Ship structure opening strength prediction method based on improved deep neural network model

Abstract

The application provides a ship structure opening strength prediction method based on an improved deep neural network model, acquires coarse grid stress distribution data before ship structure opening and fine grid stress distribution data after opening, constructs a deep neural network model, introduces labels of the coarse grid stress distribution data into a plurality of neurons of the last layer of a hidden layer for performance evaluation during model training, performs coarse grid stress supervision by limiting the difference between the output values of the neurons to which the coarse grid labels are introduced and reference stress values of the coarse grid, adopts a composite loss function composed of a fine grid stress prediction error and a coarse grid stress supervision error to train the model, and acquires an optimal training model as an opening strength prediction model. The method of the application introduces a machine learning model to perform opening strength prediction, and is improved in model application, and the overall trend of stress distribution provided by the coarse grid data before opening is considered, so that efficient prediction of the opening strength of the ship structure is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ship structure opening strength evaluation and machine learning, and particularly relates to a ship structure opening strength prediction method based on an improved deep neural network model. BACKGROUND

[0002] In ship engineering design and analysis, the structure strength prediction of openings is an important link to ensure the safety of the overall structure. However, the traditional methods are often difficult to accurately predict the stress distribution of the opening area, such as the progressive damage analysis method and the characteristic length method, which can predict the strength of the material containing holes to some extent, but these methods often rely on a large amount of experimental data and cannot simulate complex failure processes.

[0003] In recent years, machine learning and neural network methods have shown good application potential in structural stress analysis, but traditional neural network models often use "black box" end-to-end learning, lack of physical knowledge guidance, and have limited generalization ability in structural stress analysis. SUMMARY

[0004] The purpose of the present application is to provide a ship structure opening strength prediction method based on an improved deep neural network model, which improves the prediction accuracy and efficiency based on the improved deep neural network model.

[0005] To achieve the above purpose, the technical solution provided by the present application is as follows:

[0006] A ship structure opening strength prediction method based on an improved deep neural network model, the method comprising:

[0007] Obtaining historical data of ship structure openings, including parameter data affecting stress after opening and stress distribution data around the opening; the stress distribution data around the opening includes fine grid stress distribution data after opening and coarse grid stress distribution data at the corresponding position before opening;

[0008] Based on the historical data, a training set is constructed, and a deep neural network model containing an input layer, multiple hidden layers and an output layer is trained, wherein:

[0009] The training set takes the parameter data as input features and the fine grid stress distribution data as label data; the last layer of the hidden layer introduces the label of the coarse grid stress distribution data for performance evaluation, and the difference between the output value of the neuron introducing the coarse grid label and the reference stress value of the coarse grid is limited to supervise the coarse grid stress, and a composite loss function composed of the fine grid stress prediction error and the coarse grid stress supervision error is used to train the model;

[0010] Obtaining an optimal training model as a hole strength prediction model, and using the prediction model to obtain stress distribution data of a to-be-predicted hole region.

[0011] In some embodiments of the present application, the parameter data affecting the stress after the hole is formed are hole diameter, long side length, presence or absence of a pillar, hole position, and external load.

[0012] In some embodiments of the present application, the long side length is determined based on the hole shape, and for a hole in the shape of a waist circle, the long side length is taken as the input, and for a hole in the shape of a right circle, the long side length is taken as 0.

[0013] The value of the presence or absence of a pillar is 1 if the pillar exists, and 0 otherwise.

[0014] In some embodiments of the present application, the fine grid stress distribution data is preprocessed to construct a training set, including:

[0015] The area around the hole is divided into a plurality of grids along the circumferential direction with a preset grid size as the analysis unit size, and the average stress in each grid is taken as the stress value on the corresponding grid to construct a label set.

[0016] In some embodiments of the present application, when the number of divided grids is < N, the number of grids is supplemented by interpolation.

[0017] During interpolation, the grid with the maximum stress remains unchanged, and the remaining grids are divided by interpolation.

[0018] In some embodiments of the present application, the historical data is obtained by finite element simulation, and the stress distribution data around the hole is obtained in the following manner:

[0019] After modeling the ship, coarse grids are divided according to the spacing between the bones, and after applying the load, coarse grid stress distribution data at the to-be-opened position is obtained.

[0020] A hole model is constructed for the to-be-opened position, and fine grids are divided at the hole position, and after applying the same load, fine grid stress distribution data around the hole is obtained.

[0021] In some embodiments of the present application, the last layer of the hidden layer selects the labels of the coarse grid stress distribution data introduced by the first two neurons.

[0022] In some embodiments of the present application, the fine grid stress prediction error and the coarse grid stress supervision error are calculated using a mean square error function.

[0023] The weight of the fine grid stress prediction error is greater than the weight of the coarse grid stress supervision error.

