Offshore wind turbine structure health monitoring method and system

Abstract

The application discloses a kind of offshore wind turbine structure health monitoring method and system, the method comprises: offline stage, strain snapshot matrix is constructed by finite element simulation, low-dimensional feature subspace base matrix is extracted using intrinsic orthogonal decomposition, and structure grid is mapped into graph;While carrying out Fourier feature mapping of node coordinate, after splicing with reconstructed strain characteristics, train the spectral graph neural network based on chebyshev graph convolution.Online monitoring stage, according to the single-sided sparse strain data collected in real time, mode coefficient is quickly solved and full-field strain is initially reconstructed by mask matrix and pre-stored base matrix, and then input trained network model, real-time deduce high-fidelity full-field displacement field and stress field.The application can effectively adapt to the arbitrary change of sensor layout, realize the real-time, accurate reconstruction of complex structure full-field physical response under the harsh conditions of single-sided sparse measurement, and solve the dependence problem of traditional technology on fixed point and double-sided measurement.

Description

Technical Field

[0001] This invention relates to offshore wind power structural health monitoring technology, and in particular to a method and system for offshore wind turbine structural health monitoring. Background Technology

[0002] Offshore wind power is a key area of ​​global energy development. As wind turbines extend into deeper waters, the harsh environmental loads make real-time health monitoring (SHM) of the supporting structure crucial for ensuring safe operation. The key to monitoring lies in understanding the structure's full-field response (such as strain, displacement, and stress) under complex operating conditions in real time. While existing finite element numerical simulations offer high accuracy, their computational complexity is enormous, making it difficult to meet the real-time requirements of online monitoring. Conventional data-driven methods, while fast, heavily rely on pre-fixed sensor layouts and typically require sensors to be placed on both sides. However, in practical engineering, large wind turbines usually only allow for monitoring on one side (inner wall), and sensor positions often change due to installation limitations or operational failures, rendering existing technologies unsuitable.

[0003] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a method and system for monitoring the structural health of offshore wind turbines.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for monitoring the structural health of offshore wind turbines includes the following steps:

[0007] S1. Offline Training Phase: Full-field response data of the structure under typical working conditions is obtained based on finite element simulation, and a strain snapshot matrix is ​​constructed. The strain snapshot matrix is ​​decomposed and truncated using eigenorthogonal decomposition to extract and save low-dimensional feature subspace basis matrices that characterize the main deformation modes of the structure. The finite element mesh of the structure is mapped to an undirected graph, and its Laplacian matrix is ​​calculated. Fourier feature mapping is performed on the spatial coordinates of the structural nodes to enhance the ability to capture high-frequency details, and the mapped coordinate features are concatenated with the strain features reconstructed based on the feature subspace to form enhanced node features. Using the enhanced node features as input and full-field displacement and stress as labels, a spectral neural network model is trained. The network utilizes graph convolution operations based on Chebyshev polynomials to perform feature aggregation and nonlinear mapping on the undirected graph topology.

[0008] S2. Online Monitoring Phase: A mask matrix is ​​generated based on the positions of sparse strain sensors arranged on one side of the field, and a sparse basis matrix is ​​constructed using the mask matrix and a pre-stored low-dimensional feature subspace basis matrix; sparse strain measurement data is acquired in real time, and the low-dimensional modal coefficients are quickly estimated and the full-field strain distribution is initially reconstructed by solving the linear equations based on the sparse basis matrix; the initially reconstructed full-field strain distribution and the pre-processed Fourier coordinate features are input into a trained spectral neural network model, and high-fidelity full-field displacement field and stress field are output in real time through forward inference;

[0009] S3. Result Output Stage: The real-time output of the full-field displacement field and stress field is visualized to realize real-time monitoring of the structural health status.

[0010] Further, in step S1, the low-dimensional feature subspace basis matrix is ​​extracted using intrinsic orthogonal decomposition. Specifically, this includes: performing singular value decomposition on the full-field strain snapshot matrix obtained through numerical simulation to obtain a left singular vector matrix representing the orthogonal space modes of the structure; determining the truncation order based on the energy proportion of the singular values, selecting a corresponding number of left singular vectors to construct a truncated basis matrix, which constitutes a low-dimensional linear subspace that can approximately represent the full-field strain at any time; during the online monitoring phase, the basis matrix is ​​filtered based on a mask matrix constructed according to the sparse sensor positions to obtain a sparse basis matrix, and a linear equation system is constructed based on the sparse basis matrix and the collected sparse strain measurement data, and the low-dimensional modal coefficients are obtained by solving the least squares method.

[0011] Further, in step S1, the Fourier feature mapping of the spatial coordinates of the structural nodes specifically includes: mapping the normalized spatial coordinates of the nodes into high-dimensional feature vectors through a frequency control matrix constructed using deterministic logarithmic frequency sampling; the frequency control matrix contains a set of frequency components that grow geometrically; the high-dimensional feature vectors are composed of the sine and cosine function values ​​of the normalized spatial coordinates after transformation by the frequency control matrix.

[0012] Further, in step S1, the construction of the spectral graph neural network model specifically includes: mapping the structural finite element mesh into an undirected graph containing nodes, edges, and a weighted adjacency matrix, and calculating a normalized Laplacian matrix based on the weighted adjacency matrix; approximating the spectral domain convolution kernel with Chebyshev polynomials, defining a graph convolution operation on the scaled normalized Laplacian matrix, wherein the convolution operation aggregates multi-order neighborhood information of nodes by performing linear combinations of input features with Chebyshev polynomials of different orders; stacking multiple graph convolutional layers and using nonlinear activation functions for feature extraction; and finally decoding the extracted deep features into full-field displacement and stress prediction values ​​using a node-aware multilayer perceptron.

[0013] Furthermore, the Chebyshev polynomial is efficiently calculated using a recursive formula.

