System for structural feature extraction and model correction fusing convolution and self-attention mechanism

Abstract

The application discloses a structural feature extraction and model correction system fusing convolution and self-attention mechanism, relates to the field of structural engineering information technology, and introduces a convolution-self-attention fusion mechanism and a double-domain consistency constraint optimization model in the whole process of cable net structure modeling, feature extraction and parameter correction, so that the precision dynamic correction from finite element simulation data to measured response data is realized. The output of a finite element modeling subsystem is taken as a multilayer objective function, after the sensitivity principal components are extracted by a principal component screening subsystem, the system extracts local correlation features by using a convolutional neural network in the feature extraction stage, and then captures global dependence features by using a self-attention mechanism, so that the model can not only pay attention to the microscale response of a local structure, but also identify an overall dynamics mode. Compared with the traditional mode, the application introduces a disturbance consistency constraint and an adversarial training mechanism, so that the network can still maintain stable output under input disturbance.

Description

Technical Field

[0001] This invention relates to the field of structural engineering information technology, specifically to a structural feature extraction and model correction system that integrates convolution and self-attention mechanisms. Background Technology

[0002] With the widespread application of complex spatial structures in projects such as long-span buildings, cable-net roofs, and cable domes, the stress state of the structure during construction and operation differs significantly from the design model. Traditional static calculation models struggle to reflect real-time changes in the structure's dynamic characteristics. To improve the accuracy and stability of structural analysis, finite element models are typically used to solve and update the structure's dynamic behavior. However, due to complex boundary conditions, numerous node couplings, and frequent changes in cable forces during construction, significant deviations exist between the initial parameters of the finite element model and the actual structural response. Therefore, there is an urgent need to establish an adaptive model correction system based on multi-source data.

[0003] In existing technologies, finite element model correction for cable net structure construction primarily employs single sensitivity analysis or least squares inversion parameters. For example, in conventional finite element model update processes, researchers typically adjust cable lengths or node masses to fit measured frequencies, making the calculated frequencies close to the measurement results. Furthermore, with the increase in the number of sensors at the construction site, the data dimensionality and time series length significantly increase, making it difficult for traditional correction algorithms to simultaneously handle response information at different scales; for instance, the coupling relationship between local vibration modes and overall frequencies cannot be effectively modeled. Current technologies lack an intelligent system capable of automatically extracting highly sensitive features and achieving cross-domain adaptive correction under the multi-scale response characteristics of structures. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a structural feature extraction and model correction system that integrates convolution and self-attention mechanisms, thus solving the problems mentioned in the background technology.

[0005] To achieve the above objectives, this invention provides the following technical solution: a structural feature extraction and model correction system that integrates convolution and self-attention mechanisms, comprising,

[0006] The finite element modeling subsystem is used to establish a finite element model of the construction process based on the initial static parameter set of the cable net structure during construction, and to obtain finite element calculation results based on the finite element model of the construction process. The finite element calculation results are used to reflect the dynamic behavior characteristics of the cable net structure.

[0007] The objective function construction subsystem will extract the mode shape vectors at the local level, the natural frequencies at the global level, and the nodal displacements at the intermediate level based on the finite element calculation results, and form the objective function matrix.

[0008] The principal component screening subsystem is used to obtain the sensitivity tensor based on the objective function matrix and the initial static parameter set, and then obtain the sensitivity principal components and high-sensitivity combinations after matrix decomposition.

[0009] The feature extraction subsystem is used to construct a hierarchical convolutional and attention fusion feature extraction network, which converts the highly sensitive combination into a modified unified feature representation. It introduces finite amplitude perturbations based on the sensitivity tensor in the objective function domain and the modified parameter domain, and establishes consistency constraint relationships and iterative feedback to construct a training network model optimized by dual-domain perturbation and consistency constraints.

[0010] The prediction training subsystem will use the training network model optimized by dual-domain perturbation and consistency constraints to perform correction parameter prediction, error calculation and iterative update on multi-source sensor data, so as to obtain the converged correction parameter set and update the finite element model of the construction process.

[0011] Preferably, the finite element modeling subsystem includes:

[0012] The element division module is used to take the initial static parameter set of the cable net structure as input to the three-dimensional geometric model and establish a finite element model of the construction process in conjunction with the finite element analysis platform. The cable net structure includes radial cables, nodes and support ends. The radial cables are made of tension elements, the nodes are made of mass elements, and the support ends are set with variable boundary conditions. The initial static parameter set includes geometric parameters, material parameters and variable boundary conditions of the support ends.

[0013] The finite element analysis module is used to call the finite element solver to perform modal and transient analyses after the model is built, so as to obtain the finite element calculation results of the cable net structure. The finite element calculation results include several natural frequencies and corresponding mode shape vectors, as well as the displacement time history response of each node at each sampling time point.

[0014] Preferably, the objective function construction subsystem includes:

[0015] The partitioning module is used to take the first three mode shape vectors in the finite element calculation results as the local level response; take the first five natural frequencies in the finite element calculation results as the global level response; and take the program sequence of each node displacement in the finite element calculation results as the intermediate level response.

[0016] The building module is used to integrate the mode shape vectors of the local level, the natural frequencies of the global level, and the time histories of the nodal displacements of the intermediate level, and to form the objective function matrix based on the corresponding positions of the features of each level in the time and space dimensions. The objective function matrix is ​​used to describe the multidimensional coupling relationship of the objective function at each level.

[0017] Preferably, the principal component screening subsystem includes:

[0018] The sensitivity tensor module is used to take the initial static parameter set as the correction parameter set, and calculate the sensitivity tensor based on the objective function matrix and the correction parameter set using the multidimensional finite difference method. The sensitivity tensor is used to describe the joint response relationship of the objective function in the parameter space and time dimension.

[0019] The decomposition module is used to input the sensitivity tensor into the matrix decomposition program, obtain the principal components of sensitivity through singular value decomposition, find the direction with the largest singular value based on the principal component results, and take the objective function and correction parameters corresponding to the direction with the largest singular value as a high-sensitivity combination, and then form a new input-output pair by the high-sensitivity combination.

[0020] Preferably, the feature extraction subsystem includes:

[0021] The network module is used to construct a hierarchical convolutional and attention-fusion feature extraction network. Based on the input-output pairs constructed using a highly sensitive combination, it builds a feature mapping structure to generate a unified feature representation that can be mapped to the modified parameter prediction space. The hierarchical convolutional and attention-fusion feature extraction network includes an input construction unit, a convolutional feature extraction unit, a temporal modeling unit, an attention feature fusion unit, and a residual mapping unit. Specifically:

[0022] The input building unit is used to transform the objective function into an input feature tensor and pair it with the corresponding correction parameters;

[0023] The convolutional feature extraction unit and the time-series modeling unit are used to perform convolution calculations on the input feature tensor in the spatial and temporal dimensions, respectively, to extract local and dynamic features;

[0024] The attention feature fusion unit is used to perform self-attention operations on the results of convolution calculations to capture the global dependencies between principal components with different sensitivities and obtain a weighted feature matrix;

[0025] The residual mapping unit is used to perform an identity mapping between the weighted feature matrix and the input feature tensor, generating a unified feature representation by element-wise superposition.

