Intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twin

Abstract

This invention discloses a method for intelligently predicting fatigue damage evolution in aviation elastic plates based on digital twins. The method comprises the following steps: S1, constructing a finite element model and performing modal calibration to establish a digital twin model; S2, acquiring real-time monitoring data; S3, constructing a crack-induced potential energy tensor field to obtain a crack growth potential energy probability distribution map; S4, generating a probability prediction map using a potential energy-driven graph evolution coupling algorithm; S5, constructing a topological damage mapping function and predicting the dynamic evolution of topological damage with load time; S6, applying a time folding algorithm during the dynamic evolution process to obtain a latent space representation of the fatigue state; S7, inverting the current structure's topological damage state based on a contrasting twin variational autoencoder network, updating the digital twin model, and ultimately outputting the structural damage evolution trend. This invention integrates digital twins with graph evolution algorithms to achieve intelligent fatigue damage prediction in aviation elastic plates.

Description

Technical Field

[0001] The present invention relates to the technical field of aviation structure health monitoring, and in particular to an intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins. Background Art

[0002] As the structural complexity of aviation equipment continues to increase, the fatigue damage evolution of key components during service has become a core issue affecting their safety and life prediction. Aerospace elastic plates, as typical lightweight, high-load components, often operate under cyclically varying loads and coupled multi-field environments. Their fatigue damage exhibits multi-scale evolution, localized mutations, and non-unique path characteristics, posing significant challenges to traditional structural health monitoring and life prediction methods.

[0003] Currently, fatigue prediction techniques for aviation structures primarily rely on methods such as stress-life methods (SN curves), fracture mechanics methods (Paris's law), and finite element simulation analysis. These methods typically assume a fixed initial state for the material or structure and perform damage calculations based on a preset load history. These methods lack the ability to provide real-time feedback on dynamic changes in the structure's service state. Furthermore, existing methods often employ static modeling, making it difficult to incorporate real-time evolutionary characteristics such as actual flight load variations, uncertainty in crack propagation paths, and stress concentration in high-risk areas. This results in significant conservatism or bias in fatigue life assessments.

[0004] In recent years, digital twins, as a new generation of high-precision predictive modeling paradigm, have been gradually applied to the field of aviation structural health management. By dynamically mapping the physical structure with the digital model, virtual-real collaboration and dynamic evolution tracking of the structural state can be achieved. However, in existing research, digital twin modeling is mostly limited to the mapping of structural geometry, boundary conditions, and stress response, and fails to deeply integrate crack induction mechanisms, probabilistic modeling of damage evolution paths, and data-driven latent space representation methods. In addition, most digital twin frameworks lack a graphical structure modeling mechanism for "evolution trend prediction", which makes it impossible to perform high-precision, multi-scale predictions of the entire life cycle path of cracks from initiation to propagation to failure.

[0005] On the other hand, although deep learning methods have been widely introduced in data-driven models for fatigue damage identification and life prediction, they are typically used as black-box classifiers or regression tools, lacking coupling with the physical damage evolution process. This makes it difficult to meet the dual requirements of interpretability and accuracy for high-reliability aviation applications. Furthermore, currently used neural network architectures are primarily based on sequence processing (such as LSTM and CNN), and a unified modeling framework that integrates the physical evolution mechanism of cracks, structural topological characteristics, and fatigue stage characterization capabilities has not yet been established.

[0006] In summary, the existing technology still has the following shortcomings in dealing with the intelligent prediction of fatigue damage of aviation elastic sheets: first, there is a lack of modeling of crack induction mechanisms during the multi-cycle evolution of flight loads; second, it is impossible to construct a potential energy-driven crack path graph evolution model to achieve path-level evolution prediction; third, there is a lack of coupling of structural topological damage with load time series to form a dynamically evolving topological graph structure; fourth, the staged latent space representation of the fatigue state and the damage inversion mechanism have not been effectively combined with the digital twin system for closed-loop update.

[0007] Therefore, how to provide an intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose an intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins. The present invention integrates digital twin modeling, crack induced potential energy analysis and graph evolution prediction algorithm, drives structural state updates through real-time monitoring data, constructs crack evolution path probability map and topological damage dynamic evolution model, and ultimately realizes accurate prediction of fatigue damage and remaining life assessment of aviation elastic sheets. It has the advantages of high prediction accuracy, explainable evolution trend, and strong ability to adapt to dynamic working conditions.

[0009] According to an embodiment of the present invention, a method for intelligently predicting fatigue damage evolution of an aviation elastic sheet based on digital twins includes the following steps:

[0010] S1. Collect the structural parameters, material properties, and historical load spectra of the aviation elastic sheet, build a finite element model, perform modal calibration, and establish a digital twin model;

[0011] S2. Real-time collection of stress response and guided wave signals during flight to obtain real-time monitoring data;

[0012] S3. Inputting the real-time monitoring data into the digital twin model, constructing a crack-induced potential energy tensor field, and obtaining a crack development potential energy probability distribution map;

[0013] S4. Based on the crack development potential energy probability distribution map, a potential energy-driven graph evolution coupling algorithm is used to perform multiple rounds of evolutionary calculations to generate a probability prediction map of the crack evolution path;

[0014] S5. Based on the probability prediction map of the crack evolution path, a topological damage mapping function is constructed to form a structural topological damage map, and the dynamic evolution process of the topological damage with load time is predicted;

[0015] S6. Apply the time folding algorithm in the dynamic evolution process to extract the stage evolution characteristics in the multi-cycle fatigue response and obtain the fatigue state latent space representation;

[0016] S7. Compare the real-time monitoring data with the fatigue state latent space representation input into the twin variational autoencoder network, invert the topological damage state of the current structure, update the digital twin model, and finally output the structural damage evolution trend.

