A predictive operation and maintenance method and system for deep-sea wind turbine platforms driven by digital twins

Abstract

This invention discloses a digital twin-driven predictive operation and maintenance method and system for deep-sea wind turbine platforms. The method includes acquiring real-time, multi-source, heterogeneous data of a target deep-sea wind turbine platform based on a sensor network; constructing a digital twin of the target wind turbine platform with multi-physics field integration; inputting the real-time, multi-source, heterogeneous data into the digital twin for simulation processing to obtain a failure probability distribution for the target wind turbine platform; and executing an operation and maintenance strategy corresponding to the failure probability distribution. The predictive operation and maintenance method for deep-sea wind turbine platforms provided by embodiments of the present invention improves the efficiency of predictive operation and maintenance for deep-sea wind turbine platforms.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine operation and maintenance technology, and in particular to a digital twin-driven deep-sea wind turbine platform predictive operation and maintenance method and system. Background Art

[0002] As global demand for clean energy continues to grow, deep-sea wind power, with its abundant resources and minimal environmental impact, has become a key area of ​​future wind power development. However, deep-sea wind turbine platforms are constantly exposed to extreme environments such as high salt spray, strong winds and waves, and complex currents. These platforms face multiple risks, including structural fatigue, component wear, and electrical failures. Their operation and maintenance is far more difficult and costly than onshore and offshore wind turbines.

[0003] Existing models often use single physical field modeling. Although it can accurately characterize the aerodynamic characteristics and structural mechanical properties of the wind turbine, it is difficult to achieve real-time simulation when dealing with the coupling effects of multiple physical fields due to the extremely high computational complexity. In addition, the model parameters are difficult to dynamically adjust according to actual operating data, resulting in deviations between the predicted results and the actual operating conditions, thereby affecting the efficiency of the wind turbine platform's predictive operation and maintenance. Summary of the Invention

[0004] The present invention provides a digital twin-driven predictive operation and maintenance method for deep-sea wind turbine platforms to solve the technical problem of how to improve the existing predictive operation and maintenance method for wind turbine platforms, thereby achieving the effect of improving the predictive operation and maintenance efficiency of wind turbine platforms.

[0005] To solve the above technical problems, an embodiment of the present invention provides a predictive operation and maintenance method for a deep-sea wind turbine platform driven by a digital twin, comprising:

[0006] Acquire real-time multi-source heterogeneous data from deep-sea target wind turbine platforms based on sensor networks;

[0007] Constructing a digital twin of the target wind turbine platform with multi-physics integration;

[0008] Inputting the real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain a failure probability distribution of the target wind turbine platform;

[0009] Executing an operation and maintenance strategy corresponding to the failure probability distribution;

[0010] The digital twin construction process includes:

[0011] The acquired historical multi-source heterogeneous data of the target wind turbine platform are sequentially subjected to spatiotemporal calibration, multimodal feature extraction, and weighted fusion processing to obtain a fused feature vector.

[0012] Constructing a dual graph structure of the target wind turbine platform based on the fused feature vector, and performing spatiotemporal convolution processing on the dual graph structure based on a cross-graph attention mechanism to obtain a spatiotemporal feature matrix; wherein the dual graph structure includes a physical topology graph and a data association graph;

[0013] Performing distributed training on the spatiotemporal feature matrix based on a federated learning framework to obtain digital twin model parameters;

[0014] Multi-physics field integration processing is performed on the digital twin model parameters to obtain a digital twin of the target wind turbine platform, wherein the multi-physics field integration processing is designed to accelerate the construction of the digital twin based on model reduction technology and incremental learning mechanism.

[0015] As one preferred solution, constructing the dual graph structure of the target wind turbine platform according to the fused feature vector includes:

[0016] Performing node representation learning on the fused feature vector based on a graph embedding algorithm to obtain a low-dimensional vector representation of the physical component and sensor data;

[0017] Constructing the physical topology graph according to the low-dimensional vector representation, wherein the nodes of the physical topology graph are low-dimensional vector representations of wind turbine components, and the edges are spatial correlation matrices based on mechanical connection relationships;

[0018] Based on Granger causality test and information entropy theory, the dynamic causal strength between sensor data is calculated to obtain the data correlation matrix;

[0019] Constructing the data association graph according to the data association matrix, wherein the nodes of the data association graph are low-dimensional vector representations of sensor data, and the edges are the time association matrix based on the dynamic causal strength;

[0020] An adjacency tensor of the dual graph structure is constructed based on the spatial correlation matrix and the temporal correlation matrix to obtain a dual graph structure that integrates physical structure and data dependency.

[0021] As one of the preferred solutions, the dual graph structure is subjected to spatiotemporal convolution processing according to the cross-graph attention mechanism to obtain a spatiotemporal feature matrix, including:

[0022] Calculating the cross-graph attention weights between the physical topology graph and the data association graph based on a multi-head self-attention mechanism to obtain a cross-graph association matrix;

[0023] Constructing a spatiotemporal attention graph convolutional network according to the cross-graph association matrix, wherein the spatiotemporal attention graph convolutional network includes a spatial attention layer, a temporal attention layer, and a feature fusion layer;

[0024] Extracting spatiotemporal features from the dual graph structure based on the spatiotemporal attention graph convolutional network to obtain a spatial feature matrix and a temporal feature matrix;

[0025] Perform feature fusion on the spatial feature matrix and the temporal feature matrix according to a tensor decomposition algorithm to obtain a spatiotemporal feature tensor;

[0026] Performing time series modeling on the spatiotemporal feature tensor to obtain the spatiotemporal feature matrix.

[0027] As one of the preferred solutions, the cross-graph attention weights between the physical topology graph and the data association graph are calculated based on the multi-head self-attention mechanism to obtain a cross-graph association matrix, including:

[0028] Mapping node features of the physical topology graph and the data association graph to a plurality of different subspaces respectively;

[0029] Calculate the attention score between the physical topology graph nodes and the data association graph nodes in each subspace based on the dot product attention mechanism;

[0030] Normalizing the attention scores to obtain cross-graph attention weights;

[0031] Based on the multi-head weighted aggregation strategy, the cross-graph attention weights of multiple subspaces are fused to output a cross-graph association matrix that integrates multi-dimensional information.

[0032] As one of the preferred solutions, the multi-physics field integration processing of the digital twin model parameters includes:

[0033] Constructing an aerodynamic model of the target wind turbine platform based on a computational fluid dynamics method to obtain fluid load distribution data;

[0034] Constructing a structural mechanics model of the target wind turbine platform according to a finite element analysis method to obtain stress and strain distribution data;

[0035] Building a power transmission model of the target wind turbine platform based on an electrical system modeling method to obtain electrical parameter distribution data;

[0036] Performing data fusion on the fluid load distribution data, stress and strain distribution data, and electrical parameter distribution data according to a multi-physics field coupling algorithm to obtain a multi-physics field coupling model;

[0037] The multi-physics field coupling model is calibrated based on the digital twin model parameters to obtain the digital twin.

[0038] As one of the preferred solutions, the multi-physics field integration processing of the digital twin model parameters further includes:

[0039] Based on the model order reduction technology, the aerodynamic model, the structural mechanics model and the power transmission model are respectively reduced in dimension to obtain a low-dimensional aerodynamic reduced-order model, a structural mechanics response surface model and a power transmission reduced-order model;

[0040] generating fluid load distribution data, stress-strain distribution data, and electrical parameter distribution data according to the low-dimensional aerodynamics reduced-order model, the structural mechanics response surface model, and the power transmission reduced-order model;

[0041] Establishing a historical multi-physics field simulation database based on the fluid load distribution data, the stress and strain distribution data, and the electrical parameter distribution data;

[0042] When new operating condition data is input, the matching degree between the new operating condition data and the historical operating condition data in the historical multi-physics field simulation database is calculated by cosine similarity, and the historical simulation results with a matching degree lower than a preset threshold are used as difference data;

[0043] Correcting the multi-physics field coupling model based on the difference data to obtain a corrected multi-physics field coupling model;

[0044] The multi-physics field coupling model is calibrated according to the digital twin model parameters to obtain the digital twin.

[0045] As one preferred solution, inputting the real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain the failure probability distribution of the target wind turbine platform includes:

[0046] Preprocessing the real-time multi-source heterogeneous data based on edge computing technology to obtain standardized real-time data;

[0047] Performing state estimation on the standardized real-time data according to a Kalman filter algorithm to obtain a real-time state vector;

[0048] Driving the digital twin to perform real-time simulation based on the real-time state vector to obtain a predicted state vector;

[0049] Uncertainty quantification is performed on the predicted state vector according to a Bayesian inference algorithm to obtain the fault probability distribution.