[0014] Further, in step S2, a mask matrix is ​​generated based on the location of the field sensors. Specifically, a sparse matrix is ​​constructed, with the number of rows equal to the number of sensors and the number of columns equal to the total number of structural nodes. In each row of the matrix, only the element at the index position of the structural node where the sensor corresponding to that row is installed is 1, and the rest of the elements are 0.

[0015] Further, in step S2, the full-field strain distribution is initially reconstructed by solving a system of linear equations. Specifically, the mask matrix is ​​used to filter the pre-stored low-dimensional feature subspace basis matrix to obtain a sparse basis matrix corresponding to the current sensor layout. Based on the sparse basis matrix and the real-time acquired sparse strain measurement data, a linear least squares problem is constructed and solved to obtain the optimal low-dimensional modal coefficients. The optimal low-dimensional modal coefficients are linearly combined with the pre-stored low-dimensional feature subspace basis matrix to reconstruct the initial full-field strain distribution.

[0016] Further, in step S2, the strain value of each node in the initially reconstructed full-field strain distribution is concatenated with the high-dimensional coordinate feature vector obtained by Fourier feature mapping of that node to form the enhanced node feature as input to the spectral neural network model.

[0017] Furthermore, in step S3, the visualization specifically includes: mapping the predicted full-field displacement and stress data back to the three-dimensional geometric model of the structure, and displaying the spatial distribution of physical quantities through cloud map rendering technology.

[0018] A subspace graph network-based offshore wind turbine structural health monitoring system includes:

[0019] The data acquisition module is used to collect strain data transmitted back by the sparse strain sensor group deployed on one side of the inner wall of the wind turbine structure in real time.

[0020] The preprocessing and reconstruction module, connected to the data acquisition module, is used to generate a mask matrix based on the current sensor layout, and use a pre-stored low-dimensional feature subspace basis matrix extracted based on intrinsic orthogonal decomposition to map the strain data into low-dimensional modal coefficients by solving a system of linear equations, thereby initially reconstructing the full-field strain distribution.

[0021] The deep inference module, connected to the preprocessing and reconstruction module, integrates a trained spectral neural network to receive the initially reconstructed full-field strain distribution and the Fourier coordinate features of the nodes. It performs nonlinear mapping on the structural topology through graph convolution operations based on Chebyshev polynomials to deduce high-fidelity full-field displacement and stress fields in real time.

[0022] The visualization output module, connected to the deep inference module, is used to receive full-field displacement and stress field data, map them onto a three-dimensional structural model, and perform cloud map rendering and display to support real-time monitoring of structural health status and identification of dangerous areas.

[0023] An offline database is used to store the low-dimensional feature subspace basis matrix, the parameters of the trained spectral neural network, and the grid topology information of the wind turbine structure, and to provide data support for the preprocessing and reconstruction module and the deep inference module.

[0024] A computer program product includes a computer program that, when executed by a processor, implements the offshore wind turbine structural health monitoring method.

[0025] The present invention has the following beneficial effects:

[0026] This invention proposes a method and system for monitoring the structural health of offshore wind turbines, effectively solving the problem of full-field state reconstruction under conditions of non-fixed sensor layout and sparse single-sided distribution, and realizing real-time reconstruction of structural response during operation and maintenance. This invention achieves real-time high-fidelity reconstruction of the full-field state by constructing a subspace-based graph deep learning architecture and collaboratively utilizing three key technical aspects. Specifically, it utilizes intrinsic orthogonal decomposition to construct a dimensionality-reduced subspace, mapping arbitrarily distributed sparse measurement data into unified low-dimensional modal coefficients, thus overcoming the dependence on a fixed sensor layout; it introduces Fourier feature mapping technology to convert low-dimensional spatial coordinates into high-dimensional trigonometric function features, significantly enhancing the model's ability to capture high-frequency structural details (such as stress concentration areas); and it constructs a spectral graph neural network, using Chebyshev polynomials to perform convolution operations on a non-Euclidean graph topology, establishing a nonlinear mapping relationship from fused features to full-field displacement and stress field.

[0027] Through the above architecture, the advantages of this invention are specifically reflected in several aspects. The intrinsic orthogonal decomposition algorithm effectively utilizes the structural prior knowledge learned offline, transforming the complex online monitoring problem into a fast solution problem for low-dimensional linear systems. A key advantage is that when the on-site sensor layout changes or some sensors fail, only the mask matrix and the corresponding sparse basis matrix need to be updated, without retraining the feature basis or neural network model, thus perfectly adapting to scenarios with arbitrary and variable sensor layouts in practical engineering. Simultaneously, the enhanced node feature vectors generated through Fourier eigenmaps integrate the global distribution information and fine spatial location information of the strain physical field, serving as input to the spectral graph neural network, significantly improving the accuracy and convergence speed of the full-field reconstruction. The spectral graph neural network is specifically designed to process irregular, non-Euclidean complex geometric topological data such as offshore wind turbine support structures. It can deeply learn the spatial topological dependencies of the structure, achieving end-to-end accurate extrapolation from single-sided sparse input to double-sided full-field high-fidelity response.

[0028] Furthermore, the monitoring results provided by this invention are intuitive and clear. The calculated full-field displacement and stress data are mapped back to a three-dimensional structural model using a visualization module, and the magnitude of physical quantities is indicated by color using cloud maps, thereby achieving real-time and intuitive monitoring of the structural health status. The system can also automatically identify and issue warnings for dangerous areas exceeding set thresholds through numerical filtering. Visualization experimental results show that the displacement and stress fields predicted by this method are highly consistent with the results of high-fidelity finite element calculations.

[0029] In summary, this invention enables full-field state reconstruction of complex support structures of offshore wind turbines under arbitrary sensor layouts. In particular, under actual constrained working conditions where only sparse measurement data from one side can be acquired, it completes real-time high-fidelity inversion of full-field displacement and stress, thereby solving the core problem of existing monitoring technologies that heavily rely on fixed sensor layouts and are difficult to effectively capture the full-field physical response of structures.