[0026] Preferably, the feature extraction subsystem further includes:

[0027] The perturbation data generation module is used to define the objective function domain and the correction parameter domain, and introduce finite-amplitude perturbations based on the sensitivity tensor into the objective function domain and the correction parameter domain respectively. Specifically, the perturbation of the objective function domain is achieved by adding a small perturbation generated by a zero-mean Gaussian distribution to each sampling point of the unified feature representation after correction, generating the first perturbation sample, thereby expanding the perturbation of the feature layer; the perturbation of the correction parameter domain is achieved by applying a fixed percentage finite difference perturbation to the correction parameter set, generating the second perturbation sample, so as to form a dual-domain perturbation dataset of the objective function domain and the correction parameter domain.

[0028] The consistency constraint establishment module is used to define the feature domain mapping function and the parameter domain mapping function based on the dual-domain perturbation dataset. It calculates the residual value using the first perturbation sample and the second perturbation sample, uses the residual value as the perturbation feedback signal, compares the difference of the output results before and after the perturbation, and constructs the consistency constraint function. The consistency constraint function is used to realize the consistency constraint relationship between the objective function domain and the modified parameter domain.

[0029] The adversarial training and residual correction module is used to input the dual-domain perturbation dataset into the training network and set up a discriminator unit to distinguish between the original samples and the perturbation samples. By minimizing the discrimination error, the parameters of the feature domain mapping function and the parameter domain mapping function are updated. During the training iteration, the residual is fed back to the backpropagation path of the network to dynamically correct the parameter weights of the feature domain mapping function and the parameter domain mapping function. When the consistency loss function and the discriminator error converge simultaneously, the parameter update is stopped, and the training network model optimized by dual-domain perturbation and consistency constraints is obtained.

[0030] Preferably, the prediction training subsystem includes:

[0031] The measurement module is used to monitor multi-source sensor data during the construction of the cable net structure, including displacement data, cable force data, and acceleration data. The displacement data is mapped to the corresponding node degrees of freedom in the finite element model of the construction process, serving as the node displacement boundary conditions. The cable force data is converted into equivalent tension loads and applied to the endpoints of the corresponding tension elements. The acceleration data is analyzed in the frequency domain to obtain natural frequency parameters and relative displacement modes at the sensor placement locations. The relative displacement modes at the sensor placement locations represent the mode shape vector components of each mode at the node where the sensor is located.

[0032] The measured processing module is used to combine the measured natural frequency parameters, the relative displacement pattern at the sensor placement position, and the displacement data to generate the measured results. Based on the measured results, the measured objective function matrix is ​​obtained.

[0033] The modified parameter prediction module is used to calculate the modified parameter prediction values ​​through forward propagation based on the trained network model optimized by dual-domain perturbation and consistency constraints, thereby realizing the inversion process from structural representation to parameter space.

[0034] Preferably, the prediction training subsystem further includes:

[0035] The error module is used to input the predicted values ​​of the correction parameters into the finite element model of the construction process, recalculate the objective function matrix, and calculate the prediction error of the correction parameters by comparing it with the measured objective function matrix.

[0036] The iterative update module is used to iteratively update the correction parameters according to the changing trend of the prediction error. When the change of the prediction error is lower than the preset threshold after several consecutive iterations, the current correction parameter set is determined to be in a convergent state, and the converged correction parameter set is re-inputted into the finite element model of the construction process.

[0037] The present invention has the following beneficial effects:

[0038] By introducing a convolutional-self-attention fusion mechanism and a dual-domain consistency constraint optimization model throughout the cable net structure modeling, feature extraction, and parameter correction processes, high-precision dynamic correction from finite element simulation data to measured response data is achieved. The natural frequencies, mode shape vectors, and displacement time histories output by the finite element modeling subsystem are used as inputs to a multi-layer objective function. After the principal component screening subsystem extracts the sensitivity principal components, the system utilizes a convolutional neural network to extract locally correlated features during the feature extraction stage, and then uses a self-attention mechanism to capture globally dependent features. This allows the model to focus on both the microscale response of the local structure and the identification of the overall dynamic mode. Compared to traditional parameter correction methods based on unidirectional regression or static fitting, this invention, by introducing perturbation consistency constraints and adversarial training mechanisms, enables the network to maintain output stability even under input perturbations, thereby improving the model's robustness, generalization ability, and structural response prediction accuracy. This effectively achieves adaptive correction of cable net structure parameters and dynamic convergence of the model.

[0039] The perturbation data generation module and consistency constraint establishment module set in the feature extraction subsystem construct a dual-domain perturbation dataset in the objective function domain and the modified parameter domain. The perturbation amplitude is controlled by the maximum gradient of the sensitivity tensor and the finite difference step size, respectively, so that the perturbation introduction has physical meaning and controllability. The zero-mean Gaussian perturbation in the objective function domain is used to expand the response distribution range of the feature layer, and the finite amplitude perturbation in the modified parameter domain is used to enhance the local discriminability of the input parameters. By comparing the mapping results before and after perturbation through the consistency constraint function, and combining the adversarial training mechanism to minimize the discrimination error, the collaborative optimization of the feature domain mapping function and the parameter domain mapping function is realized, thereby ensuring the consistency of the network output under perturbation input and enabling the model to have perturbation invariance. This mechanism overcomes the problem of overfitting or instability of the traditional finite element-neural network coupled model under perturbation input, and enhances the self-stability and adaptability of the model under different construction stages and different environmental conditions. Attached Figure Description

[0040] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0042] Please see Figure 1 This invention provides a structural feature extraction and model correction system that integrates convolution and self-attention mechanisms, including a finite element modeling subsystem, used to establish a finite element model of the construction process based on the initial static parameter set of the cable net structure during construction, and to obtain finite element calculation results based on the finite element model of the construction process, the finite element calculation results being used to reflect the dynamic behavior characteristics of the cable net structure;

[0043] The objective function construction subsystem will extract the mode shape vectors at the local level, the natural frequencies at the global level, and the nodal displacements at the intermediate level based on the finite element calculation results, and form the objective function matrix.

[0044] The principal component screening subsystem is used to obtain the sensitivity tensor based on the objective function matrix and the initial static parameter set, and then obtain the sensitivity principal components and high-sensitivity combinations after matrix decomposition.

[0045] The feature extraction subsystem is used to construct a hierarchical convolutional and attention fusion feature extraction network, which converts the highly sensitive combination into a modified unified feature representation. It introduces finite amplitude perturbations based on the sensitivity tensor in the objective function domain and the modified parameter domain, and establishes consistency constraint relationships and iterative feedback to construct a training network model optimized by dual-domain perturbation and consistency constraints.

[0046] The prediction training subsystem will use the training network model optimized by dual-domain perturbation and consistency constraints to perform correction parameter prediction, error calculation and iterative update on multi-source sensor data, so as to obtain the converged correction parameter set and update the finite element model of the construction process.

[0047] In this embodiment of the invention, by combining finite element analysis with deep feature learning, multi-source data-driven modeling and dynamic parameter correction of cable-net structures during construction are achieved. The finite element modeling subsystem is used to establish the basic physical model of the structure. Its core is to generate a finite element model of the construction process through an initial static parameter set (including cable length, node coordinates, material parameters, etc.). For example, when the tension of a certain diameter cable changes during construction, this module can output the modal frequencies and mode shapes of the structure under that cable change in real time, reflecting the overall dynamic characteristics.