[0050] As one preferred solution, the uncertainty quantification of the predicted state vector based on the Bayesian inference algorithm includes:

[0051] Determining a priori probability of a fault state of the target wind turbine platform based on historical fault data and expert experience;

[0052] The predicted state vector is used as observation data, and the likelihood probability of the observation data under different fault states is calculated based on the probability model;

[0053] Calculate the posterior probability distribution by combining the prior probability and the likelihood probability according to the Bayesian formula;

[0054] The probability density function and cumulative distribution function of the fault state are calculated based on the posterior probability distribution to obtain a fault probability distribution containing uncertainty information.

[0055] As one preferred solution, executing an operation and maintenance strategy corresponding to the failure probability distribution includes:

[0056] Classify the fault probability distribution into risk levels to obtain the fault risk level;

[0057] Optimize the decision making of the fault risk level according to the reinforcement learning algorithm to generate an optimal operation and maintenance strategy set;

[0058] Performing simulation verification on the optimal operation and maintenance strategy set based on the digital twin to obtain strategy execution effect evaluation results;

[0059] Comprehensively sort the strategy execution effect evaluation results according to the multi-objective decision-making algorithm to determine the target operation and maintenance strategy;

[0060] The target operation and maintenance strategy is sent to the execution terminal, and the execution status is fed back in real time.

[0061] Another embodiment of the present invention provides a digital twin-driven deep-sea wind turbine platform predictive operation and maintenance system, including:

[0062] Acquisition module, used to acquire real-time multi-source heterogeneous data of deep-sea target wind turbine platforms based on sensor networks;

[0063] A construction module for constructing a digital twin of the target wind turbine platform with multi-physics field integration;

[0064] A simulation module, configured to input the real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain a failure probability distribution of the target wind turbine platform;

[0065] An execution module, configured to execute an operation and maintenance strategy corresponding to the failure probability distribution;

[0066] The digital twin construction process includes:

[0067] The acquired historical multi-source heterogeneous data of the target wind turbine platform are sequentially subjected to spatiotemporal calibration, multimodal feature extraction, and weighted fusion processing to obtain a fused feature vector.

[0068] Constructing a dual graph structure of the target wind turbine platform based on the fused feature vector, and performing spatiotemporal convolution processing on the dual graph structure based on a cross-graph attention mechanism to obtain a spatiotemporal feature matrix; wherein the dual graph structure includes a physical topology graph and a data association graph;

[0069] Performing distributed training on the spatiotemporal feature matrix based on a federated learning framework to obtain digital twin model parameters;

[0070] Multi-physics field integration processing is performed on the digital twin model parameters to obtain a digital twin of the target wind turbine platform, wherein the multi-physics field integration processing is designed to accelerate the construction of the digital twin based on model reduction technology and incremental learning mechanism.

[0071] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0072] 1) The present invention uses digital twin technology to construct a complete predictive operation and maintenance system for deep-sea wind turbine platforms. Its beneficial effects are mainly reflected in the following aspects: real-time collection of multi-source heterogeneous data through sensor networks, and the use of dual graph structures combined with cross-graph attention mechanisms to process spatiotemporal features. At the same time, it integrates the federated learning framework and multi-physics field integration technology, which not only realizes the accurate modeling of the complex physical characteristics of the wind turbine platform, but also accelerates the construction efficiency of the digital twin through model reduction and incremental learning, effectively solving the problems of data dispersion and limited computing resources in deep-sea environments.

[0073] 2) At the level of operation and maintenance prediction and strategy execution, the present invention achieves accurate quantification of failure probability through edge computing, Kalman filtering and Bayesian reasoning, and then combines reinforcement learning with multi-objective decision-making to generate the optimal operation and maintenance strategy. It can not only predict the risk of equipment failure in advance, but also dynamically optimize the maintenance plan based on the simulation verification results, significantly improving the intelligence and refinement of operation and maintenance, greatly reducing the operation and maintenance costs of deep-sea wind turbine platforms, improving equipment operation reliability, and providing technical support for the safe and efficient development of deep-sea wind power resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 1 is a flow chart of a predictive operation and maintenance method for a deep-sea wind turbine platform driven by a digital twin in one embodiment of the present invention;

[0075] Figure 2 A schematic diagram of a dual graph structure construction process in one embodiment of the present invention;

[0076] Figure 3 Schematic diagram of a predictive operation and maintenance system for a deep-sea wind turbine platform driven by a digital twin according to one embodiment of the present invention;

[0077] Reference numerals:

[0078] Among them, 11, acquisition module; 12, construction module; 13, simulation module; 14, execution module. DETAILED DESCRIPTION

[0079] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0080] In the description of the present invention, the terms "first," "second," "third," etc. are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first," "second," "third," etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0081] In describing the present invention, it should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. Those skilled in the art will understand the specific meanings of the above terms in the present invention in specific circumstances.

[0082] One embodiment of the present invention provides a predictive operation and maintenance method for a deep-sea wind turbine platform driven by a digital twin. For details, see Figure 1 , Figure 1 The figure shows a flow chart of a method for predicting operation and maintenance of a deep-sea wind turbine platform driven by a digital twin in one embodiment of the present invention, which includes steps S1-S4:

[0083] S1: Acquire real-time multi-source heterogeneous data from deep-sea target wind turbine platforms based on sensor networks;

[0084] In one embodiment of the present invention, various sensors, including vibration sensors, temperature sensors, pressure sensors, strain gauges, and current / voltage sensors, are deployed at key locations on a deep-sea wind turbine platform, such as blades, hubs, gearboxes, towers, generators, and foundation structures, to collect data on the equipment's operating status. Furthermore, meteorological and ocean current sensors are deployed around the platform to monitor environmental data such as wind speed, direction, wave height, and current speed and direction in real time. This is shown in Table 1, which shows the multi-source heterogeneous data types for the deep-sea wind turbine platform.

[0085] Table 1

[0086]

[0087] The data collected by these sensor networks exhibit significant differences in time frequency, data format, and physical significance, resulting in multi-source heterogeneous data. After collection, edge computing nodes perform preliminary processing on the raw data, including outlier removal and data completion. The processed data is then transmitted to the data center via communication links such as satellite communications or submarine optical cables.

[0088] S2: Build a digital twin of the target wind turbine platform with multi-physics integration;

[0089] It should be noted that the digital twin is a digital mirror of the physical entity of the deep-sea wind turbine platform. It is built based on historical data and real-time monitoring data through modeling and simulation technology. It can reflect the status and performance of the physical entity in real time and predict future operating trends and failures.

[0090] This invention builds a multi-physics integrated digital twin, aiming to provide a highly accurate and reliable virtual model for predictive operation and maintenance of deep-sea wind turbine platforms. This digital twin can simulate wind turbine operation under various operating conditions, analyze component stress and wear, and predict potential failures, thus providing a basis for developing scientific and rational operation and maintenance strategies.

[0091] In a preferred embodiment of the present invention, the process of constructing a digital twin includes:

[0092] S21: performing spatiotemporal calibration, multimodal feature extraction, and weighted fusion processing on the acquired historical multi-source heterogeneous data of the target wind turbine platform to obtain a fused feature vector;

[0093] It should be noted that this step aims to transform raw, chaotic, and heterogeneous historical data from multiple sources into a unified, efficient, and information-rich fusion feature vector. This provides a solid data foundation for the subsequent construction of a dual graph structure that accurately reflects the physical structure and operating principles of the wind turbine platform, as well as for training a high-performance digital twin model. Through spatiotemporal data calibration, feature extraction, and fusion, data accuracy, consistency, and availability are ensured, improving data quality.

[0094] In this embodiment, first, a dynamic time warping algorithm is used for sensor data with precise timestamps. While allowing the time axis to be stretched, the optimal matching path between different time series data is found to achieve time series alignment. For sensors with clock deviations, the time delay is determined by calculating the cross-correlation function between the time series of each sensor, and then the timestamp is corrected. At the same time, the sensor clock is regularly synchronized with the help of the global positioning system or the network time protocol to ensure that the time error is within an acceptable range.

[0095] Based on the wind turbine platform's 3D design drawings and sensor installation layout, a unified 3D spatial coordinate system was constructed. An inertial measurement unit (IMU) was used to obtain the platform's attitude information in real time, and the sensor data was converted to a unified coordinate system using a coordinate transformation matrix. Furthermore, laser scanning and photogrammetry technologies were used to regularly measure the actual sensor installation positions. These were then compared to the theoretical positions and error corrections were performed to ensure accurate correspondence between the data and the physical spatial locations.