[0017] Optionally, the S1 specifically includes:

[0018] S11. Collecting structural parameters of the aviation elastic sheet, wherein the structural parameters include geometric dimensions, connection configuration, boundary conditions, and component distribution;

[0019] S12. Collecting material properties of the aviation elastic sheet, wherein the material properties include elastic modulus, Poisson's ratio, density, fatigue strength, and crack growth parameter;

[0020] S13. Based on the structural parameters and material properties, a finite element model is constructed using a three-dimensional solid modeling method. High-risk areas are identified by analyzing the initial stress distribution of the structure. The high-risk areas include hole edges, component corners, and connection boundaries. Stress gradients are used to perform local mesh refinement in the high-risk areas, and the size of the refined elements varies proportionally with the stress gradient threshold. The overall model is divided into multiple material property sections according to the process connection type, and a fatigue damage evolution parameter set is assigned to each section.

[0021] S14, applying fixed boundary conditions and load conditions under the mission profile to the finite element model, and performing modal analysis, wherein the modal analysis includes extracting first-order to high-order modal frequencies and corresponding vibration shapes of the structure to obtain a reference modal result;

[0022] S15, comparing the benchmark modal result with the historical measured modal data, and modifying the finite element model by adjusting the elastic modulus, connection stiffness, and damping parameters until the difference in the calculated results meets a set error threshold;

[0023] S16. Collect the historical load spectrum of the aviation elastic sheet under the mission profile and construct a digital twin model together with the modified finite element model.

[0024] Optionally, the S3 specifically includes:

[0025] S31, synchronously calibrating the acquired real-time monitoring data with the historical load spectrum to form a composite stress response input of the structure under the current working condition;

[0026] S32. Inputting the composite stress response into the digital twin model to generate a time-varying stress tensor distribution within the structural unit;

[0027] S33. Calculate the dynamic potential energy density of the structural unit based on the dynamic crack induced energy weighted by the task spectrum:

[0028] ;

[0029] in, Represents structural unit The dynamic potential energy density, Indicates the total number of task segments, represents the task weight, Represents structural unit In the The equivalent stress under each task stage is represents the exponential function, represents the time decay factor, Represents structural unit In the The damage evolution depth under each mission stage, represents the norm;

[0030] S34, organizing the dynamic potential energy density within the structural unit into a crack-induced potential energy tensor field, wherein the crack-induced potential energy tensor field covers all structural units and forms a spatial distribution;

[0031] S35. Based on the crack-induced potential energy tensor field and in combination with the structural topological adjacency relationship, a crack development probability value of each structural unit is generated:

[0032] ;

[0033] in, Represents structural unit The crack growth probability value, represents the normalization constant, represents the topological guidance coefficient, Represents structural unit The set of adjacent units of Represents structural unit The dynamic potential energy density;

[0034] S36. Outputting a crack growth potential energy probability distribution diagram, wherein the crack growth potential energy probability distribution diagram is based on the structural coordinate system and uses a color gradient to represent the crack growth probability value of each structural unit.

[0035] Optionally, the S4 specifically includes:

[0036] S41, discretizing the obtained crack growth potential energy probability distribution graph into a graph structure, wherein the vertex set is a structural unit and the edge set is a physical connection between the structural units;

[0037] S42. Obtain the dynamic potential energy density of each structural unit vertex and calculate the connection tension coefficient of each edge , the connection tension coefficient is calculated based on the structural stiffness difference and geometric distance weighting between adjacent structural unit vertices;

[0038] S43. Calculate the dynamic potential energy density difference of all edges in the graph structure , forming a local potential energy gradient map, which is used to describe the direction and intensity of energy inhomogeneity between adjacent structural unit vertices;

[0039] S44. Establish a potential-driven graph evolution coupling algorithm and use the coupling mechanism to perform multiple rounds of evolutionary calculations:

[0040] In each round of iteration, the evolution state value of each structural unit vertex is calculated based on the current evolution state value, dynamic potential energy density difference value, and the dynamic potential energy density difference value. And the connection tension coefficient is updated, and the update rule is:

[0041] ;

[0042] in, Indicates the Structural unit during round iteration The evolutionary state value of Indicates the Structural unit during round iteration The evolutionary state value of Represents structural unit The set of adjacent units of represents the hyperbolic tangent function, represents the regulating factor;

[0043] S45. During the graph evolution iteration process, the updated results of the evolution state values ​​of each structural unit are recorded to form a state distribution set on the graph structure; when the evolution state change amplitude of all structural units is lower than the set threshold or the maximum number of iterations is reached, the iteration is terminated;

[0044] S46, normalizing the evolution state value of the final structural unit to form a state distribution diagram representing the crack growth trend intensity;

[0045] S47. Based on the state distribution diagram, identifying connection paths with increasing evolution state values ​​of adjacent structural units, and constructing a set of candidate crack evolution paths;

[0046] S48. For each candidate crack evolution path, calculate the average value of the evolution state values ​​of the structural units included as the path score, and normalize all path scores to obtain the crack evolution probability value of each candidate crack evolution path, where the crack evolution probability value represents the relative tendency intensity of the crack extending along the corresponding candidate crack evolution path; based on the set of candidate crack evolution paths and the corresponding crack evolution probability values, output a probability prediction map of the crack evolution path.

