Physical constraint neural network prediction method and system for service life of offshore photovoltaic steel structure
By employing a cascaded physical constraint neural network prediction method at the micro, meso, and macro levels, combined with a physics-driven knowledge graph and a Bayesian physics-neural hybrid model, the contradiction between complexity and interpretability in the life prediction of offshore photovoltaic steel structures is resolved, achieving high-precision and reliable small-sample life prediction and critical failure path identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MCC (SHANGHAI) STEEL STRUCTURE TECHNOLOGY CORP LTD
- Filing Date
- 2025-12-12
- Publication Date
- 2026-04-24
AI Technical Summary
Existing technologies struggle to achieve high precision, strong physical interpretability, low data dependence, and adaptability to complex environmental coupling effects in offshore photovoltaic steel structures, resulting in inaccurate and unreliable predictions, especially under small sample conditions where reliable lifetime predictions are difficult to provide.
A physical constraint neural network prediction method with a three-level cascaded structure of micro, meso, and macro levels is adopted. The corrosion-fatigue coupling mechanism is modeled by a physical knowledge graph driven by physical principles. Combined with a Bayesian physical-neural hybrid model and a physical augmentation data generator, the physical model and deep learning are integrated to achieve high-precision, highly interpretable, and reliable lifetime prediction with small samples.
It significantly improves prediction accuracy and reliability, providing reliable predictions with a 95% confidence interval based on limited measured data, and can identify critical failure paths and sensitive components, supporting engineering maintenance.
Smart Images
Figure CN121919947A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of building structure health monitoring and remaining life prediction, and in particular to a physical constraint neural network prediction method for the life of offshore photovoltaic steel structures. Background Technology
[0002] With the rapid development of the offshore photovoltaic industry, its supporting steel structures are subjected to long-term service in harsh marine environments, facing the combined effects of various environmental factors such as salt spray corrosion, wave impact, and wind loads. These factors can trigger multiple failure modes in the steel structure materials, including corrosion, fatigue crack initiation and propagation, ultimately leading to structural performance degradation or even sudden fracture, causing huge economic losses and safety risks. Therefore, accurate prediction of the remaining life of offshore photovoltaic steel structures is of great significance for ensuring their safe operation and optimizing maintenance strategies.
[0003] In recent years, Predictive Fault and Health Management (PHM) technology has been widely applied in the industrial field. In PHM, prediction is the foundation of health management, aiming to make scientific maintenance decisions by sensing equipment status and combining predictive information. However, existing prediction methods have significant shortcomings when applied to offshore photovoltaic steel structures: physical model-based methods, while interpretable, struggle to accurately characterize complex nonlinear dynamic relationships and multiphysics coupling effects; while data-driven methods based on deep learning, though adept at handling complex data, generally lack physical constraints, exhibiting a "black box" problem, resulting in low physical reliability of prediction results. Furthermore, in scenarios like marine engineering where measured data is scarce, the model's generalization ability and reliability are difficult to guarantee.
[0004] Specifically, when it comes to predicting the lifespan of offshore steel structures, existing technologies struggle to simultaneously meet the four major requirements of high accuracy, strong interpretability, low data dependence, and adaptability to the coupling effects of complex environments.
[0005] Specifically as follows:
[0006] 1. Traditional statistical and physical modeling methods can often only provide deterministic point estimates, but cannot quantify the uncertainty of predictions, making it difficult to meet the reliability assessment requirements of engineering decisions;
[0007] 2. Purely data-driven deep learning methods lack physical mechanism constraints, which may lead to prediction results that violate physical laws, poor interpretability, difficulty for engineers to trust, and are prone to overfitting under small sample conditions, resulting in insufficient prediction stability.
[0008] 3. Existing methods are difficult to effectively model the coupling mechanism of multiple physical fields and multiple failure modes in marine environments, such as salt spray corrosion, wave cyclic loading, and wind-induced vibration, leading to inaccurate understanding of the structural health status;
[0009] 4. For newly built or marine photovoltaic structures lacking a large amount of historical monitoring data, existing methods generally suffer from the problem of "data hunger" and cannot provide reliable early lifetime predictions under limited measured data conditions.
