonly used neural network in the art, such as a convolutional neural network (CNN). The specific type of neural network can be selected as needed to ensure that the basic model for total longitudinal stress prediction (CSFP) can be obtained through training.

[0057] In one embodiment of the present invention, a standard stress field data set under sea conditions is preset for the calculation domain, and the following is obtained:

[0058]

[0059] Where D is the basic data set of stress field under the preset sea conditions in the computational domain, is the basic sample of stress field data under preset sea conditions in the i-th computational domain in the basic dataset of stress field under preset sea conditions, X low is the minimum value of the finite element mesh node of the longitudinal strength member, X up is the maximum value of the finite element mesh node of the longitudinal strength member, Mesh is the node coordinate generation function of the longitudinal strength member; Patran is the node coordinate X of the longitudinal strength member (i) Calculation function of total longitudinal stress under preset sea conditions;

[0060] D L ′ is the mean normalized total longitudinal stress of the longitudinal strength member; is the mean value of the total longitudinal stress at all node coordinates of the longitudinal strength member, D Ls is the variance of the total longitudinal stress at all node coordinates of the longitudinal strength member;

[0061] D Z ′ is the mean normalized node coordinate of the longitudinal strength member; is the mean value of the node coordinates of all the node coordinates of the longitudinal strength members, D Zs is the node coordinate variance of all node coordinates of longitudinal strength members;

[0062] Using D Z ′ is the mean normalized node coordinate D of the longitudinal strength member Z ′ and the node coordinates normalized with the mean D Z ′ corresponds to the mean normalized total longitudinal stress D L ’Form a stress field data sample under the preset sea conditions in the calculation domain.

[0063] As can be seen from the above description, the total longitudinal stress of a longitudinal strength member is related to the sea conditions of the vessel in which the longitudinal strength member is located. Therefore, when reconstructing the total longitudinal stress field of a longitudinal strength member, it is necessary to determine the sea conditions in which the longitudinal strength member is located, that is, to determine a preset sea condition, which includes wave load conditions. After determining the preset sea condition, the computational domain specifically refers to performing finite element calculations based on the finite element model of the longitudinal strength member under the preset sea condition, that is, using finite element calculations to generate a basic stress field dataset for the computational domain under the preset sea condition.

[0064] It can be seen from the above description that after the type of longitudinal strength member is selected, the finite element tools commonly used in this technical field can be used to obtain the finite element model of the longitudinal strength member. Thereafter, the finite element model is analyzed and calculated based on the finite element theory to obtain a standard data set of stress fields under the preset sea conditions in the calculation domain.

[0065] For the finite element model of the longitudinal strength member, the number of grid nodes in the longitudinal strength member can be determined, thereby determining the stress field data sample N1 under the preset sea condition in the computational domain. In specific implementation, for any grid node, the node coordinates of the grid node and the total longitudinal stress value of the grid node under the preset sea condition are determined. At this point, a basic sample of stress field data under the preset sea condition in the computational domain is obtained. Using the node coordinates of all grid nodes and the total longitudinal stress values ​​of the grid nodes under the preset sea condition, a basic stress field data set D under the preset sea condition in the computational domain is formed.

[0066] For the finite element model of longitudinal strength members, the The node coordinates of each grid node are determined by The total longitudinal stress value of the mesh node is determined by this method. (i) , generally including the three-dimensional coordinates of the grid nodes, the three-dimensional coordinates are the coordinates in the spatial coordinate system established based on the ship, such as the X-axis can be constructed by the length of the ship, the Y-axis can be constructed by the width of the ship, and the Z-axis can be constructed by the height direction of the ship.

[0067] When using the node coordinate generation function Mesh to generate the node coordinates of the mesh node, it is generally necessary to count the minimum value of the mesh node and the maximum value of the grid node Minimum value of grid nodes and the maximum value of the grid node The finite element software can be used to obtain the coordinates using statistical methods commonly used in the art. The node coordinate generation function Mesh and the total longitudinal stress calculation function Patran can both adopt existing commonly used forms. For example, the total longitudinal stress calculation function Patran can adopt the calculation function form included in existing commonly used total longitudinal stress calculation tools.

