Method for detecting offshore wind turbine jacket based on physical information condition reversible network

Abstract

The present application relates to the technical field of offshore wind turbine, and particularly relates to a method for detecting offshore wind turbine jacket based on physical information reversible network. The method comprises the following steps: obtaining damage factor training data of a jacket structure model; constructing a conditional reversible neural network model for structure damage identification; training the conditional reversible neural network model based on physical information consistency; applying the trained conditional reversible neural network model to damage detection of a real jacket structure, and outputting damage position and damage degree of the jacket structure. The method can quickly and accurately detect the damage position and damage degree of the jacket, and timely repair the jacket according to the damage result, thereby effectively ensuring stable operation of the offshore wind turbine in the whole life cycle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of offshore wind turbines, in particular to a method for detecting the jacket of an offshore wind turbine based on a physically consistent invertible network. BACKGROUND

[0002] The jacket structure has stronger rigidity and stability, and has become the main support structure of the current offshore wind turbine equipment in the medium and deep water area. The safety of the structure determines the continuous operation ability of the offshore wind turbine equipment. However, under the action of the marine environment load, corrosion, cracks and other damages will gradually occur at the local structure position, which seriously weakens the rigidity of the overall structure. Therefore, the research on the damage detection method of the jacket support structure is the key to ensure the stable operation of the offshore wind turbine in the whole life cycle.

[0003] In the current research on the damage detection of the support structure of the offshore wind turbine, the sensitivity-based and heuristic-based deterministic methods are widely used. However, the application of these methods inevitably introduces additional errors, and the deterministic method cannot effectively evaluate the uncertainty of the model update result. The Bayesian inference method constructs the posterior probability distribution of the damage factor based on the real response state monitoring information of the structure, and converts the damage detection problem into a probability reasoning problem, so as to realize the quantitative evaluation of the uncertainty of the model damage state. However, since the posterior probability distribution function needs to be integrated in high dimension, the Markov chain-Monte Carlo sampling, Gibbs sampling method and other methods are usually used to obtain the posterior distribution, but these methods all face the problem of low computational efficiency caused by a large number of samples, and the modal characteristic parameters of the support structure are usually obtained by calling finite element simulation in the sampling process, which further increases the computational burden.

[0004] As a bidirectional inference model, the conditional reversible neural network can directly sample according to the latent distribution of the real observed data of the structure after the forward training of the model, and then generate sample data through the reverse inference process to obtain the posterior probability distribution of the damage factor. This method can effectively solve the problem of a large number of iterative sampling in the Bayesian inference method. However, the current research on the training process of the conditional reversible neural network model only focuses on the difference between the probability distributions, but ignores the consistency of the physical information of the structure.

[0005] Therefore, for the damage detection problem of the jacket support structure of the offshore wind turbine, there is an urgent need for a conditional reversible neural network model construction and training method that can guarantee the physical consistency of the structure, so as to realize the rapid identification and analysis of the structural damage. SUMMARY

[0006] The purpose of this invention is to overcome the above-mentioned defects in the existing technology and propose an offshore wind turbine jacket inspection method based on physical information condition reversible network. This method can quickly and accurately detect the damage location and degree of the jacket and repair the jacket in a timely manner according to the damage results, effectively ensuring the stable operation of the offshore wind turbine throughout its entire life cycle.

[0007] The technical solution of this invention is: a method for detecting offshore wind turbine jackets based on physical information condition reversible networks, comprising the following steps:

[0008] S1. Obtain training data on damage factors for the duct stent structure model;

[0009] S2. Construct a conditionally invertible neural network model for structural damage identification;

[0010] S3. Based on the consistency of physical information, train the conditionally reversible neural network model;

[0011] S4. Apply the trained conditional reversible neural network model to damage detection of real duct stent structures, and output the damage location and degree of the duct stent structure.

[0012] In step S1 of the present invention, a fully automated modal analysis calculation framework for the duct stent structure model is established based on Python, and the damage factor training data of the duct stent structure model is obtained through the fully automated modal analysis calculation framework.

[0013] The damage factor training data includes the damage factor and the modal response feature parameters corresponding to the damage factor. The modal response feature parameters include the natural frequencies and incomplete mode shape vectors under the current damage state.