[0010] Therefore, developing a physical constraint neural network prediction method for the lifespan of marine photovoltaic steel structures is one of the important problems that urgently need to be solved. Summary of the Invention
[0011] To address the aforementioned problems in existing technologies, the present invention aims to provide a physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures. By deeply integrating physical models with deep learning, this method resolves the key contradictions between model complexity and interpretability, and data dependence and reliability in the lifespan prediction of offshore photovoltaic steel structures, ultimately achieving high-precision, physically reliable, and small-sample-reliable remaining lifespan prediction.
[0012] Another objective of this invention is to provide a physical constraint neural network prediction system for the lifespan of marine photovoltaic steel structures.
[0013] To address the aforementioned problems, this invention employs the following technical solution: a physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures. This method integrates a physical model with deep learning through a three-layer cascaded structure of micro, meso, and macroscopic levels. The method includes the following steps:
[0014] Step 1: Microscopic level. Construct a degradation model of offshore photovoltaic steel structures at the microscale. For the environment in which the offshore photovoltaic steel structures are located, use a physics-driven knowledge graph to model the coupling mechanism of marine environmental corrosion and fatigue. Map the steel structure material properties to the corrosion rate and fatigue crack propagation rate of the steel material in the marine environment. Construct a degradation behavior model of the steel material at the microscale.
[0015] Step 2, at the meso-level, construct a physical constraint graph neural network to abstract the connection relationships of each node in the offshore photovoltaic support structure into a graph structure, where nodes represent the structure and edges represent the connection relationships;
[0016] Step 3: At the macro level, a Bayesian physical-neural hybrid model is used to integrate multi-scale information. A physical augmentation data generator is used to generate synthetic data covering different operating conditions. The synthetic data is then fused with the measured data to output the probability distribution and confidence interval of the remaining lifespan.
[0017] Furthermore, it also includes step four, which involves back-mapping the lifetime prediction distribution data described in step three to the meso-level, calculating and tracing the equivalent stress and strain at the meso-level, and identifying key failure paths and sensitive components.
[0018] Furthermore, the corrosion-fatigue coupling mechanism described in step one specifically involves modeling the total crack propagation rate under each stress cycle as a linear superposition of fatigue-driven propagation and corrosion-enhanced propagation; specifically as follows:
[0019]
[0020] In the formula,
[0021] : Fatigue crack propagation rate per stress cycle (mm / cycle);
[0022] : with temperature pH Relevant material parameters, ;
[0023] : Stress intensity factor range (MPa·m¹ / ²);
[0024] Paris law exponent, dimensionless;
[0025] Corrosion-fatigue coupling coefficient, dimensionless;
[0026] Corrosion rate of materials in a purely corrosive environment (mm / year).
[0027] Furthermore, in the physical constraint graph neural network described in step two, the node state information update mechanism is to fuse the current node's historical state with its neighborhood interaction information to generate a new state representation for the node, specifically:
[0028] (21) Node feature update rules
[0029]
[0030] In the formula:
[0031] : No. Nodes in a layered network eigenvectors;
[0032] Activation functions, such as the SiLU function;
[0033] : No. The learnable parameter matrix of the layer;
[0034] :node The set of neighboring nodes;
[0035] : Edge weight normalization coefficient, which can be taken as , For node degree;
[0036] : No. Layer from node Transmitted to node Edge information;
[0037] (22) "Edge information aggregation" integrates physical priors:
[0038]
[0039] In the formula:
[0040] Multilayer perceptron;
[0041] Vector concatenation operation;
[0042] The original attribute characteristics of the edges, including connection type and geometric relationship;
[0043] : Enhancement terms based on physical partial differential equations;
[0044] : Weighting coefficient of the physical enhancement term;
[0045] Among them, edge attributes It specifically includes key attributes that affect fatigue life, such as the condition of corrosion-protective coatings designed for marine environments and the type of weld details.
[0046] (23) By designing the "physical conservation law" as the regularization term of the loss function, the network is forced to learn solutions that conform to the laws of physics:
[0047]
[0048] In the formula:
[0049] Physical constraint loss;
[0050] : The feature matrix of all nodes;
[0051] Stress tensor;
[0052] External force vector;
[0053] : Strain tensor;
[0054] Young's modulus;
[0055] Poisson's ratio;
[0056] : Hyperparameters that balance the weights of the loss term.
