A method for modifying a stochastic model of an offshore engineering structure
By training variational autoencoders and conditionally invertible neural network models, a mapping relationship between the finite element model parameters of marine engineering structures and their actual acceleration responses is established. This solves the problems of low correction accuracy and long processing time in marine engineering structures, and achieves efficient model correction and real-time evaluation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- OCEAN UNIV OF CHINA
- Filing Date
- 2026-01-09
- Publication Date
- 2026-06-02
AI Technical Summary
Existing stochastic model correction methods have low accuracy and are too time-consuming in correcting marine engineering structures, making it difficult to meet the needs of real-time assessment and emergency response.
By employing variational autoencoder neural network models and conditional invertible neural network models, and by training the probability density distribution characteristics of the acceleration response signal, a mapping relationship between the finite element model parameters and the actual acceleration response is established, thereby enabling direct correction of the finite element model parameters.
It improved the accuracy of model correction, shortened the correction time, and met the needs of real-time assessment and emergency response for health monitoring of marine engineering structures.
Smart Images
Figure CN122133364A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marine engineering technology, and more specifically, to a method for correcting stochastic models of marine engineering structures. Background Technology
[0002] Marine engineering structures are the core support for marine resource development, and their mechanical properties directly affect the safety of engineering operations. To ensure structural safety, numerical models are commonly used to calculate the mechanical properties of structures. However, the initial numerical models established based on design parameters often differ significantly from the actual state of the structure. This difference stems from multiple factors: model simplification errors, construction deviations, and material performance degradation during long-term service. The superposition of these uncertainties prevents the initial model from accurately reflecting the true mechanical response of the structure, thus affecting the reliability of monitoring work such as damage identification and life assessment, becoming a key bottleneck restricting the improvement of the accuracy of structural health monitoring.
[0003] To eliminate discrepancies between models and actual structures, model correction techniques have emerged. Unlike traditional deterministic model correction, which treats parameters such as loads and materials as fixed values, stochastic model correction quantifies the probability distribution of various uncertainties and incorporates them into the model correction process. By probabilistically adjusting the parameters of the initial model, stochastic model correction can construct an accurate model that highly matches the actual mechanical properties of the structure. This not only improves the accuracy of structural response prediction but also provides more reliable model support for structural health monitoring, effectively reducing the risk of misjudgments caused by model errors.
[0004] Current stochastic model adjustments still have the following problems in their application to marine engineering structures:
[0005] (1) It is difficult to construct the parameter mapping relationship. There is a complex nonlinear relationship between the model parameters and the actual acceleration response of large marine engineering structures. Existing methods are difficult to accurately capture this mapping relationship, which leads to parameter estimation deviations during the correction process and affects the accuracy of model correction.
[0006] (2) The correction process takes too long. Existing methods often rely on tedious iterative optimization or a large number of sample calculations. Even under normal parameter scales, it often takes several hours or even several days to complete a complete correction, which is far from meeting the timeliness requirements of real-time assessment and emergency response in marine engineering structure health monitoring. Summary of the Invention
[0007] In order to solve the technical problems of low correction accuracy and excessive time consumption in the application of existing stochastic model correction methods for marine engineering structures, this invention proposes a stochastic model correction method for marine engineering structures, which can solve the above problems.
[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for correcting a stochastic model of marine engineering structures includes: Step 1: Obtain the initial finite element model and select m model parameters as parameters to be corrected; Step 2: Determine the range of variation of the parameter to be corrected, and sample the parameter to be corrected n times to obtain n sets of m-dimensional parameter samples; n sets of parameter samples are input into the initial finite element model to perform transient analysis under wave excitation. d positions in the initial finite element model are selected for observation. Each set of parameter samples yields an acceleration response signal in t×d dimensions, where t is the total time step and the total dimension of the response is n×t×d. Step 3: Use the acceleration response signal to train the variational autoencoder neural network model and the conditional invertible neural network model; Step four: Deploy d acceleration sensors on the marine engineering structure, with their locations corresponding one-to-one with the d observation locations in the initial finite element model, and monitor for a duration of t. * The acceleration response was obtained by measuring the actual acceleration response signal. The measured acceleration response signal is input into the variational autoencoder neural network model to obtain the first measured latent variable z1 representing the probability density distribution characteristics of the measured response. * ; The first measured latent variable z1 * The input is fed into the conditional invertible neural network model, the latent variable z2 is sampled, and the posterior predicted value of the parameter to be corrected is obtained through the inverse calculation of the invertible coupling block. The mean of the posterior predicted value is the corrected value of the parameter to be corrected. The correction value is input into the initial finite element model to obtain the corrected finite element model.