[0068] When calculating the total longitudinal stress of the grid nodes based on the finite element theory, it is generally necessary to configure the finite element calculation conditions. The configured finite element calculation conditions are also the preset sea conditions mentioned above, which generally include the wave loads on the structure and the bow and stern boundary conditions. That is, after configuring the finite element theory calculation conditions, the total longitudinal stress calculation function Patran can be used according to the finite element theory to calculate the tensile and compressive stresses along the x direction at all nodes of the ship structure. Also called total longitudinal stress.

[0069] After obtaining the basic dataset D of stress field under the preset sea conditions in the computational domain using the above method, mean normalization is required to facilitate transmission between neurons during neural network training. When performing mean normalization, both the node coordinates of the grid nodes and the total longitudinal stress values ​​of the grid nodes need to be mean normalized. In the above mean normalization, the corresponding homogenization and variance can be determined based on the basic dataset D of stress field under the preset sea conditions in the computational domain, so that the mean normalized total longitudinal stress D of any node can be obtained respectively. L ′ and the mean normalized node coordinate D of any node Z ′.

[0070] In specific implementation, using D Z ′ is the mean normalized node coordinate D of the longitudinal strength member Z ′ and the node coordinates normalized with the mean D Z ′ corresponds to the mean normalized total longitudinal stress D L 'Form a stress field data sample under the preset sea condition in the computational domain. Using N1 stress field data samples under the preset sea condition in the computational domain, a standard stress field data set under the preset sea condition in the computational domain can be obtained. Mean normalization of the node coordinates specifically refers to mean normalization of the corresponding coordinate positions of the node coordinates on the x-axis, y-axis, and z-axis.

[0071] In one embodiment of the present invention, when a neural network model is trained using a standard stress field dataset under a preset sea condition in the computational domain to obtain a basic model for total longitudinal stress prediction CSFP, the following is true:

[0072] The mean normalized node coordinates of the longitudinal strength components in the stress field data sample under the preset sea conditions in the computational domain are used as the input of the neural network, and the mean normalized total longitudinal stress values ​​of the corresponding nodes in the stress field data sample under the preset sea conditions in the computational domain are used as the output of the neural network.

[0073] The convergence of the damage function of the neural network is taken as the target state of the basic model; during training, the Adam algorithm is used for gradient descent of the loss function, and the Sigmoid function is used for the activation function.

[0074] From the above description, it can be seen that the basic model for total longitudinal stress prediction CSFP can be based on the convolutional neural network CNN. That is, after the convolutional neural network CNN is trained using the standard stress field data set under the preset sea conditions in the computational domain to reach the target training state of the basic model, the basic model for total longitudinal stress prediction CSFP can be obtained.

[0075] Figure 2An embodiment of training to obtain the basic model CSFP for total longitudinal stress prediction is shown in FIG. Specifically, when training the basic model CSFP for total longitudinal stress prediction, the above-mentioned mean normalized node coordinates are used as the input of the neural network, and the mean normalized total longitudinal stress value is used as the output.

[0076] During specific training, the convergence of the loss function is used as the basic model target training state. When the training state is reached, θ can be obtained. CSFP For the basic model CSFP for total longitudinal stress prediction, Figure 2 In the network structure, the network structure consists of N convolutional blocks and K1 fully connected layers, where the convolutional block consists of M convolutional layers and b pooling layers. The weights and biases between neurons in the entire network structure are w ij and b j During training, the training conditions can be configured as follows: the number of iterations is 500, the loss function gradient descent uses the Adam algorithm, the activation function uses the Sigmoid function, the initial learning rate α is 0.0001, M = 1 / b = 1 / N = 1 / K1 = 192*96*48*24.