[0096] In multimodal feature extraction, for time series data such as vibration, current, and voltage in multi-source heterogeneous data, the Fourier transform is used to convert the time domain signal into a frequency domain signal to extract the energy, amplitude and other features of different frequency components; the wavelet transform is used to perform multi-resolution analysis on the signal to capture the local change characteristics of the signal at different time scales; and the characteristic parameters reflecting the overall characteristics of the signal can be obtained by calculating the mean, variance, kurtosis, skewness and other statistical quantities of the signal.

[0097] Convolutional neural networks are used to extract features from multi-source, heterogeneous surveillance video data. By building multiple layers of convolutional, pooling, and fully connected layers, they automatically learn image features such as texture, shape, and edges. Object detection algorithms are used to identify key components in the image (such as blades and towers) and extract features such as their location, size, and state. Furthermore, image segmentation techniques can be used to divide the image into different regions and analyze the characteristics of each region.

[0098] For environmental data such as wind speed, wind direction, and wave height in multi-source heterogeneous data, the statistical characteristics such as mean, maximum, minimum, and standard deviation are calculated to describe the changing trends of environmental factors. The environmental data are modeled using a time series prediction algorithm to predict future environmental changes. At the same time, the correlation characteristics between environmental data and wind turbine operating parameters are extracted, such as the relationship between wind speed and power generation.

[0099] After obtaining the multimodal feature data, the principal component analysis method is first used to reduce the dimensionality of the features extracted from each mode, analyze the variance contribution rate of each principal component, and use it as the initial weight; then, by calculating the mutual information between each modal feature and the equipment fault, the importance of the feature to fault diagnosis is evaluated, and the weight is further adjusted; in addition, combined with the experience of domain experts, the weight is manually modified, for example, giving higher weights to features that reflect the failure of key components.

[0100] The standardized feature vectors of each modality are weighted and summed according to the calculated weights to obtain a fused feature vector; or the features of each modality are combined into a high-dimensional tensor by tensor splicing, and then dimensionality reduction and feature fusion are performed through fully connected layers and attention mechanisms, finally generating a fused feature vector that contains multi-source information and highlights key features.

[0101] S22: constructing a dual graph structure of the target wind turbine platform based on the fused feature vectors, and performing spatiotemporal convolution processing on the dual graph structure based on the cross-graph attention mechanism to obtain a spatiotemporal feature matrix; wherein the dual graph structure includes a physical topology graph and a data association graph;

[0102] It should be noted that the dual graph structure includes a physical topology graph and a data association graph. The physical topology graph describes the wind turbine platform from the physical structure level, with nodes representing the various physical components of the wind turbine. The weight of the edge reflects the tightness of the mechanical connection between the components or the force transmission relationship, and is used to characterize the spatial structure layout and mechanical association of the wind turbine; the data association graph focuses on the data collected by the sensors, with the nodes being the feature representation of the sensor data, and the edge weight calculated through Granger causality test, information entropy theory, etc., reflecting the causal dependence and mutual influence relationship between different sensor data in the time series.

[0103] This step constructs a dual graph structure to intuitively express the physical structure information and sensor data information of the wind turbine platform in the form of a graph. It also uses the cross-graph attention mechanism and spatiotemporal convolution processing to explore the complex relationship between the physical structure and data, and extract the spatiotemporal feature matrix that can reflect the operating status of the wind turbine. This provides structured, high-value feature data for subsequent digital twin model training, thereby more accurately simulating the operation process of the wind turbine platform and predicting potential faults.

[0104] For details, see Figure 2, Figure 2 A schematic diagram of a dual graph structure construction process provided by an embodiment of the present invention. In a preferred embodiment of the present invention, the dual graph structure of the target wind turbine platform is constructed based on the fused feature vectors, including:

[0105] S220: Perform node representation learning on the fused feature vector based on a graph embedding algorithm to obtain a low-dimensional vector representation of the physical component and sensor data;

[0106] It should be noted that graph embedding algorithms are a technique for mapping high-dimensional graph data into a low-dimensional space. By learning the vector representations of nodes in this low-dimensional space, they preserve the topological structure and semantic information between nodes in the graph, facilitating subsequent graph analysis and processing. In this paper, they are used to convert fused feature vectors into low-dimensional vectors suitable for constructing graph structures.

[0107] In this embodiment, appropriate graph embedding algorithms are selected, including DeepWalk and Node2Vec. Taking Node2Vec as an example, the length and number of random walks are first defined. Random walks are then performed on a virtually constructed initial graph structure containing all physical components and sensor data nodes to generate a node sequence.

[0108] The fused feature vector is used as the initial feature of the node, and the generated node sequence is trained using the Skip-Gram model. During training, a low-dimensional vector representation of each node is learned by optimizing the objective function, so that nodes with similar distances in the low-dimensional space have similar structural or semantic relationships in the original graph. Ultimately, a low-dimensional vector representation of the physical component and sensor data is obtained.

[0109] S221: Constructing a physical topology graph based on the low-dimensional vector representation, wherein the nodes of the physical topology graph are low-dimensional vector representations of wind turbine components, and the edges are spatial correlation matrices based on mechanical connection relationships;

[0110] In this embodiment, the obtained low-dimensional vector representation of the physical component is used as a node of the physical topology graph, and each node corresponds to a physical component of the wind turbine.

[0111] Based on the wind turbine's design drawings and mechanical engineering knowledge, the mechanical connections between the various physical components are determined. For example, the blades and hub are bolted together, and the tower and foundation are welded together. A spatial correlation matrix is ​​constructed based on these connections. For components with direct mechanical connections, non-zero weights are assigned to the corresponding positions in the matrix. The weights are determined by the tightness of the connection or the force transfer coefficient. For components without direct connections, the corresponding matrix elements are set to 0.

[0112] Based on the node and spatial association matrix, a physical topology diagram is constructed to clearly present the spatial layout and connection relationship of the wind turbine physical components.

[0113] S222: Calculate the dynamic causal strength between sensor data based on Granger causality test and information entropy theory to obtain the data correlation matrix;

[0114] The Granger causality test is a statistical method used to determine whether a causal relationship exists between two time series data sets. Specifically, it determines whether changes in one series can statistically predict changes in the other. Information entropy theory, used to measure the uncertainty or information content of data, assists the Granger causality test in this paper by quantifying the degree of information transfer between sensor data and thereby calculating dynamic causal strength. Dynamic causal strength is a quantitative indicator of the mutual influence and causal relationship between sensor data in a time series and is used to determine edge weights in a data association graph.

[0115] In this example, a Granger causality test is applied to each set of sensor data time series. Specifically, the lag order for the test is set. A regression model is constructed that includes the two time series and their lags. An F-test is then performed to determine whether one series Granger causes the other. In other words, this test determines whether the historical information of one series significantly improves the predictive power of the other.

[0116] By combining information entropy theory, we can calculate the amount of information transferred between two sensor data sets. For example, by calculating conditional entropy and mutual information, we can quantify the extent to which the information in one data sequence reduces the uncertainty of another data sequence, further assisting in determining the strength of the causal relationship.

[0117] Combining the Granger causality test results and information entropy calculation results, we calculated the dynamic causal strength for each pair of sensor data and constructed a data association matrix. The value of the matrix element corresponds to the dynamic causal strength between the two sensor data points, with larger values ​​indicating stronger causal relationships.

[0118] S223: Constructing a data association graph based on the data association matrix, wherein the nodes of the data association graph are low-dimensional vector representations of sensor data, and the edges are a time association matrix based on dynamic causal strength;

[0119] In this embodiment, the acquired low-dimensional vector representation of the sensor data is used as a node of the data association graph, and each node corresponds to data collected by a sensor.

[0120] According to the data association matrix, for sensor data pairs whose dynamic causal strength is greater than the set threshold, edges are added to connect the corresponding nodes in the data association graph, and the weights of the edges are set to the corresponding dynamic causal strength values, thereby constructing a data association graph that reflects the temporal dependence and causal relationship of the sensor data.

[0121] S224: Construct an adjacency tensor of a dual graph structure based on the spatial correlation matrix and the temporal correlation matrix to obtain a dual graph structure that integrates the physical structure and the data dependency.

[0122] In this embodiment, the spatial correlation matrix and the temporal correlation matrix are used as two dimensional information of the adjacency tensor to construct a high-order tensor, such as a three-dimensional tensor. Among them, the first dimension corresponds to the nodes of the physical topology graph, the second dimension corresponds to the nodes of the data association graph, and the third dimension is used to distinguish between spatial correlation and temporal correlation information. In this way, the structural information of the physical topology graph and the data association graph are integrated into an adjacency tensor, realizing the deep fusion of physical structure and data dependency, and obtaining the final dual graph structure, which provides a structured data foundation for subsequent feature extraction and model training.