[0048] The objective function construction subsystem is used to organize the finite element calculation results into an objective function matrix that can calculate the structural response relationship. For example, the first three mode shape vectors are regarded as local responses, the first five natural frequencies as global responses, and the nodal time history responses as intermediate responses, thereby constructing a comprehensive representation of multi-dimensional structural behavior and providing standardized input for subsequent sensitivity analysis.

[0049] The principal component screening subsystem analyzes the objective function matrix and the initial static parameter set jointly, using the sensitivity tensor to determine the intensity of the influence of different objective functions on the correction parameters, and extracts sensitive principal components and highly sensitive combinations through matrix decomposition. For example, when it is found that the cable length variation has the greatest impact on the first mode shape, the system automatically identifies this combination as a highly sensitive input-output pair, reducing redundant parameter interference.

[0050] The feature extraction subsystem constructs a hierarchical convolutional and self-attention fusion network to transform highly sensitive combinations into a unified feature representation. The convolutional part is responsible for extracting local spatial features (such as the variation of node displacement), while the self-attention part captures global dependencies (such as the correlation between frequencies). The integrity of the input information is maintained through identity mapping, achieving deep fusion between feature layers.

[0051] The prediction and training subsystem utilizes a network model optimized with dual-domain perturbation and consistency constraints to predict parameters and correct errors in multi-source sensor data (displacement, cable force, and acceleration) during the construction process, and automatically updates the model through iterative feedback. For example, when a shift in the acceleration spectrum of a node is detected, the system can automatically correct the boundary stiffness parameters of that region. In summary, the collaborative operation of all subsystems in this system enables a closed-loop process from finite element model establishment, objective function extraction, feature learning to parameter prediction and correction, making the modeling and correction process of cable net structures data-driven, intelligent, and dynamic.

[0052] In a preferred embodiment of the present invention, the finite element modeling subsystem includes:

[0053] The element partitioning module is used to take the initial static parameter set of the cable net structure as input to the three-dimensional geometric model and combine it with a finite element analysis platform (such as ANSYS) to establish a finite element model of the construction process. The cable net structure includes radial cables, nodes and support ends. The three-dimensional geometric model is the initial structural morphology model established in the finite element analysis, that is, the digital shape description of the structure, used to initially define the spatial distribution relationship of the cable net structure. In other words, it is the spatial blueprint of the entire finite element calculation.

[0054] Radial cables refer to cables that are radially distributed from the center of the structure or certain core nodes; nodes refer to the intersection of radial cables, loop cables or other components; support ends refer to the boundary positions where the cable net connects to the support structure, such as cable end anchor points or supports.

[0055] The radial cables are constructed using tension elements (Cable280 elements), and the nodes are constructed using mass elements (Mass21 elements). Variable boundary conditions are set at the support ends to reflect the inertial characteristics of the cable nodes without increasing the geometric complexity of the model.

[0056] The initial static parameter set includes geometric parameters (cable length, node coordinates), material parameters (elastic modulus, cross-sectional area, mass density), and variable boundary conditions at the support ends (support stiffness, constraint conditions). These parameters are given based on design drawings or initial construction measurement data and have not been corrected on-site, hence they are called static parameters.

[0057] The variable boundary conditions at the support end mean that at the cable net support points (such as anchor points and tower support points), some degrees of freedom (such as horizontal displacement) are constrained and fixed, while some degrees of freedom (such as vertical displacement or rotation) remain released; at the same time, these boundary conditions can be dynamically adjusted as the construction stage progresses.

[0058] The initial static parameter set will be identified as candidate parameters that can be used for correction after the modeling is completed.

[0059] This invention establishes a finite element model of a cable net structure based on conceptual modeling in the finite element analysis platform ANSYS, using Cable280 elements to simulate the cables. The elastic modulus, density, and diameter of each cable are based on measured data. The mass of each cable net node is simplified to point mass and set at each cable connection using mass21 elements. Specifically, in the finite element platform, parametric constraints are set for the degrees of freedom of the support points. Then, in the loading conditions, displacement constraint values ​​that change with time or loading condition number are defined (e.g., releasing the X-direction degree of freedom in stage 1 and applying a certain displacement in stage 2). Using load step control or a script interface, the boundary conditions during the construction tensioning process are gradually applied, so that the model can simulate the tensioning or release behavior during construction in stages, rather than a fixed support.

[0060] During modeling, a line model was first drawn based on the actual cable net structure's geometric parameters using ANSYS's DesignModeler module, and cross-sectional parameters were specified for each cable. Next, material properties were assigned to each cable and nodal masses were set in the static structure, along with corresponding long-range displacements based on the tensioning scheme. Finally, this static structure was used as the prestressing environment for modal analysis, completing the establishment of a finite element model of the construction process. The purpose was to construct a realistic numerical model of the cable net structure within the finite element platform ANSYS, serving as the foundational data source for objective function calculation and subsequent machine learning corrections in this invention.

[0061] The Cable280 element in ANSYS is specifically designed to simulate three-dimensional tension elements that only bear tensile forces. Using the Cable280 element accurately simulates the stress behavior of cables. Specifically, the Cable280 only bears tensile forces (not compressive forces). Once the element is under compression, the program automatically disables it, meaning it no longer transmits force. This perfectly matches the physical characteristics of cable, rope, and cord structures: they can only be pulled, not pushed, reflecting the geometric nonlinear effects of cables. This makes the model more closely approximate real cable structures when solving for tension distribution, modal frequencies, and mode responses, allowing for reliable modal analysis and dynamic response data at the physical level. For example, using Cable280, the system can accurately simulate cable sag, cable force changes, vibration modes, and other characteristics, providing accurate response curves for subsequent sensitivity tensor calculations. In a cable net structure, each radial cable and loop cable is discretized from several Cable280 elements; these tension elements together form the structural framework. In the analysis, the tension changes of these elements directly affect the overall stiffness matrix, which in turn affects the modal frequencies and mode responses. For example, if the length of a radial cable in a cable net roof changes slightly, the tension change calculated by the Cable280 elements will cause a shift in the natural frequency. This shift data will then be used by subsequent modules for sensitivity tensor calculation and parameter correction.

[0062] The Mass21 element is a lumped mass element in ANSYS. It defines the mass at nodes. Specifically, the intersections of a cable net can typically be considered "mass concentration points," but their mass originates from connectors, clamps, or localized reinforcement components. Using the Mass21 element concentrates this mass at the nodes without affecting the definition of the overall stiffness matrix. The aim is to allow the model to accurately reflect the structural inertial characteristics without increasing geometric complexity. For example, during modal analysis, the Mass21 definition enables the system to correctly calculate the nodal acceleration response and natural frequencies. In cable net structures, it is used to simulate the localized concentrated mass at nodes, such as connectors, clamps, and fasteners. During modal analysis, these Mass21 elements ensure that the calculated mode shapes and frequencies accurately reflect the true inertial distribution of the structure. For instance, if a metal connector (mass 2 kg) is installed at a cable net node, its geometry does not need to be drawn in the finite element method. Simply define a Mass21 element at the node and assign it a mass of 2 kg; the model will then accurately reflect the nodal inertia during modal analysis.