[0077] Therefore, after reaching the target training state, the loss function R of the stress field prediction benchmark model CSFP is obtained. D′ (X (i) ,θ CSFP ). In addition, the commonly used technical means in this technical field can be used to determine whether the loss function converges, that is, whether the basic model target training is obtained. The specific method and process for determining whether the basic model target training state is achieved can be selected according to actual needs, so as to meet the requirements of obtaining the stress field prediction benchmark model CSFP through training.

[0078] In one embodiment of the present invention, for the constructed total longitudinal stress prediction basic model CSFP, the following is obtained:

[0079]

[0080] Among them, R D′ (X (i) ,θ CSFP ) is the loss function of the basic model CSFP for total longitudinal stress prediction; N1 is the number of stress field data samples under the preset sea conditions in the standard data set of stress field under the preset sea conditions in the computational domain; θ CSFP is the weight w trained between neurons in the basic model CSFP for total longitudinal stress prediction ij and bias b j The set of f(X (i) θ CSFP ) is the output function of the output layer of the CSFP basic model for total longitudinal stress prediction, is the total longitudinal stress value of the stress field data sample under the preset sea conditions in the i-th computational domain in the standard dataset of the stress field under the preset sea conditions in the computational domain, X (i) The node coordinates of the i-th data sample in the standard data set of stress field under preset sea conditions for the computational domain, It is the predicted value of the total longitudinal stress of the basic stress prediction model CSFP under the preset sea conditions.

[0081] Specifically, the total longitudinal stress prediction value of the basic model CSFP is That is, the output function f(X (i) θ CSFP )The predicted value output. Weight w ij and bias b j The set θ CSFP , which can be determined based on the basic model target state achieved through neural network training, and is specifically consistent with the existing situation after neural network training.

[0082] In one embodiment of the present invention, when performing knowledge transfer learning on the basic total longitudinal stress prediction model CSFP, the method includes:

[0083] Two layers of neurons are added to the output layer of the basic model for total longitudinal stress prediction CSFP in sequence. The standard data set of actual ship test is used as the target domain, and the weight and bias set θ between neurons in the basic model for total longitudinal stress prediction CSFP is frozen. CSFP Under the above conditions, the two added layers of neurons are trained until the target training state of the prediction model is reached to generate the total longitudinal stress prediction model RMMSF of the longitudinal strength member.

[0084] Figure 2 In the longitudinal strength member total longitudinal stress prediction model RMMSF, the total longitudinal stress prediction basic model CSFP and two layers of neurons are added. The two layers of neurons are directly added to the output layer of the total longitudinal stress prediction basic model CSFP to serve as the output layer of the longitudinal strength member total longitudinal stress prediction model RMMSF. In specific implementation, after adding the two layers of neurons, the two layers of neurons need to be trained. During training, the weights w between the neurons in the total longitudinal stress prediction basic model CSFP are maintained. ij and bias b j The set θ CSFP , that is, the weights w between neurons in the basic model CSFP for freezing total longitudinal stress prediction ij and bias b j The set θ CSFP , only the weights and bias θ between the two added layers of neurons are changed exp .

[0085] When the two added layers of neurons are trained to the target training state for the prediction model, the RMMSF prediction model for the total longitudinal stress of longitudinal strength members is obtained. In specific implementation, the training conditions for the two added layers of neurons can be configured as follows: 500 iterations, using the Adam algorithm for gradient descent of the loss function, the Sigmoid function for the activation function, an initial learning rate of 0.0001, and K2 = 32*32. The target training state of the prediction model can generally be set by the number of iterations, or by meeting the test set accuracy. The specific settings for test accuracy and number of iterations should be based on meeting actual needs.

[0086] In one embodiment of the present invention, for the longitudinal strength member total longitudinal stress prediction model RMMSF, the following is obtained:

[0087]

[0088] in, is the loss function of the total longitudinal stress prediction model RMMSF of longitudinal strength members; θ exp To increase the weights and biases between two layers of neurons; is the output function of the output layer of the RMMSF model for predicting the total longitudinal stress of longitudinal strength components, is the total longitudinal stress value of the ith actual ship test in the actual ship test standard data set, is the predicted value of the total longitudinal stress of the longitudinal strength member total longitudinal stress prediction model RMMSF.