[0047] Optionally, the S5 specifically includes:

[0048] S51. Based on the probability prediction map of the crack evolution path, extract all candidate crack evolution paths and corresponding crack evolution probability values;

[0049] S52. Construct a topological damage mapping function for weighted accumulation of the frequency of occurrence of the structural unit in multiple candidate crack evolution paths and the crack evolution probability value to calculate the topological damage value:

[0050] ;

[0051] in, Represents structural unit The topological damage value of represents the exponential function, represents the adjustment factor, represents the candidate crack evolution path The crack evolution probability value is summed up over all structural units. The set of candidate crack evolution paths;

[0052] S53, calculating the topological damage value of all structural units to form a structural topological damage graph, wherein the structural topological damage graph has the structural units as vertices, the physical connections as edges, and the topological damage value as the attribute weight of each vertex;

[0053] S54, combining the structural topological damage map with the historical load time series, and using a time stepping mechanism to predict the iterative change process of the topological damage value with the flight load cycle;

[0054] S55. At each time step, the topology damage value is dynamically updated according to the current topology damage value, the connection relationship between adjacent structural units, and the propagation direction of the historical crack path to form a topology damage evolution state sequence, and the dynamic evolution result of the structural topology damage diagram with load time is output.

[0055] Optionally, the S6 specifically includes:

[0056] S61. Using the dynamic evolution results of the structural topology damage diagram along with the load time as the input of the time folding algorithm, and organizing it into a fatigue damage evolution data sequence according to the time dimension;

[0057] S62, performing window division on the fatigue damage evolution data sequence, extracting multiple evolution segments of fixed period length, and forming a set of periodic fatigue response segments;

[0058] S63, applying a temporal feature compression mechanism to each cycle fatigue response segment to fold the dynamic change pattern into a low-dimensional representation, wherein the folding operation preserves the dynamic evolution characteristics through feature fusion, wherein the dynamic evolution characteristics include stage turning points, change slopes, and incremental mutations;

[0059] S64. Compare and analyze the folded representations of all cyclic fatigue response fragments to identify the stage boundaries, state mutation nodes and similar evolution patterns in the fatigue evolution process, and form stage evolution labels;

[0060] S65. Establish a correspondence between the folded representation of each cycle fatigue response segment and the corresponding stage label, and construct a fatigue state latent space representation to describe the representative fatigue state characteristics of the structure at different evolution stages.

[0061] Optionally, the contrastive twin variational autoencoder network includes two input channels, respectively for receiving a structural feature vector from real-time monitoring data and a representative fatigue state feature vector from a fatigue state latent space representation;

[0062] Each input channel is equipped with an encoder for extracting temporal features. Each encoder is connected to a variational reparameterization module, which maps the encoder output to a mean vector and a variance vector, and uses a differentiable sampling method to generate latent variable samples. The cosine similarity of the latent variable samples of the two input channels is calculated as a contrast loss to enhance the representation consistency of the latent space. The contrastive twin variational autoencoder network parameter training is completed through joint optimization.

[0063] The contrast twin variational autoencoder network includes a decoder, which is composed of multiple layers of deconvolution layers and activation function layers and is used to invert the topological damage state of the current structure.

[0064] Optionally, the updating of the digital twin model specifically includes mapping the structural topological damage state obtained by inversion to the finite element structure of the digital twin model, adjusting the stiffness parameters of the damaged unit, correcting the boundary conditions of crack evolution, and finally outputting the structural damage evolution trend.

[0065] The beneficial effects of the present invention are:

[0066] First, the present invention introduces crack-induced potential energy tensor field modeling, integrates historical load spectra and real-time monitoring stress responses based on the idea of ​​non-local mechanics, and realizes the mapping from original monitoring data to physical crack risk areas, effectively reflecting the energy driving mechanism of damage initiation and laying a high-resolution physical foundation for subsequent path evolution modeling.

[0067] Secondly, the present invention employs a potential-driven graph evolution coupling algorithm to construct a crack path prediction map based on the topological connectivity of structural units and the distribution of potential energy gradients. This allows for probabilistic modeling and directional identification of crack propagation paths. Compared to traditional single-point crack life prediction methods, this graph structure modeling can reflect the global evolutionary trends of possible crack propagation in complex structures, improving the reliability and engineering applicability of damage path prediction.

[0068] In addition, the present invention constructs a topological damage mapping function and a structural topological damage map, couples them with the flight load time series, introduces a time-stepping mechanism to simulate the dynamic evolution process of topological damage, and realizes the time-series prediction from structural response data to topological damage evolution trend, which significantly enhances the adaptability and dynamic feedback capability of the digital twin model to the actual service process.

[0069] This invention also incorporates a time-folding algorithm to extract the phased characteristics of multi-cycle fatigue response and constructs a latent space representation of the fatigue state. This, combined with a contrastive twin variational autoencoder network for structural state inversion, enables the digital twin model to continuously correct and update the structural state based on monitoring data. Ultimately, the system outputs structural damage evolution trends and remaining fatigue life predictions, providing strong technical support for proactive maintenance and life management of aviation equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 This is the overall flow chart of the intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins proposed in the present invention;

[0072] Figure 2 This is a flowchart for constructing the crack-induced potential energy tensor field for the intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins proposed in the present invention;

[0073] Figure 3 This is a schematic diagram of the structure of the comparative twin variational autoencoder network for the intelligent prediction method of fatigue damage evolution of aviation elastic sheets based on digital twins proposed in the present invention. DETAILED DESCRIPTION

[0074] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0075] refer to Figure 1-Figure 3 The intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins includes the following steps:

[0076] S1. Collect the structural parameters, material properties, and historical load spectra of the aviation elastic sheet, build a finite element model, perform modal calibration, and establish a digital twin model;

[0077] S2. Real-time collection of stress response and guided wave signals during flight to obtain real-time monitoring data;

[0078] S3. Inputting the real-time monitoring data into the digital twin model, constructing a crack-induced potential energy tensor field, and obtaining a crack development potential energy probability distribution map;

[0079] S4. Based on the crack development potential energy probability distribution map, a potential energy-driven graph evolution coupling algorithm is used to perform multiple rounds of evolutionary calculations to generate a probability prediction map of the crack evolution path;

[0080] S5. Based on the probability prediction map of the crack evolution path, a topological damage mapping function is constructed to form a structural topological damage map, and the dynamic evolution process of the topological damage with load time is predicted;

[0081] S6. Apply the time folding algorithm in the dynamic evolution process to extract the stage evolution characteristics in the multi-cycle fatigue response and obtain the fatigue state latent space representation;

[0082] S7. Compare the real-time monitoring data with the fatigue state latent space representation input into the twin variational autoencoder network, invert the topological damage state of the current structure, update the digital twin model, and finally output the structural damage evolution trend.

[0083] By systematically integrating structural parameters, material properties, historical load spectra, and real-time monitoring data, a digital twin model with real-time response capabilities was constructed, accurately reflecting the dynamic mechanical behavior of aviation elastic sheets under complex service conditions. Through a closed-loop data-driven model feedback loop throughout the entire process, an intelligent prediction chain from crack induction, path evolution, topological damage extension, to fatigue life assessment was achieved. This effectively overcomes the shortcomings of traditional fatigue analysis, such as prediction lag, uncontrollable paths, and non-updated models, significantly improving the accuracy, robustness, and engineering applicability of damage prediction.

[0084] In this embodiment, S1 specifically includes:

[0085] S11. Collecting structural parameters of the aviation elastic sheet, wherein the structural parameters include geometric dimensions, connection configuration, boundary conditions, and component distribution;

[0086] S12. Collecting material properties of the aviation elastic sheet, wherein the material properties include elastic modulus, Poisson's ratio, density, fatigue strength, and crack growth parameter;

[0087] S13. Based on the structural parameters and material properties, a finite element model is constructed using a three-dimensional solid modeling method. High-risk areas are identified by analyzing the initial stress distribution of the structure. The high-risk areas include hole edges, component corners, and connection boundaries. Stress gradients are used to perform local mesh refinement in the high-risk areas, and the size of the refined elements varies proportionally with the stress gradient threshold. The overall model is divided into multiple material property sections according to the process connection type, and a fatigue damage evolution parameter set is assigned to each section.

[0088] S14, applying fixed boundary conditions and load conditions under the mission profile to the finite element model, and performing modal analysis, wherein the modal analysis includes extracting first-order to high-order modal frequencies and corresponding vibration shapes of the structure to obtain a reference modal result;

[0089] S15, comparing the benchmark modal result with the historical measured modal data, and modifying the finite element model by adjusting the elastic modulus, connection stiffness, and damping parameters until the difference in the calculated results meets a set error threshold;

[0090] S16. Collect the historical load spectrum of the aviation elastic sheet under the mission profile and construct a digital twin model together with the modified finite element model.

[0091] This step established a highly refined and physically reliable finite element model, and through a modal calibration process, the model's calculated results were highly consistent with the measured data. The use of stress gradient-driven local mesh refinement significantly improved the modeling accuracy of crack initiation in high-risk areas. Multi-material property partitioning and fatigue parameter assignment enabled the model to truly reflect the connection characteristics and fatigue differences between components, providing a high-fidelity structural foundation for subsequent crack-induced potential energy modeling and enhancing the predictive accuracy and stability of the digital twin model.

[0092] In this embodiment, S3 specifically includes:

[0093] S31, synchronously calibrating the acquired real-time monitoring data with the historical load spectrum to form a composite stress response input of the structure under the current working condition;

[0094] S32. Inputting the composite stress response into the digital twin model to generate a time-varying stress tensor distribution within the structural unit;

[0095] S33. Calculate the dynamic potential energy density of the structural unit based on the dynamic crack induced energy weighted by the task spectrum:

[0096] ;

[0097] in, Represents structural unit The dynamic potential energy density, Indicates the total number of task segments, represents the task weight, Represents structural unit In the The equivalent stress under each task stage is represents the exponential function, represents the time decay factor, Represents structural unit In the The damage evolution depth under each mission stage, represents the norm;

[0098] S34, organizing the dynamic potential energy density within the structural unit into a crack-induced potential energy tensor field, wherein the crack-induced potential energy tensor field covers all structural units and forms a spatial distribution;

[0099] S35. Based on the crack-induced potential energy tensor field and in combination with the structural topological adjacency relationship, a crack development probability value of each structural unit is generated:

[0100] ;

[0101] in, Represents structural unit The crack growth probability value, represents the normalization constant, represents the topological guidance coefficient, Represents structural unit The set of adjacent units of Represents structural unit The dynamic potential energy density;

[0102] S36. Outputting a crack growth potential energy probability distribution diagram, wherein the crack growth potential energy probability distribution diagram is based on the structural coordinate system and uses a color gradient to represent the crack growth probability value of each structural unit.