[0014] The establishment of a fully automated computational framework for modal analysis includes the following steps:

[0015] S1.1 Customize the local element coordinate information, boundary condition constraints, material properties of the structure, pipe node size information and element type of the jacket structure, and simulate the damage state by reducing the elastic modulus, multiplying the elastic modulus of a specific element by a damage factor.

[0016] S1.2. Iterate through and calculate the stiffness matrix and uniform mass matrix of all local elements. Then, based on the corresponding transformation relationship between the local coordinate system and the global coordinate system of each element, complete the assembly of the global stiffness matrix and global mass matrix of the jacket structure model.

[0017] S1.3 Apply boundary constraints to the jacket structure. Based on the division of fixed degrees of freedom and free degrees of freedom, extract the sub-matrices corresponding to the free degrees of freedom from the overall stiffness matrix and the overall mass matrix, and obtain the modal response characteristic parameters by using the eigenvalue iterative solution method.

[0018] By generating random combinations of damage factors at local units, a training dataset for a conditionally invertible neural network is obtained, which takes the damage factors at local units as input and modal response feature parameters as output.

[0019] The implementation process of step S2 is as follows:

[0020] S2.1 Data preloading and preprocessing:

[0021] Read the training data of the conditionally invertible neural network model, verify the consistency of the input dimension, and use a normalizer to perform scale normalization on the damage factor and modal response feature parameters.

[0022] S2.2 Design of reversible coupling layers in reversible neural networks:

[0023] random variable Follows probability distribution function The function is transformed into another variable x through an invertible and differentiable transformation function f, and the probability density function after a single transformation is... Represented as:

[0024] ,

[0025] in, This is the inverse transformation process; Let be the determinant of the matrix; This is the Jacobian matrix in the inverse transformation;

[0026] When there are multiple transformation steps, the total log probability is expressed as the sum of the contribution values ​​at each step:

[0027] ,

[0028] in, Points based on the fundamental distribution are subjected to a series of inverse transformation processes. Transform into the final point ;

[0029] Will Dimensional input variables Divided into having Vie and Vie If a part of it is changed while the other part remains unchanged, the forward change form is:

[0030] ,

[0031] in, A scaling function parameterized by a fully connected neural network; A translation function parameterized by a fully connected neural network; This is represented as element-wise multiplication; For input vectors Subcomponents that have undergone affine transformation after being split; For input vectors Subcomponents that have not undergone affine transformation after being split;

[0032] The Jacobian matrix remains unchanged during forward propagation, thus ensuring that the Jacobian matrix remains constant. It has a block triangular structure:

[0033] ,

[0034] in, It is the identity matrix; It is a diagonal matrix with diagonal elements as follows: ;

[0035] After logarithmic transformation, we can obtain a formula to simplify the computational complexity of the Jacobian matrix:

[0036] ,

[0037] By learning the parameter values ​​of the scaling and translation functions during model training, these values ​​are then applied to the inverse inference process. and To achieve the original input variables The recovery process:

[0038] ,

[0039] S2.3 Conditional Network Construction:

[0040] A conditional network is integrated into the affine coupling layer unit. By inputting the modal response feature parameters output in step S1 into this conditional network, the corresponding conditional variables are automatically extracted. Bidirectional transformation used in conditional affine coupled layer units, fusing condition variables The forward transformation process is as follows:

[0041] ,

[0042] Fusion condition variables The reverse transformation process is as follows:

[0043] ,

[0044] in, These are the feature factors of the training data obtained in step S1 during the forward training process. These are characteristic factors of the modal response data of the actual monitored jacket structure during backward inference;

[0045] The determinant of the Jacobian matrix required for the post-transformation of the fused conditional network is expressed as:

[0046] .