[0063] Variable boundary conditions are set to simulate the construction tensioning process. Specifically, the definition of remote displacement means simulating the control action of the tensioning machine or hydraulic jack to cause the cable end to produce a specified displacement. At the same time, setting partial degree of freedom release at the support end (such as freedom around the Y-axis) ensures that the model is consistent with the actual installation connection method. The purpose is to reproduce the process of the cable net gradually tensioning to form a force balance during the construction stage, so that subsequent analysis is based on the real boundary.

[0064] The finite element analysis module is used to call the finite element solver to perform modal and transient analyses after the model (construction process finite element model) is built, so as to obtain the finite element calculation results of the cable net structure. The finite element calculation results include several natural frequencies and corresponding mode shape vectors, as well as the displacement time history response of each node at each sampling time point. The finite element calculation results are used to characterize the dynamic behavior of the cable net structure at multiple scales and serve as the basic input data for subsequent objective function construction, sensitivity analysis and model correction.

[0065] Specifically, regarding the modal analysis section, in solving the eigenvalues ​​( On the issue: Where K is the structural stiffness matrix, describing the structure's resistance to deformation, and M is the mass matrix, describing the structure's mass distribution. The natural frequency (the quantity to be solved). The mode shape vector represents the structural deformation mode at the natural frequency. This analysis outputs several natural frequencies of the structure and their corresponding mode shape vectors. The meaning of the equation is: at a certain specific frequency... Below, the stiffness and inertia of the structure are mediated through... Equilibrium. The calculation involves solving |K- M|=0, find the frequency and corresponding mode shape; where, (K- M) =0, which indicates that the system of equations is homogeneous and has no non-zero solutions. =0, but this has no physical meaning (all nodes do not move). For a non-zero solution to exist, the following must be satisfied: |K- M|=0, this is the characteristic equation;

[0066] A finite element solver is a computational program or core algorithm module used to solve discretized finite element equations.

[0067] Mode shape vectors describe the displacement distribution of a structure in a specific mode (i.e., the relative motion of each node in that mode). Lower-order modes (such as 1st to 3rd) correspond to the main deformation modes of the structure as a whole, have a directional effect on correction parameters, and are usually most sensitive to the tension state during construction. Higher-order modes correspond to local high-frequency vibrations, have limited impact on large-scale construction, and are difficult to measure. Specifically, for the first-order mode among the first three modes, it usually corresponds to the overall rigid body or overall translational / oscillating mode; the second and third-order modes often correspond to the main bending, torsional, or overall vibration modes. These three modes are sufficient to reflect the response characteristics of the structure in the low-frequency region (the part most easily excited). The reason they are considered local responses is that the mode shape vector is nodal-level displacement information, containing detailed local displacement modes, and is therefore classified as a local level. Selecting the first three modes can balance representativeness and computational controllability.

[0068] Natural frequencies reflect the overall dynamic characteristics of a structure under different modes. They are affected by the overall stiffness and mass distribution; changing any component parameter (such as cable length or tension) will cause a change in the global natural frequency. In general engineering vibration analysis, the first five natural frequencies can cover the main dynamic characteristics of the structure; information from lower orders (such as only considering the first 1-2 orders) is insufficient and cannot distinguish vibration characteristics in multiple directions; frequencies from higher orders (above the 10th order) often correspond to local, small-scale high-frequency modes, are not sensitive to overall stiffness and construction conditions, and instead increase the computational burden. The reason they are considered as the overall response is that natural frequencies are scalar data, directly corresponding to the overall modal stiffness level, and are therefore used as a feature at the overall level; the first five frequencies can reflect the global stiffness while avoiding redundancy.

[0069] For the transient analysis, after setting the displacement and cable force boundary conditions during the loading construction process, the displacement time history response of key nodes at each sampling time point is extracted.

[0070] The objective function construction subsystem includes:

[0071] The partitioning module is used to take the first three mode shape vectors (modal analysis results) from the finite element calculation results as the local level response; the first five natural frequencies (modal analysis results) from the finite element calculation results as the global level response; and the sequence of nodal displacements (transient analysis results) from the finite element calculation results as the intermediate level response. This yields numerical results of the structure at the local, global, and intermediate levels. In other words, these three types of results can reflect the dynamic response characteristics of the structure at different scales. Then, a sensitivity matrix is ​​constructed based on these inputs for sensitivity analysis.

[0072] The local response is the result of finite element modal analysis, describing the relative displacement direction and magnitude of each node of the structure under a certain mode. It is obtained by performing modal analysis on the finite element platform ANSYS and extracting the eigenvectors of the first three modes.

[0073] The overall response is the natural frequency of the structure's free vibration under different modes, which can reflect the overall stiffness and mass distribution characteristics. It is also obtained by performing modal analysis on the finite element platform ANSYS to read the first five natural frequencies.

[0074] The intermediate level response is the displacement curve of a node during construction as it changes over time. It is obtained by selecting representative key nodes in the construction (such as support points, mid-span nodes, and tension control nodes), performing transient calculations on a finite element platform, and outputting the displacement of these nodes at each time sampling point. The reason it serves as the intermediate level response is to connect the local and the overall, reflecting the transitional behavior during the construction loading process through the time response of the nodes.

[0075] The building module is used to integrate the mode shape vectors at the local level, the natural frequencies at the global level, and the nodal displacement time histories at the intermediate level. The objective function matrix is ​​formed by the corresponding positions of the features at each level in the time and space dimensions. The objective function matrix is ​​used to describe the multidimensional coupling relationship of the objective function at each level (local, global, and intermediate levels), that is, the dynamic response relationship of the structure at multiple scales.

[0076] The objective function matrix is ​​a three-dimensional matrix structure. In this matrix, the rows represent different types of objective functions (mode difference, frequency difference, displacement difference); the columns represent different correction parameters (cable length, elastic modulus, density, etc.); and the layers represent different time history sampling points (T1, T2, etc.). Therefore, the objective function matrix is ​​actually a multi-dimensional response mapping structure based on "objective function type × correction parameter × time sampling point". In other words, it organizes the finite element calculation results into a numerical structure that can be referenced in sensitivity calculation.

[0077] The objective functions in this invention include frequency-type objective functions, mode-type objective functions, and displacement-time-history objective functions. The frequency-type objective function is the difference between the frequency calculated by the finite element method and the measured frequency. The mode-type objective function is the difference between the finite element mode vector at the sensor position and the measured mode. The displacement-time-history objective function is the difference between each node over the entire time history.

[0078] In this embodiment of the invention, the finite element modeling subsystem and the objective function construction subsystem operate collaboratively, realizing the computable expression of the dynamic response of the cable net structure and the reconstruction of multi-scale features. The element partitioning module in the finite element modeling subsystem is used to establish the basic computational model. It transforms the initial static parameter set of the cable net structure (such as cable length, node coordinates, elastic modulus, density, etc.) into the geometric and material inputs required for finite element solution, enabling the structure to possess realistic physical properties in the computational domain. For example, when tension elements are used for radial cables and mass elements are used for nodes, the nodal inertia and the coupling relationship between cables can be accurately described. After setting variable boundary conditions at the support ends, the structural response changes corresponding to changes in boundary stiffness during the construction stage can be simulated. This step serves to provide a physically consistent basic structural model for subsequent model correction.