[0024] In some embodiments of the present application, the deep neural network model is a multi-layer perceptron (MLP).

[0025] In some embodiments of the present application, the method of obtaining the optimal training model is to use an early stopping strategy during model training, and when the validation set loss does not appear to decrease within a continuous number of rounds, the training is automatically stopped and the current optimal model is output.

[0026] In some embodiments of the present application, the method further comprises performing inverse normalization processing on the plurality of grid stress value data output by the opening strength prediction model, and displaying the stress value data through a cloud chart visualization technology to show the stress distribution of the opening area.

[0027] The present application has the following beneficial effects:

[0028] The present application uses a deep neural network model to learn and implement ship structure opening strength prediction, and improves the deep neural network learning model. In the last hidden layer, the neurons are selected to introduce the physical label of the coarse grid stress before opening, and the stress mapping from the coarse grid to the fine grid is studied. The present application considers the embodiment of the overall trend of the coarse grid data before opening on the stress distribution, explicitly distinguishes the coarse grid and fine grid data input channels in the network architecture, so that the model can learn the stress conversion characteristics across scales, thereby improving the explainability and reliability of the beam opening stress evaluation neural network model, and the improved model further improves the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0029] The present application and its features and advantages will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The same reference numbers in all the drawings indicate the same parts. The drawings are not necessarily drawn to scale, the emphasis being on embodying the principle of the application.

[0030] Figure 1 is a deep neural network architecture diagram based on a multi-layer perceptron (MLP) in an embodiment.

[0031] Figure 2 is a coarse grid model diagram of a certain deck of a roll-on / roll-off ship in an embodiment.

[0032] Figure 3 is a uniform load working condition diagram in an embodiment.

[0033] Figure 4 is a local opening fine grid model diagram in an embodiment.

[0034] Figure 5 is a stress cloud chart diagram in an embodiment.

[0035] Figure 6is an improved MLP and an MLP prediction error angle distribution in an embodiment. DETAILED DESCRIPTION

[0036] It should be noted that the embodiments and features in the embodiments in the present application can be combined with each other without conflict. It should be noted that the terms used in the present application are only for describing the specific embodiments, and are not intended to limit the exemplary embodiments according to the present application.

[0037] Embodiment 1

[0038] The present embodiment provides an improved MLP-based efficient prediction method for the opening strength of a ship structure, comprising the following steps:

[0039] S101: Obtain the historical data of the opening, including external load (uniform load), opening diameter, long side length, number of struts (whether there is a strut), opening position and stress distribution around the opening, wherein the stress distribution around the opening includes the stress distribution of the coarse grid before the opening and the stress distribution of the fine grid after the opening;

[0040] S102: Preprocess according to the historical data, including:

[0041] S201: According to the shape of the opening, it is divided into a right circle and a waist circle, the waist circle is input with its long side length, and the right circle is input with 0;

[0042] According to whether there is a strut at the beam where the opening is located, the number of struts is valued as: 1 if there is a strut, otherwise 0.

[0043] The opening position is according to the coordinates (x, y, z), and the coordinate value of each axis is taken as an input. The selection of the origin point has no effect on modeling. In the present embodiment, the center of the cross beam is selected as the origin point.

[0044] S202: Divide the fine grid stress distribution around the opening into 28 grids, and take the average stress in each grid as the stress value on the corresponding grid.

[0045] When dividing the grid, in the present embodiment, the fine grid stress data around the opening within 50mm × 50mm is extracted by finite element analysis as the key analysis area. Based on the size of the analysis area, when the diameter of the opening is the largest, the grid division around the opening is exactly 28. For smaller diameter holes, the surrounding grid may not be enough. 28, in order to ensure the consistency of the output dimension of the machine learning model, the grid data is supplemented to 28 using an interpolation method. One optional operation mode is: if the grid is less than 28, select the grid with the maximum stress in the fine grid as the starting point, and perform interpolation and division between two grids counterclockwise until the number of grids reaches 28. When interpolating, the size of the grid with the maximum stress does not change, and only the interpolation and division between other grids are performed.

[0046] S103: Construct a data set according to the processed historical data, comprising:

[0047] S203: Take the external load, opening diameter, long side length, number of struts, and opening position contained in the historical data as input features to construct an input data set X; take the stress values on the corresponding 28 grids as output features to construct an output data set Y (label set).

[0048] In some embodiments of the present application, the finite element method is used to simulate the stress distribution at the opening of the ship beam. First, the finite element model of the ship beam is established using the widely used finite element software Patran in the ship industry, and the simulation results are obtained by calculating through Nastran. In the coarse grid model, all the main structures of the whole ship are modeled, the deck, the outer plate, the longitudinal and transverse bulkheads, and the main supporting members (stringers), web plates and other components are modeled using 2D plate elements, and the components such as the bone material and the reinforcing rib are modeled using 1D beam elements, which have no significant effect on the calculation results. The coarse grid division is carried out according to the spacing of the bone material, as shown in Figure 2 The load condition is set as the typical load in the actual operation of the roll-on / roll-off ship, mainly as the uniform load. Figure 3 The load view for a certain condition.