[0089] Specifically, when the longitudinal strength member total longitudinal stress prediction model RMMSF is obtained through training, the loss function of the longitudinal strength member total longitudinal stress prediction model RMMSF can be obtained. Prediction value of the total longitudinal stress prediction model RMMSF for longitudinal strength members This is the output function The corresponding output value.

[0090] In one embodiment of the present invention, for producing a standard data set for a real ship test of a ship under a preset sea condition, there are:

[0091]

[0092] Among them, D exp It is the basic data set for the actual ship test; is the boundary node set of the longitudinal strength member, is the set of calculated values ​​of total longitudinal stress at the boundary nodes of longitudinal strength members, and N2 is the number of boundary nodes of longitudinal strength members; is the measured node set of the longitudinal strength member, is the set of total longitudinal stress measured values ​​corresponding to the measured nodes of the longitudinal strength member, N3 is the number of measured nodes of the longitudinal strength member; {y exp} is the measured value of the total longitudinal stress at the measured node;

[0093] Basic data set for real ship test

[0094] D e ' xp D is the basic data set for the actual ship test exp The mean normalization process is performed to form a standard data set for actual ship test.

[0095] Specifically, the basic data set D of the actual ship test of longitudinal strength components exp , including a number of measured node data groups located at the longitudinal strength members and a number of boundary node data groups located at the longitudinal strength members, Figure 3 In the description, the longitudinal strength member is taken as an example of the deck. Several measured nodes are arranged on the deck, and several boundary nodes are selected. The boundary nodes are the nodes on the boundary contour of the deck. The number of measured nodes N3 and the number of boundary nodes N2 are related to the longitudinal strength member and can be selected according to needs.

[0096] After arranging the measured nodes, the total longitudinal stress value of the measured nodes can be obtained by using existing commonly used technical means. At this time, the measured node set of the longitudinal strength member can be obtained. And the total longitudinal stress measured value set corresponding to the measured nodes of the longitudinal strength components

[0097] Similarly, after selecting the boundary nodes, the above-mentioned total longitudinal stress prediction method can be used to obtain the boundary node set of the longitudinal strength member And the total longitudinal stress calculation value set of the boundary nodes of the longitudinal strength members

[0098] Measured node collection Including node coordinates of measured nodes and boundary node sets Including the node coordinates of the boundary nodes. Therefore, for the basic data set D exp Specifically, mean normalization processing is performed on the above-mentioned node coordinates and total longitudinal stress values. The specific process of mean normalization processing can refer to the above description.

[0099] Therefore, the total longitudinal stress value of the ith actual ship test in the actual ship test standard data set is That is the measured value of the total longitudinal stress or the calculated value of the total longitudinal stress after the above-mentioned mean normalization processing.

[0100] For the standard data set of actual ship test, it is generally necessary to divide the standard data set of actual ship test into proportions to form training sets, test sets and validation sets after division. The situation of proportional division to obtain training sets, test sets and validation sets may be related to the total amount of the standard data set of actual ship test and the division ratio, so as to meet the training of the two added layers of neurons.

[0101] In one embodiment of the present invention, for any non-measured node of the longitudinal strength member, the coordinates of the non-measured node are determined, and the coordinates are processed in sequence through the total longitudinal stress prediction basic model CSFP and two layers of neurons added to the output layer of the total longitudinal stress prediction basic model CSFP, so as to output the total longitudinal stress prediction value of the non-measured node under the preset sea conditions through the output layer of the two added layers of neurons.

[0102] Specifically, non-measured nodes refer to nodes on the longitudinal strength components where no stress detection sensors are arranged, and measured nodes are nodes on the longitudinal strength components where stress detection sensors are arranged. For non-measured nodes, the predicted value of the total longitudinal stress can be obtained through the longitudinal strength component total longitudinal stress prediction model RMMSF; for measured nodes, the measured value of the total longitudinal stress under the preset sea conditions can be obtained through the arranged stress detection sensors.