[0079] The finite element analysis module is used to solve the dynamic characteristics of a structure. It calculates the structure's natural frequencies and mode shapes through modal analysis to reflect the overall and local vibration characteristics; simultaneously, it calculates the displacement time history response of each node through transient analysis to describe the dynamic displacement process of the structure under external loads and boundary changes. For example, when the tension of a cable increases, the system can automatically output the frequency increase and the change in node displacement curves. Compared to existing models that rely on single static analysis, this module can obtain complete time-domain and frequency-domain response information, laying the foundation for sensitivity analysis and feature extraction.

[0080] The partitioning module in the objective function construction subsystem is used to layer the finite element calculation results according to different scales: the first three mode shape vectors represent the local response, the first five natural frequencies represent the global response, and the program sequence at nodal displacements represents the intermediate response. This layering method can reflect the multi-scale dynamic characteristics of the structure from local to global.

[0081] The construction module integrates these hierarchical results into a unified objective function matrix, where rows represent objective function types, columns represent correction parameters, and layers represent time sampling points, forming a three-dimensional numerical structure that can be referenced in sensitivity analysis. For example, when cable length variations cause mode shape deviations and frequency drifts, this matrix can quantify these response differences. Compared to traditional single-dimensional parameter fitting methods, this invention, by constructing an objective function matrix, explicitly expresses the structural dynamics in spatial, temporal, and parameter dimensions, thereby achieving a high-precision mapping from finite element modeling to feature analysis. This provides a stable and reliable computational foundation for subsequent sensitivity tensor construction and model correction.

[0082] In a preferred embodiment of the present invention, the principal component screening subsystem includes:

[0083] The sensitivity tensor module is used to take the initial static parameter set as the correction parameter set, and calculate the sensitivity tensor based on the objective function matrix and the correction parameter set using a multidimensional finite difference method. The sensitivity tensor describes the joint response relationship of the objective function in the parameter space and time dimension, and its calculation formula is as follows:

[0084] ,in, The elements of the sensitivity tensor represent the objective function. At the sampling point For the correction parameters Sensitivity Let i be the objective function (such as the nodal displacement of a certain mode shape in the first order, or the third order frequency). This is the j-th correction parameter (e.g., the length of the first radial cable). To correct for the disturbance in the parameters (usually taken as 5% of the original value). Let i be the k-th time history sampling point, i be the objective function number, j be the correction parameter number, and k be the time history sampling point number.

[0085] Obtaining the sensitivity tensor involves selecting the j-th correction parameter and applying a small perturbation to it. Then, the objective function is recalculated at each time history sampling point. The value of is used to compare the magnitude of the change in the objective function before and after the perturbation. The result is divided by the perturbation amount to obtain the rate of change of the objective function in the direction of that parameter, which is the sensitivity. By repeating this process on all objective functions (i-th dimension), all correction parameters (j-th dimension), and all time sampling points (k-th dimension), a three-dimensional sensitivity tensor is formed to characterize the joint response relationship of the objective function in the parameter space and time dimension.

[0086] The purpose of calculating the sensitivity tensor is to determine which objective functions are sensitive to which correction parameters, thus avoiding the introduction of irrelevant or low-sensitivity features into subsequent neural networks.

[0087] The modified parameter set is used as the input variable for applying perturbation in sensitivity calculation. The objective function matrix and the modified parameter set are used together to establish the sensitivity tensor mapping of the objective function in the parameter space and time dimension.

[0088] The decomposition module is used to input the sensitivity tensor into the matrix decomposition program, obtain the principal components of sensitivity through singular value decomposition, find the direction with the largest singular value based on the principal component results, and take the objective function and correction parameters corresponding to the direction with the largest singular value as a high-sensitivity combination. Then, the high-sensitivity combination is used to form a new input-output pair for subsequent model correction process.

[0089] Matrix factorization is a mathematical computation process used to decompose a high-dimensional matrix or tensor into several low-dimensional matrices or vectors to reveal the main information structure. Sensitivity tensors are three-dimensional data. To analyze the correlation between parameters using mathematical methods, they need to be transformed into a two-dimensional matrix. For example, rows represent combinations of different objective functions and sampling points; columns represent different correction parameters (cable length, elastic modulus, etc.). This dimensionality reduction and expansion process is called matrix transformation. Then, it is input into a matrix factorization program to perform singular value decomposition. Its function is to transform complex three-dimensional sensitivity data into a form that can be processed by numerical algorithms, preparing for subsequent extraction of principal components (main influencing factors).

[0090] The matrix decomposition program refers to converting the sensitivity tensor into a two-dimensional matrix and inputting it into the singular value decomposition algorithm (SVD algorithm). It can be implemented using MATLAB, Python, or the built-in matrix analysis tools of the finite element analysis platform to ensure the universality of the calculation and platform independence.

[0091] Singular Value Decomposition (SVD) is a classic matrix factorization algorithm. After executing SVD, the system outputs a sequence of singular values ​​(representing the strength of each coupling mode) and left and right singular vectors (corresponding to the main directions of change in the objective function and correction parameters, respectively). Each singular value represents the coupling strength in a "dominant direction." The larger the singular value, the more sensitive the objective function is to parameter changes in that direction. SVD performs "feature analysis" on the sensitivity tensor, extracting its dominant coupling direction. This avoids a large amount of redundant or low-sensitivity data in subsequent neural network inputs, improving training efficiency and stability.

[0092] The direction of the maximum singular value refers to the pair of left and right singular vectors corresponding to the maximum singular value in the SVD decomposition results. It represents the dominant coupling mode in the entire sensitivity matrix where "changes in the objective function are most sensitive to changes in the correction parameters." For example, assuming the sensitivity matrix shows that the first-order mode frequency is most sensitive to changes in cable length, followed by the third-order mode to changes in nodal mass, then the direction of the maximum singular value may correspond to the relationship between "cable length and the first-order mode frequency." By identifying this direction, the system can determine the most critical input-output combination for subsequent model corrections, i.e., the highly sensitive combination.

[0093] Sensitivity principal components are the main singular values ​​obtained after SVD decomposition and their corresponding vectors. They represent the strongest linear mapping direction between the objective function and the correction parameters, i.e., which objective functions have the greatest impact on which parameters. Principal components correspond to the larger singular values ​​in the singular value diagonal matrix (representing the magnitude of the sensitivity principal components). Physically, they represent the response direction that contributes the most in the high-dimensional objective function-correction parameter space. This is used to reduce low-sensitivity features, decrease computational scale, retain main features, improve the prediction accuracy of the correction parameters, and enhance the stability and convergence speed of neural network training.