[0047] The implementation process of step S3 is as follows:

[0048] The parameters to be trained in a conditionally invertible neural network include a set of parameters used to obtain the scaling and translation functions. And the conditional network parameter set used to obtain the modal response feature parameters obtained in the first step. ;

[0049] To obtain parameters The optimal value, the training objective of a conditionally invertible neural network is to make the conditional distribution generated by the model... With training data distribution To maintain consistency, KL divergence is used to quantify the difference between the two distributions, and the loss function is defined as:

[0050] ,

[0051] in, For the first The latent variables of each sample are defined by a standard Gaussian distribution, i.e.:

[0052] ;

[0053] Let be the Jacobian determinant of the m-th training sample; These are the original modal response feature parameters without a multilayer perceptron model; These are the corresponding feature factors after processing by the multilayer perceptron model; Let m be the latent variable of the m-th training sample;

[0054] The loss function based on physical information consistency includes the nonnegativity penalty function L. neg Frequency error penalty function L freq and mode shape MAC value penalty function L mac The nonnegativity penalty function is:

[0055] ,

[0056] The frequency error penalty function is:

[0057] ,

[0058] The mode shape MAC value penalty function is:

[0059] ,

[0060] in, This indicates that the average value of each batch of training data is taken during the model training process; The damage factor to be corrected is the duct structure model predicted by the conditionally invertible neural network model. The natural frequencies of the modified jacket structure model; The mode shapes of the corrected jacket structure model; The original intrinsic frequency corresponding to the damage factor before correction; The original mode shape corresponding to the damage factor before correction;

[0061] By adding dynamic weight coefficients to the physical information loss function, the training loss function of the conditionally invertible neural network model that integrates physical information consistency becomes:

[0062] ,

[0063] in, The dynamic weighting coefficients of the nonnegativity penalty function; For the frequency error penalty function, the dynamic weighting coefficients are used. The dynamic weighting coefficients of the mode shape MAC value penalty function;

[0064] The formula for the dynamic weighting coefficient is expressed as:

[0065] ,

[0066] in, include , , ; These are the initial weighting coefficients for the three types of dynamic weighting coefficients mentioned above; This is the scaling factor for the weights; This refers to the training epoch of the current model; This refers to the total number of training rounds.

[0067] The specific steps of step S4 are as follows:

[0068] S4.1. Accelerometers are placed at the pipe node locations of the actual duct stent structure to collect the actual response data of the duct stent under external environmental excitation, thereby obtaining the original duct stent acceleration response data under the current damage state.

[0069] S4.2 Utilize the random subspace identification algorithm to analyze the original jacket acceleration response data and obtain the true modal response data of the jacket structure, including the natural frequencies and mode shapes of the jacket structure;

[0070] S4.3 Input the real modal response data under the current state into the trained conditional reversible neural network model, and output the identification results of the damage location and damage degree of the duct stent structure to complete the damage detection process of the duct stent structure.

[0071] The beneficial effects of this invention are:

[0072] (1) This application constructs modal response feature parameters for calculating the duct support structure under different damage factors, which can realize fully automatic and fast acquisition of training datasets for conditional invertible neural networks;

[0073] (2) This application constructs a conditional invertible neural network model for identifying damage to duct support structures. The conditional affine coupling layer unit in this model can realize a bidirectional mapping relationship between input data and latent variables, and improves the expressive power for complex posterior probability distributions by cascading multiple units.

[0074] (3) This application proposes a model training loss function based on physical information consistency, which balances the dynamic balance between data fitting and physical constraints during model training by defining dynamic weight coefficients;

[0075] (4) The method proposed in this application can obtain the posterior distribution of the damage location and damage degree of the current state of the duct stent structure based on the original duct stent acceleration response data under the current damage state, and can quickly and accurately output the identification results of the damage location and damage degree of the duct stent structure, thus completing the damage detection process of the duct stent structure.

[0076] The method proposed in this application can detect damage to the jacket structure in a timely manner, and repair the jacket in a timely manner based on the location and extent of damage obtained by the method, effectively ensuring the stable operation of the offshore wind turbine throughout its entire life cycle. Attached Figure Description

[0077] Figure 1 This is a flowchart of the forward process of an affine coupling layer unit in a conditionally invertible neural network.

[0078] Figure 2This is a flowchart of the backward process of an affine coupling layer unit in a conditionally reversible neural network.

[0079] Figure 3 This is a unidirectional process flow diagram of the conditional affine coupling layer unit;

[0080] Figure 4 This is a comparison chart of the calculated modal natural frequencies of the ductwork structure using ABAQUS simulation software and the Python calculation program constructed in this application.