[0030] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall process of the offshore wind turbine structural health monitoring method of the present invention.

[0032] Figure 2 This is a detailed flowchart of the structural health method according to an embodiment of the present invention.

[0033] Figure 3 This is a structural health monitoring system architecture diagram according to an embodiment of the present invention.

[0034] Figure 4 This is a cloud diagram showing the overall displacement and deformation of the semi-submersible wind turbine foundation according to an embodiment of the present invention.

[0035] Figure 5 This is a von Mises stress distribution cloud diagram of a semi-submersible wind turbine foundation according to an embodiment of the present invention. Detailed Implementation

[0036] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0037] It should be noted that when a component is referred to as "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as "connected to" another component, it can be directly connected to or indirectly connected to that other component. Furthermore, a connection can be used for fixing, coupling, or communication.

[0038] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0039] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0040] This invention aims to address the challenge of reconstructing the full-field physical state of offshore wind turbines in real time and with high fidelity under engineering conditions where sensor layouts are not fixed and only sparse measurements can be performed on one side. It proposes a method and system for monitoring the structural health of offshore wind turbines that integrates subspace dimensionality reduction, high-frequency coordinate encoding, and graph convolutional inference. By adapting to arbitrary measurement point layouts through intrinsic orthogonal decomposition, enhancing detail perception through Fourier feature mapping, and mining complex topological relationships using spectral neural networks, it can accurately and in real time infer the full-field displacement and stress distribution based solely on sparse strain data from one side, effectively overcoming the heavy reliance of traditional technologies on fixed, dual-side sensor layouts.

[0041] See Figure 1 and Figure 2 This invention provides a method for monitoring the structural health of offshore wind turbines, comprising the following steps:

[0042] Step S1. Offline Training Phase: Obtain full-field response data of the structure under typical working conditions based on finite element simulation, and construct a strain snapshot matrix; use Proper Orthogonal Decomposition (POD) to perform eigenbase decomposition and truncation on the strain snapshot matrix, extract and save the low-dimensional feature subspace basis matrix that can characterize the main deformation modes of the structure; map the finite element mesh of the structure to an undirected graph and calculate its Laplacian matrix; perform Fourier Feature Mapping (FFM) on the spatial coordinates of the structural nodes to enhance the ability to capture high-frequency details, and concatenate the mapped coordinate features with the strain features reconstructed based on the feature subspace to form enhanced node features; use the enhanced node features as input, and full-field displacement and stress as labels to train a Spectral Graph Network (SGN), which uses graph convolution operations based on Chebyshev polynomials to perform feature aggregation and nonlinear mapping on the undirected graph topology.

[0043] In some embodiments, step S1, extracting the low-dimensional feature subspace basis matrix using intrinsic orthogonal decomposition, specifically includes: performing singular value decomposition on the full-field strain snapshot matrix obtained through numerical simulation to obtain a left singular vector matrix characterizing the orthogonal space modes of the structure; determining the truncation order based on the energy proportion of the singular values, selecting a corresponding number of left singular vectors to construct a truncated basis matrix, the basis matrix forming a low-dimensional linear subspace that can approximately characterize the full-field strain at any time; during the online monitoring phase, constructing a mask matrix based on the sparse sensor positions to filter the basis matrix, obtaining a sparse basis matrix, and constructing a linear equation system based on the sparse basis matrix and the collected sparse strain measurement data, and solving it using the least squares method to obtain the low-dimensional modal coefficients.

[0044] In some embodiments, step S1, which involves performing Fourier feature mapping on the spatial coordinates of the structural nodes, specifically includes: mapping the normalized spatial coordinates of the nodes into high-dimensional feature vectors through a frequency control matrix constructed using deterministic logarithmic frequency sampling; the frequency control matrix contains a set of frequency components that grow geometrically; and the high-dimensional feature vectors are composed of a combination of sine and cosine function values ​​of the normalized spatial coordinates after transformation by the frequency control matrix.

[0045] In some embodiments, step S1, the construction of the spectral graph neural network model specifically includes: mapping the structural finite element mesh into an undirected graph containing nodes, edges, and a weighted adjacency matrix, and calculating a normalized Laplacian matrix based on the weighted adjacency matrix; approximating the spectral domain convolution kernel with Chebyshev polynomials, defining a graph convolution operation on the scaled normalized Laplacian matrix, wherein the convolution operation aggregates multi-order neighborhood information of nodes by performing linear combinations of input features with Chebyshev polynomials of different orders; stacking multiple graph convolutional layers and using nonlinear activation functions for feature extraction; and finally decoding the extracted deep features into full-field displacement and stress prediction values ​​using a node-aware multilayer perceptron. Preferably, the Chebyshev polynomials can be efficiently calculated using recursive formulas.

[0046] Step S2. Online Monitoring Stage: Generate a mask matrix based on the location of the sparse strain sensors arranged on one side of the field, and construct a sparse basis matrix using the mask matrix and the pre-stored low-dimensional feature subspace basis matrix; collect sparse strain measurement data in real time, quickly estimate the low-dimensional modal coefficients and preliminarily reconstruct the full-field strain distribution by solving the linear equation system based on the sparse basis matrix; input the preliminarily reconstructed full-field strain distribution and the preprocessed Fourier coordinate features into the trained spectral neural network model, and output high-fidelity full-field displacement field and stress field in real time through forward inference.

[0047] In some embodiments, in step S2, generating a mask matrix based on the location of the field sensors specifically involves: constructing a sparse matrix with the number of rows equal to the number of sensors and the number of columns equal to the total number of structural nodes. In each row of the matrix, only the element at the index position of the structural node where the sensor corresponding to that row is installed is 1, and the rest of the elements are 0.