[0103] This step introduces a task-spectrum-weighted dynamic potential energy modeling approach for the first time in fatigue damage modeling, constructing a crack-induced potential energy tensor field that effectively captures the energy concentration trend of the structure under multi-task conditions. By constructing a probabilistic model based on the potential energy gradient, the crack sensitivity distribution of each structural unit can be physically characterized, enabling the prediction of crack initiation probability based on the non-local stress field. Compared to the traditional equivalent stress method, this method has the advantages of strong time resolution and clear energy physics interpretation, providing an accurate driving source for subsequent path prediction.

[0104] In this embodiment, the S4 specifically includes:

[0105] S41, discretizing the obtained crack growth potential energy probability distribution graph into a graph structure, wherein the vertex set is a structural unit and the edge set is a physical connection between the structural units;

[0106] S42. Obtain the dynamic potential energy density of each structural unit vertex and calculate the connection tension coefficient of each edge , the connection tension coefficient is calculated based on the structural stiffness difference and geometric distance weighting between adjacent structural unit vertices;

[0107] S43. Calculate the dynamic potential energy density difference of all edges in the graph structure , forming a local potential energy gradient map, which is used to describe the direction and intensity of energy inhomogeneity between adjacent structural unit vertices;

[0108] S44. Establish a potential-driven graph evolution coupling algorithm and use the coupling mechanism to perform multiple rounds of evolutionary calculations:

[0109] In each round of iteration, the evolution state value of each structural unit vertex is calculated based on the current evolution state value, dynamic potential energy density difference value, and the dynamic potential energy density difference value. And the connection tension coefficient is updated, and the update rule is:

[0110] ;

[0111] in, Indicates the Structural unit during round iteration The evolutionary state value of Indicates the Structural unit during round iteration The evolutionary state value of Represents structural unit The set of adjacent units of represents the hyperbolic tangent function, represents the regulating factor;

[0112] S45. During the graph evolution iteration process, the updated results of the evolution state values ​​of each structural unit are recorded to form a state distribution set on the graph structure; when the evolution state change amplitude of all structural units is lower than the set threshold or the maximum number of iterations is reached, the iteration is terminated;

[0113] S46, normalizing the evolution state value of the final structural unit to form a state distribution diagram representing the crack growth trend intensity;

[0114] S47. Based on the state distribution diagram, identifying connection paths with increasing evolution state values ​​of adjacent structural units, and constructing a set of candidate crack evolution paths;

[0115] S48. For each candidate crack evolution path, calculate the average value of the evolution state values ​​of the structural units included as the path score, and normalize all path scores to obtain the crack evolution probability value of each candidate crack evolution path, where the crack evolution probability value represents the relative tendency intensity of the crack extending along the corresponding candidate crack evolution path; based on the set of candidate crack evolution paths and the corresponding crack evolution probability values, output a probability prediction map of the crack evolution path.

[0116] A structural graph model is constructed using crack development probability maps, integrating dynamic potential energy differences with structural tension to establish an iteratively evolving graph coupling mechanism. By manipulating crack propagation directionality through nonlinear functions, the team predicts crack path deviation trends in complex structures. The resulting crack evolution path probability map exhibits strong directionality, high risk concentration, and path visualization, effectively enhancing the confidence of predicted paths and the refinement of structural health management.

[0117] In this embodiment, the S5 specifically includes:

[0118] S51. Based on the probability prediction map of the crack evolution path, extract all candidate crack evolution paths and corresponding crack evolution probability values;

[0119] S52. Construct a topological damage mapping function for weighted accumulation of the frequency of occurrence of the structural unit in multiple candidate crack evolution paths and the crack evolution probability value to calculate the topological damage value:

[0120] ;

[0121] in, Represents structural unit The topological damage value of represents the exponential function, represents the adjustment factor, represents the candidate crack evolution path The crack evolution probability value is summed up over all structural units. The set of candidate crack evolution paths;

[0122] S53, calculating the topological damage value of all structural units to form a structural topological damage graph, wherein the structural topological damage graph has the structural units as vertices, the physical connections as edges, and the topological damage value as the attribute weight of each vertex;

[0123] S54, combining the structural topological damage map with the historical load time series, and using a time stepping mechanism to predict the iterative change process of the topological damage value with the flight load cycle;

[0124] S55. At each time step, the topology damage value is dynamically updated according to the current topology damage value, the connection relationship between adjacent structural units, and the propagation direction of the historical crack path to form a topology damage evolution state sequence, and the dynamic evolution result of the structural topology damage diagram with load time is output.

[0125] This step proposes a topological damage mapping function, integrating the multi-path crack evolution probability into the topological damage value of the structural unit. This creates a structural topological damage map and combines it with the flight load spectrum for dynamic evolution modeling, enabling accurate prediction of the spatial trend of damage expansion over time. Compared to traditional single-moment prediction methods based on stress fields, this method introduces a topological map and a time-stepping mechanism to form a traceable, evolvable, and feedback-based damage prediction system, enhancing the temporal continuity and early warning capabilities of structural state evolution.

[0126] In this embodiment, S6 specifically includes:

[0127] S61. Using the dynamic evolution results of the structural topology damage diagram along with the load time as the input of the time folding algorithm, and organizing it into a fatigue damage evolution data sequence according to the time dimension;

[0128] S62, performing window division on the fatigue damage evolution data sequence, extracting multiple evolution segments of fixed period length, and forming a set of periodic fatigue response segments;

[0129] S63, applying a temporal feature compression mechanism to each cycle fatigue response segment to fold the dynamic change pattern into a low-dimensional representation, wherein the folding operation preserves the dynamic evolution characteristics through feature fusion, wherein the dynamic evolution characteristics include stage turning points, change slopes, and incremental mutations;

[0130] S64. Compare and analyze the folded representations of all cyclic fatigue response fragments to identify the stage boundaries, state mutation nodes and similar evolution patterns in the fatigue evolution process, and form stage evolution labels;

[0131] S65. Establish a correspondence between the folded representation of each cycle fatigue response segment and the corresponding stage label, and construct a fatigue state latent space representation to describe the representative fatigue state characteristics of the structure at different evolution stages.