[0081] Figure 5 It is a conditionally reversible neural network model targeting damage factors. The prediction results;

[0082] Figure 6 It is a conditionally reversible neural network model targeting damage factors. The prediction results;

[0083] Figure 7 It is a conditionally reversible neural network model targeting damage factors. The prediction results;

[0084] Figure 8 This is a graph showing the effect of noise on a conditionally invertible neural network model.

[0085] Figure 9 This is a performance comparison chart between this application and the existing methods below. Detailed Implementation

[0086] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0087] Specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many ways other than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0088] The present invention describes a method for detecting offshore wind turbine jackets based on physical information condition reversible networks, which includes the following steps.

[0089] The first step, to achieve rapid calculation of the modal response characteristics of the jacket structure under different damage states, is to establish a fully automated modal analysis calculation framework for the jacket structure model based on Python, and to obtain the damage factor training data of the jacket structure model. The damage factor training data includes damage factors and modal response characteristic parameters, which include the natural frequencies and incomplete mode shape vectors of the structure under the current damage state.

[0090] The establishment of a fully automated computational framework for modal analysis of the jacket structure model mainly includes the following steps.

[0091] First, the local element coordinate information, boundary condition constraints, material properties, pipe node size information, and element type of the jacket structure are defined. In this embodiment, Euler-Bernoulli beam elements are used to define the local elements. The damage state is simulated by reducing the elastic modulus. The elastic modulus of a specific element is multiplied by a damage factor. The damage factor ranges from 0 to 1. The smaller the damage factor, the more severe the damage to the element.

[0092] Second, the stiffness matrix and uniform mass matrix of all local elements are calculated. Then, based on the transformation relationship between the local coordinate system and the global coordinate system of each element, the global stiffness matrix and global mass matrix of the jacket structure model are assembled.

[0093] Third, boundary constraints are applied to the jacket structure. Based on the division of fixed and free degrees of freedom, submatrices corresponding to the free degrees of freedom are extracted from the global stiffness matrix and global mass matrix. The generalized eigenvalue problem is transformed into a standard eigenvalue problem, and the modal response characteristic parameters are obtained by using an eigenvalue iterative solution method.

[0094] The eigenvalue iterative solution method involves setting an initial vector, defining the relative error between modal eigenvalues ​​and mode shape vectors as a convergence criterion, and gradually optimizing the eigenvalues ​​and mode shape vectors through multiple rounds of training iterations. The iteration terminates when the error satisfies the convergence condition.

[0095] By generating random combinations of damage factors at local units, a training dataset for a conditionally invertible neural network is obtained, which takes the damage factors at local units as input and modal response feature parameters as output.

[0096] The second step is to construct a conditionally reversible neural network model for structural damage identification.

[0097] Read the damage factor training data obtained in the first step, verify the consistency of the input dimension, and use a normalizer to perform scale normalization on the input damage factors and the output modal response feature parameters. Also, reduce the dimensionality of high-dimensional input parameters or modal response feature parameters to reduce the complexity of model computation and training.

[0098] The core of a reversible neural network lies in achieving a bidirectional mapping between a Gaussian normal distribution and a complex target distribution through a series of transformations. This process is based on the variable substitution theorem in probability theory. Consider random variables. Follows probability distribution function This function can be transformed into another variable x through an invertible and differentiable transformation function f, and the probability density function after a single transformation is... Represented as:

[0099] ,

[0100] in, This is the inverse transformation process; Let be the determinant of the matrix; It is the Jacobian matrix in the inverse transformation, and also an important parameter for the volume change between the two during the quantization transformation.

[0101] To make the calculated values ​​more stable and easier to optimize using deep learning frameworks, the result is changed to a logarithmic probability form:

[0102] ,

[0103] When there are multiple transformation steps, the total log probability can be expressed as the sum of the contribution values ​​at each step:

[0104] ,

[0105] in, Points based on the fundamental distribution are subjected to a series of inverse transformation processes. Transform into the final point .