[0094] The decomposition module, after obtaining the sensitivity tensor, inputs it into a matrix decomposition program for feature analysis. This program performs singular value decomposition (SVD) on the sensitivity tensor, decomposing the original high-dimensional sensitivity data into several sets of eigenvectors and corresponding singular values. Each singular value characterizes the coupling strength between the objective function and the correction parameters, while each eigenvector describes the directional distribution of sensitivity changes. The decomposition results determine the master sequence of singular values, sorting them by size. The direction with the largest singular value is selected as the principal direction of sensitivity, and the corresponding objective function and correction parameters are extracted along this principal direction, forming the objective function-parameter combination with the highest sensitivity. This highest-sensitivity combination is output as a new input-output correspondence set, serving as the basis for subsequent model correction and parameter adjustment calculations, thus achieving dimensionality reduction of high-dimensional sensitivity information and extraction of highly correlated features.

[0095] In this embodiment of the invention, the principal component screening subsystem transforms the structural model correction process from empirical parameter selection to data-driven quantitative sensitivity screening. This subsystem consists of a sensitivity tensor module and a decomposition module; their cooperation enables the automatic identification of highly sensitive feature combinations from high-dimensional response data. The sensitivity tensor module establishes a quantitative mapping relationship between the objective function and the correction parameters. By using the initial static parameter set as the input correction parameter set and calculating the sensitivity tensor based on the objective function matrix using a multidimensional finite difference method, the response rate description of the objective function under different parameter directions and time sampling points is achieved.

[0096] Here, the sensitivity tensor reflects the strength of the impact of each parameter change on the objective function output, and is the core indicator characterizing the structural response sensitivity. For example, when the radial cable length is increased by 5%, if the first-order mode displacement changes significantly while the fifth-order frequency changes only slightly, the sensitivity corresponding to the former is higher than that of the latter. In this way, the system can intuitively determine the degree of contribution of different parameters to the structural dynamic characteristics.

[0097] The decomposition module is used for feature analysis of the obtained high-dimensional sensitivity tensor. This module calls a matrix factorization program to perform singular value decomposition (SVD) on the sensitivity tensor, decomposing the complex three-dimensional sensitivity data into several eigenvectors and singular value sets. The larger the singular value, the stronger the coupling between the objective function and the correction parameters. The system determines the direction of the largest singular value by sorting and extracts the corresponding objective function and correction parameters to form a highly sensitive combination. For example, among multiple sets of parameters, if the singular value corresponding to the combination of support stiffness and second-order frequency is the largest, then this combination is considered the key factor dominating the structural response.

[0098] Through the above steps, this invention can not only effectively reduce the redundancy of high-dimensional sensitivity data, but also automatically identify the most influential feature parameter pairs. Compared with the existing model update method that relies on experience judgment of a single parameter, this solution realizes an intelligent process of extracting core features from multi-parameter, high-dimensional time-series responses through joint analysis of sensitivity tensor and singular value decomposition, thereby providing a more accurate and physically consistent input basis for subsequent feature learning and model correction.

[0099] In a preferred embodiment of the present invention, the feature extraction subsystem includes:

[0100] The network module is used to construct a hierarchical convolutional and attention-fusion feature extraction network. Based on the input-output pairs constructed using a highly sensitive combination, it builds a feature mapping structure to generate a unified feature representation that can be mapped to the modified parameter prediction space. The hierarchical convolutional and attention-fusion feature extraction network includes an input construction unit, a convolutional feature extraction unit, a temporal modeling unit, an attention feature fusion unit, and a residual mapping unit. Specifically:

[0101] The input building unit is used to transform the objective function into an input feature tensor and pair it with the corresponding correction parameters;

[0102] The convolutional feature extraction unit and the time-series modeling unit are used to perform convolution calculations on the input feature tensor in the spatial and temporal dimensions, respectively, to extract local and dynamic features;

[0103] The attention feature fusion unit is used to perform self-attention operations on the results of convolution calculations to capture the global dependencies between principal components with different sensitivities and obtain a weighted feature matrix;

[0104] The residual mapping unit is used to perform an identity mapping between the weighted feature matrix (i.e., fused features) and the input feature tensor, generating a unified feature representation by element-wise superposition;

[0105] Specifically, the input construction unit receives the input-output pairs composed of highly sensitive combinations output by the decomposition module and transforms them into input feature tensors. The input feature tensors consist of local mode shape features, global frequency features, and intermediate time history features, which are derived from the mode shape vector, natural frequencies, and key node displacement time sequence output by the preceding objective function construction subsystem, respectively. The input construction module normalizes and aligns the input feature tensors in terms of time dimension to form a unified feature input tensor structure, so as to organize the numerical features into a data structure that can be processed by the deep network.

[0106] The convolutional feature extraction unit performs two-dimensional convolution operations on the input feature tensor. The dimension of the convolution kernel is determined based on the distribution range of the principal components in the sensitivity analysis results, in order to extract the correlation patterns between local features in the spatial dimension and obtain the convolutional feature map obtained through convolution calculation. The feature map is used to reflect the highly correlated feature relationships in the local range of the objective function matrix, so as to extract local correlated features. Convolution is used because the mode shape vector and frequency sensitivity distribution have local features in the "parameter and node" space, and convolution can extract these local patterns (such as the strong response of the length change of a certain cable to the displacement of a certain mode shape node).

[0107] Two-dimensional convolution is a feature extraction method in convolutional neural networks (CNNs), which is used to extract local feature patterns by sliding a small window (i.e., the convolution kernel) on two-dimensional data.

[0108] The dimension of the convolution kernel is determined based on the distribution range of the principal components in the sensitivity analysis results to ensure that the convolution kernel can cover the objective function component most sensitive to the correction parameters. The convolution operation formula is as follows:

[0109] The specific logic is as follows: taking the input matrix X(i, k) as the center, a local region of B×N is extracted, which includes several rows (objective function components) and several columns (correction parameters or time sampling points) around the point. Each value in this local region is multiplied by the corresponding weight in the convolution kernel, and then all weighted results are summed to obtain a single output value. , representing the local response intensity of the local region, change the center position of the window, that is, adjust r and s, and repeat the above operation to obtain the entire convolutional feature map.

[0110] in, The output value at position (r, s) of the convolutional feature map represents the local highly sensitive feature response extracted under the current objective function-parameter combination, which is used to reflect the local sensitive feature patterns of the input matrix in different regions;

[0111] The convolution kernel weights represent the weight coefficients at position (m, n) within the convolution window. They are learned by the network during training and are used to control the importance of different feature points in the current neighborhood, reflecting the coupling weights between different parameters. Specifically, the acquisition of convolution kernel weights is divided into two stages (initialization stage and training stage):

[0112] Initialization Phase: At the initial stage, the system employs a sensitivity-weighted initialization strategy based on the energy distribution of the principal components in the sensitivity tensor. Specifically, if a component of the objective function has a high sensitivity to a certain correction parameter, the corresponding convolutional kernel weights are initialized with higher values; conversely, they are initialized with lower values. This ensures that the convolutional kernels are initially "biased" towards highly sensitive regions, focusing on physically critical response locations from the outset, thus avoiding training oscillations caused by random initialization.