[0057] Furthermore, the Bayesian physical-neural hybrid model described in step three specifically refers to...
[0058]
[0059] In the formula:
[0060] : Predicted remaining lifespan;
[0061] Input features, specifically multi-scale information.
[0062] Merging datasets ;
[0063] Model parameter set This includes neural network parameters and finite element model parameters;
[0064] The posterior distribution of the parameters is approximated through variational inference.
[0065] The Bayesian physics-neural hybrid model combines the uncertainty of neural prediction with the physical analytical posterior of the finite element model to output a lifetime prediction distribution with confidence intervals.
[0066] Furthermore, the physical augmentation data generator shown in step three is implemented in the following way:
[0067] (31) A physical model is established based on the parametric finite element model, including a geometric model, a material model, a boundary condition model simulating the marine corrosion environment, and a load model including wave cyclic load and wind-induced vibration for typical forms of marine photovoltaic support structures;
[0068] (32) Make the synthesized data With real data To verify and calibrate the consistency in statistical distribution, the following steps are used:
[0069] (321) Feature space distribution alignment: using maximum mean difference or domain adversarial training techniques to reduce the distribution difference between synthetic data and a small amount of measured data at the feature level;
[0070] (322) Statistical consistency test of key physical quantities: compare the statistical characteristics of synthetic data and measured data on key physical quantities, and make the error range within ±5%;
[0071] (33) Cross-validation of model generalization performance on fused datasets After training the model, evaluate its performance on a pure measured data test set. If the generalization performance is not up to standard, adjust the parameters of the finite element model and regenerate the data, iterating until the model reaches the predetermined accuracy requirement on the measured data.
[0072] Furthermore, step (33) specifically involves,
[0073]
[0074] In the formula:
[0075] Synthetic datasets generated using the finite element method;
[0076] : Scale of synthetic data;
[0077] Parametric finite element simulator;
[0078] Random input parameters, following a distribution .
[0079] Furthermore, the identification of critical failure paths and sensitive components specifically involves,
[0080]
[0081] In the formula:
[0082] Based on node features Calculated von Mises equivalent stress;
[0083] : The yield strength of the material;
[0084] The current node's feature vector Mapped to von Mises equivalent stress, and the equivalent stress is compared with the yield strength of the steel material. Compare the output Boolean values to determine whether the steel structure at the current node has yielded.
[0085] This invention also provides a system for predicting the lifespan of the aforementioned marine photovoltaic steel structure using a physical constraint neural network, comprising a micro-layer, a meso-layer, and a macro-layer; wherein, the micro-layer uses a physics-driven knowledge graph to model the corrosion-fatigue coupling mechanism, mapping the steel structure material properties and environmental factors to its corrosion rate and fatigue crack propagation rate in a marine environment, and constructing a basic characteristic model of the material at different scales;
[0086] At the meso-level, a physical constraint graph neural network is constructed to represent the offshore photovoltaic steel structure as a graph structure composed of nodes and edges. The embedded physical conservation law is used as the regularization term of the loss function to achieve accurate prediction of the stress distribution of the structure.
[0087] The macroscopic layer uses a Bayesian physical-neural hybrid model to integrate multi-scale information. It generates synthetic data covering different operating conditions through a physical augmentation data generator. The synthetic data is then fused with the measured data to output the probability distribution and confidence interval of the remaining lifespan.
[0088] Compared with the prior art, the beneficial technical effects of the present invention are as follows:
[0089] 1. Deep integration of physics and data: Through a hybrid architecture with three cascaded layers of micro, meso and macro levels, the organic integration of physical prior knowledge and data-driven learning is achieved. This not only overcomes the limitations of pure physical models, but also solves the problem of the lack of physical constraints in pure deep learning models, significantly improving prediction accuracy and reliability.
[0090] 2. Multiphysics Coupled Failure Modeling: At the microscopic level, a physics-driven knowledge graph is used to accurately model the corrosion-fatigue coupling mechanism, effectively characterizing the synergistic effects of multiple destructive factors in the marine environment, and laying a solid physical foundation for the entire prediction system.