[0106] To ensure efficient computation while maintaining the model's expressive power, a special structural unit for the affine coupling layer was designed. Dimensional input variables Divided into having Vie and Vie Then, only a portion of it is changed, while the other portion remains unchanged, such as... Figure 1 and Figure 2 As shown. Therefore, its forward transformation form is:

[0107] ,

[0108] in, A scaling function parameterized by a fully connected neural network; A translation function parameterized by a fully connected neural network; This is represented as element-wise multiplication; For input vectors Subcomponents that have undergone affine transformation after being split; For input vectors Subcomponents that have not undergone affine transformation after splitting. This design makes... The Jacobian matrix remains unchanged during forward propagation, thus ensuring that the Jacobian matrix remains constant. It has a block triangular structure, that is:

[0109] ,

[0110] in, It is the identity matrix; It is a diagonal matrix with diagonal elements as follows: Therefore, the determinant of the Jacobian matrix can be calculated using the following formula:

[0111] ,

[0112] After a logarithmic transformation, we can obtain the formula for simplifying the computational complexity of the Jacobian matrix:

[0113] .

[0114] By learning the parameter values ​​of the scaling and translation functions during model training, these values ​​can then be applied to the inverse reasoning process, i.e., through... and To achieve the original input variables The recovery process, namely:

[0115] ,

[0116] Traditional reversible neural networks can only obtain edge distributions. To obtain the modal response characteristic parameters in the first step Conditional distribution Incorporating conditional networks into affine coupling layers, such as... Figure 3 As shown.

[0117] This conditional network is constructed using a multilayer perceptron model. By inputting the modal response feature parameters obtained in the first step into this conditional network, the corresponding conditional variables are automatically extracted. Used for bidirectional transformations in conditional affine coupled layer units. Incorporates condition variables. The forward transformation process in the affine coupling layer is as follows:

[0118] ,

[0119] Incorporating condition variables The inverse transformation process in the affine coupling layer is as follows:

[0120] ,

[0121] in, These are the feature factors of the training data obtained in the first step of the forward training process. These are characteristic factors of the modal response data of the actual monitored jacket structure during backward inference.

[0122] The determinant of the Jacobian matrix required for the post-transformation of the fused conditional network can be expressed as:

[0123] .

[0124] To enhance the ability of conditionally invertible neural networks to represent complex posterior probability distributions, multiple conditional affine coupling layers are typically stacked. To prevent the coupling layers from learning only local feature dependencies, feature permutation operations are used between adjacent coupling layers to shuffle the feature order.

[0125] The third step is to train the conditionally reversible neural network model constructed in the second step based on the consistency of physical information.

[0126] The training parameters of a conditionally invertible neural network model typically consist of two parts, including the parameter set used to obtain the scaling and translation functions in the model. And the conditional network parameter set used to obtain the modal response feature parameters obtained in the first step. Therefore, in order to obtain the parameters The optimal value is typically the training objective of a conditionally invertible neural network model, which aims to generate a conditional distribution that satisfies the conditional distribution. With training data distribution The two distributions should remain consistent. Therefore, KL divergence is used to quantify the difference between the two distributions, and the loss function is typically defined as:

[0127] ,

[0128] in, For training data distribution Expectations; The log probability of the true distribution of the training data; Estimate the log probability of the posterior distribution for the training data.

[0129] consider It relates only to the training data itself, and not to the parameter set to be trained in the conditionally invertible neural network model. Since it is irrelevant, it can be considered a constant, and the above equation can be viewed as maximizing... Considering the simulated modal response data obtained from training... It is usually generated under different conditions and needs to be... Since the distribution takes the expectation, the optimization objective function can be expressed as:

[0130] ,

[0131] Training data exists training samples At this point, the expected value of the data can be approximated by sample evaluation. Therefore, the loss function considering distributional consistency is ultimately:

[0132] ,

[0133] in, It is the first The latent variables of each training sample Defined as a standard Gaussian distribution of the latent variable, i.e.:

[0134] ;

[0135] Let be the Jacobian determinant of the m-th sample; These are the original modal response feature parameters without a multilayer perceptron model; These are the corresponding feature factors after processing by the multilayer perceptron model.

[0136] To ensure that the posterior distribution of damage factors obtained by the conditionally reversible neural network model conforms to the laws of structural dynamics, this application introduces a loss function based on the consistency of physical information into the training process of the conditionally reversible neural network model, thereby guiding the model learning process to obtain a solution consistent with the physical modal response of the ductwork.