[0123] In a preferred embodiment of the present invention, a spatiotemporal convolution process is performed on the dual graph structure according to the cross-graph attention mechanism to obtain a spatiotemporal feature matrix, including:

[0124] S225: Calculate the cross-graph attention weights between the physical topology graph and the data association graph based on the multi-head self-attention mechanism to obtain a cross-graph association matrix;

[0125] Among them, the multi-head self-attention mechanism is a deep learning mechanism that uses multiple independent attention heads to parallelly calculate the feature representation of the input sequence. Each head focuses on different subspace information of the input, and finally splices or weightedly fuses the multi-head results to enhance the model's ability to capture complex correlation relationships.

[0126] The cross-graph attention weight is used to measure the correlation strength between the node features of the physical topology graph and the data association graph. The dependency weights between different graph structures are dynamically calculated through the attention mechanism to reflect the interaction importance of cross-graph nodes.

[0127] The cross-graph association matrix is ​​a matrix based on the cross-graph attention weights, which records the interaction intensity between the nodes of the physical topology graph and the data association graph, and is used to guide the cross-graph transmission and fusion of spatiotemporal features.

[0128] This step uses a multi-head attention mechanism to establish semantic associations between physical components and sensor data, breaking the limitations of traditional methods that separate physical structure from monitoring data. Furthermore, by leveraging subspace decomposition and multi-head aggregation, it simultaneously captures the spatial structure of the physical topology and the temporal dynamics of the sensor data, generating a more comprehensive feature representation.

[0129] In a preferred embodiment of the present invention, the cross-graph attention weights between the physical topology graph and the data association graph are calculated based on the multi-head self-attention mechanism to obtain a cross-graph association matrix, including:

[0130] S2251: Mapping node features of the physical topology graph and the data association graph to multiple different subspaces respectively;

[0131] In this embodiment, node feature matrices are obtained from the physical topology graph and the data association graph, representing the raw features of physical components and sensor data, respectively. A separate linear transformation matrix is ​​designed for each attention head to project the raw features into multiple low-dimensional subspaces. Each subspace corresponds to a specific association pattern, such as mechanical connection, temperature change, or vibration response. Through matrix operations, the raw features are decomposed into multiple parallel subspace representations, each focusing on a different type of feature relationship.

[0132] S2252: Calculate the attention score between the physical topology graph nodes and the data association graph nodes in each subspace based on the dot product attention mechanism;

[0133] In this embodiment, in each subspace, the node features of the physical topology graph are used as queries, and the node features of the data association graph are used as keys. A dot product operation is used to calculate the similarity between the nodes in the physical topology graph and the nodes in the data association graph in each subspace. A higher dot product result indicates more similar feature representations of the two nodes in that subspace. To stabilize gradient training, the dot product result is divided by the square root of the key vector dimension to avoid vanishing or exploding gradients.

[0134] S2253: Normalize the attention scores to obtain cross-graph attention weights;

[0135] In this embodiment, a Softmax function is applied to the scaled attention scores to convert them into a weight matrix in the form of a probability distribution. The normalized weights represent the degree of attention that the physical topology map nodes pay to the data association map nodes. A higher weight indicates a higher priority for information transfer.

[0136] S2254: Based on the multi-head weighted aggregation strategy, the cross-graph attention weights of multiple subspaces are fused to output a cross-graph correlation matrix that integrates multi-dimensional information.

[0137] In this embodiment, the cross-graph attention weight matrices of all subspaces are concatenated column by column to form a composite matrix containing multi-dimensional correlation information. Using a learnable linear transformation matrix, the concatenated composite matrix is ​​projected back to the original dimensions, achieving weighted aggregation of multi-head information. The final cross-graph correlation matrix is ​​output, whose elements represent the combined correlation strength between physical components and sensor data, guiding subsequent cross-graph information transfer and feature fusion.

[0138] S226: Constructing a spatiotemporal attention graph convolutional network according to the cross-graph association matrix, wherein the spatiotemporal attention graph convolutional network includes a spatial attention layer, a temporal attention layer, and a feature fusion layer;

[0139] It should be noted that the spatiotemporal attention graph convolutional network is a network architecture that integrates spatial attention, temporal attention and graph convolution operations to process spatiotemporal feature extraction of graph structure data.

[0140] Spatial attention layer: focuses on the spatial associations of nodes in the graph, such as physical topological relationships;

[0141] Temporal attention layer: focuses on the temporal dependencies of node features, such as dynamic causal relationships in data association graphs;

[0142] Feature fusion layer: Integrates spatiotemporal features to generate a unified representation.

[0143] In this embodiment, a three-layer network architecture is designed based on a cross-graph association matrix. Graph convolution is used to aggregate the neighbor features of nodes in the physical topology graph. At the same time, a cross-graph association matrix is ​​introduced to dynamically adjust the aggregation weights of spatial features, highlighting physical components that are strongly related to the data association graph. Temporal convolution or an attention mechanism is used to process the temporal features in the data association graph. Physical structure information is introduced through the cross-graph association matrix to enhance the physical semantics of the temporal features. A nonlinear transformation function is designed to fuse the outputs of the spatial attention layer and the temporal attention layer to generate a joint spatiotemporal feature representation.

[0144] S227: Extracting spatiotemporal features of the dual graph structure based on the spatiotemporal attention graph convolutional network to obtain a spatial feature matrix and a temporal feature matrix;

[0145] In this embodiment, the physical topology graph is fed into the spatial attention layer, where node features are aggregated through graph convolution and an attention mechanism to generate a spatial feature matrix reflecting the spatial relationships between components. The data association graph is fed into the temporal attention layer, where temporal dependencies of the data are captured through time series modeling and an attention mechanism to generate a temporal feature matrix reflecting operational state changes. The cross-graph association matrix is ​​used to transfer information between the two feature matrices, further enhancing the temporal dynamics of spatial features and the physical correlation of temporal features.

[0146] S228: Perform feature fusion on the spatial feature matrix and the temporal feature matrix according to the tensor decomposition algorithm to obtain a spatiotemporal feature tensor;

[0147] Among them, the spatiotemporal feature tensor is a multidimensional array containing spatial dimension, time dimension and feature dimension, which is used to represent the fused spatiotemporal features and can be further reduced to a matrix through tensor decomposition.

[0148] In this embodiment, the spatial feature matrix and the temporal feature matrix are concatenated by dimension to construct a high-order tensor containing three dimensions: space, time, and features. Algorithms such as CP decomposition or Tucker decomposition are used to decompose the high-order tensor into the product of multiple low-dimensional tensors, extracting the main characteristic components and removing noise and redundant information. The space-time feature tensor is reconstructed from the decomposed low-dimensional tensors, achieving an organic fusion of spatial and temporal features and generating a more compact and expressive joint space-time representation.

[0149] S229: Perform time series modeling on the spatiotemporal feature tensor to obtain a spatiotemporal feature matrix.

[0150] In this embodiment, the spatiotemporal feature tensor is expanded along the time dimension to generate a feature sequence, with each time step corresponding to a multidimensional feature vector. Time series models such as LSTM, GRU, or Transformer are used to process the feature sequence, capturing the long-term dependencies and changing trends of the data and learning time series patterns and regularities. The output of the time series model is organized into a matrix form, with each row representing a comprehensive feature vector for a time step and each column representing a spatiotemporal feature of a specific dimension. This ultimately forms a complete spatiotemporal feature matrix for subsequent state prediction and fault diagnosis of the digital twin.

[0151] S23: Distributed training of the spatiotemporal feature matrix is ​​performed based on the federated learning framework to obtain the parameters of the digital twin model;

[0152] It's important to note that the federated learning framework is a distributed machine learning paradigm that allows participants (such as geographically dispersed wind farms) to collaboratively train models without sharing raw data. Data remains locally, and only model parameters or gradient information are exchanged through encryption, achieving a balance between privacy protection and data utilization.

[0153] Distributed training is the process of assigning training tasks to multiple computing nodes (such as servers and edge devices) for parallel execution. This accelerates model training and improves resource utilization through collaborative optimization.

[0154] Digital twin model parameters are the learnable weights and biases in the digital twin model, which are continuously optimized through the training process so that the model can accurately simulate the behavior and performance of the physical system.

[0155] In wind power scenarios, data from various wind farms contains sensitive operational information. Federated learning prevents the external transmission of raw data, complying with privacy regulations. Furthermore, data from different wind farms is siloed due to factors such as geography and ownership. Federated learning enables cross-site collaborative modeling, integrating distributed data to improve model generalization. Furthermore, local training on edge devices reduces data transmission latency, enabling real-time decision-making and fault warning.