[0049] After the loading calculation of each condition is completed, the beam with high stress is selected as the typical condition for analysis according to the coarse grid results, the grid refinement analysis and calculation are carried out on the selected strong beam opening, and the size and position of the opening are parameterized designed. The simulated opening size covers the round hole of the commonly used aperture of the ship, and the opening position is selected as the typical arrangement of the center and edge of the beam to comprehensively analyze the variation law of the stress distribution. According to the standard of the Norwegian Ship Classification Society, the stress data of the fine grid of 50mm x 50mm around the opening are extracted as the key analysis area through the finite element analysis, mainly focusing on the center value of the Mises stress of the upper and lower layers. The fine grid round hole modeling is as shown in Figure 4 .

[0050] The input features that significantly affect the stress after the opening are selected, including the geometric parameters of the beam (opening diameter, opening position), load parameters (uniform load, etc.), and the number of struts, and the stress values of the corresponding coarse grid and fine grid are recorded as the data labels of the model. These data points cover the stress values at the key positions around the opening and their corresponding geometric coordinates, which constitute a high-dimensional feature data set for training the neural network.

[0051] S104: Train the artificial intelligence model using the constructed training set to obtain an optimal model;

[0052] Specifically, the dataset is divided into a training set and a test set according to a proportional method, with the training set accounting for 80% of the total data and the test set accounting for 20%. The model is trained using the divided dataset, and the accuracy is validated using the test set to obtain the optimal model.

[0053] To improve overall prediction performance and enhance forecast reliability, this embodiment employs an improved deep neural network model to train the forecast model, such as... Figure 1 As shown. This embodiment uses an MLP model as the deep neural network model. The improved MLP model's input layer accepts seven dimensions of input features, including external load, opening diameter, long side length, number of supports, and opening location (X, Y, and Z axis coordinates each occupy one input dimension). It contains five hidden layers. The outputs of the first and second neurons in the last hidden layer serve as the coarse-grid stress prediction value. The hidden layers use the Sigmoid activation function, and the output layer uses the Softplus activation function. The loss function is a weighted average of the fine-grid stress error and the coarse-grid stress error, as detailed below:

[0054] The model incorporates a coarse-grid stress prediction loss into its loss function by calling the output values ​​of the first two neurons in the last hidden layer. These output values ​​correspond to the physical labels of the coarse-grid stress data. Its form is:

[0055]

[0056]

[0057] in, and These are the prediction losses for fine-grid and coarse-grid predictions, respectively. and The first in the dataset The first sample One predicted value and one actual value. and These are hyperparameters used to balance the weights of the two parts of the loss. In this paper... , .

[0058] In the last hidden layer of the model, the network reflects the coarse mesh stress distribution of the structure through the calculation of the output values ​​of the first two neurons. We define the last hidden layer of the network as... ,in Among them, the output values ​​of the first two neurons Compare the label values ​​with the coarse mesh stress data.

[0059]

[0060] where, and are the output values of the last two neurons of the last hidden layer for the i-th sample in the dataset, and are the corresponding coarse grid stress data labels.

[0061] During the training process of the artificial intelligence model, the Adam optimizer is adopted, and the mean square error (MSE) is taken as the loss function. The parameters of the model are optimized according to the training set at each iteration, and the training is ended after 10,000 iterations.

[0062] S105: Using the trained model to predict the opening strength under actual working conditions, and outputting stress distribution data;

[0063] S106: Visualizing the output data to obtain a data cloud map;

[0064] Specifically, in order to facilitate the processing and understanding of the output data, and to enable the user to more intuitively understand the complex data analysis results, based on the processing of S205, the 28 grid stress value data output by the model is subjected to inverse normalization processing, and these stress value data are displayed through cloud map visualization technology to show the stress distribution of the opening area. A schematic effect diagram of the stress cloud map is shown in FIG. 2. Figure 5

[0065] Through the above steps, the historical data obtained is used to train the artificial intelligence model, and an optimal model capable of predicting the stress distribution of the opening area in the structure is obtained. The stress distribution data output by the optimal model is subjected to visual processing to obtain a stress distribution cloud map.

[0066] Example 2

[0067] This embodiment compares the performance of the improved deep neural network (improved MLP) of the present application and the traditional deep neural network (MLP) in the stress prediction of the opening area of the ship beam.