[0048] In some embodiments, in step S2, the full-field strain distribution is initially reconstructed by solving a system of linear equations. Specifically, the mask matrix is ​​used to filter the pre-stored low-dimensional feature subspace basis matrix to obtain a sparse basis matrix corresponding to the current sensor layout; based on the sparse basis matrix and the real-time acquired sparse strain measurement data, a linear least squares problem is constructed and solved to obtain the optimal low-dimensional modal coefficients; the optimal low-dimensional modal coefficients are linearly combined with the pre-stored low-dimensional feature subspace basis matrix to reconstruct the initial full-field strain distribution.

[0049] In some embodiments, in step S2, the strain value of each node in the initially reconstructed full-field strain distribution is concatenated with the high-dimensional coordinate feature vector obtained by Fourier feature mapping of the node to form an enhanced node feature that serves as input to the spectral neural network model.

[0050] Step S3. Result Output Stage: The real-time output of the full-field displacement field and stress field is visualized to realize real-time monitoring of the structural health status.

[0051] In some embodiments, step S3, the visualization display specifically includes: mapping the predicted full-field displacement and stress data back to the three-dimensional geometric model of the structure, and displaying the spatial distribution of physical quantities through cloud map rendering technology.

[0052] See Figure 3 This invention also provides a structural health monitoring system for offshore wind turbines based on subspace graph networks, including a data acquisition module, a preprocessing and reconstruction module, a deep inference module, a visualization output module, and an offline database. Specifically, the data acquisition module is used to collect strain data transmitted in real time from a sparse strain sensor array deployed on the inner wall of one side of the wind turbine structure. The preprocessing and reconstruction module, connected to the data acquisition module, is used to generate a mask matrix based on the current sensor layout and, using a pre-stored low-dimensional feature subspace basis matrix extracted based on intrinsic orthogonal decomposition, map the strain data into low-dimensional modal coefficients by solving a system of linear equations, thereby initially reconstructing the full-field strain distribution. The deep inference module, connected to the preprocessing and reconstruction module, integrates a trained spectral graph neural network to receive the initially reconstructed full-field strain distribution and the Fourier transforms of the nodes. The coordinate features are nonlinearly mapped onto the structural topology using graph convolution operations based on Chebyshev polynomials, enabling real-time derivation of high-fidelity full-field displacement and stress fields. The visualization output module, connected to the deep inference module, receives the full-field displacement and stress field data, maps them to a 3D structural model, and renders and displays cloud maps to support real-time monitoring of structural health and identification of hazardous areas. The offline database stores the low-dimensional feature subspace basis matrix, the parameters of the trained spectral neural network, and the grid topology information of the wind turbine structure, providing data support for the preprocessing and reconstruction module and the deep inference module.

[0053] The proposed method and system for monitoring the structural health of offshore wind turbines addresses the core challenges of insufficient real-time performance in existing finite element simulations and the difficulty of adapting traditional data-driven methods to the single-sided, sparse, and variable sensor layouts required in real-world engineering projects. This invention innovatively integrates intrinsic orthogonal decomposition, Fourier eigenmaps, and spectral neural networks to construct a highly efficient subspace graph deep learning architecture. This architecture maps arbitrary sparse measurement point data to a unified low-dimensional subspace through intrinsic orthogonal decomposition, achieving strong adaptability to sensor layout changes and rapid online solution. Fourier eigenmaps enhance the ability to capture high-frequency structural details (such as stress concentrations). Furthermore, spectral neural networks directly perform convolutional inference on the non-Euclidean topology of the structure, enabling end-to-end real-time inversion from single-sided sparse strain input to high-fidelity full-field displacement and stress responses. This significantly improves computational efficiency while maintaining computational accuracy. Visualization provides real-time, intuitive monitoring of the structural state and identification of hazardous areas, thus offering an effective solution for monitoring the structural health of offshore wind turbines under complex engineering constraints.

[0054] The following further describes specific embodiments of the present invention, algorithm examples, and experimental verification.

[0055] A method and system for monitoring the structural health of offshore wind turbines aims to solve the challenge of reconstructing the full-field state of offshore wind turbines under conditions of non-fixed sensor layout and sparse single-sided sensor distribution. To this end, a subspace-based graph deep learning architecture is constructed, achieving real-time, high-fidelity reconstruction of the full-field state through the synergistic effect of three key technical components. Specifically, firstly, an intrinsic orthogonal decomposition is used to construct a dimensionality-reduced subspace, mapping arbitrarily distributed sparse measurement data into unified low-dimensional modal coefficients, thereby overcoming the dependence on a fixed sensor layout. Secondly, Fourier feature mapping technology is introduced to convert low-dimensional spatial coordinates into high-dimensional trigonometric function features, enhancing the model's ability to capture high-frequency details of the structure. Finally, a spectral graph neural network is constructed, using Chebyshev polynomials to perform convolution operations on a non-Euclidean graph topology, establishing a nonlinear mapping relationship from fused features to full-field displacement and stress fields. The implementation methods and principles of each part are explained in detail below.

[0056] Intrinsic orthogonal decomposition

[0057] This invention utilizes the intrinsic orthogonal decomposition algorithm to extract the low-dimensional feature subspace of the structural response in the offline stage, and to achieve rapid estimation from sparse measurements to full-field modal coefficients in the online monitoring stage. The implementation process includes three steps: constructing the full-field snapshot matrix, eigenbase decomposition, and solving for the sparse coefficients.

[0058] ① Construction of the full-field snapshot matrix

[0059] First, high-fidelity numerical simulations are used to obtain full-field physical quantity data (such as the full-field strain field) of the structure at different times. Assume the structural model has... Each node, in Data is collected at discrete time steps, and the full-field surface strain vector at each time step is ( As one column, construct the full-field snapshot matrix. ,in The total number of nodes. This represents the time step. This matrix not only contains the spatial topological information of the structure, but also its time-varying characteristics under dynamic loads.

[0060] ② Characteristic basis decomposition and subspace truncation

[0061] To extract the main deformation modes of the structure, the snapshot matrix is... Perform Singular Value Decomposition (SVD):

[0062]

[0063] in, It is a left singular vector matrix, whose column vectors represent the orthogonal spatial modes (i.e., POD modes) of the structure. It is a diagonal matrix containing singular values; It is a right singular vector matrix.