[0132] By proposing a time folding analysis strategy for fatigue cyclic evolution, we can effectively identify the stage characteristics and potential mutation nodes in the fatigue response process. By forming a latent space representation through multi-cycle compression expression, we provide a structural basis for the stage division, cluster analysis and change trend modeling of fatigue state. Compared with the single sequence model, this method has the advantages of strong stage recognition ability, high feature compression rate, and sensitivity to evolution trends, providing a refined and interpretable time series representation for subsequent damage state identification.

[0133] In this embodiment, the contrastive twin variational autoencoder network includes two input channels, respectively for receiving a structural feature vector from real-time monitoring data and a representative fatigue state feature vector from a fatigue state latent space representation;

[0134] Each input channel is equipped with an encoder for extracting temporal features. Each encoder is connected to a variational reparameterization module, which maps the encoder output to a mean vector and a variance vector, and uses a differentiable sampling method to generate latent variable samples. The cosine similarity of the latent variable samples of the two input channels is calculated as a contrast loss to enhance the representation consistency of the latent space. The contrastive twin variational autoencoder network parameter training is completed through joint optimization.

[0135] The contrast twin variational autoencoder network includes a decoder, which is composed of multiple layers of deconvolution layers and activation function layers and is used to invert the topological damage state of the current structure.

[0136] This network architecture uses dual input channels to process real-time monitoring data and latent space representations, respectively. It achieves differential enhancement and spatial fitting between fatigue states through variational coding and a contrastive loss mechanism. Unlike traditional neural networks, this architecture achieves compressed representation and similarity matching of structural states in the latent space, while also possessing strong damage inversion capabilities and trainability. A decoder inverts the topological damage map, enabling visual output of the damage state and providing an accurate input source for real-time updates of the digital twin model.

[0137] In this embodiment, the updating of the digital twin model specifically includes mapping the structural topological damage state obtained by inversion to the finite element structure of the digital twin model, adjusting the stiffness parameters of the damaged unit, correcting the boundary conditions of crack evolution, and finally outputting the structural damage evolution trend.

[0138] The structural topological damage state obtained by inversion is mapped to the digital twin finite element model, and the stiffness parameters and boundary conditions of the damaged elements are modified, achieving a closed-loop update of the digital twin model's structural state. By re-driving the model for crack propagation simulation and response analysis, the structural damage evolution trend and remaining fatigue life prediction results are ultimately output. This method integrates the entire chain from data monitoring to damage identification to model modification to prediction output, improving the real-time and adaptability of life assessment.

[0139] Example 1:

[0140] To verify the feasibility of the present invention, it was applied to an aviation elastic sheet structure inside the main wing of a medium-range unmanned aerial platform. As one of the main load-bearing components, this structure is subjected to long-term cyclic bending, torsion, and impact loads. Its service environment is complex, and microcracks caused by multi-source stress coupling often initiate and gradually expand during the flight mission cycle, creating a multi-path damage risk. Traditional fatigue life prediction methods have limited recognition capabilities in such scenarios, especially in predicting crack propagation paths and updating topological structure states, where there is a significant lag.

[0141] In this example, component geometry, connection configurations, and boundary conditions are first extracted through 3D laser scanning and assembly drawings. An initial finite element model is then constructed using material experimental data. Modal test data is then used to refine the model, resulting in a high-fidelity digital twin. Multiple sets of strain gauges and guided wave sensors are deployed in vulnerable areas of the component to collect real-time stress response and guided wave data during flight, which serve as input signals to drive the twin model's evolution.

[0142] The method first calculates the crack-induced potential energy tensor field and establishes a coupled graph model between the energy density distribution and the structural topological connectivity. Subsequently, a probabilistic prediction map of the structural crack evolution path is iteratively generated through a graph evolution mechanism, effectively identifying high-risk paths and estimating propagation trends. A topological damage mapping function is constructed in conjunction with the path map, forming a structural topological damage map. This is then combined with the load spectrum to perform a time-step evolution simulation, outputting the dynamic damage distribution process over the load cycle.

[0143] In order to improve the accuracy of fatigue state identification, a time folding algorithm is introduced to extract the stage features in multiple fatigue cycles. After forming a latent space representation, it is input into the twin variational autoencoder network together with real-time monitoring data to invert the current topological damage state. Finally, the digital twin model is updated and the structural damage evolution trend and remaining life prediction results are output.

[0144] By tracking the fatigue status of multiple flight components, it was found that this method has significantly higher spatial accuracy in crack path prediction than traditional methods, especially in terms of crack starting point identification offset, life prediction accuracy and prediction advance time.

[0145] Table 1 Performance comparison analysis of the traditional method and the method of the present invention in fatigue prediction of aviation elastic sheets

[0146]

[0147] Analyzing the data in Table 1 above, in actual flight missions of multiple aviation elastic sheet components, by comparing the traditional fatigue life prediction method with the intelligent prediction method based on digital twins and graph evolution proposed in this invention, it can be clearly seen that the present invention has significant advantages in damage identification accuracy and life prediction accuracy. Taking component A-01 as an example, its actual crack starts at the edge of the hole. The traditional method has an identification offset of 6.2 mm at this location, while this method is only 1.1 mm, and the spatial error is reduced by more than 80%. This trend is also clearly reflected in components A-02 and A-05, where the identification offset is reduced from 5.1 mm and 6.8 mm to 0.8 mm and 1.3 mm, respectively, demonstrating the high sensitivity and stability of this method in identifying the crack starting position.