[0091] 3. Physical constraints ensure credibility: At the meso level, by designing a physical constraint graph neural network, physical laws such as stress balance and energy conservation are embedded as regularization terms in the loss function during the network training process. This ensures that the model's predicted output strictly follows basic physical laws, greatly enhancing the interpretability and credibility of the results.
[0092] 4. High-Reliability Prediction with Small Sample Size: At the macroscopic level, a Bayesian physics-neural hybrid model is employed, combined with a physics-enhanced data generator, to integrate large-scale synthetic data generated from finite element simulations with a small amount of measured data. This design enables the system to provide reliable predictions with a 95% confidence interval even with only 20-30 sets of measured data, effectively solving the problem of scarce data for newly constructed structures.
[0093] 5. Traceable critical failure paths: Through an interpretable backpropagation mechanism, the final life prediction results of the system can be mapped back to the physical space, automatically identifying critical failure nodes and dominant failure paths in the structure, providing intuitive and specific guidance for engineering maintenance. Attached Figure Description
[0094] Figure 1 This is a diagram illustrating the microscopic, mesoscopic, and macroscopic cascaded structure of an embodiment of this application.
[0095] Figure 2 This is a flowchart of the method described in the embodiments of this application;
[0096] Figure 3 This is a system structure diagram of an embodiment of this application. Detailed Implementation
[0097] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0098] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "vertical," "horizontal," and "inner," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, unless otherwise explicitly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0099] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.
[0100] Example 1
[0101] like Figures 1-3As shown, this invention provides a physical constraint neural network prediction system for the lifespan of offshore photovoltaic steel structures, comprising a micro-layer, a meso-layer, and a macro-layer. The micro-layer employs a physics-driven knowledge graph model to model the corrosion-fatigue coupling mechanism, mapping the steel structure material properties and environmental factors (e.g., salinity 3.5%, pH 8.2, temperature 35℃) to a corrosion rate of 0.2 mm / year and a fatigue crack propagation rate of 5 × 10⁻⁵ mm / cycle, thus constructing a basic characteristic model of the material at different scales. This mapping process considers not only the chemical properties of the material but also the influence of environmental conditions, thereby ensuring the accuracy and comprehensiveness of the characteristic model.
[0102] At the meso-level, a physically constrained graph neural network is constructed to represent the offshore photovoltaic steel structure as a graph structure of nodes (components) and edges (connections). By embedding physical conservation laws (stress balance, energy conservation) as regularization terms in the loss function (with a weight of 0.4), accurate prediction of the stress distribution of the structure is achieved. By transforming the complex structure into a simplified graph structure, the system can be more easily calculated and analyzed, while ensuring the integrity and accuracy of the physical meaning.
[0103] The macroscopic layer uses a Bayesian physical-neural hybrid model to integrate multi-scale information and combines it with a physical augmentation data generator to generate synthetic data.
[0104] The physical augmentation data generator is implemented through the following steps:
[0105] (31) A physical model is established based on the parametric finite element model, including a geometric model, a material model (using E690 steel), a boundary condition model (one end fixed, the other end free), and a load model (self-weight and wind load); the establishment of the model ensures that the system can accurately simulate the physical behavior of the actual structure.
[0106] (32) Random samples are generated using the Monte Carlo method to collect response data under different working conditions; this method can generate a large amount of sample data, providing a foundation for subsequent data processing and analysis. The generated data is preprocessed through data cleaning and feature extraction; thus, the synthetic data is synthesized. With real data Consistency in statistical distribution; the preprocessing step ensures data quality and consistency, providing a stable foundation for model training. Specifically, it includes the following steps:
[0107] (321) Feature space distribution alignment: using maximum mean difference or domain adversarial training techniques to reduce the distribution difference between synthetic data and a small amount of measured data at the feature level;
[0108] (322) Statistical consistency test of key physical quantities: compare the statistical characteristics of synthetic data and measured data on key physical quantities, and make the error range within ±5%;
[0109] (33) The processed data is fused with the measured data (20 sets), and the fused dataset is... After training the model, its performance is evaluated on a pure measured data test set. If the generalization performance is not up to standard, the parameters of the finite element model are adjusted and the data is regenerated, and this process is repeated iteratively until the model achieves the predetermined accuracy requirement on the measured data. The fusion method fully utilizes the accuracy of the measured data and combines it with the richness of the generated data to form a high-quality training set.