[0064] Based on the magnitude of the singular values, select the first... Construct the truncated POD basis matrix using the left singular vectors corresponding to the largest singular values. ,in The total number of nodes. This is the modality cutoff number. This matrix... This constitutes a low-dimensional linear subspace capable of describing most of the structural response characteristics. At this point, the total strain at any given time... It can be approximated as the basis matrix and a set of low-dimensional modal coefficients. Linear combination:

[0065] .

[0066] ② Coefficient solving based on sparse measurements

[0067] In actual monitoring, sensors cannot cover all nodes; they can only acquire sparse measurement data whose locations may change. Let the current number of sensors be... Its position is determined by a mask matrix (or index vector). Confirmed, among which For the number of sensors, This represents the total number of nodes. It is a sparse matrix, where each row has an element of 1 only at the node position corresponding to the sensor, and the rest are 0.

[0068] At this point, the sparse measurement vector The relationship with the full-field mode can be expressed as:

[0069]

[0070] in, Let be the sparse basis matrix after filtering by the mask matrix, where For the number of sensors, is the modality cutoff number.

[0071] This is an overdetermined or well-posed system of equations (usually...) This invention employs the least squares method to solve this system of equations to obtain the optimal modal coefficients. :

[0072] .

[0073] Once the modal coefficients are obtained The formula can be used Quickly reconstruct the full-field strain distribution at that moment. .

[0074] Through the steps described above, the intrinsic orthogonal decomposition algorithm can effectively utilize the structural prior knowledge (POD basis) learned offline, transforming the online monitoring problem into a problem of solving a low-dimensional linear system. A key advantage of this method is that when the sensor layout changes, only the mask matrix needs to be updated. and the corresponding sparse basis matrix without retraining the feature base This allows for a perfect fit to any sensor layout scenario required by the present invention.

[0075] Fourier feature mapping

[0076] After obtaining the strain field features reconstructed by intrinsic orthogonal decomposition, in order to further improve the model's learning ability for high gradient regions (such as stress concentration regions), this invention introduces Fourier feature mapping technology to encode the spatial coordinates of the structure.

[0077] Deep neural networks often exhibit spectral bias, meaning they tend to learn low-frequency functions but struggle to capture high-frequency details. This is due to the node coordinates of offshore wind turbine structures. Low-dimensional data, when directly used as input to a neural network, makes it difficult for the model to learn complex, high-frequency spatial changes. Therefore, this invention maps low-dimensional coordinates to a high-dimensional feature space using Fourier feature mapping.

[0078] The specific implementation steps are as follows:

[0079] Suppose any node in the structure The normalized spatial coordinates are (Include (Coordinates in three directions). This invention defines a mapping function. This coordinate is mapped to a high-dimensional vector containing sine and cosine terms:

[0080]

[0081] in, The frequency control matrix determines the spectral coverage of the feature map. To avoid the instability of random sampling and to capture multi-scale features in a balanced manner, this invention employs deterministic logarithmic frequency sampling to construct... Specifically, It contains a set of frequency components that grow exponentially. Preferably, the hyperparameter m can be set to 8, which allows the mapped features to cover a wide spectrum from low-frequency contours to high-frequency details, thereby achieving the best reconstruction results.

[0082] After the above mapping, the original coordinates of each node are transformed into a high-dimensional Fourier feature vector. Feature Dimension From frequency group number Decision, that is Due to the number of frequency groups The corresponding feature mapping dimension is Subsequently, this coordinate feature vector is concatenated with the previously reconstructed nodal strain features to form an enhanced nodal feature vector. :

[0083]

[0084] This feature vector It contains both global distribution information of the strain physical field and fine spatial location information. As input to the subsequent spectral neural network, it can significantly improve the accuracy and convergence speed of the full-field reconstruction.

[0085] Spectral Neural Networks

[0086] To handle the complex geometric topology data of offshore wind turbine support structures, which are irregular and non-Euclidean, this invention constructs a spectral neural network. It uses spectral theory to directly define convolution operations on the structural topology, thereby establishing a nonlinear mapping from node features to the overall physical response.

[0087] ① Graph mapping of structural topology

[0088] First, the aforementioned finite element mesh of the wind turbine structure is mapped as an undirected graph. .in, For the diagram A set of nodes, corresponding to the nodes of a finite element mesh; It is the set of edges, corresponding to the connection relationships of the grid cells; It is a weighted adjacency matrix used to describe the connection strength between nodes.

[0089] To perform spectral analysis, this invention introduces a normalized Laplace matrix. The definition is as follows:

[0090]

[0091] in, for An identity matrix of order 1. This is a degree matrix (diagonal matrix), whose diagonal elements Laplace matrix It contains the inherent topological information of the wind turbine structure, which is the basis for graph convolution operations.

[0092] ② Spectral convolution based on Chebyshev polynomials

[0093] To avoid the expensive eigenvalue decomposition of the Laplacian matrix and to achieve spatial locality in convolution operations, this invention employs Chebyshev polynomials to approximate the convolution kernel in the spectral domain.

[0094] For the input signal (enhanced node features) The graph convolution operation is defined as a linear combination of Chebyshev polynomials:

[0095]

[0096] in:

[0097] Represents graph convolution operation;

[0098] For the first The learnable coefficients (i.e., convolution kernel parameters) of the Chebyshev polynomial. From 0 to ;

[0099] The size of the convolution kernel determines the receptive field of the convolution operation (i.e., the area that each node can aggregate). (Information within the order neighborhood), which is taken as 8 in this example;

[0100] The scaled Laplace matrix is ​​intended to satisfy the domain of the Chebyshev polynomial. ;

[0101] for Chebyshev polynomials can be efficiently calculated using recursive formulas: ,and .