[0156] In this embodiment, a central server first defines a unified digital twin model architecture, such as an LSTM or Transformer network, and distributes the initial model parameters to edge devices (clients) at each wind farm. Each client adjusts input layer parameters, such as the number of sensors and time step, based on the local data dimensions to ensure compatibility between the model structure and local data.

[0157] Each client trains a model using locally stored spatiotemporal feature matrix data. During training, the client adds differential privacy noise to the model parameters or gradients to protect the privacy of the original data. After local training is complete, the client sends the encrypted model update, rather than the original data, to the central server.

[0158] The central server collects model updates from all clients and generates global model parameters using a federated averaging or adaptive aggregation algorithm. During this aggregation process, the server dynamically adjusts weights based on the amount of data or performance of each client to ensure that the global model balances all data sources. The updated global model parameters are then distributed to all clients, and the next round of iterative training begins.

[0159] After each round of training, the client evaluates model performance on a local validation set and reports these metrics to the server. The server monitors global performance metrics (such as prediction accuracy and loss function value) and terminates training when these metrics reach preset thresholds or when the training round is complete. The final global model parameters are deployed to each client, which processes the new spatiotemporal feature matrix in real time to achieve wind turbine status prediction and fault warning.

[0160] After training is complete, the central server analyzes each client's local model, extracting common and individual characteristics to further optimize the global model structure. The system also supports continuous learning. When new wind farms are added or the operating environment changes, the model can be updated through incremental training to maintain the accuracy and adaptability of the digital twin model.

[0161] S24: performing multi-physics field integration processing on the digital twin model parameters to obtain a digital twin of the target wind turbine platform, wherein the multi-physics field integration processing is designed to accelerate the construction of the digital twin based on model order reduction technology and incremental learning mechanism;

[0162] Multi-physics integration involves fusing models and data from multiple physical fields (such as fluid mechanics, structural mechanics, and thermodynamics) to comprehensively describe the complex physical behavior of wind turbine platforms. By integrating the interactions between these different physical fields, the simulation accuracy of the digital twin is improved.

[0163] Model order reduction technology simplifies the computational complexity of complex physical models, significantly reducing the amount of computation while preserving key physical properties, thereby accelerating the real-time simulation capabilities of digital twins. Common methods include projection and kriging interpolation.

[0164] Incremental learning is a machine learning strategy that allows a model to continuously update itself by learning from new data, building upon existing knowledge without having to retrain the entire model. This is particularly useful for real-time optimization of digital twins in dynamic environments.

[0165] This step integrates multi-physics models, including fluid dynamics and structural mechanics, to fully capture the complex physical interactions during wind turbine operation. Model order reduction techniques reduce the computational burden, enabling the digital twin to rapidly respond to operational state changes. Incremental learning mechanisms are used to continuously optimize model parameters to adapt to dynamic factors such as turbine aging and environmental changes.

[0166] In a preferred embodiment of the present invention, multi-physics field integration processing is performed on the digital twin model parameters, including:

[0167] S2411: Construct an aerodynamic model of the target wind turbine platform based on computational fluid dynamics methods to obtain fluid load distribution data;

[0168] Computational fluid dynamics (CFD) is a technique that numerically solves fluid mechanics equations to simulate the flow characteristics of air or liquids. In wind turbine applications, CFD is used to analyze the interaction between airflow and wind turbine blades and calculate fluid load distribution.

[0169] The aerodynamic model is a mathematical model that describes the aerodynamic characteristics of the fan when it operates in an airflow. It is constructed using the CFD method and outputs the distribution data of the fluid load on each component of the fan.

[0170] S2412: Construct a structural mechanics model of the target wind turbine platform according to the finite element analysis method to obtain stress and strain distribution data;

[0171] Finite element analysis (FEA) is a numerical calculation method that discretizes complex structures into a finite number of elements. By solving the mechanical equilibrium equations between these elements, the stress and strain distribution of the structure under load is analyzed. Structural mechanics models are used to analyze the mechanical response of wind turbine structures under various loads and output stress and strain distribution data.

[0172] S2413: Constructing a power transmission model of the target wind turbine platform based on an electrical system modeling method to obtain electrical parameter distribution data;

[0173] The power transmission model describes the energy transmission process from the generator to the grid in the wind turbine electrical system. It outputs electrical parameter distribution data for evaluating power quality and system efficiency.

[0174] In this embodiment, specialized models are first constructed for different physical domains of the wind turbine platform. Computational fluid dynamics methods are used to establish an aerodynamic model, simulating the flow of air around the wind turbine blades and tower and calculating the fluid load distribution. Simultaneously, a structural mechanics model is constructed using finite element analysis methods, discretizing the wind turbine structure into a finite number of elements and analyzing its stress and strain distribution under various loads. Furthermore, an electrical system modeling method is used to construct a power transmission model, describing the transmission of electrical energy from the generator to the grid and analyzing the distribution of electrical parameters such as voltage and current.

[0175] S2414: Performing data fusion on the fluid load distribution data, the stress and strain distribution data, and the electrical parameter distribution data according to a multi-physics coupling algorithm to obtain a multi-physics coupling model;

[0176] In this embodiment, the output data from the three physical field models described above is fed into a multiphysics coupling algorithm. This algorithm accounts for the interdependencies between the various physical fields. For example, the fluid loads calculated by the aerodynamic model are fed into the structural mechanics model as boundary conditions, while also considering the effects of structural deformation on fluid flow, achieving bidirectional coupling. Through iterative solution, a unified multiphysics coupling model is ultimately obtained, fully describing the behavior of the wind turbine in complex physical environments.

[0177] S2415: Calibrate the parameters of the multi-physics field coupling model based on the digital twin model parameters to obtain a digital twin.

[0178] In this example, actual wind turbine operating data was used to calibrate the parameters of the multi-physics coupling model. By comparing the model's predicted values ​​with the actual measured values, key model parameters were adjusted to ensure that the model output closely matches the physical behavior. After multiple iterations of optimization, a highly accurate digital twin was ultimately obtained, capable of reflecting the operating status of the target wind turbine platform in real time and predicting future performance changes.

[0179] In another preferred embodiment of the present invention, performing multi-physics field integration processing on the digital twin model parameters further includes:

[0180] S2421: Perform dimensionality reduction processing on the aerodynamic model, the structural mechanics model, and the power transmission model based on the model order reduction technology to obtain a low-dimensional aerodynamic reduced-order model, a structural mechanics response surface model, and a power transmission reduced-order model;

[0181] Among them, the low-dimensional aerodynamic reduced-order model is an aerodynamic model simplified by applying model reduction technology. By retaining the dominant modes or characteristic parameters, the fluid load distribution is quickly calculated, reducing the consumption of computing resources.

[0182] The structural mechanics response surface model is an approximate model constructed based on the results of finite element analysis. By fitting the relationship between input parameters (such as load) and output responses (such as stress and strain), it replaces time-consuming finite element calculations and achieves rapid response prediction.

[0183] The power transmission reduced-order model is a simplified power system model. By identifying key electrical parameters and transmission paths, it can quickly calculate the parameter distribution during the power transmission process and improve simulation efficiency.

[0184] In this example, model reduction techniques were first applied to simplify the three physical field models of aerodynamics, structural mechanics, and power transmission. For the aerodynamics model, key features were extracted by identifying the dominant flow modes and constructing a low-dimensional reduced-order model. For the structural mechanics model, a response surface relationship between input load and output stress and strain was fitted based on finite element analysis results. For the power transmission model, a simplified reduced-order model was established by analyzing key electrical paths and parameters. These reduced-order models significantly improve computational efficiency while retaining the core physical properties of the original models.

[0185] S2422: Generate fluid load distribution data, stress and strain distribution data, and electrical parameter distribution data based on the low-dimensional aerodynamics reduced-order model, the structural mechanics response surface model, and the power transmission reduced-order model;

[0186] S2423: Establishing a historical multi-physics field simulation database based on the fluid load distribution data, stress and strain distribution data, and electrical parameter distribution data;

[0187] Among them, the historical multi-physics field simulation database stores the database of multi-physics field simulation results under historical working conditions, including fluid load, stress strain and electrical parameter distribution data under different operating conditions, which is used as a reference benchmark for new working condition analysis.

[0188] S2424: When new operating condition data is input, the matching degree between the new operating condition data and the historical operating condition data in the historical multi-physics field simulation database is calculated by cosine similarity, and the historical simulation results with matching degrees below a preset threshold are used as difference data;

[0189] Among them, the difference data refers to the part where the matching degree between the new operating condition data and the historical data is lower than the preset threshold, reflecting the prediction deviation of the model under the new operating condition and is used to guide model correction.

[0190] In this embodiment, a reduced-order model is used to rapidly generate multiphysics simulation results under different operating conditions, building a historical multiphysics simulation database. When new operating condition data is input, the cosine similarity between the new condition and each condition in the historical database is calculated to assess the degree of match. Historical simulation results with a match below a preset threshold are marked as difference data, reflecting potential deviations in the model's predictions under the new condition.