[0103] In specific implementations, the predicted total longitudinal stress values ​​at non-measured nodes under preset sea conditions include the tensile and compressive stresses at the non-measured nodes in the x-direction. Similarly, the measured total longitudinal stress values ​​include the tensile and compressive stresses at the measured nodes in the x-direction, where the x-direction is the aforementioned length of the ship.

[0104] In one embodiment of the present invention, for a given longitudinal strength member, after determining the number of measured nodes, the number of unmeasured nodes can be determined. At this point, a predicted total longitudinal stress value and a measured total longitudinal stress value can be obtained. Subsequently, based on the measured total longitudinal stress values ​​at the measured nodes of the longitudinal strength member under a preset sea condition and the predicted total longitudinal stress values ​​at the unmeasured nodes of the longitudinal strength member under the preset sea condition, a total longitudinal stress field reconstruction method commonly used in the art can be employed to reconstruct the total longitudinal stress field of the longitudinal strength member under the preset sea condition.

[0105] In summary, a total longitudinal stress field reconstruction device suitable for ships in actual sea conditions can be obtained. In one embodiment of the present invention, a total longitudinal stress field reconstruction processor is included, wherein:

[0106] For any longitudinal strength member, the total longitudinal stress field of the longitudinal strength member under the preset sea condition is reconstructed using the above-described method.

[0107] Specifically, the total longitudinal stress field reconstruction processor can generally be a common computer device, that is, a longitudinal strength member total longitudinal stress prediction model RMMSFF is constructed in the computer device, and the longitudinal strength member total longitudinal stress prediction model RMMSF can output the total longitudinal stress prediction value of the non-measured node.

[0108] Based on the measured values ​​of the total longitudinal stress at the measured nodes of the longitudinal strength components under the preset sea conditions and the predicted values ​​of the total longitudinal stress at the non-measured nodes of the longitudinal strength components under the preset sea conditions, the total longitudinal stress field reconstruction processor can reconstruct the total longitudinal stress field of the longitudinal strength components under the preset sea conditions. For the specific construction of the total longitudinal stress prediction model RMMSFF of the longitudinal strength components and the method of reconstructing the total longitudinal stress field of the longitudinal strength components under the preset sea conditions, please refer to the above description.

[0109] In practice, when predicting the total longitudinal stress at unmeasured nodes under pre-set sea conditions, it is generally necessary to perform mean normalization on the node coordinates of these unmeasured nodes. The specific method and process for performing mean normalization can be referenced above. During mean normalization, the mean and variance required for mean normalization can be calculated based on the corresponding coordinate values ​​of all unmeasured nodes and measured nodes throughout the longitudinal strength member.

[0110] In summary, the neural network-based total longitudinal stress prediction basic model CSFP in the present invention replaces the traditional finite element theory. From the perspective of machine learning, the finite element theoretical knowledge is learned through the standard data set of stress field under preset sea conditions in the calculation domain, and after verification, the total longitudinal stress prediction basic model CSFP that satisfies the smaller loss function is obtained. Therefore, the total longitudinal stress value of any node in the longitudinal strength component space can be obtained.

[0111] For the same longitudinal strength member, considering the stress error between the actual ship test and the finite element calculation, a total longitudinal stress prediction model RMMSFF of longitudinal strength members is established based on transfer learning. That is, two layers are added on the basis of the CSFP network to fine-tune the model, and finally the total longitudinal stress prediction model RMMSFF of longitudinal strength members is obtained. The total longitudinal stress prediction model RMMSFF of longitudinal strength members can be used to obtain the predicted value of the total longitudinal stress of non-measured nodes, which provides support for the reconstruction of the total longitudinal stress field, so that the total longitudinal stress of longitudinal strength members under actual sea conditions can be predicted quickly and accurately, and the reconstruction of the total longitudinal stress field of the ship under actual sea conditions can be realized, meeting the needs of the hull total longitudinal strength verification.