[0009] In some embodiments, the variational autoencoder neural network model described in step three includes an encoder and a decoder, and the training method for the variational autoencoder neural network model includes:
[0010] The acceleration response signals of each group are input to the encoder. The encoder encodes the input signals according to the encoder parameters and outputs the encoded signals to the decoder. The decoder decodes the input signals to obtain the regenerated signals of the acceleration response signals. The error between the regenerated signals and the acceleration response signals is calculated. The encoder parameters are adjusted, and the error is minimized during training to obtain a trained variational autoencoder neural network model.
[0011] In some embodiments, the training method for the conditionally invertible neural network model in step three includes: The encoder calculates the first latent variable z1 representing the probability density distribution characteristics of the acceleration response signal based on each set of acceleration response signals and their corresponding encoder parameters, and obtains n×p-dimensional data, where p is the data length of the first latent variable z1. A conditional invertible neural network model is constructed. Each set of parameter samples is input into the conditional invertible neural network model as training samples. The first hidden variable z1 corresponding to the parameter samples is used as a condition, and the second hidden variable z2 is obtained through forward learning of the invertible coupling block. The second latent variable z2 is inversely calculated through the invertible coupling block to obtain the inverse calculation signal of the parameter sample. The loss function is constructed with the goal of maximizing the log-likelihood of the training sample, and the training of the conditional invertible neural network model is completed.
[0012] In some embodiments, the step of constructing a conditionally invertible neural network model further includes the step of initializing the second latent variable z2, by selecting a standard normal distribution as the initial second latent variable z2.
[0013] In some embodiments, the initial finite element model mentioned in step one is obtained based on the design parameters of the marine engineering structure.
[0014] In some embodiments, the method for determining the parameter to be corrected in step one includes: Select some model parameters as initial correction parameters, and calculate the contribution of each initial correction parameter to the structural response using a sensitivity analysis algorithm; Select the top m initial correction parameters with the largest contribution as the parameters to be corrected, and determine the range of variation of the parameters to be corrected.
[0015] In some embodiments, the sensitivity analysis algorithm is the Sobol method.
[0016] In some embodiments, the contribution of each initially selected correction parameter to the structural response is calculated as follows: Calculate the first-order sensitivity index of each initial selection correction parameter. and overall sensitivity index :
[0017] ;
[0018] ; in, Indicates the parameter adjustment based on the initial selection. The variance caused by uncertainty This represents the total variance of the model output. This represents the variance caused by the uncertainty of the remaining input parameters; Select one and The initial correction parameter whose maximum value is greater than 0.1 is selected as the parameter to be corrected.
[0019] In some embodiments, in step two, the parameter to be corrected is sampled n times within the range of variation using the Latin hypercube sampling method.
[0020] In some embodiments, the wave excitation in step two uses environmental excitation measured on marine engineering structures.
[0021] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The stochastic model correction method for marine engineering structures of the present invention trains a variational autoencoder neural network model and a conditional invertible neural network model. The variational autoencoder neural network model is used to extract the probability density distribution characteristics of the acceleration response, and the conditional invertible neural network model is used to establish the mapping relationship between the model parameters of the finite element model and the actual acceleration response of the marine engineering structure. It can directly correct the model parameters of the finite element model based on the measured acceleration response, avoiding the problem of low correction accuracy caused by parameter estimation deviation during the correction process. The correction accuracy of this scheme is high.
[0022] This method only requires iterative calculations during the model training step. In the actual application stage, it only requires reverse calculation using the trained model. There are no more iterative calculations during the correction process, and the structural parameters can be corrected in a short time, meeting the real-time correction needs during structural operation and maintenance.
[0023] Other features and advantages of the present invention will become clearer after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. Attached Figure Description
[0024] Figure 1 This is a training block diagram of the variational autoencoder neural network model and the conditional invertible neural network model in one embodiment of the stochastic model correction method for marine engineering structures proposed in this invention. Figure 2 This is an application block diagram of the variational autoencoder neural network model and the conditional invertible neural network model in one embodiment of the stochastic model correction method for marine engineering structures proposed in this invention; Figure 3 This is a schematic diagram of the initial finite element model constructed in one embodiment of the stochastic model correction method for marine engineering structures proposed in this invention; Figure 4 This is a standardized posterior prediction map of each model parameter in one embodiment of the stochastic model correction method for marine engineering structures proposed in this invention. Figure 5 This is a schematic diagram of the fitted line between the predicted value and the target value in one embodiment of the stochastic model correction method for marine engineering structures proposed in this invention. Detailed Implementation
[0025] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] Example 1, see Figure 1 As shown in the figure, this embodiment proposes a method for correcting a stochastic model of marine engineering structures, including: Step 1: Obtain the initial finite element model and select m model parameters of the initial finite element model as parameters to be corrected.