[0110] The system described in this application also includes an interpretable backpropagation mechanism to map the prediction results back to the physical space and identify key failure paths and influencing factors, specifically:
[0111] 41) The prediction results are transmitted back to each node through the backpropagation path; this method ensures that the prediction results can accurately correspond to the physical model.
[0112] 42) Calculate the equivalent stress and strain at each node to determine if they exceed the material's yield strength (σs = 390 MPa); through this step, the system can identify potential failure risks; specifically...
[0113] The identification of critical failure paths and sensitive components specifically involves...
[0114]
[0115] In the formula:
[0116] Based on node features Calculated von Mises equivalent stress;
[0117] : The yield strength of the material;
[0118] The feature vector of the current node Mapped to von Mises equivalent stress, and the equivalent stress is compared with the yield strength of the steel material. Compare the output Boolean values to determine whether the steel structure at the current node has yielded.
[0119] 43) If the yield strength is exceeded, mark the node as a failed node and return to the node in the previous layer for further analysis; this method can trace back to the root cause of the problem.
[0120] 44) If the yield strength is not exceeded, return to the mesoscopic level for the next step of analysis. This cyclical mechanism ensures that the system can be analyzed in depth to determine the state of the entire structure.
[0121] The system described in this application has an inference time of 180 seconds on an NVIDIA T4 GPU and achieves efficient real-time monitoring through three-layer cascaded parallel computing. This parallel computing method greatly improves the system's efficiency, enabling it to respond quickly to changes in actual conditions.
[0122] The entire system employs the Adam optimization algorithm during the training phase, with a learning rate of 0.001, a batch size of 64, and 1000 iterations. These parameter choices ensure rapid model convergence and high accuracy.
[0123] Example 2:
[0124] like Figures 1-3 As shown, the physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures provided by this invention integrates physical models with deep learning through a three-layer cascaded structure of micro, meso, and macro levels. The method includes the following steps:
[0125] Step 1: Microscopic Level. A degradation model of the offshore photovoltaic steel structure at the microscopic scale is constructed. Considering the environment in which the offshore photovoltaic steel structure is located, a physics-driven knowledge graph is used to model the coupling mechanism of marine environmental corrosion and fatigue. The steel structure material properties and environmental factors (salinity 3.0%, pH 8.0, temperature 30℃) are mapped to a corrosion rate of 0.1 mm / year and a fatigue crack propagation rate of 2 × 10^-4 mm / cycle. A degradation behavior model of the steel material at the microscopic scale is constructed. By comprehensively considering environmental and material properties, the accuracy and reliability of the microscopic layer are ensured.
[0126] The corrosion-fatigue coupling mechanism is specifically modeled as a linear superposition of two parts: fatigue-driven propagation and corrosion-enhanced propagation; as follows:
[0127]
[0128] In the formula,
[0129] : Fatigue crack propagation rate per stress cycle (mm / cycle);
[0130] : with temperature pH Relevant material parameters, ;
[0131] : Stress intensity factor range (MPa·m¹ / ²);
[0132] Paris law exponent, dimensionless;
[0133] Corrosion-fatigue coupling coefficient, dimensionless;
[0134] Corrosion rate of materials in a purely corrosive environment (mm / year).
[0135] Step 2, at the meso-level, a physically constrained graph neural network is constructed. The offshore photovoltaic steel structure is represented as a graph structure of nodes (components) and edges (connections). By embedding physical conservation laws (stress balance, energy conservation) as the regularization term of the loss function (with a weight of 0.45), accurate prediction of the structural stress distribution is achieved. The method described in this application transforms the complex structure into a simplified graph structure, making the system easier to calculate and analyze, while ensuring the integrity and accuracy of the physical meaning.