Claims

1. A method for reconstructing the total longitudinal stress field of a ship under actual sea conditions, characterized by: For a longitudinal strength member, the total longitudinal stress field reconstruction method of the longitudinal strength member includes: A longitudinal strength member total longitudinal stress prediction model RMMSF is constructed for predicting the total longitudinal stress of the longitudinal strength member under a preset sea condition, wherein: When constructing the longitudinal strength member total longitudinal stress prediction model RMMSF, a neural network-based total longitudinal stress prediction basic model CSFP is first constructed. When constructing the total longitudinal stress prediction basic model CSFP, a stress field standard data set under a preset sea condition in a calculation domain generated based on the longitudinal strength member finite element model is used for training, so as to obtain the total longitudinal stress prediction basic model CSFP after the training reaches the basic model target state; The standard data set of stress field under the preset sea condition of the computational domain includes a plurality of stress field data samples under the preset sea condition of the computational domain, and any stress field data sample under the preset sea condition of the computational domain includes the node coordinates of a node of the longitudinal strength member and the total longitudinal stress value of the node under the preset sea condition; Producing a standard data set of actual ship tests on the longitudinal strength members under preset sea conditions, configuring the produced standard data set of actual ship tests as a target domain, and performing knowledge transfer learning on the total longitudinal stress prediction basic model CSFP based on the configured target domain to generate the longitudinal strength member total longitudinal stress prediction model RMMSF after the knowledge transfer learning; For any non-measured node of the longitudinal strength member, a total longitudinal stress prediction model RMMSF of the longitudinal strength member is used to predict the total longitudinal stress predicted value of the non-measured node under the preset sea condition; Based on the measured values ​​of the total longitudinal stress of the measured nodes of the longitudinal strength components under the preset sea conditions and the predicted values ​​of the total longitudinal stress of the non-measured nodes of the longitudinal strength components under the preset sea conditions, the total longitudinal stress field of the longitudinal strength components under the preset sea conditions is reconstructed.

2. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to claim 1, characterized in that: For the constructed total longitudinal stress prediction basic model CSFP, we have: Among them, R D′ (X (i) ,θ CSFP ) is the loss function of the basic model CSFP for total longitudinal stress prediction; N1 is the number of stress field data samples under the preset sea conditions in the standard data set of stress field under the preset sea conditions in the computational domain; θ CSFP is the weight w trained between neurons in the basic model CSFP for total longitudinal stress prediction ij and bias b j The set of f(X (i) θ CSFP ) is the output function of the output layer of the CSFP basic model for total longitudinal stress prediction, is the total longitudinal stress value of the stress field data sample under the preset sea conditions in the i-th computational domain in the standard dataset of the stress field under the preset sea conditions in the computational domain, X (i) is the node coordinate of the stress field data sample under the preset sea condition of the i-th computational domain in the standard dataset of stress field under the preset sea condition of the computational domain, It is the predicted value of the total longitudinal stress of the basic stress prediction model CSFP under the preset sea conditions.

3. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to claim 2, characterized in that: For the standard stress field data set under the preset sea conditions in the computational domain, we have: Where D is the basic data set of stress field under the preset sea conditions in the computational domain, is the basic sample of stress field data under preset sea conditions in the i-th computational domain in the basic dataset of stress field under preset sea conditions, X low is the minimum value of the finite element mesh node of the longitudinal strength member, X up is the maximum value of the finite element mesh node of the longitudinal strength member, Mesh is the node coordinate generation function of the longitudinal strength member; Patran is the node coordinate X of the longitudinal strength member (i) Calculation function of total longitudinal stress under preset sea conditions; D L ′ is the mean normalized total longitudinal stress of the longitudinal strength member; is the mean value of the total longitudinal stress at all node coordinates of the longitudinal strength member, D Ls is the variance of the total longitudinal stress at all node coordinates of the longitudinal strength member; D Z ′ is the mean normalized node coordinate of the longitudinal strength member; is the mean value of the node coordinates of all the node coordinates of the longitudinal strength members, D Zs is the node coordinate variance of all node coordinates of longitudinal strength members; Using D Z ′ is the mean normalized node coordinate D of the longitudinal strength member Z ′ and the node coordinates normalized with the mean D Z ′ corresponds to the mean normalized total longitudinal stress D L ’Form a stress field data sample under the preset sea conditions in the calculation domain.

4. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to claim 3 is characterized in that: When the neural network model is trained using the standard stress field data set under the preset sea conditions in the computational domain to obtain the basic model for total longitudinal stress prediction CSFP, we have: The mean normalized node coordinates of the longitudinal strength components in the stress field data sample under the preset sea conditions in the computational domain are used as the input of the neural network, and the mean normalized total longitudinal stress values ​​of the corresponding nodes in the stress field data sample under the preset sea conditions in the computational domain are used as the output of the neural network. The convergence of the damage function of the neural network is taken as the target state of the basic model; during training, the Adam algorithm is used for gradient descent of the loss function, and the Sigmoid function is used for the activation function.

5. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to any one of claims 1 to 4, characterized in that: The knowledge transfer learning of the basic total longitudinal stress prediction model CSFP includes: Two layers of neurons are added to the output layer of the basic model for total longitudinal stress prediction CSFP in sequence. The standard data set of actual ship test is used as the target domain, and the weight and bias set θ between neurons in the basic model for total longitudinal stress prediction CSFP is frozen. CSFP Under the above conditions, the two added layers of neurons are trained until the target training state of the prediction model is reached to generate the total longitudinal stress prediction model RMMSF of the longitudinal strength member.

6. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to claim 5, characterized in that: For the total longitudinal stress prediction model RMMSF of longitudinal strength members, we have: in, is the loss function of the total longitudinal stress prediction model RMMSF of longitudinal strength members; θ exp To increase the weights and biases between two layers of neurons; is the output function of the output layer of the RMMSF model for predicting the total longitudinal stress of longitudinal strength components, is the total longitudinal stress value of the ith actual ship test in the actual ship test standard data set, is the predicted value of the total longitudinal stress of the longitudinal strength member total longitudinal stress prediction model RMMSF.

7. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to claim 6, characterized in that: For the production of standard data sets for actual ship tests under preset sea conditions, there are: Among them, D exp It is the basic data set for the actual ship test; is the boundary node set of the longitudinal strength member, is the set of calculated values ​​of total longitudinal stress at the boundary nodes of longitudinal strength members, and N2 is the number of boundary nodes of longitudinal strength members; is the measured node set of the longitudinal strength member, is the set of total longitudinal stress measured values ​​corresponding to the measured nodes of the longitudinal strength member, N3 is the number of measured nodes of the longitudinal strength member; {y exp } is the measured value of the total longitudinal stress at the measured node; Basic data set for real ship test D e ' xp D is the basic data set for the actual ship test exp The mean normalization process is performed to form a standard data set for actual ship test.

8. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to claim 5, characterized in that: The longitudinal strength members include decks, outer plates or inner bottom plates; For any non-measured node of the longitudinal strength member, the coordinates of the non-measured node are determined, and the coordinates are processed in sequence through the total longitudinal stress prediction basic model CSFP and the two layers of neurons added to the output layer of the total longitudinal stress prediction basic model CSFP, so as to output the total longitudinal stress prediction value of the non-measured node under the preset sea conditions through the output layer of the two added layers of neurons.

9. The method for reconstructing the total longitudinal stress field suitable for ships under actual sea conditions according to any one of claims 1 to 4, characterized in that: The predicted value of the total longitudinal stress of the non-measured node under the preset sea condition includes the tensile and compressive stresses of the non-measured node in the x-direction.

10. A device for reconstructing the total longitudinal stress field of a ship under actual sea conditions, characterized by: Includes a total longitudinal stress field reconstruction processor, where For any longitudinal strength member, the total longitudinal stress field of the longitudinal strength member under a preset sea condition is reconstructed using the method described in any one of claims 1 to 9.

Images