[0028] The initial finite element model has many parameters. In order to reduce the amount of calculation, this scheme selects some of the model parameters for modification.
[0029] Step 2: Determine the range of variation of the parameter to be corrected, and sample the parameter to be corrected n times to obtain n sets of m-dimensional parameter samples.
[0030] The range of variation of the parameter to be corrected depends on the elastic modulus of the material of the marine engineering structure, and is equal to the standard value ± the preset value.
[0031] n sets of parameter samples are input into the initial finite element model to perform transient analysis under wave excitation. d positions in the initial finite element model are selected for observation. Each set of parameter samples yields a t×d dimension acceleration response signal, where t is the total time step and the total response dimension is n×t×d.
[0032] Step 3: Train the variational autoencoder neural network model and the conditional invertible neural network model using the acceleration response signal. The variational autoencoder neural network model is used to extract the probability density distribution characteristics of the acceleration response, while the conditional invertible neural network model uses the probability density distribution characteristics of the acceleration response as conditions to establish the mapping relationship between the model parameters of the finite element model and the actual acceleration response of the marine engineering structure.
[0033] Step four: Deploy d acceleration sensors on the marine engineering structure, with their locations corresponding one-to-one with the d observation locations in the initial finite element model, and monitor for a duration of t. *The acceleration response is measured to obtain the actual acceleration response signal. This step matches the acceleration response monitoring location of the entity with the acceleration response monitoring location in the finite element model, making the digital finite element model closer to the actual marine engineering structure. The digital model can more accurately reflect the true mechanical response of the structure, thus enabling subsequent application steps to calculate the probability density distribution characteristics of the measured response using a conditional reversible neural network model. These probability density distribution characteristics are then used as conditions for the conditional reversible neural network model, which inversely calculates the model parameters based on these conditions. These model parameters are used to correct the finite element model.
[0034] In some embodiments, the selection of d locations is made as dispersed as possible to cover different areas of the marine engineering structure and capture acceleration signals at different directions and heights.
[0035] like Figure 2 As shown, the measured acceleration response signal is input into the variational autoencoder neural network model to obtain the first measured latent variable z1 representing the probability density distribution characteristics of the measured response. * .
[0036] The first measured latent variable z1 * The input is fed into the conditional invertible neural network model, the latent variable z2 is sampled, and the posterior predicted value of the parameter to be corrected is obtained through the inverse calculation of the invertible coupling block. The mean of the posterior predicted value is the corrected value of the parameter to be corrected.
[0037] Input the correction value into the initial finite element model to obtain the corrected finite element model.
[0038] The stochastic model correction method for marine engineering structures in this embodiment trains a variational autoencoder neural network model and a conditional invertible neural network model. The variational autoencoder neural network model is used to extract the probability density distribution characteristics of the acceleration response, and the conditional invertible neural network model is used to establish the mapping relationship between the model parameters of the finite element model and the actual acceleration response of the marine engineering structure. It can directly correct the model parameters of the finite element model based on the measured acceleration response, avoiding the problem of low correction accuracy caused by parameter estimation deviation during the correction process.
[0039] This method only requires iterative calculations during the model training step. In the actual application stage, it only requires reverse calculation using the trained model. There are no more iterative calculations during the correction process, and the structural parameters can be corrected in a short time, meeting the real-time correction needs during structural operation and maintenance.
[0040] A variational autoencoder (VAE) is a generative model that generates new data by learning the latent representation of the data. Its core idea is to model the data distribution as a probability distribution and use variational inference for approximate inference. A variational autoencoder neural network model includes an encoder, a decoder, and a sampler. The encoder maps the input data to distribution parameters (such as mean and variance) in the latent space, typically implemented by a fully connected network. The sampler samples from the latent space based on the distribution parameters output by the encoder, for example, sampling latent variables from a Gaussian distribution. The decoder reconstructs the sampled latent variables into the output in the data space, attempting to recover the original input. In some embodiments, such as... Figure 1 As shown, the training method for the variational autoencoder neural network model in step three includes: The acceleration response signals of each group are input to the encoder. The encoder encodes the input signals according to the encoder parameters and outputs the encoded signals to the decoder. The decoder decodes the input signals to obtain the regenerated signals of the acceleration response signals. The error between the regenerated signals and the acceleration response signals is calculated. The encoder parameters are adjusted, and the error is minimized during training to obtain a trained variational autoencoder neural network model.