[0148] In terms of life prediction, the present invention also significantly improves accuracy. The traditional method achieved a life deviation of 23.8% for the A-02 component, but this method reduced it to 7.0%. The error for the A-03 component dropped from 17.2% to 5.3%. This difference demonstrates that the traditional method is limited by the static stress assumption and linear expansion model, making it difficult to handle the nonlinear characteristics of the structural state evolving over time. The present invention, however, integrates the structural topology, crack propagation path probability, and fatigue latent space characteristics to achieve dynamic damage modeling and trend prediction, making the evaluation results more closely aligned with the actual service state.

[0149] The lead time for life prediction, another key indicator, fully demonstrates the practical engineering value of this invention. The life prediction for the A-02 component was 59 hours earlier than traditional methods, effectively improving proactive maintenance preparation and replacement scheduling. The lead time for the A-05 component was 56 hours, providing the user with ample processing time.

[0150] Overall, the present invention significantly outperforms traditional technologies in multiple dimensions, including spatial identification of component fatigue damage, life prediction deviation control, and prediction timeliness, verifying its application feasibility and practical engineering value under complex flight loads.

[0151] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins, characterized by: The steps include: S1. Collect the structural parameters, material properties, and historical load spectra of the aviation elastic sheet, build a finite element model, perform modal calibration, and establish a digital twin model; S2. Real-time collection of stress response and guided wave signals during flight to obtain real-time monitoring data; S3. Inputting the real-time monitoring data into the digital twin model, constructing a crack-induced potential energy tensor field, and obtaining a crack development potential energy probability distribution map; S4. Based on the crack development potential energy probability distribution map, a potential energy-driven graph evolution coupling algorithm is used to perform multiple rounds of evolutionary calculations to generate a probability prediction map of the crack evolution path; S5. Based on the probability prediction map of the crack evolution path, a topological damage mapping function is constructed to form a structural topological damage map, and the dynamic evolution process of the topological damage with load time is predicted; S6. Apply the time folding algorithm in the dynamic evolution process to extract the stage evolution characteristics in the multi-cycle fatigue response and obtain the fatigue state latent space representation; S7. Compare the real-time monitoring data with the fatigue state latent space representation input into the twin variational autoencoder network, invert the topological damage state of the current structure, update the digital twin model, and finally output the structural damage evolution trend; The S3 specifically includes: S31, synchronously calibrating the acquired real-time monitoring data with the historical load spectrum to form a composite stress response input of the structure under the current working condition; S32. Inputting the composite stress response into the digital twin model to generate a time-varying stress tensor distribution within the structural unit; S33. Calculate the dynamic potential energy density of the structural unit based on the dynamic crack induced energy weighted by the task spectrum: ； in, Represents structural unit The dynamic potential energy density, Indicates the total number of task segments, represents the task weight, Represents structural unit In the The equivalent stress under each task stage is represents the exponential function, represents the time decay factor, Represents structural unit In the The damage evolution depth under each mission stage, represents the norm; S34, organizing the dynamic potential energy density within the structural unit into a crack-induced potential energy tensor field, wherein the crack-induced potential energy tensor field covers all structural units and forms a spatial distribution; S35. Based on the crack-induced potential energy tensor field and in combination with the structural topological adjacency relationship, a crack development probability value of each structural unit is generated: ； in, Represents structural unit The crack growth probability value, represents the normalization constant, represents the topological guidance coefficient, Represents structural unit The set of adjacent units of Represents structural unit The dynamic potential energy density; S36. Outputting a crack growth potential energy probability distribution diagram, wherein the crack growth potential energy probability distribution diagram is based on the structural coordinate system and uses a color gradient to represent the crack growth probability value of each structural unit.

2. The intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins according to claim 1 is characterized in that: Said S1 specifically includes: S11. Collecting structural parameters of the aviation elastic sheet, wherein the structural parameters include geometric dimensions, connection configuration, boundary conditions, and component distribution; S12. Collecting material properties of the aviation elastic sheet, wherein the material properties include elastic modulus, Poisson's ratio, density, fatigue strength, and crack growth parameter; S13. Based on the structural parameters and material properties, a finite element model is constructed using a three-dimensional solid modeling method. High-risk areas are identified by analyzing the initial stress distribution of the structure. The high-risk areas include hole edges, component corners, and connection boundaries. Stress gradients are used to perform local mesh refinement in the high-risk areas, and the size of the refined elements varies proportionally with the stress gradient threshold. The overall model is divided into multiple material property sections according to the process connection type, and a fatigue damage evolution parameter set is assigned to each section. S14, applying fixed boundary conditions and load conditions under the mission profile to the finite element model, and performing modal analysis, wherein the modal analysis includes extracting first-order to high-order modal frequencies and corresponding vibration shapes of the structure to obtain a reference modal result; S15, comparing the benchmark modal result with the historical measured modal data, and modifying the finite element model by adjusting the elastic modulus, connection stiffness, and damping parameters until the difference in the calculated results meets a set error threshold; S16. Collect the historical load spectrum of the aviation elastic sheet under the mission profile and construct a digital twin model together with the modified finite element model.