[0191] S2425: Correcting the multi-physics field coupling model based on the difference data to obtain a corrected multi-physics field coupling model;

[0192] S2426: Calibrate the parameters of the multi-physics field coupling model according to the digital twin model parameters to obtain a digital twin.

[0193] In this embodiment, the multi-physics coupling model is modified based on the difference data. By adjusting model parameters, updating boundary conditions, or improving sub-model structures, the model's predictions under the new operating conditions are made closer to reality. The modified model is then calibrated using the digital twin model parameters to further optimize model accuracy. During the calibration process, actual sensor data is compared with the model's predictions, and model parameters are iteratively adjusted until a satisfactory match is achieved.

[0194] S3: Input real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain the failure probability distribution of the target wind turbine platform;

[0195] Among them, simulation processing refers to the process of inputting real-time data into the digital twin, simulating the operating behavior of the wind turbine under current working conditions through model calculation, and predicting future state changes.

[0196] The failure probability distribution is used to describe the probability distribution of failure of each wind turbine component in the future. It is usually expressed in the form of a probability value or probability density function to quantify the failure risk.

[0197] In a preferred embodiment of the present invention, real-time multi-source heterogeneous data is input into the digital twin for simulation processing to obtain the failure probability distribution of the target wind turbine platform, including:

[0198] S31: Preprocessing real-time multi-source heterogeneous data based on edge computing technology to obtain standardized real-time data;

[0199] Among them, edge computing technology is a computing mode that processes and analyzes data close to the data source (such as sensors), reducing data transmission delays, improving system response speed, and protecting data privacy.

[0200] In this embodiment, real-time multi-source heterogeneous data is first preprocessed on the edge device at the wind turbine site. For different types of sensor data, corresponding filtering algorithms are used to remove noise and perform standardization to unify the data format and dimension.

[0201] S32: performing state estimation on the standardized real-time data according to the Kalman filter algorithm to obtain a real-time state vector;

[0202] Among them, the Kalman filter algorithm is a recursive optimal estimation algorithm that estimates the true state of the system by fusing current measurement values ​​and historical prediction values, and effectively handles noise and uncertainty.

[0203] In this embodiment, a Kalman filter algorithm is used to perform state estimation on standardized real-time data. The algorithm combines the wind turbine's physical model with historical data to predict the current state and fuses the predicted value with the actual measured value to obtain the optimal state estimate. For example, by fusing data from multiple vibration sensors, the true vibration state of the wind turbine blades is estimated, filtering out false signals caused by environmental interference. The state estimation result is represented as a real-time state vector, which contains the key operating parameters of each wind turbine component.

[0204] S33: driving the digital twin to perform real-time simulation based on the real-time state vector to obtain a predicted state vector;

[0205] In this embodiment, a real-time state vector is input into the digital twin, driving the model for real-time simulation. Based on the laws of physics and historical parameters, the digital twin simulates the turbine's current operating behavior and predicts future state changes. For example, based on the current wind speed, load, and temperature distribution, the stress distribution and fatigue damage trends of the blades can be predicted. The simulation results are output as a predicted state vector, which contains the future state parameters of each component.

[0206] S34: Quantify the uncertainty of the predicted state vector according to the Bayesian reasoning algorithm to obtain the fault probability distribution.

[0207] The Bayesian inference algorithm is a statistical inference method based on Bayes' theorem. It combines prior knowledge and observed data to calculate the probability distribution of uncertain events. Uncertainty quantification is used to assess the degree of uncertainty in model prediction results. In this step, it is used to quantify the probability distribution of fault predictions, reflecting the credibility of the predictions.

[0208] In this embodiment, a Bayesian inference algorithm is used to quantify the uncertainty of the predicted state vector. The algorithm combines historical failure data with domain knowledge to establish a failure probability model and calculate the failure probability distribution of each component at different time points. For example, based on the predicted stress values ​​and material fatigue properties of the blade, the probability distribution of a crack occurring in the blade within the next month is calculated. For complex failure scenarios, the algorithm considers the interplay of multiple failure modes and generates a multidimensional failure probability distribution, providing a comprehensive basis for operation and maintenance decision-making.

[0209] In a preferred embodiment of the present invention, uncertainty quantification of the predicted state vector is performed based on a Bayesian inference algorithm, including:

[0210] S341: Determine the prior probability of a target wind turbine platform failure state based on historical failure data and expert experience;

[0211] Among them, the prior probability is a preliminary estimate of the possibility of a fault state occurring based on historical data or expert experience before observing new data, reflecting domain knowledge and historical trends.

[0212] First, based on historical failure data and expert experience, we determined the prior probability of each fault state for the target wind turbine platform. We collected historical failure records for similar wind turbines and calculated the frequency of failures of different components. We also invited domain experts to assess the likelihood of each fault state in the absence of new data, focusing on rare but critical failure modes. By combining historical statistical results with expert opinion, we constructed a prior probability distribution, which provided a foundation for subsequent reasoning.

[0213] S342: Using the predicted state vector as observation data, and calculating the likelihood probability of the observation data under different fault states based on the probability model;

[0214] The predicted state vector is the prediction output by the digital twin, containing the state parameters of each wind turbine component at a future moment. Observational data is measurable data collected in real time or predicted by the model. In this step, it serves as the predicted state vector and is used to update the fault state judgment. The probability model is a mathematical model that describes the statistical relationship between the fault state and the observed data. It is used to calculate the probability of the observed data occurring under different fault conditions.

[0215] In this embodiment, the predicted state vectors output by the digital twin are used as observation data, and a probabilistic model is used to calculate the likelihood of these observations occurring under different fault conditions. For example, a statistical relationship model is established between parameters such as temperature and vibration and bearing faults, calculating the probability of observing the current temperature and vibration values ​​when the bearing is at different fault levels. For complex fault scenarios, the joint probability distribution of multiple state parameters is considered to more comprehensively capture the relationship between fault characteristics and observation data.

[0216] S343: Calculate the posterior probability distribution based on the Bayesian formula by combining the prior probability and the likelihood probability;

[0217] The likelihood probability is the conditional probability of the observed data occurring under a given fault state, reflecting the degree of support for different fault states from the observed data. The posterior probability distribution is the updated probability distribution of the fault state after considering the new observation data, which integrates prior knowledge and current evidence.

[0218] In this embodiment, the prior probability and likelihood probability are combined according to the Bayesian formula to calculate the posterior probability distribution. This step reverses the probability distribution from "fault state causes observation data" to "observation data infers fault state." For example, given the prior probability of a fault state and the likelihood of the observation data under that state, the Bayesian formula is used to calculate the posterior probability of that fault state given the current observation data. As new observation data is continuously input, the posterior probability is continuously updated to reflect the latest fault probability.

[0219] S344: Calculate the probability density function and cumulative distribution function of the fault state based on the posterior probability distribution to obtain a fault probability distribution containing uncertainty information.

[0220] The probability density function (PDF) describes the probability distribution of a continuous random variable. In this step, it represents the distribution density of the fault probability at different values. The cumulative distribution function (CDF) represents the probability that a random variable is less than or equal to a certain value. It is used to quickly query the cumulative value of the fault probability and assist in setting decision thresholds.

[0221] In this embodiment, the probability density function (PDF) and cumulative distribution function (CDF) of the fault state are calculated based on the posterior probability distribution. The PDF intuitively displays the distribution of the fault probability at different values, while the CDF allows for quick query of the cumulative value of the fault probability. These probability distributions serve as the final fault prediction results for risk assessment and decision support. For example, different risk thresholds can be set. When the CDF shows that the probability of failure of a component exceeds 80%, an advanced warning is triggered; when it exceeds 50%, a close attention prompt is issued.

[0222] S4: Execute the operation and maintenance strategy corresponding to the failure probability distribution.

[0223] In a preferred embodiment of the present invention, executing an operation and maintenance strategy corresponding to the failure probability distribution includes:

[0224] S41: Classify the fault probability distribution into risk levels to obtain a fault risk level;

[0225] Among them, the fault risk level is to divide the fault risk into different levels according to the characteristics of the fault probability distribution, such as low, medium and high risk, to facilitate the formulation of differentiated operation and maintenance strategies.

[0226] In this embodiment, the failure probability distribution is classified into risk levels. Risk levels are categorized based on the probability of failure, impact (e.g., downtime losses, repair costs), and uncertainty. For example, low risk: probability <10% and minimal impact; medium risk: probability 10%-50% or moderate impact; and high risk: probability >50% or significant impact. Clear classification criteria and corresponding initial response strategies are set for each risk level, forming a risk level matrix.