[0041] In this embodiment, the variational autoencoder neural network model is trained based on a probabilistic graphical model. It assumes that the encoder output data distribution follows a Gaussian distribution and approximates the posterior distribution through variational inference. Error is used to measure the difference between the output and input data, ensuring reconstruction quality.
[0042] The Conditional Invertible Neural Network (CIIN) is an extended model based on invertible neural networks (INNs), which enhances its flexibility and robustness in specific tasks by introducing conditional variables. In this embodiment, during the training process of the CIIN model, the probability density distribution characteristics of the acceleration response output by the variational autoencoder neural network model are used as conditions to establish a mapping relationship between the model parameters of the finite element model and the actual acceleration response of the marine engineering structure. Therefore, in practical applications, by inputting the measured response probability density distribution characteristics as conditions and utilizing the established mapping relationship, the model parameters corresponding to the measured response can be calculated inversely. These model parameters then serve as correction parameters to adjust the corresponding model parameters of the finite element model.
[0043] The probability density distribution characteristics of acceleration response reflect the high-dimensional characteristics of acceleration response. By using it as a condition for conditional reversible neural network models, it is beneficial to reduce the amount of conditional data in the conditional reversible neural network models.
[0044] In some embodiments, the training method for the conditionally invertible neural network model in step three includes: The encoder calculates the first latent variable z1, representing the probability density distribution characteristics of each acceleration response signal, based on each set of acceleration response signals and their corresponding encoder parameters, resulting in n×p dimensional data, where p is the data length of the first latent variable z1. For example... Figure 1 As shown, the first latent variable z1 includes eigenvalues μ and eigenvalues σ.
[0045] A conditional invertible neural network model is constructed. Each set of parameter samples is input into the conditional invertible neural network model as training samples. The first hidden variable z1 corresponding to the parameter sample is used as a condition, and the second hidden variable z2 is obtained through forward learning of the invertible coupling block.
[0046] In some embodiments, the step of constructing the conditionally invertible neural network model further includes initializing the second hidden variable z2, selecting a standard normal distribution as the initial second hidden variable z2. The conditionally invertible neural network realizes the distribution transformation between the conditional probability distribution of the parameter to be corrected under known acceleration response characteristics and the standard normal distribution.
[0047] The second latent variable z2 is inversely calculated through the invertible coupling block to obtain the inverse calculation signal of the parameter sample. The loss function is constructed with the goal of maximizing the log likelihood of the training sample, and the training of the conditional invertible neural network model is completed.
[0048] The loss is calculated by maximizing the log-likelihood of the training samples. The parameters of the previously trained model are then corrected by reverse calculation. The forward and reverse calculations are then performed again until the loss is minimized.
[0049] In some embodiments, the initial finite element model in step one is obtained based on the design parameters of the marine engineering structure.
[0050] In some embodiments, an initial finite element model of the marine engineering structure is established based on the design and construction data of the actual marine engineering structure.
[0051] In some embodiments, the method for determining the parameter to be corrected in step one includes: Select some model parameters as initial correction parameters, and calculate the contribution of each initial correction parameter to the structural response using a sensitivity analysis algorithm.
[0052] Select the top m parameters with the largest contributions as the initial correction parameters and determine the range of variation for these parameters. It is understood that m is a positive integer.
[0053] The contribution reflects the degree of influence of the acceleration response of marine engineering structures. Therefore, in this scheme, several parameters with large contributions are selected from the initial selection of correction parameters as parameters to be corrected, which further reduces the amount of calculation without affecting the accuracy.
[0054] In some embodiments, the sensitivity analysis algorithm is the Sobol method.
[0055] In some embodiments, the contribution of each initially selected correction parameter to the structural response is calculated as follows: Calculate the first-order sensitivity index of each initial selection correction parameter. and overall sensitivity index :
[0056] ;
[0057] ; in, Indicates the parameter adjustment based on the initial selection. The variance caused by uncertainty This represents the total variance of the model output. This represents the variance caused by the uncertainty of the remaining input parameters.