3. The intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins according to claim 1 is characterized in that: The S4 specifically includes: S41, discretizing the obtained crack growth potential energy probability distribution graph into a graph structure, wherein the vertex set is a structural unit and the edge set is a physical connection between the structural units; S42. Obtain the dynamic potential energy density of each structural unit vertex and calculate the connection tension coefficient of each edge , the connection tension coefficient is calculated based on the structural stiffness difference and geometric distance weighting between adjacent structural unit vertices; S43. Calculate the dynamic potential energy density difference of all edges in the graph structure , forming a local potential energy gradient map, which is used to describe the direction and intensity of energy inhomogeneity between adjacent structural unit vertices; S44. Establish a potential-driven graph evolution coupling algorithm and use the coupling mechanism to perform multiple rounds of evolutionary calculations: In each round of iteration, the evolution state value of each structural unit vertex is calculated based on the current evolution state value, dynamic potential energy density difference value, and the dynamic potential energy density difference value. And the connection tension coefficient is updated, and the update rule is: ； in, Indicates the Structural unit during round iteration The evolutionary state value of Indicates the Structural unit during round iteration The evolutionary state value of Represents structural unit The set of adjacent units of represents the hyperbolic tangent function, represents the regulating factor; S45. During the graph evolution iteration process, the updated results of the evolution state values ​​of each structural unit are recorded to form a state distribution set on the graph structure; when the evolution state change amplitude of all structural units is lower than the set threshold or the maximum number of iterations is reached, the iteration is terminated; S46, normalizing the evolution state value of the final structural unit to form a state distribution diagram representing the crack growth trend intensity; S47. Based on the state distribution diagram, identifying connection paths with increasing evolution state values ​​of adjacent structural units, and constructing a set of candidate crack evolution paths; S48. For each candidate crack evolution path, calculate the average value of the evolution state values ​​of the structural units included as the path score, and normalize all path scores to obtain the crack evolution probability value of each candidate crack evolution path, where the crack evolution probability value represents the relative tendency intensity of the crack extending along the corresponding candidate crack evolution path; based on the set of candidate crack evolution paths and the corresponding crack evolution probability values, output a probability prediction map of the crack evolution path.

4. The intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins according to claim 1 is characterized in that: The S5 specifically includes: S51. Based on the probability prediction map of the crack evolution path, extract all candidate crack evolution paths and corresponding crack evolution probability values; S52. Construct a topological damage mapping function for weighted accumulation of the frequency of occurrence of the structural unit in multiple candidate crack evolution paths and the crack evolution probability value to calculate the topological damage value: ； in, Represents structural unit The topological damage value of represents the exponential function, represents the adjustment factor, represents the candidate crack evolution path The crack evolution probability value is summed up over all structural units. The set of candidate crack evolution paths; S53, calculating the topological damage value of all structural units to form a structural topological damage graph, wherein the structural topological damage graph has the structural units as vertices, the physical connections as edges, and the topological damage value as the attribute weight of each vertex; S54, combining the structural topological damage map with the historical load time series, and using a time stepping mechanism to predict the iterative change process of the topological damage value with the flight load cycle; S55. At each time step, the topology damage value is dynamically updated according to the current topology damage value, the connection relationship between adjacent structural units, and the propagation direction of the historical crack path to form a topology damage evolution state sequence, and the dynamic evolution result of the structural topology damage diagram with load time is output.

5. The intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins according to claim 1 is characterized in that: The S6 specifically includes: S61. Using the dynamic evolution results of the structural topology damage diagram along with the load time as the input of the time folding algorithm, and organizing it into a fatigue damage evolution data sequence according to the time dimension; S62, performing window division on the fatigue damage evolution data sequence, extracting multiple evolution segments of fixed period length, and forming a set of periodic fatigue response segments; S63, applying a temporal feature compression mechanism to each cycle fatigue response segment to fold the dynamic change pattern into a low-dimensional representation, wherein the folding operation preserves the dynamic evolution characteristics through feature fusion, wherein the dynamic evolution characteristics include stage turning points, change slopes, and incremental mutations; S64. Compare and analyze the folded representations of all cyclic fatigue response fragments to identify the stage boundaries, state mutation nodes and similar evolution patterns in the fatigue evolution process, and form stage evolution labels; S65. Establish a correspondence between the folded representation of each cycle fatigue response segment and the corresponding stage label, and construct a fatigue state latent space representation to describe the representative fatigue state characteristics of the structure at different evolution stages.

6. The intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins according to claim 1 is characterized in that: The contrastive twin variational autoencoder network includes two input channels, respectively for receiving a structural feature vector from real-time monitoring data and a representative fatigue state feature vector from a fatigue state latent space representation; Each input channel is equipped with an encoder for extracting temporal features. Each encoder is connected to a variational reparameterization module, which maps the encoder output to a mean vector and a variance vector, and uses a differentiable sampling method to generate latent variable samples. The cosine similarity of the latent variable samples of the two input channels is calculated as a contrast loss to enhance the representation consistency of the latent space. The contrastive twin variational autoencoder network parameter training is completed through joint optimization. The contrast twin variational autoencoder network includes a decoder, which is composed of multiple layers of deconvolution layers and activation function layers and is used to invert the topological damage state of the current structure.

7. The intelligent prediction method for fatigue damage evolution of aviation elastic sheets based on digital twins according to claim 1 is characterized in that: The updating of the digital twin model specifically includes mapping the structural topological damage state obtained by inversion to the finite element structure of the digital twin model, adjusting the stiffness parameters of the damaged unit, correcting the boundary conditions of crack evolution, and finally outputting the structural damage evolution trend.

Images