[0102] ③ Overall response prediction

[0103] The network architecture of this invention comprises multiple stacked Chebyshev spectral convolutional layers. Each layer aggregates the feature information of neighboring nodes through the aforementioned convolutional operations and extracts high-dimensional abstract features through a nonlinear activation function (such as ReLU).

[0104] go through After layer graph convolution processing, each node obtains a deep feature representation containing information about itself and its multi-order neighborhood. Finally, this feature is input into a node-wise MLP with shared weights for decoding, outputting the predicted value of the global physical quantity corresponding to that node.

[0105]

[0106] in, The predicted full-field response matrix includes the full-field displacement vector ( ) and total stress components ( wait).

[0107] Through the above mechanism, the spectral graph neural network can learn the spatial topological dependencies of the offshore wind turbine structure, thereby realizing end-to-end inference from sparse input on one side to high-fidelity response across the entire field on both sides.

[0108] The specific implementation process of this method is as follows:

[0109] (1) Construct a high-fidelity offline dataset. The basic information of the structure to be analyzed includes: the geometric dimensions of the offshore wind turbine support structure, material properties (Young's modulus, Poisson's ratio, density, etc.) and the marine environmental parameters. Using finite element analysis software and hydrodynamic analysis tools (such as Nemoh), time-varying environmental loads (including wave force, wind load and hydrostatic pressure) are applied to the wind turbine structure to perform high-fidelity quasi-static mechanical simulation.

[0110] During the simulation, a fixed time step is used. Record the response data of all finite element nodes of the structure. Extract the full-field strain vector, full-field displacement vector, and full-field stress vector. The responses at each time step are arranged in columns to construct a full-field strain snapshot matrix. Displacement snapshot matrix and Stress snapshot matrix ,in This represents the total number of nodes in the structure. This dataset covers the physical behavior characteristics of the structure under typical operating conditions, providing a data foundation for subsequent model learning.

[0111] (2) Extracting the dimensionality-reduced feature subspace. Following the intrinsic orthogonal decomposition method, the full-field strain snapshot matrix is ​​processed. Perform singular value decomposition (SVD). First, calculate the decomposition... The cutoff order is determined based on the proportion of singular value energy. (Usually, an order with an energy percentage greater than 99.9% is chosen), and the first part of the left singular matrix is ​​truncated. The columns form the POD basis matrix The linear subspace defined by this basis matrix can express the main deformation modes of the structure with extremely low data dimensionality. At this point, the full-field strain at any given time... A unique modal coefficient vector Approximate representation, i.e. The basis matrix is ​​stored as a prior knowledge base for online reconstruction.

[0112] (3) Constructing the structural graph topology and enhancing node features. To meet the computational requirements of the spectral graph neural network, the finite element mesh of the wind turbine structure is first transformed into an undirected graph. The node set Each edge set corresponds one-to-one with a finite element node. The connection relationship of the grid cells is determined (two nodes are connected if they are in the same cell). The normalized Laplace matrix is ​​calculated. This is used for subsequent graph convolution operations. Simultaneously, following the Fourier feature mapping method, the normalized spatial coordinates of each node are... Position encoding is performed using a random Gaussian matrix. Mapping low-dimensional coordinates to high-dimensional trigonometric function features The location feature is then concatenated with the approximate strain value reconstructed in step (2) to form the enhanced feature input vector for each node. This feature vector integrates physical field information with fine geometric position information. For each node... Its input feature vector It is by It is pieced together, among which This is the strain value of the node reconstructed by POD. It is the Fourier coordinate feature of this node.

[0113] (4) Training the spectral neural network model. A spectral neural network is constructed, whose structure includes several Chebyshev graph convolutional layers and multi-head multilayer perceptrons (MLPs) at the ends. Each MLP is configured independently for each variable, with a total of 6 MLPs (6-head perceptrons). Each MLP has 128 neurons in its first layer using the ReLU activation function, and its second layer outputs the predicted displacement or stress value. Finally, the structures of the 6-head perceptrons are concatenated to obtain 6-dimensional predicted values.

[0114] The offline dataset generated in step (1) is used as training samples to enhance node features. For network input, with real displacement across the entire field. and the actual stress of the whole field The network parameters are labeled. Mean squared error (MSE) is used as the loss function, and the Adam optimizer is applied to adjust the network parameters. (Including convolution kernel coefficients and MLP weights) are iteratively updated, with the learning rate set to 0.01. , The batch size is 8, and the training run is 2000 epochs. The training objective is to minimize the difference between the predicted and actual values. After training, the network model parameters are saved. At this point, the model has mastered the ability to deduce high-fidelity displacement and stress fields from a rough strain field.

[0115] (5) Configure the on-site sensor layout and mask matrix. During the actual online monitoring phase, determine the number of single-sided strain sensors based on the actual on-site conditions of the offshore wind turbine (such as installation space limitations, cable routing, etc.). and its installation location.

[0116] Construct a sensor layout mask matrix based on the node index where the sensor is located. This matrix is ​​sparse, with each row corresponding to a sensor. Elements are 1 only at the corresponding node column position, and 0 for the rest. The corresponding sparse POD basis matrix is ​​extracted using this mask matrix. It is worth noting that when the location of the field sensors changes or some sensors fail, only the mask matrix needs to be regenerated. And update the sparse basis matrix. This eliminates the need to retrain the neural network model, thus enabling adaptation to arbitrary sensor layouts.

[0117] (6) Real-time full-field status reconstruction and visualization. During wind turbine operation, data is collected in real time. Sparse strain data transmitted back by each sensor First, by solving the linear least squares problem... Quickly estimate the current modal coefficients and reconstruct the initial full-field strain. Subsequently, The Fourier positional features are input into a trained spectral neural network, which then outputs high-fidelity full-field displacement predictions through graph convolution operations. and predicted values ​​of total stress .