[0058] Select one and The initial correction parameter whose maximum value is greater than 0.1 is selected as the parameter to be corrected.
[0059] In some embodiments, step two involves sampling the parameters to be corrected n times within a range of variation using the Latin hypercube sampling method. The range of variation can be set based on empirical values. This method covers the entire sampling space with as few samples as possible, which helps the subsequent conditionally invertible neural network learn an accurate posterior distribution of the parameters.
[0060] In some embodiments, the wave excitation in step two uses environmental excitation measured on marine engineering structures.
[0061] Example 2, in this example, the following is taken as an example: Figure 3 The model of the marine platform structure shown is modified to illustrate this scheme.
[0062] The first step is to establish Figure 3 The finite element model of the offshore platform shown includes a leg 1-1, a jacket column 1-2, an upper block 1-3, horizontal braces 3-1, 3-2, and 3-3 arranged sequentially from bottom to top, and four corner points 4-1, 4-2, 4-3, and 4-4 at the top. The initial selection and correction parameters selected in this embodiment are: the elastic modulus E1 of the leg 1-1, the elastic modulus E2 of the jacket column 1-2, the elastic modulus E3 of the upper block 1-3, the elastic moduli E4-E6 of the three horizontal braces 2-1, 2-2, and 2-3, the elastic moduli E7-E9 of the three inner horizontal braces 3-1, 3-2, and 3-3, and the concentrated masses m1-m4 of the four corner points 4-1, 4-2, 4-3, and 4-4.
[0063] The elastic modulus was set to vary from -10% to 10%, and the lumped mass to vary from -20% to 20%. The Sobol sensitivity analysis method was used to calculate the contribution of each initially selected correction parameter to the structural frequency. Parameters with a contribution greater than 0.1 were selected, including E1, E2, E3, m1, m2, m3, and m4 (a total of 7 parameters). The Lading hypercube sampling method was used to sample 10,000 sets of these 7 parameters, resulting in a 10,000 × 7 parameter sample. Each set of parameters was substituted into the finite element model to calculate the acceleration response of the structure under each set of parameters, forming the response sample. The parameter sample and the response sample together constitute the model correction dataset.
[0064] The second step is to construct a variational autoencoder neural network, where both the encoder and decoder use three fully connected layers. Each set of response samples is then fed into the variational autoencoder for training. Figure 1 As shown in the upper part, by minimizing the acceleration response signal x and the regenerated signal of the acceleration response signal... The error is used for training, and the trained first latent variable z1 is extracted as the response feature. A conditionally invertible neural network is constructed, taking the parameter samples as input and the first latent variable z1 as the condition. The loss function is constructed with the objective of minimizing the negative log-likelihood of the training samples for training, as shown below. Figure 1 As shown, the standard normal distribution is selected as the second latent variable z2.
[0065] The third step involves setting seven structural parameters as target values, calculating the structural acceleration response under these parameters as the measured acceleration response, and substituting this response into the variational autoencoder. Figure 2 As shown, the first measured latent variable z1 is used to obtain the probability distribution characteristics of the true response. * Then, this is used as a conditional input into a trained conditional invertible neural network. By sampling the second latent variable z2 and performing inverse calculations, the standardized posterior distribution of the seven model parameters is obtained, such as... Figure 4 As shown in the diagram. The horizontal line represents the difference between the mean of the posterior distribution and the target value; the closer to 0, the closer to the target value. The vertical axis represents the variance of the posterior distribution; the smaller the variance, the higher the certainty of the predicted value. Figure 4 As shown, the mean of the posterior distribution of the seven parameters is used as the predicted value. The difference between the predicted value and the target value is close to 0, indicating that the overall prediction effect is very good.
[0066] The fourth step involves further pre-setting one thousand sets of structural parameters. Using the process described in the third step, the posterior distribution of each set of corrected parameters is calculated. The mean of the posterior distribution is used as the predicted parameter value. The one thousand predicted values of E1 are then plotted together with the target value, as shown below. Figure 5As shown, the slope k of the straight line fitting the predicted and target values is 0.99929, indicating a good linear relationship between the predicted and target values. Further calculation of the coefficient of determination R² between the predicted and target values yields a value of 0.99963, indicating very high accuracy of the predictions.