[0136] In the physical constraint graph neural network, the node state information update mechanism is to fuse the current node's historical state with its neighborhood interaction information to generate a new state representation for the node, specifically:
[0137] (21) Node feature update rules
[0138]
[0139] In the formula:
[0140] : No. Nodes in a layered network eigenvectors;
[0141] Activation functions, such as the SiLU function;
[0142] : No. The learnable parameter matrix of the layer;
[0143] :node The set of neighboring nodes;
[0144] : Edge weight normalization coefficient, which can be taken as , For node degree;
[0145] : No. Layer from node Transmitted to node Edge information;
[0146] (22) "Edge information aggregation" integrates physical priors:
[0147]
[0148] In the formula:
[0149] Multilayer perceptron;
[0150] Vector concatenation operation;
[0151] The original attribute characteristics of the edges, including connection type and geometric relationship;
[0152] : Enhancement terms based on physical partial differential equations;
[0153] : Weighting coefficient of the physical enhancement term;
[0154] Among them, edge attributes It specifically includes key attributes that affect fatigue life, such as the condition of corrosion-protective coatings designed for marine environments and the type of weld details.
[0155] (23) By designing the "physical conservation law" as the regularization term of the loss function, the network is forced to learn solutions that conform to the laws of physics:
[0156]
[0157] In the formula:
[0158] Physical constraint loss;
[0159] : The feature matrix of all nodes;
[0160] Physical constraint loss;
[0161] : The feature matrix of all nodes;
[0162] Stress tensor;
[0163] External force vector;
[0164] : Strain tensor;
[0165] Young's modulus;
[0166] Poisson's ratio;
[0167] : Hyperparameters that balance the weights of the loss term.
[0168] Step 3: At the macro level, a Bayesian physical-neural hybrid model is used to integrate multi-scale information. A physical augmentation data generator is used to generate synthetic data covering different working conditions. The synthetic data is then fused with the measured data to output the probability distribution and confidence interval of the remaining lifespan.
[0169] The Bayesian physical-neural hybrid model is specifically as follows:
[0170]
[0171] In the formula:
[0172] : Predicted remaining lifespan;
[0173] Input features, specifically multi-scale information.
[0174] Merging datasets ;
[0175] Model parameter set This includes neural network parameters and finite element model parameters;
[0176] The posterior distribution of the parameters is approximated through variational inference.
[0177] The Bayesian physics-neural hybrid model combines the uncertainty of neural prediction with the physical analytical posterior of the finite element model to output a lifetime prediction distribution with confidence intervals.
[0178] The physical augmentation data generator shown is implemented in the following way:
[0179] (31) A physical model is established based on the parametric finite element model, including a geometric model, a material model, a boundary condition model simulating the marine corrosion environment, and a load model including wave cyclic load and wind-induced vibration for typical forms of marine photovoltaic support structures;
[0180] (32) Make the synthesized data With real data To verify and calibrate the consistency in statistical distribution, the following steps are used:
[0181] (321) Feature space distribution alignment: using maximum mean difference or domain adversarial training techniques to reduce the distribution difference between synthetic data and a small amount of measured data at the feature level;
[0182] (322) Statistical consistency test of key physical quantities: compare the statistical characteristics of synthetic data and measured data on key physical quantities, and make the error range within ±5%;
[0183] (33) Cross-validation of model generalization performance on fused datasets After training the model, evaluate its performance on a pure measured data test set. If the generalization performance is not up to standard, adjust the parameters of the finite element model and regenerate the data, iterating until the model reaches the predetermined accuracy requirement on the measured data.
[0184] Furthermore, step (33) specifically involves,
[0185]
[0186] In the formula:
[0187] Synthetic datasets generated using the finite element method;
[0188] : Scale of synthetic data;
[0189] Parametric finite element simulator;
[0190] Random input parameters, following a distribution .
[0191] Furthermore, the identification of critical failure paths and sensitive components specifically involves,
[0192]
[0193] In the formula:
[0194] Based on node features Calculated von Mises equivalent stress;
[0195] : The yield strength of the material;
[0196] The current node's feature vector Mapped to von Mises equivalent stress, and the equivalent stress is compared with the yield strength of the steel material. Compare the output Boolean values to determine whether the steel structure at the current node has yielded.