[0137] The loss function based on physical information consistency includes three physical information loss functions, namely, non-negativity penalty functions. Frequency error penalty function Mode shape MAC value penalty function ,in, This is used to ensure that the model correction parameters predicted by the conditionally invertible neural network model are not negative. and This is used to ensure that the predicted modal response of the jacket is consistent with the actual state.

[0138] The nonnegativity penalty function is:

[0139] ,

[0140] The frequency error penalty function is:

[0141] ,

[0142] The mode shape MAC value penalty function is:

[0143] ,

[0144] in, This indicates that the average value of each batch of training data is taken during the model training process; The damage factor to be corrected is the duct structure model predicted by the conditionally invertible neural network model. The natural frequencies of the corrected jacket structure model are The corrected natural frequency is calculated and output after using the eigenvalue iterative solution method. The mode shapes of the corrected jacket structure model are The corrected mode shape is calculated and output after using the eigenvalue iterative solution method; The original intrinsic frequency corresponding to the damage factor before correction; This is the original mode shape corresponding to the damage factor before correction.

[0145] To prevent excessively strong initial physical information constraints from affecting model convergence during the early stages of training a conditionally invertible neural network model, dynamic weight coefficients are added to the physical information loss function. Therefore, the training loss function for a conditionally invertible neural network model incorporating physical information consistency is:

[0146] ,

[0147] in, The dynamic weighting coefficients of the nonnegativity penalty function; For the frequency error penalty function, the dynamic weighting coefficients are used. The dynamic weighting coefficients are the penalty function for the mode shape MAC value.

[0148] The dynamic weighting coefficients of the three physical information loss functions mentioned above can all be expressed by the following formula:

[0149] ,

[0150] in, include , , ; These are the initial weighting coefficients for the three types of dynamic weighting coefficients mentioned above; This is the scaling factor for the weights; This refers to the training epoch of the current model; This refers to the total number of training rounds.

[0151] The fourth step involves applying the conditionally reversible neural network model trained in the third step to the damage detection process of a real duct stent structure.

[0152] The trained conditional reversible neural network model can be applied to the damage detection process of real duct stents.

[0153] First, acceleration sensors need to be installed at the pipe node locations of the actual jacket structure to collect the actual response data of the jacket under external environmental excitation, thereby obtaining the original jacket acceleration response data under the current damage state.

[0154] Subsequently, the random subspace identification algorithm was used to analyze the original jacket acceleration response data to obtain the true modal response data of the jacket structure, including the natural frequencies and mode shapes of the jacket structure.

[0155] The random subspace identification algorithm constructs a state-space model, organizes acceleration response data into a Hankel matrix, extracts system features using singular value decomposition, and finally obtains the true natural frequencies and mode shapes of the jacket structure through eigenvalue decomposition.

[0156] The actual modal response data of the current duct stent structure is used as the input data of the trained conditional reversible neural network model. Random sampling is performed in the Gaussian normal distribution, and the data is combined with the actual modal response data and input into the trained conditional reversible neural network model to obtain the posterior distribution of the damage location and damage degree of the duct stent structure in the current state. The output is the identification result of the damage location and damage degree of the duct stent structure, thus completing the damage detection process of the duct stent structure.

[0157] In verifying the method described in this application, the modal analysis results of the ductwork system were first compared between the ABAQUS simulation software and the Python computational program constructed in this application. Figure 4 As shown in the figure, the calculation results of the first six natural frequencies show that the maximum error is 5.54%, occurring in the second mode, while the errors of the remaining modes are all less than 2%, thus meeting the requirements for engineering calculations. Furthermore, the differences between the mode shapes obtained by the two methods were compared, and the results show good consistency between them.

[0158] During the verification process, units 18, 32, and 35 were randomly defined as damage locations, and damage factors were... , , The prior distributions of all are uniform distributions U(0.2, 1), and the target distributions are respectively... , , The x-axis and y-axis displacement degrees of freedom of seven nodes are selected as the input of the conditionally invertible neural network model. Figures 5 to 7 The results show the posterior probability distributions of the obtained damage factors. From the results, it can be seen that the posterior probability distributions of each damage factor are as follows: , , This demonstrates that the trained conditionally reversible neural network model can effectively identify the degree of duct stent damage and analyze uncertainties.