[0113] Training Phase: During the training of the feature extraction network, the convolutional kernel weights are automatically updated through the backpropagation algorithm. The system calculates the gradient based on the prediction error of the network output (such as the prediction error of the corrected parameters) and corrects each weight. Therefore, the convolutional kernel weights are gradually optimized through the local dependency relationship between the objective function and the corrected parameters in the learning data. Since the sensitivity tensor describes the rate of change of the objective function (i.e., the partial derivative of the objective function with respect to the corrected parameters), its value is a dimensionless or weakly dimensionless normalized gradient, which is between -1 and +1 after normalization. Therefore, the convolutional kernel weights inherit this distribution feature during initialization, and their range is also between -1 and +1. Positive weights mean that increasing the parameters will lead to an increase in the objective function, negative weights mean that increasing the parameters will lead to a decrease in the objective function, weights close to zero mean that the local region has a weak influence on the objective function, and the larger the absolute value of the weight, the stronger the coupling or the stronger the sensitivity region.

[0114] B and N are the row and column dimensions of the convolution kernel, which are the height and width of the convolution kernel, respectively. They represent the range of feature regions covered by the convolution operation and are determined by the spatial distribution range of the principal components of sensitivity. The larger B×N is, the wider the range of influence considered.

[0115] In ordinary image convolution, the kernel size (such as 3×3 or 5×5) is a fixed empirical setting. However, in this system, the input is not an image, but an objective function matrix (after the objective function, finite element method × is called, after the finite element method correction parameters are called, after the finite element method × is called, and after the finite element method × is called, finite element method time sampling points are called). Different objective functions have different sensitivity distribution ranges to the parameters, so an adaptive convolution window is required. Therefore, the height and width of the convolution kernel are determined by the spatial distribution range of the principal components of sensitivity. The spatial distribution range of the principal components of sensitivity refers to the extended range of sensitivity coupling between different components of the objective function and the correction parameters. For example, if one principal component is concentrated in a local area (only affecting 3 adjacent branches) → then finite element local sensitivity is called; if another principal component is more widely distributed (affecting the entire branch group) → then finite element global sensitivity is called. This indicates that the effective area size of different principal components is different.

[0116] After determining the row and column dimensions of the convolution kernel, the finite element method B×N is used to directly set the distribution span of the principal components in the direction of the objective function and the direction of the correction parameters, so as to achieve a one-to-one correspondence between the spatial scale of the convolution kernel size and the distribution of the sensitivity principal components.

[0117] For the input tensor at position The local segment, i.e. a local region in the input feature tensor (e.g., the value of a certain objective function at a certain correction parameter), represents the value of a certain objective function component (such as mode difference or frequency difference) at the corresponding correction parameter sampling point. This convolution process extracts the joint pattern of mode shape, frequency and time history data within the local range, so that the highly sensitive region in the objective function matrix is ​​locally characterized.

[0118] The double summation symbol performs a weighted summation operation on all neighboring points covered by the convolution kernel, representing the aggregation of local information within a two-dimensional region to form a single output response;

[0119] m and n represent the offsets of the convolution kernel in the direction of the objective function component and the direction of the parameters, respectively; r and s represent the row and column positions in the convolution feature map, respectively.

[0120] The temporal modeling unit is used to input the convolutional feature map into the temporal convolutional network (TCN) and perform one-dimensional dilated convolution calculations on the features in the time dimension to capture the dynamic correlation of features evolving over time. The temporal convolution output is superimposed with the convolutional feature map to generate a joint feature representation containing both spatial and temporal dimensions. Temporal convolution is used because the temporal displacement sequence contains dynamic information (node ​​displacement changes over time), and temporal convolution can capture the time-dependent relationships.

[0121] One-dimensional dilated convolution computation refers to introducing a dilation coefficient in the time dimension. This means that during convolution computation, input points on the time series are sampled discontinuously to expand the receptive field without increasing computational cost. Each convolution output point can then perceive input data over a longer time range. For example, when the dilation coefficient is 1, adjacent sampling points (ordinary convolution) capture short-term dependencies. When the dilation coefficient is 4, sampling is performed at intervals of 3 points, followed by a finite element method (FEM) call to capture long-term dependencies. In this invention, the change in node displacement over time (i.e., the time-series response) typically includes dynamic behaviors at different time scales: short-term local fluctuations (reflecting micro-vibrations) and long-term slow trends (reflecting stress redistribution during construction). Dilated convolution can capture both of these temporal characteristics simultaneously.

[0122] The superposition operation refers to adding two feature maps (matrices) element by element. The essence of this superposition operation is to fuse spatial and temporal features to achieve joint modeling of features in both spatial and temporal domains. This structure enables the subsequent attention module to establish connections between features across multiple scales and time.

[0123] The attention feature fusion unit is used to receive the joint feature representation and perform self-attention operation on the feature dimension. By calculating the correlation weights between features, it realizes the dependency modeling of principal components with different sensitivities in the global scope to generate a weighted feature matrix, which is used to characterize the cross-feature correlation between each principal component. The self-attention mechanism is used because there may be non-local coupling relationships between different vibration modes, different frequencies, and different nodal displacements, allowing the system to pay attention to multiple global features at the same time.

[0124] The formula for calculating the attention weights when performing self-attention operations is: ,in, This is the weighted feature matrix, i.e., the attention weight matrix;

[0125] Q, U, and V represent the query, key, and value mappings of the input feature tensor, respectively, obtained through a linear transformation of the convolutional-temporal convolutional joint features. Specifically, Q indicates that the current feature vector wants to "query" the relevance of other features, obtained through a linear transformation of the feature representation output by the convolutional-temporal joint features, used to describe the focus of each principal component feature on other features; U represents the "feature of interest" possessed by each feature, obtained through another set of weight matrices mapping the feature representation that is homologous to Q, used to measure the attribute expression of the queried feature; V represents the weighted aggregated feature content, which in the cable net structure corresponds to the response feature values ​​(such as mode amplitude, nodal displacement, etc.) of each sensitivity principal component, and is the final target of the weighted summation;

[0126] The similarity between features is reflected in cable net structures, which indicates the correlation of dynamic responses between different nodes or different modes of vibration;

[0127] d is the dimension of the key vector, often taken as the square root of the feature dimension to ensure a stable numerical distribution. It is a normalization factor to prevent the gradient from becoming unstable due to excessively large dot product values; through this mechanism, global feature dependencies can be explicitly modeled, enabling the extraction of associated features between principal components with different sensitivities.

[0128] It is a correlation matrix (unnormalized attention), representing the similarity between principal components with different sensitivities, Softmax( ) is a normalization function that transforms the relevance value of each row into a probability distribution (i.e., weights), ensuring that the sum of all attention weights is 1, which facilitates subsequent weighted averaging. This is an attention weight matrix, where each element represents the attention weight of each mode shape or frequency feature on the nodal displacement feature. It involves performing a weighted average of each feature, fusing global feature information to obtain a weighted feature matrix, and achieving global feature fusion output;

[0129] The residual mapping unit is used to perform an identity mapping (i.e., element-wise superposition) between the weighted feature matrix and the input feature tensor in the input construction unit to form a corrected unified feature representation. This is used to achieve synchronous preservation of the original input information and the fused feature information during deep feature propagation. That is, when the network updates through backpropagation, it can simultaneously perceive the initial input and the deep abstract features, making the model more stable and more physically constrained.

[0130] Residual mapping is used because deep feature extraction gradually loses consistency with the original input. Residual units can "pull a straight line" to avoid network degradation and keep the original input and extracted features consistent.