[0197] Step 4: Reverse map the lifetime prediction distribution data described in Step 3 to the meso-level, calculate and trace the equivalent stress and strain at the meso-level, and identify key failure paths and sensitive components.
[0198] Through an interpretable backpropagation mechanism, the prediction results are mapped back to the physical space to identify key failure paths and influencing factors; specifically including:
[0199] 41) The prediction results are transmitted back to each node through the backpropagation path; this method ensures that the prediction results can accurately correspond to the physical model;
[0200] 42) Calculate the equivalent stress and strain of each node to determine whether they exceed the material's yield strength (σs=390MPa); through this step, the system can identify potential failure risks.
[0201] The identification of critical failure paths and sensitive components specifically involves:
[0202]
[0203] In the formula:
[0204] Based on node features Calculated von Mises equivalent stress;
[0205] : The yield strength of the material;
[0206] The current node's feature vector Mapped to von Mises equivalent stress, and the equivalent stress is compared with the yield strength of the steel material. Compare the output Boolean values to determine whether the steel structure at the current node has yielded.
[0207] 43) If the yield strength is exceeded, mark the node as a failed node and return to the node in the previous layer for further analysis; this method can trace back to the root cause of the problem.
[0208] 44) If the yield strength is not exceeded, return to the mesoscopic level for the next step of analysis. This cyclical mechanism ensures that the system can be analyzed in depth to determine the state of the entire structure.
[0209] The system, running on an NVIDIA T4 GPU with an inference time of 150 seconds, achieves efficient real-time monitoring through three layers of cascaded parallel computing. This parallel computing approach significantly improves system efficiency, enabling it to respond quickly to changes in real-world conditions.
[0210] The entire system employs the Adam optimization algorithm during the training phase, with a learning rate of 0.0005, a batch size of 128, and 2000 iterations. These parameter choices ensure rapid model convergence and high accuracy.
[0211] Finally, it should be pointed out that the above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures, characterized in that, The method integrates physical models with deep learning through a three-tiered cascaded structure of micro, meso, and macro levels. The method includes the following steps: Step 1: Construct a degradation model of offshore photovoltaic steel structures at the micro-level. For the environment in which the offshore photovoltaic steel structures are located, use a physics-driven knowledge graph to model the coupling mechanism of marine environmental corrosion and fatigue; map the steel structure material properties to the corrosion rate and fatigue crack propagation rate of the steel material in the marine environment; and construct a degradation behavior model of the steel material at the micro-level. Step 2: Construct a Physically Constrained Graph Neural Network (GNN) at the meso-level to abstract the connection relationships of each node in the offshore photovoltaic support structure into a graph structure, where nodes represent the structure and edges represent the connection relationships; Step 3: At the macro level, a Bayesian physical-neural hybrid model is used to integrate multi-scale information. A physical augmentation data generator is used to generate synthetic data covering different operating conditions. The synthetic data is then fused with the measured data to output the probability distribution and confidence interval of the remaining lifespan.
2. The physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures according to claim 1, characterized in that, It also includes step four, which involves back-mapping the lifetime prediction distribution data described in step three to the meso-level, calculating and tracing the equivalent stress and strain at the meso-level, and identifying key failure paths and sensitive components.
3. The physical constraint application network prediction method for the lifespan of offshore photovoltaic steel structures according to claim 1, characterized in that, The corrosion-fatigue coupling mechanism described in step one specifically involves modeling the total crack propagation rate under each stress cycle as a linear superposition of fatigue-driven propagation and corrosion-enhanced propagation; as detailed below: In the formula, : Fatigue crack propagation rate per stress cycle (mm / cycle); : with temperature pH Relevant material parameters, ; : Stress intensity factor range (MPa·m¹ / ²); Paris law exponent, dimensionless; Corrosion-fatigue coupling coefficient, dimensionless; Corrosion rate of materials in a purely corrosive environment (mm / year).