[0159] Subsequently, to verify the robustness and applicability of the conditionally invertible neural network model in updating its parameters under noise interference, white Gaussian noise was added to the modal response features obtained from the numerical simulation, with noise amplitudes of 4%, 8%, 10%, 15%, 20%, 25%, and 30%, respectively. The results are as follows: Figure 8 As shown, it can be seen that as the noise level increases, although the sampled model parameters still follow a Gaussian distribution, the standard deviation of the posterior probability distribution gradually increases. This indicates that the increase in noise level can lead to misjudgment of the uncertainty of the structural parameters to be updated. However, relatively speaking, although the error of the mean of the posterior distribution of the model parameters also gradually increases, it will exhibit dynamic fluctuations around the true value.

[0160] Finally, the method described in this application is compared with the SQP damage identification method based on modal compliance and the MCMC method based on Bayesian inference, as follows: Figure 9 As shown, although the SQP method has the fastest computation speed, noise has the greatest impact on it. The MCMC method has better model update performance and noise resistance, but its computation time is also the longest. Although the pre-training process of the conditionally invertible neural network model proposed in this application takes too long, once the model is trained, the time required for reverse inference of its parameters is almost negligible.

[0161] The above provides a detailed description of the offshore wind turbine structural damage detection method based on physically reversible networks. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of these embodiments are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the scope of protection of the claims. The above description of the disclosed embodiments enables those skilled in the art to implement or use this invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, this invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for inspecting offshore wind turbine jackets based on physically reversible networks, characterized in that, Includes the following steps: S1. Obtain training data on damage factors for the duct stent structure model; S2. Construct a conditionally invertible neural network model for structural damage identification; S3. Based on the consistency of physical information, train the conditionally reversible neural network model; S4. Apply the trained conditional reversible neural network model to the damage detection of a real duct stent structure, and output the damage location and damage extent of the duct stent structure. The implementation process of step S2 is as follows: S2.1 Data preloading and preprocessing: Read the training data of the conditionally invertible neural network model, verify the consistency of the input dimension, and use a normalizer to perform scale normalization on the damage factor and modal response feature parameters. S2.2 Design of reversible coupling layers in reversible neural networks: random variable Follows probability distribution function The function is transformed into another variable x through an invertible and differentiable transformation function f, and the probability density function after a single transformation is... Represented as: , in, This is the inverse transformation process; Let be the determinant of the matrix; This is the Jacobian matrix in the inverse transformation; When there are multiple transformation steps, the total log probability is expressed as the sum of the contribution values ​​at each step: ， in, Points based on the fundamental distribution are subjected to a series of inverse transformation processes. Transform into the final point ; Will Dimensional input variables Divided into having Vie and Vie If a part of it is changed while the other part remains unchanged, the forward change form is: , in, A scaling function parameterized by a fully connected neural network; A translation function parameterized by a fully connected neural network; This is represented as element-wise multiplication; For input vectors Subcomponents that have undergone affine transformation after being split; For input vectors Subcomponents that have not undergone affine transformation after being split; The Jacobian matrix remains unchanged during forward propagation, thus ensuring that the Jacobian matrix remains constant. It has a block triangular structure: ， in, It is the identity matrix; It is a diagonal matrix with diagonal elements as follows: ; After logarithmic transformation, we can obtain a formula to simplify the computational complexity of the Jacobian matrix: ， By learning the parameter values ​​of the scaling and translation functions during model training, these values ​​are then applied to the inverse inference process. and To achieve the original input variables The recovery process: ， S2.3 Conditional Network Construction: A conditional network is integrated into the affine coupling layer unit. By inputting the modal response feature parameters output from step S1 into this conditional network, the corresponding conditional variables are automatically extracted. Bidirectional transformation used in conditional affine coupled layer units, fusing condition variables The forward transformation process is as follows: , Fusion condition variables The reverse transformation process is as follows: , in, These are the feature factors of the training data obtained in step S1 during the forward training process. These are characteristic factors of the modal response data of the actual monitored jacket structure during backward inference; The determinant of the Jacobian matrix required for the post-transformation of the fused conditional network is expressed as: 。 2. The method for detecting offshore wind turbine jackets based on physically reversible networks according to claim 1, characterized in that, In step S1, a fully automated modal analysis calculation framework for the duct stent structure model is established based on Python, and the damage factor training data of the duct stent structure model is obtained through the fully automated modal analysis calculation framework. The damage factor training data includes the damage factor and the modal response feature parameters corresponding to the damage factor. The modal response feature parameters include the natural frequencies and incomplete mode shape vectors under the current damage state.