[0118] Finally, the visualization module maps the calculation results back to the 3D structural model, using cloud map colors to indicate the magnitude of stress and displacement, thereby achieving real-time and intuitive monitoring of the structural health status of offshore wind turbines. The complete process is as follows: Figure 2 As shown.

[0119] Construction of Offshore Wind Turbine Structural Health Monitoring System

[0120] This invention also constructs a marine wind turbine structural health monitoring system based on unilateral sparse measurements. The system mainly consists of four core parts: a data acquisition module, a preprocessing and reconstruction module, a deep inference module, and a visualization output module. The system architecture is as follows: Figure 3 As shown. The specific functions of each module are as follows:

[0121] (1) Data acquisition module: It is responsible for connecting with the strain sensor group installed on the inner wall of one side of the offshore wind turbine, acquiring sparse strain time series data in real time, and performing noise reduction and synchronization processing.

[0122] (2) Preprocessing and reconstruction module: The built-in sensor layout management unit generates a mask matrix based on the current sensor position, and uses the preset POD basis matrix to map sparse strain data into low-dimensional modal coefficients by solving the linear least squares problem, and outputs the preliminary full-field strain field.

[0123] (3) Deep reasoning module: The core integrates a pre-trained spectral neural network. This module receives the preliminary strain field and the Fourier position features of the nodes, uses the Chebyshev diagram convolution algorithm to mine the topological correlation of the structure, and deduce the high-fidelity full-field displacement field and stress field in real time.

[0124] (4) Visualization Output Module: This module receives the full-field displacement and stress data output by the deep inference module, maps it to structural mesh nodes, and encodes and encapsulates it according to the Visualization Toolkit (VTK) standard format to generate a VTK format file (such as a .vtu file) containing geometric topology and physical information. The generated VTK file can be read by general visualization software to observe the full-field dynamic cloud map, and supports the identification of dangerous areas exceeding the stress threshold through numerical filtering.

[0125] Experimental verification

[0126] Single-sided sparse full-field reconfiguration of semi-submersible offshore wind turbine foundation

[0127] Taking a typical three-floating semi-submersible offshore wind turbine foundation as an example, this invention verifies its ability to reconstruct the entire field of complex topology under unilateral sparse measurement conditions.

[0128] The structure consists of three cylindrical pontoons, a pontoon connecting the pontoons, and a top tower connecting to the support structure. The overall finite element model of the structure contains approximately 20,000 nodes. In actual working conditions, this structure is subjected to the combined effects of random wave loads and wind loads over a long period. Due to limitations in installation conditions and corrosion protection requirements, sensors can only be installed on the inner surface of the cylindrical pontoons (i.e., single-sided monitoring), and the number of sensors is limited. In this embodiment, 180 sensors are randomly distributed on the inner wall of the pontoons to simulate the uncertainty of sensor layout in actual engineering. Under the aforementioned implementation device, the structure to be solved and the load conditions are constructed. This embodiment considers the once-in-a-century extreme sea state in the South China Sea as the external environmental load. Specifically, the significant wave height is set to 13.6 m, the spectral peak period is set to 14.2 s, and the wind speed at the hub height is set to the rated operating wind speed of 11.4 m / s. To verify the model's generalization ability to unknown conditions, the experiment strictly uses time-series data outside the training dataset for testing.

[0129] Based on the integrated spectral neural network inference module, an independent time step not seen during the training phase is selected. The strain data collected by sparse sensors at that time step is used as input to calculate the overall displacement vector and the full-field stress tensor in real time. Based on this, the three-dimensional deformation of the structure and the stress distribution on the outer surface are further calculated. The visualization results are shown below. Figure 4 and Figure 5 As shown. Figure 4 The image shows the overall displacement and deformation cloud map of the semi-submersible wind turbine foundation, with the predicted displacement on the left and the actual displacement on the right. Figure 5 The von Mises stress distribution cloud map of a semi-submersible wind turbine foundation is shown, with the predicted stress on the left and the actual stress on the right. Experimental results show that the displacement and stress predicted by this method are highly consistent with the high-fidelity response calculated by the finite element method.

[0130] In summary, this invention proposes a method and system for monitoring the structural health of offshore wind turbines based on subspace graph networks. Compared with traditional technologies, this invention can effectively improve computational efficiency while maintaining computational accuracy. Its core advantage lies in the ability to achieve real-time high-fidelity reconstruction of multiple physical quantities across the entire structural field, relying only on sparse measurement points on one side and with arbitrary changes in sensor layout, thus solving the monitoring problem under complex engineering constraints.

[0131] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0132] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0133] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0134] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc or CD-ROM; magnetic surface memory can be disk storage or magnetic tape storage. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0135] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0136] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0137] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0138] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0140] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0141] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0142] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0143] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring the structural health of offshore wind turbines, characterized in that, Includes the following steps: S1. Offline training phase: Obtain full-field response data of the structure under typical working conditions based on finite element simulation, and construct strain snapshot matrix; The strain snapshot matrix is ​​decomposed and truncated using intrinsic orthogonal decomposition to extract and save the low-dimensional feature subspace basis matrix that can characterize the main deformation modes of the structure; the finite element mesh of the structure is mapped to an undirected graph and its Laplacian matrix is ​​calculated; Fourier feature mapping is performed on the spatial coordinates of the structural nodes to enhance the ability to capture high-frequency details, and the mapped coordinate features are spliced ​​with the strain features reconstructed based on the feature subspace to form enhanced node features; Using the enhanced node features as input and full-field displacement and stress as labels, a spectral neural network is trained. The spectral neural network uses graph convolution operations based on Chebyshev polynomials to perform feature aggregation and nonlinear mapping on the undirected graph topology. Chebyshev polynomials are used to approximate the spectral domain convolution kernel to aggregate multi-order neighborhood information of nodes. S2. Online Monitoring Phase: A mask matrix is ​​generated based on the positions of sparse strain sensors arranged on one side of the field, and a sparse basis matrix is ​​constructed using the mask matrix and a pre-stored low-dimensional feature subspace basis matrix; sparse strain measurement data is acquired in real time, and the low-dimensional modal coefficients are quickly estimated and the full-field strain distribution is initially reconstructed by solving the linear equations based on the sparse basis matrix; the initially reconstructed full-field strain distribution and the pre-processed Fourier coordinate features are input into a trained spectral neural network model, and high-fidelity full-field displacement field and stress field are output in real time through forward inference; S3. Result Output Stage: The real-time output of the full-field displacement field and stress field is visualized to realize real-time monitoring of the structural health status.