[0068] As shown in Table 1, on the test set, the improved MLP model exhibits more excellent performance: the average relative error of the fine grid stress is 13.09%, which is lower than the 23.38% of the MLP model; the average relative error of the maximum stress is 8.08%, which is slightly lower than the 14.5% of the MLP model. It is worth noting that the improved MLP model has an average relative error (MAPE) of 18.28% in the coarse grid stress prediction which cannot be handled by the MLP model, which reflects its comprehensive advantage in the multi-scale prediction of the opening stress.

[0069] Table 1 Comparison of prediction accuracy of improved MLP and MLP for ship opening strength ​​

[0070]

[0071] Meanwhile, by analyzing the prediction error of the stress at different angles of the opening area, it is found that the average relative error distribution of the improved MLP and MLP model presents the same characteristics. The angle of the grid is calculated from the top of the circular hole, and recorded in an increasing manner in a counterclockwise direction. Figure 6 The average relative error distribution characteristics of the improved MLP and MLP model at different angles of the opening area are shown, and the results show that the angle distribution of the average relative error of the improved MLP and MLP presents two obvious peaks, which appear in the first 30% and the last 30% of the angle area, and the peak value is about 20%, which corresponds to the complex fluctuation area of the stress concentration of the opening. This is mainly because in the complex fluctuation area of the stress concentration, due to the sharp change of the stress value, it is more difficult for the model to capture the nonlinear characteristics, thereby leading to the decrease of the prediction accuracy. In other angle areas, the stress changes relatively gently, and the average relative error maintains at a level of about 11%. Even in another area of stress concentration, the area near the lower end of the opening, the relative error of the model prediction remains at a low level. The improved MLP and MLP model in the symmetrical area of the stress distribution shows similar prediction error characteristics, indicating that in the symmetrical area of the stress distribution, the model encounters similar difficulties in the stress concentration area of the opening, but the improved MLP performs better overall.

Claims

1. A method for predicting the strength of a cutout in a ship structure based on an improved deep neural network model, characterized by, The method comprises: Obtaining historical data of the opening of the ship structure by finite element simulation, including parameter data affecting stress after opening and stress distribution data around the opening; the stress distribution data around the opening includes fine-grid stress distribution data after opening and coarse-grid stress distribution data at the corresponding position before opening; the stress distribution data around the opening is obtained in the following manner: after modeling the ship, coarse grids are divided according to the spacing of the bone material, and after applying the load, the coarse-grid stress distribution data at the position to be opened is obtained; a fine-grid is divided at the opening position, and after applying the same load, the fine-grid stress distribution data around the opening is obtained Based on the historical data, a training set is constructed, and a deep neural network model containing an input layer, multiple hidden layers and an output layer is trained, wherein: The training set is constructed with the parameter data as input features and the fine-grid stress distribution data as label data; the last layer of the hidden layer introduces the label of the coarse-grid stress distribution data for performance evaluation, and the difference between the output value of the neuron introducing the coarse-grid label and the reference stress value of the coarse grid is limited to supervise the coarse-grid stress, and a composite loss function composed of the fine-grid stress prediction error and the coarse-grid stress supervision error is used to train the model; An optimal training model is obtained as an opening strength prediction model, and the stress distribution data of the opening region to be predicted is obtained by using the prediction model.

2. The open-hole strength prediction method of claim 1, wherein, The parameter data affecting stress after opening includes opening diameter, long side length, whether there is a support, opening position and external load.

3. The open-hole strength prediction method of claim 2, wherein, The long side length is determined based on the shape of the opening, and for a waist-round opening, the long side length is taken as the input, and for a circular opening, the long side length is taken as 0. The value of whether there is a support is 1 if there is a support, and 0 otherwise.

4. The open-hole strength prediction method of claim 1, wherein, The fine-grid stress distribution data is preprocessed before being used to construct the training set, including: Divide the opening region into several grids along the circumferential direction with a preset grid size as the analysis unit size, and take the average stress in each grid as the stress value on the corresponding grid to construct the label set.

5. The open-hole strength prediction method of claim 4, wherein, When the number of grids is less than N, the number of grids is supplemented by interpolation. When interpolating, the grid with the maximum stress remains unchanged, and the remaining grids are interpolated and divided.

6. The open-hole strength prediction method of claim 1, wherein, The last layer of the hidden layer selects the first two neurons to introduce the label of the coarse-grid stress distribution data.

7. The open-hole strength prediction method of claim 1, wherein, The fine-grid stress prediction error and the coarse-grid stress supervision error are calculated using the mean square error function. The weight of the fine-grid stress prediction error is greater than that of the coarse-grid stress supervision error.

8. The open-hole strength prediction method of claim 1, wherein, The deep neural network model is a multi-layer perceptron (MLP).

9. The open-hole strength prediction method of claim 1, wherein, The method further comprises performing inverse normalization on the stress value data of several grids output by the opening strength prediction model, and displaying the stress value data through cloud chart visualization technology to show the stress distribution of the opening region.

Images