4. The physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures according to claim 1, characterized in that, In the physical constraint graph neural network described in step two, the node state information update mechanism is to fuse the current node's historical state with its neighborhood interaction information to generate a new state representation for the node, specifically: (21) Node feature update rules In the formula: : No. Nodes in a layered network eigenvectors; Activation functions, such as the SiLU function; : No. The learnable parameter matrix of the layer; :node The set of neighboring nodes; : Edge weight normalization coefficient, which can be taken as... , For node degree; : No. Layer from node Transmitted to node Edge information; (22) "Edge information aggregation" integrates physical priors: In the formula: Multilayer perceptron; Vector concatenation operation; The original attribute characteristics of the edges, including connection type and geometric relationship; : Enhancement terms based on physical partial differential equations; : Weighting coefficient of the physical enhancement term; Among them, edge attributes It specifically includes key attributes that affect fatigue life, such as the condition of corrosion-protective coatings designed for marine environments and the type of weld details. (23) By designing the "physical conservation law" as the regularization term of the loss function, the network is forced to learn solutions that conform to the laws of physics: In the formula: Physical constraint loss; : The feature matrix of all nodes; Stress tensor; External force vector; : Strain tensor; Young's modulus; Poisson's ratio; : Hyperparameters that balance the weights of the loss term.
5. The physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures according to claim 1, characterized in that, The Bayesian physical-neural hybrid model described in step three is specifically as follows: In the formula: : Predicted remaining lifespan; Input features, specifically multi-scale information. Merging datasets ; Model parameter set This includes neural network parameters and finite element model parameters; The posterior distribution of the parameters is approximated through variational inference. The Bayesian physics-neural hybrid model combines the uncertainty of neural prediction with the physical analytical posterior of the finite element model to output a lifetime prediction distribution with confidence intervals.
6. The physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures according to claim 1, characterized in that, The physical augmentation data generator shown in step three is implemented in the following way: (31) A physical model is established based on the parametric finite element model, including a geometric model, a material model, a boundary condition model simulating the marine corrosion environment, and a load model including wave cyclic load and wind-induced vibration for typical forms of marine photovoltaic support structures; (32) Make the synthesized data With real data To verify and calibrate the consistency in statistical distribution, the following steps are used: (321) Feature space distribution alignment: using maximum mean difference or domain adversarial training techniques to reduce the distribution difference between synthetic data and a small amount of measured data at the feature level; (322) Statistical consistency test of key physical quantities: compare the statistical characteristics of synthetic data and measured data on key physical quantities, and make the error range within ±5%; (33) Cross-validation of model generalization performance on fused datasets After training the model, evaluate its performance on a pure measured data test set. If the generalization performance is not up to standard, adjust the parameters of the finite element model and regenerate the data, iterating until the model reaches the predetermined accuracy requirement on the measured data.
7. The physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures according to claim 6, characterized in that, Step (33) specifically involves: In the formula: Synthetic datasets generated using the finite element method; : Scale of synthetic data; Parametric finite element simulator; Random input parameters, following a distribution .
8. The physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures according to claim 2, characterized in that, The identification of critical failure paths and sensitive components specifically involves... In the formula: Based on node features Calculated von Mises equivalent stress; : The yield strength of the material; The feature vector of the current node Mapped to von Mises equivalent stress, and the equivalent stress is compared with the yield strength of the steel material. Compare the output Boolean values to determine whether the steel structure at the current node has yielded.
9. A system for implementing the physical constraint neural network prediction method for the lifespan of offshore photovoltaic steel structures as described in claim 1, characterized in that, It includes a micro-layer, a meso-layer, and a macro-layer; among them, the micro-layer uses a physics-driven knowledge graph to model the corrosion-fatigue coupling mechanism, mapping the steel structure material properties and environmental factors to its corrosion rate and fatigue crack propagation rate in a marine environment, and constructing a basic characteristic model of the material at different scales; At the meso-level, a physical constraint graph neural network is constructed to represent the offshore photovoltaic steel structure as a graph structure composed of nodes and edges. The embedded physical conservation law is used as the regularization term of the loss function to achieve accurate prediction of the stress distribution of the structure. The macroscopic layer uses a Bayesian physical-neural hybrid model to integrate multi-scale information. It generates synthetic data covering different operating conditions through a physical augmentation data generator. The synthetic data is then fused with the measured data to output the probability distribution and confidence interval of the remaining lifespan.