2. The method for monitoring the structural health of offshore wind turbines as described in claim 1, characterized in that, In step S1, the low-dimensional feature subspace basis matrix is ​​extracted using intrinsic orthogonal decomposition. Specifically, this includes: performing singular value decomposition on the full-field strain snapshot matrix obtained through numerical simulation to obtain a left singular vector matrix representing the orthogonal space modes of the structure; determining the truncation order based on the energy proportion of the singular values, selecting a corresponding number of left singular vectors to construct a truncated basis matrix, which constitutes a low-dimensional linear subspace that can approximately represent the full-field strain at any time; during the online monitoring phase, the basis matrix is ​​filtered based on a mask matrix constructed according to the sparse sensor positions to obtain a sparse basis matrix, and a linear equation system is constructed based on the sparse basis matrix and the collected sparse strain measurement data, and the low-dimensional modal coefficients are obtained by solving the least squares method.

3. The method for monitoring the structural health of offshore wind turbines as described in claim 1, characterized in that, In step S1, the Fourier feature mapping of the spatial coordinates of the structural nodes specifically includes: mapping the normalized spatial coordinates of the nodes into high-dimensional feature vectors through a frequency control matrix constructed using deterministic logarithmic frequency sampling; the frequency control matrix contains a set of frequency components that grow geometrically; the high-dimensional feature vectors are composed of the sine and cosine function values ​​of the normalized spatial coordinates after transformation by the frequency control matrix.

4. The method for monitoring the structural health of offshore wind turbines as described in claim 1, characterized in that, In step S1, the construction of the spectral graph neural network specifically includes: mapping the structural finite element mesh into an undirected graph containing nodes, edges, and a weighted adjacency matrix, and calculating a normalized Laplacian matrix based on the weighted adjacency matrix; defining a graph convolution operation on the scaled normalized Laplacian matrix using the Chebyshev polynomials, wherein the convolution operation is performed by linearly combining the input features with Chebyshev polynomials of different orders; stacking multiple graph convolutional layers and using a nonlinear activation function for feature extraction; and finally decoding the extracted deep features into full-field displacement and stress prediction values ​​using a node-aware multilayer perceptron.

5. The method for monitoring the structural health of offshore wind turbines as described in claim 1, characterized in that, In step S2, a mask matrix is ​​generated based on the location of the field sensors. Specifically, a sparse matrix is ​​constructed, with the number of rows equal to the number of sensors and the number of columns equal to the total number of structural nodes. In each row of the matrix, only the element at the index position of the structural node where the sensor corresponding to that row is installed is 1, and the rest of the elements are 0.

6. The method for monitoring the structural health of offshore wind turbines as described in claim 5, characterized in that, In step S2, the full-field strain distribution is initially reconstructed by solving a system of linear equations. Specifically, the mask matrix is ​​used to filter the pre-stored low-dimensional feature subspace basis matrix to obtain a sparse basis matrix corresponding to the current sensor layout. Based on the sparse basis matrix and the real-time acquired sparse strain measurement data, a linear least squares problem is constructed and solved to obtain the optimal low-dimensional modal coefficients. The optimal low-dimensional modal coefficients are linearly combined with the pre-stored low-dimensional feature subspace basis matrix to reconstruct the initial full-field strain distribution.

7. The method for monitoring the structural health of offshore wind turbines as described in claim 1, characterized in that, In step S2, the strain value of each node in the initially reconstructed full-field strain distribution is concatenated with the high-dimensional coordinate feature vector obtained by Fourier feature mapping of that node to form the enhanced node feature as input to the spectral neural network model.

8. The method for monitoring the structural health of offshore wind turbines as described in claim 1, characterized in that, In step S3, the visualization specifically includes: mapping the predicted full-field displacement and stress data back to the three-dimensional geometric model of the structure, and displaying the spatial distribution of physical quantities through cloud map rendering technology.

9. A structural health monitoring system for offshore wind turbines, characterized in that, include: The data acquisition module is used to collect strain data transmitted back by the sparse strain sensor group deployed on one side of the inner wall of the wind turbine structure in real time. The preprocessing and reconstruction module, connected to the data acquisition module, is used to generate a mask matrix based on the current sensor layout, and use a pre-stored low-dimensional feature subspace basis matrix extracted based on intrinsic orthogonal decomposition to map the strain data into low-dimensional modal coefficients by solving a system of linear equations, thereby initially reconstructing the full-field strain distribution. The deep inference module, connected to the preprocessing and reconstruction module, integrates a trained spectral neural network to receive the initially reconstructed full-field strain distribution and the Fourier coordinate features of the nodes. It performs nonlinear mapping on the structural topology through graph convolution operations based on Chebyshev polynomials to deduce high-fidelity full-field displacement and stress fields in real time. Chebyshev polynomials are used to approximate the spectral domain convolution kernel to aggregate multi-order neighborhood information of the nodes. The visualization output module, connected to the deep inference module, is used to receive full-field displacement and stress field data, map them onto a three-dimensional structural model, and perform cloud map rendering and display to support real-time monitoring of structural health status and identification of dangerous areas. An offline database is used to store the low-dimensional feature subspace basis matrix, the parameters of the trained spectral neural network, and the grid topology information of the wind turbine structure, and to provide data support for the preprocessing and reconstruction module and the deep inference module.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the offshore wind turbine structural health monitoring method as described in any one of claims 1 to 8.

Images