[0227] S42: Optimize the decision-making of the fault risk level according to the reinforcement learning algorithm to generate the optimal operation and maintenance strategy set;

[0228] Reinforcement learning is a machine learning method that uses a reward mechanism to learn optimal decision-making strategies through interaction between an agent and its environment. This algorithm is used to optimize O&M decisions in this step. The optimal O&M strategy set is a set of O&M strategies generated by the reinforcement learning algorithm, each corresponding to a different failure risk scenario, designed to maximize long-term O&M benefits.

[0229] In this example, a reinforcement learning algorithm is used to optimize O&M strategies for different risk levels. Risk level, device status, and O&M history are used as inputs to construct a decision-making environment. A reward function is defined as the optimization objective, such as reducing failure probability, reducing costs, and improving safety. Through repeated interactions between the agent and the environment, the optimal decision-making strategy is learned, generating a set of optimal O&M strategies encompassing various scenarios. For example, for high-risk faults, the algorithm might recommend immediate downtime for maintenance; for low-risk faults, it might suggest extending the monitoring period.

[0230] S43: Simulate and verify the optimal operation and maintenance strategy set based on the digital twin to obtain the strategy execution effect evaluation result;

[0231] In this example, the optimal set of O&M strategies is input into a digital twin for simulation verification. The digital twin simulates the operational state changes of the wind turbine under different strategies and evaluates the effectiveness of the strategies. For example, it simulates the improvement in equipment performance caused by maintenance operations and predicts the impact of downtime on power generation. Key evaluation indicators are collected and a strategy execution evaluation report is generated to provide data support for subsequent decision-making.

[0232] S44: comprehensively sorting the strategy execution effect evaluation results according to the multi-objective decision-making algorithm to determine the target operation and maintenance strategy;

[0233] In this embodiment, a multi-objective decision-making algorithm is used to comprehensively rank the evaluation results. Taking into account multiple objectives, such as minimizing cost, maximizing reliability, and reducing downtime, strategies are quantitatively compared using weighted scoring or Pareto optimality methods. For example, different weights are assigned to cost, reliability, and security, and a comprehensive score is calculated for each strategy. Based on the ranking results, the target O&M strategy that best meets business needs is determined, ensuring an optimal balance between multiple objectives.

[0234] S45: Send the target operation and maintenance strategy to the execution terminal and provide real-time feedback on the execution status.

[0235] In this embodiment, the target O&M strategy is distributed to the execution terminal. The execution terminal executes specific operations according to the strategy and provides real-time feedback on the execution status. The execution results are compared with the simulation predictions of the digital twin to assess the discrepancy between the actual results and the expected results. If deviations are found, the causes are analyzed and the decision model is updated, forming a closed-loop optimization loop. For example, if the actual maintenance time exceeds the simulation prediction, the time estimation parameters in the subsequent strategy are adjusted to improve the accuracy of future decisions.

[0236] Establish a continuous learning mechanism to continuously optimize reinforcement learning models and multi-objective decision-making algorithms as O&M data accumulates. Analyze historical policy execution results, summarize lessons learned, and update reward functions and weight distribution. Transform successful O&M strategies into a knowledge base to provide reference for new wind turbines and similar scenarios. Through continuous evolution, O&M strategies are increasingly aligned with actual needs, improving overall wind turbine O&M efficiency and reliability.

[0237] Another embodiment of the present invention provides a digital twin driven deep sea wind turbine platform predictive operation and maintenance system. For details, please refer to Figure 3 , Figure 3 The figure shows a schematic diagram of a predictive operation and maintenance system for a deep-sea wind turbine platform driven by a digital twin in one embodiment of the present invention, which includes an acquisition module 11, a construction module 12, a simulation module 13, and an execution module 14, wherein:

[0238] An acquisition module 11 is used to acquire real-time multi-source heterogeneous data of a deep-sea target wind turbine platform based on a sensor network;

[0239] Building module 12, for building a digital twin of the target wind turbine platform with multi-physics integration;

[0240] The simulation module 13 is used to input real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain the failure probability distribution of the target wind turbine platform;

[0241] An execution module 14 is configured to execute an operation and maintenance strategy corresponding to the failure probability distribution;

[0242] The digital twin construction process includes:

[0243] The acquired historical multi-source heterogeneous data of the target wind turbine platform are sequentially subjected to spatiotemporal calibration, multimodal feature extraction, and weighted fusion processing to obtain a fused feature vector.

[0244] Based on the fused feature vectors, a dual graph structure of the target wind turbine platform is constructed. This dual graph structure is then subjected to spatiotemporal convolution processing using a cross-graph attention mechanism to obtain a spatiotemporal feature matrix. The dual graph structure includes a physical topology graph and a data association graph.

[0245] Based on the federated learning framework, the spatiotemporal feature matrix is ​​distributedly trained to obtain the parameters of the digital twin model;

[0246] Multi-physics field integration processing is performed on the digital twin model parameters to obtain the digital twin of the target wind turbine platform. The multi-physics field integration processing is designed to accelerate the construction of the digital twin based on model reduction technology and incremental learning mechanism.

[0247] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0248] 1) The present invention uses digital twin technology to construct a complete predictive operation and maintenance system for deep-sea wind turbine platforms. Its beneficial effects are mainly reflected in the following aspects: real-time collection of multi-source heterogeneous data through sensor networks, and the use of dual graph structures combined with cross-graph attention mechanisms to process spatiotemporal features. At the same time, it integrates the federated learning framework and multi-physics field integration technology, which not only realizes the accurate modeling of the complex physical characteristics of the wind turbine platform, but also accelerates the construction efficiency of the digital twin through model reduction and incremental learning, effectively solving the problems of data dispersion and limited computing resources in deep-sea environments.

[0249] 2) At the level of operation and maintenance prediction and strategy execution, the present invention achieves accurate quantification of failure probability through edge computing, Kalman filtering and Bayesian reasoning, and then combines reinforcement learning with multi-objective decision-making to generate the optimal operation and maintenance strategy. It can not only predict the risk of equipment failure in advance, but also dynamically optimize the maintenance plan based on the simulation verification results, significantly improving the intelligence and refinement of operation and maintenance, greatly reducing the operation and maintenance costs of deep-sea wind turbine platforms, improving equipment operation reliability, and providing technical support for the safe and efficient development of deep-sea wind power resources.

[0250] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A predictive operation and maintenance method for a deep-sea wind turbine platform driven by a digital twin, characterized in that: The following steps are involved: Acquire real-time multi-source heterogeneous data from deep-sea target wind turbine platforms based on sensor networks; Constructing a digital twin of the target wind turbine platform with multi-physics integration; Inputting the real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain a failure probability distribution of the target wind turbine platform; Executing an operation and maintenance strategy corresponding to the failure probability distribution; The digital twin construction process includes: The acquired historical multi-source heterogeneous data of the target wind turbine platform are sequentially subjected to spatiotemporal calibration, multimodal feature extraction, and weighted fusion processing to obtain a fused feature vector. Constructing a dual graph structure of the target wind turbine platform based on the fused feature vector, and performing spatiotemporal convolution processing on the dual graph structure based on a cross-graph attention mechanism to obtain a spatiotemporal feature matrix; wherein the dual graph structure includes a physical topology graph and a data association graph; Performing distributed training on the spatiotemporal feature matrix based on a federated learning framework to obtain digital twin model parameters; performing multi-physics field integration processing on the digital twin model parameters to obtain a digital twin of the target wind turbine platform, wherein the multi-physics field integration processing is designed to accelerate the construction of the digital twin based on model order reduction technology and incremental learning mechanism; The step of constructing the dual graph structure of the target wind turbine platform according to the fused feature vector includes: Performing node representation learning on the fused feature vector based on a graph embedding algorithm to obtain a low-dimensional vector representation of the physical component and sensor data; Constructing the physical topology graph according to the low-dimensional vector representation, wherein the nodes of the physical topology graph are low-dimensional vector representations of wind turbine components, and the edges are spatial correlation matrices based on mechanical connection relationships; Based on Granger causality test and information entropy theory, the dynamic causal strength between sensor data is calculated to obtain the data correlation matrix; Constructing the data association graph according to the data association matrix, wherein the nodes of the data association graph are low-dimensional vector representations of sensor data, and the edges are the time association matrix based on the dynamic causal strength; Constructing an adjacency tensor of the dual graph structure based on the spatial correlation matrix and the temporal correlation matrix to obtain a dual graph structure that integrates physical structure and data dependency; The multi-physics field integration processing of the digital twin model parameters includes: Constructing an aerodynamic model of the target wind turbine platform based on a computational fluid dynamics method to obtain fluid load distribution data; Constructing a structural mechanics model of the target wind turbine platform according to a finite element analysis method to obtain stress and strain distribution data; Building a power transmission model of the target wind turbine platform based on an electrical system modeling method to obtain electrical parameter distribution data; Performing data fusion on the fluid load distribution data, stress and strain distribution data, and electrical parameter distribution data according to a multi-physics field coupling algorithm to obtain a multi-physics field coupling model; Calibrate the parameters of the multi-physics field coupling model based on the digital twin model parameters to obtain the digital twin; Inputting the real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain the failure probability distribution of the target wind turbine platform includes: Preprocessing the real-time multi-source heterogeneous data based on edge computing technology to obtain standardized real-time data; Performing state estimation on the standardized real-time data according to a Kalman filter algorithm to obtain a real-time state vector; Driving the digital twin to perform real-time simulation based on the real-time state vector to obtain a predicted state vector; Uncertainty quantification is performed on the predicted state vector according to a Bayesian inference algorithm to obtain the fault probability distribution.