3. The method for detecting offshore wind turbine jackets based on physically reversible networks according to claim 2, characterized in that, The establishment of a fully automated computational framework for modal analysis includes the following steps: S1.1 Customize the local element coordinate information, boundary condition constraints, material properties of the structure, pipe node size information and element type of the jacket structure, and simulate the damage state by reducing the elastic modulus, multiplying the elastic modulus of a specific element by a damage factor. S1.

2. Iterate through and calculate the stiffness matrix and uniform mass matrix of all local elements. Then, based on the corresponding transformation relationship between the local coordinate system and the global coordinate system of each element, complete the assembly of the global stiffness matrix and global mass matrix of the jacket structure model. S1.3 Apply boundary constraints to the jacket structure. Based on the division of fixed degrees of freedom and free degrees of freedom, extract the sub-matrices corresponding to the free degrees of freedom from the overall stiffness matrix and the overall mass matrix, and obtain the modal response characteristic parameters by using the eigenvalue iterative solution method. By generating random combinations of damage factors at local units, a training dataset for a conditionally invertible neural network is obtained, which takes the damage factors at local units as input and modal response feature parameters as output.

4. The method for detecting offshore wind turbine jackets based on physically reversible networks according to claim 1, characterized in that, The implementation process of step S3 is as follows: The parameters to be trained in a conditionally invertible neural network include a set of parameters used to obtain the scaling and translation functions. And the conditional network parameter set used to obtain the modal response feature parameters obtained in the first step. ; To obtain parameters The optimal value, the training objective of a conditionally invertible neural network is to make the conditional distribution generated by the model... With training data distribution To maintain consistency, KL divergence is used to quantify the difference between the two distributions, and the loss function is defined as: , in, For the first The latent variables of each sample are defined by a standard Gaussian distribution, i.e.: ； Let be the Jacobian determinant of the m-th training sample; These are the original modal response feature parameters without a multilayer perceptron model; These are the corresponding feature factors after processing by the multilayer perceptron model; Let m be the latent variable of the m-th training sample; The loss function based on physical information consistency includes the nonnegativity penalty function L. neg Frequency error penalty function L freq and mode shape MAC value penalty function L mac The nonnegativity penalty function is: , The frequency error penalty function is: , The mode shape MAC value penalty function is: , in, This indicates that the average value of each batch of training data is taken during the model training process; The damage factor to be corrected is the duct structure model predicted by the conditionally invertible neural network model. The natural frequencies of the modified jacket structure model; The mode shapes of the corrected jacket structure model; The original intrinsic frequency corresponding to the damage factor before correction; The original mode shape corresponding to the damage factor before correction; By adding dynamic weight coefficients to the physical information loss function, the training loss function of the conditionally invertible neural network model that integrates physical information consistency becomes: ， in, The dynamic weighting coefficients of the nonnegativity penalty function; For the frequency error penalty function, the dynamic weighting coefficients are used. The dynamic weighting coefficients of the mode shape MAC value penalty function; The formula for the dynamic weighting coefficient is expressed as: ， in, include , , ; These are the initial weighting coefficients for the three types of dynamic weighting coefficients mentioned above; This is the scaling factor for the weights; This refers to the training epoch of the current model; This refers to the total number of training rounds.

5. The method for detecting offshore wind turbine jackets based on physically reversible networks according to claim 1, characterized in that, The specific steps of step S4 are as follows: S4.

1. Accelerometers are placed at the pipe node locations of the actual duct stent structure to collect the actual response data of the duct stent under external environmental excitation, thereby obtaining the original duct stent acceleration response data under the current damage state. S4.2 Utilize the random subspace identification algorithm to analyze the original jacket acceleration response data and obtain the true modal response data of the jacket structure, including the natural frequencies and mode shapes of the jacket structure; S4.3 Input the real modal response data under the current state into the trained conditional reversible neural network model, and output the identification results of the damage location and damage degree of the duct stent structure to complete the damage detection process of the duct stent structure.

Images