[0067] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A method for correcting a stochastic model of a marine engineering structure, characterized in that, include: Step 1: Obtain the initial finite element model and select m model parameters as parameters to be corrected; Step 2: Determine the range of variation of the parameter to be corrected, and sample the parameter to be corrected n times to obtain n sets of parameter samples; n sets of parameter samples are input into the initial finite element model to perform transient analysis under wave excitation. d positions in the initial finite element model are selected for observation. Each set of parameter samples yields an acceleration response signal in t×d dimensions, where t is the total time step and the total dimension of the response is n×t×d. Step 3: Use the acceleration response signal to train the variational autoencoder neural network model and the conditional invertible neural network model; Step four: Deploy d acceleration sensors on the marine engineering structure, with their locations corresponding one-to-one with the d observation locations in the initial finite element model, and monitor for a duration of t. * The acceleration response was obtained by measuring the actual acceleration response signal. The measured acceleration response signal is input into the variational autoencoder neural network model to obtain the first measured latent variable z1 representing the probability density distribution characteristics of the measured response. * ; The first measured latent variable z1 * The input is fed into the conditional invertible neural network model, the latent variable z2 is sampled, and the posterior predicted value of the parameter to be corrected is obtained through the inverse calculation of the invertible coupling block. The mean of the posterior predicted value is the corrected value of the parameter to be corrected. The correction value is input into the initial finite element model to obtain the corrected finite element model.
2. The method for correcting stochastic models of marine engineering structures according to claim 1, characterized in that, The variational autoencoder neural network model described in step three includes an encoder and a decoder. The training methods for the variational autoencoder neural network model include: The acceleration response signals of each group are input to the encoder. The encoder encodes the input signals according to the encoder parameters and outputs the encoded signals to the decoder. The decoder decodes the input signals to obtain the regenerated signals of the acceleration response signals. The error between the regenerated signals and the acceleration response signals is calculated. The encoder parameters are adjusted, and the error is minimized during training to obtain a trained variational autoencoder neural network model.
3. The method for correcting stochastic models of marine engineering structures according to claim 2, characterized in that, The training method for the conditionally invertible neural network model described in step three includes: The encoder calculates the first latent variable z1 representing the probability density distribution characteristics of the acceleration response signal based on each set of acceleration response signals and their corresponding encoder parameters, and obtains n×p-dimensional data, where p is the data length of the first latent variable z1. A conditional invertible neural network model is constructed. Each set of parameter samples is input into the conditional invertible neural network model as training samples. The first hidden variable z1 corresponding to the parameter samples is used as a condition, and the second hidden variable z2 is obtained through forward learning of the invertible coupling block. The second latent variable z2 is inversely calculated through the invertible coupling block to obtain the inverse calculation signal of the parameter sample. The loss function is constructed with the goal of maximizing the log-likelihood of the training sample, and the training of the conditional invertible neural network model is completed.
4. The method for correcting stochastic models of marine engineering structures according to claim 3, characterized in that, The process of constructing a conditionally invertible neural network model also includes the step of initializing the second hidden variable z2, by selecting a standard normal distribution as the initial second hidden variable z2.
5. The method for correcting stochastic models of marine engineering structures according to claim 1, characterized in that, The initial finite element model mentioned in step one is obtained based on the design parameters of the marine engineering structure.
6. The method for correcting stochastic models of marine engineering structures according to claim 1, characterized in that, The methods for determining the parameters to be corrected in step one include: Select some model parameters as initial correction parameters, and calculate the contribution of each initial correction parameter to the structural response using a sensitivity analysis algorithm; Select the top m initial correction parameters with the largest contribution as the parameters to be corrected, and determine the range of variation of the parameters to be corrected.
7. The method for correcting stochastic models of marine engineering structures according to claim 6, characterized in that, The sensitivity analysis algorithm is the Sobol method.
8. The method for correcting stochastic models of marine engineering structures according to claim 6, characterized in that, The contribution of each initially selected correction parameter to the structural response is calculated as follows: Calculate the first-order sensitivity index of each initial selection correction parameter. and overall sensitivity index : ； ； in, Indicates the parameter adjustment based on the initial selection. The variance caused by uncertainty This represents the total variance of the model output. This represents the variance caused by the uncertainty of the remaining input parameters; Select one and The initial correction parameter whose maximum value is greater than 0.1 is selected as the parameter to be corrected.
9. The method for correcting stochastic models of marine engineering structures according to any one of claims 1-8, characterized in that, In step two, the parameter to be corrected is sampled n times within the range of variation using the Latin hypercube sampling method.
10. The method for correcting stochastic models of marine engineering structures according to any one of claims 1-8, characterized in that, The wave excitation described in step two uses environmental excitation measured on marine engineering structures.