2. The predictive operation and maintenance method for deep-sea wind turbine platforms driven by digital twins according to claim 1, characterized in that: The dual graph structure is subjected to spatiotemporal convolution processing according to the cross-graph attention mechanism to obtain a spatiotemporal feature matrix, including: Calculating the cross-graph attention weights between the physical topology graph and the data association graph based on a multi-head self-attention mechanism to obtain a cross-graph association matrix; Constructing a spatiotemporal attention graph convolutional network according to the cross-graph association matrix, wherein the spatiotemporal attention graph convolutional network includes a spatial attention layer, a temporal attention layer, and a feature fusion layer; Extracting spatiotemporal features from the dual graph structure based on the spatiotemporal attention graph convolutional network to obtain a spatial feature matrix and a temporal feature matrix; Perform feature fusion on the spatial feature matrix and the temporal feature matrix according to a tensor decomposition algorithm to obtain a spatiotemporal feature tensor; Performing time series modeling on the spatiotemporal feature tensor to obtain the spatiotemporal feature matrix.

3. The predictive operation and maintenance method for deep sea wind turbine platforms driven by digital twins according to claim 2, characterized in that: The cross-graph attention weights between the physical topology graph and the data association graph are calculated based on the multi-head self-attention mechanism to obtain a cross-graph association matrix, including: Mapping node features of the physical topology graph and the data association graph to a plurality of different subspaces respectively; Calculate the attention score between the physical topology graph nodes and the data association graph nodes in each subspace based on the dot product attention mechanism; Normalizing the attention scores to obtain cross-graph attention weights; Based on the multi-head weighted aggregation strategy, the cross-graph attention weights of multiple subspaces are fused to output a cross-graph association matrix that integrates multi-dimensional information.

4. The predictive operation and maintenance method for deep sea wind turbine platforms driven by digital twins according to claim 1, characterized in that: The multi-physics field integration processing of the digital twin model parameters further includes: Based on the model order reduction technology, the aerodynamic model, the structural mechanics model and the power transmission model are respectively reduced in dimension to obtain a low-dimensional aerodynamic reduced-order model, a structural mechanics response surface model and a power transmission reduced-order model; generating fluid load distribution data, stress-strain distribution data, and electrical parameter distribution data according to the low-dimensional aerodynamics reduced-order model, the structural mechanics response surface model, and the power transmission reduced-order model; Establishing a historical multi-physics field simulation database based on the fluid load distribution data, the stress and strain distribution data, and the electrical parameter distribution data; When new operating condition data is input, the matching degree between the new operating condition data and the historical operating condition data in the historical multi-physics field simulation database is calculated by cosine similarity, and the historical simulation results with a matching degree lower than a preset threshold are used as difference data; Correcting the multi-physics field coupling model based on the difference data to obtain a corrected multi-physics field coupling model; The multi-physics field coupling model is calibrated according to the digital twin model parameters to obtain the digital twin.

5. The predictive operation and maintenance method for deep sea wind turbine platforms driven by digital twins according to claim 1, characterized in that: The uncertainty quantification of the predicted state vector based on the Bayesian inference algorithm includes: Determining a priori probability of a fault state of the target wind turbine platform based on historical fault data and expert experience; The predicted state vector is used as observation data, and the likelihood probability of the observation data under different fault states is calculated based on the probability model; Calculate the posterior probability distribution by combining the prior probability and the likelihood probability according to the Bayesian formula; The probability density function and cumulative distribution function of the fault state are calculated based on the posterior probability distribution to obtain a fault probability distribution containing uncertainty information.

6. The predictive operation and maintenance method for deep-sea wind turbine platforms driven by digital twins according to claim 1, characterized in that: The executing an operation and maintenance strategy corresponding to the failure probability distribution includes: Classify the fault probability distribution into risk levels to obtain the fault risk level; Optimize the decision making of the fault risk level according to the reinforcement learning algorithm to generate an optimal operation and maintenance strategy set; Performing simulation verification on the optimal operation and maintenance strategy set based on the digital twin to obtain strategy execution effect evaluation results; Comprehensively sort the strategy execution effect evaluation results according to the multi-objective decision-making algorithm to determine the target operation and maintenance strategy; The target operation and maintenance strategy is sent to the execution terminal, and the execution status is fed back in real time.

7. A digital twin-driven deep-sea wind turbine platform predictive operation and maintenance system, characterized by: include: Acquisition module, used to acquire real-time multi-source heterogeneous data of deep-sea target wind turbine platforms based on sensor networks; A construction module for constructing a digital twin of the target wind turbine platform with multi-physics integration; A simulation module, configured to input the real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain a failure probability distribution of the target wind turbine platform; An execution module, configured to execute an operation and maintenance strategy corresponding to the failure probability distribution; The digital twin construction process includes: The acquired historical multi-source heterogeneous data of the target wind turbine platform are sequentially subjected to spatiotemporal calibration, multimodal feature extraction, and weighted fusion processing to obtain a fused feature vector. Constructing a dual graph structure of the target wind turbine platform based on the fused feature vector, and performing spatiotemporal convolution processing on the dual graph structure based on a cross-graph attention mechanism to obtain a spatiotemporal feature matrix; wherein the dual graph structure includes a physical topology graph and a data association graph; Performing distributed training on the spatiotemporal feature matrix based on a federated learning framework to obtain digital twin model parameters; performing multi-physics field integration processing on the digital twin model parameters to obtain a digital twin of the target wind turbine platform, wherein the multi-physics field integration processing is designed to accelerate the construction of the digital twin based on model order reduction technology and incremental learning mechanism; The step of constructing the dual graph structure of the target wind turbine platform according to the fused feature vector includes: Performing node representation learning on the fused feature vector based on a graph embedding algorithm to obtain a low-dimensional vector representation of the physical component and sensor data; Constructing the physical topology graph according to the low-dimensional vector representation, wherein the nodes of the physical topology graph are low-dimensional vector representations of wind turbine components, and the edges are spatial correlation matrices based on mechanical connection relationships; Based on Granger causality test and information entropy theory, the dynamic causal strength between sensor data is calculated to obtain the data correlation matrix; Constructing the data association graph according to the data association matrix, wherein the nodes of the data association graph are low-dimensional vector representations of sensor data, and the edges are the time association matrix based on the dynamic causal strength; Constructing an adjacency tensor of the dual graph structure based on the spatial correlation matrix and the temporal correlation matrix to obtain a dual graph structure that integrates physical structure and data dependency; The multi-physics field integration processing of the digital twin model parameters includes: Constructing an aerodynamic model of the target wind turbine platform based on a computational fluid dynamics method to obtain fluid load distribution data; Constructing a structural mechanics model of the target wind turbine platform according to a finite element analysis method to obtain stress and strain distribution data; Building a power transmission model of the target wind turbine platform based on an electrical system modeling method to obtain electrical parameter distribution data; Performing data fusion on the fluid load distribution data, stress and strain distribution data, and electrical parameter distribution data according to a multi-physics field coupling algorithm to obtain a multi-physics field coupling model; Calibrate the parameters of the multi-physics field coupling model based on the digital twin model parameters to obtain the digital twin; Inputting the real-time multi-source heterogeneous data into the digital twin for simulation processing to obtain the failure probability distribution of the target wind turbine platform includes: Preprocessing the real-time multi-source heterogeneous data based on edge computing technology to obtain standardized real-time data; Performing state estimation on the standardized real-time data according to a Kalman filter algorithm to obtain a real-time state vector; Driving the digital twin to perform real-time simulation based on the real-time state vector to obtain a predicted state vector; Uncertainty quantification is performed on the predicted state vector according to a Bayesian inference algorithm to obtain the fault probability distribution.

Images