[0131] This unit maintains feature consistency between input and output through identity mapping, preventing information loss during transmission;

[0132] Finally, the corrected unified feature representation serves as the output of the hierarchical convolutional and attention fusion feature extraction network, providing the input foundation for subsequent parameter correction learning and prediction. This enables the construction of an intermediate layer that maps from highly sensitive feature combinations to the parameter correction space. Specifically, the corrected unified feature representation is used to adaptively fuse local convolutional features and global attention features at spatial scales, forming a unified structural response representation across frequency, temporal, and geometric topological dimensions. This feature representation is used as a high-dimensional input vector in subsequent dual-domain mapping and parameter prediction. Through mapping training between the feature domain and the parameter domain, the mapping relationship between the structural response domain and the parameter domain is established. It also serves as a reference benchmark during adversarial training and residual correction to maintain output consistency under perturbation inputs. Ultimately, it provides corrected structural feature input in the parameter correction prediction stage to support the inversion of model parameters and the iterative convergence of the finite element model.

[0133] In traditional finite element model correction, the structural response (such as frequency and mode shape) and model parameters (such as cable length and elastic modulus) often exhibit a strongly nonlinear, multi-scale, and multi-dimensional coupling relationship. Hierarchical convolutional and attention-based feature extraction networks are designed to automatically learn this complex mapping. Composed of multiple layers (hierarchical structure), each layer undertakes different types of feature processing tasks: convolutional layers handle spatial correlation, temporal layers handle temporal evolution, attention layers handle feature weights and global dependencies, and residual layers handle information fidelity and stable propagation. Through multi-layered fusion, a unified feature representation reflecting the relationship between the objective function and the correction parameters is ultimately obtained.

[0134] In this embodiment of the invention, the feature extraction subsystem constructs a hierarchical convolutional and self-attention fusion feature extraction network, which realizes a unified mapping from highly sensitive feature combination to the modified parameter prediction space, effectively improving the feature extraction accuracy and generalization ability of the cable net structure model.

[0135] The input building unit transforms the highly sensitive combinations (objective function and correction parameter pairs) obtained from the preceding modules into a computable input feature tensor. This input tensor consists of local mode shape features (reflecting local nodal motion patterns), global frequency features (reflecting global stiffness changes), and intermediate time history features (reflecting the dynamic response process). Normalization and time-dimensional alignment ensure consistency across different feature sources in terms of spatiotemporal scales. For example, when nodal displacement and frequency change scales are asynchronous, this unit can achieve data alignment through normalization mapping, providing a unified input for subsequent convolutional processing.

[0136] The convolutional feature extraction unit is used to extract highly correlated patterns between local features in the spatial dimension. By setting the convolution kernel dimension to correspond with the sensitivity principal component distribution, the system can focus on the objective function region most sensitive to the correction parameters. For example, when sensitivity analysis shows that cable length variations mainly affect the modal response of a specific node, the convolution operation will focus on capturing the structural features of that region, thereby improving the targeting of feature selection.

[0137] The temporal modeling unit is used to analyze the dynamic evolution of features over time. It models the time series through one-dimensional dilated convolution and superimposes the results with spatial convolution to generate a joint feature representation. For example, during construction loading, the nodal response fluctuates over time. This unit can capture the temporal response features caused by changes in cable force, enabling the model to have dynamic recognition capabilities.

[0138] The attention feature fusion unit further performs self-attention operations on the joint features to learn the global dependencies between principal components with different sensitivities, forming a weighted feature matrix to represent the interaction strength between different features. For example, the system can automatically identify the nonlinear dependency between the third-order frequency and the second-node displacement, improving the model's ability to understand complex coupled behaviors.

[0139] The residual mapping unit element-wise superimposes the weighted feature matrix and the original input feature tensor using an identity mapping method to generate a corrected unified feature representation. This ensures that original information is not lost during deep propagation and prevents feature degradation caused by excessive network abstraction. For example, when the convolutional layer extracts too much local detail, the residual path can retain the original frequency information to balance the overall representation. In summary, this invention, through multi-level collaboration of hierarchical convolution, self-attention fusion, and residual mapping, enables the feature extraction process to simultaneously possess local accuracy and global relevance. Compared to traditional single convolutional or fully connected structures, it can more efficiently represent the spatiotemporal dynamic features of the cable network structure, providing highly robust feature input for corrected parameter prediction.

[0140] In a preferred embodiment of the present invention, the feature extraction subsystem further includes:

[0141] The perturbation data generation module is used to define the objective function domain and the correction parameter domain, and introduce finite-amplitude perturbations based on the sensitivity tensor into both domains. Specifically, the objective function domain perturbation adds a small perturbation generated by a zero-mean Gaussian distribution to each sampling point of the corrected unified feature representation. The amplitude of this perturbation is determined by the maximum gradient value of the sensitivity tensor (the perturbation intensity will not exceed the change amplitude at the most drastic point of sensitivity, thus simulating extreme perturbation situations while maintaining the physical stability of the model), generating the first perturbation sample to expand the perturbation of the feature layer. The correction parameter domain perturbation applies a fixed percentage of finite difference perturbation to the correction parameter set. This perturbation value is taken as a fixed percentage (e.g., 5%) of the original parameter value, specifically determined by the step size setting of the sensitivity calculation formula, generating the second perturbation sample to form a dual-domain perturbation dataset for both the objective function domain and the correction parameter domain, achieving synchronous robust enhancement of the feature domain and parameter domain.

[0142] The perturbation generated by a zero-mean Gaussian distribution refers to a random disturbance signal that follows a normal distribution. This means that a small segment of noise or disturbance is added to each sampling point. The overall mean of these disturbances is zero (they will not systematically deviate from the original characteristic distribution), but they can simulate the small uncertainties that exist in the system during actual construction or measurement. For example, if the node displacement is 0.015m, it may become 0.015±0.0003 after adding the disturbance; if the frequency is 3.00Hz, it may become 3.00±0.02Hz after adding the disturbance.

[0143] The sensitivity tensor describes the strength of the objective function's response to the parameters. The maximum gradient value represents the most sensitive direction in the system, that is, the part that is "most responsive to disturbances". Therefore, if the disturbance amplitude is set according to the maximum gradient, it can be ensured that the disturbance is sufficient to cause changes in the model response without compromising the structural stability.

[0144] The process of determining the maximum gradient value of the sensitivity tensor involves traversing all sensitivity data to identify the portion of the system response where changes are most pronounced in both the parameter space and time series. This represents the position of the objective function that is "most susceptible to perturbations" among all parameters. This maximum gradient value reflects both the model's numerical sensitivity to input perturbations and indirectly illustrates the physical vulnerability of the structure to parameter adjustments. Therefore, using this maximum gradient value as a benchmark in subsequent perturbation magnitude settings ensures that the introduced perturbations are sufficiently representative without compromising the model's stability.

[0145] In the modified parameter domain, the perturbation amplitude is determined based on the step size setting of the sensitivity calculation formula. The step size is used to calculate the finite difference increment of the objective function with respect to the rate of change of the parameter (i.e., the small increment of the parameter in the sensitivity calculation). Its value is usually 1% to 5% of the original value of the parameter, so as to simulate the physical feasible range of parameter perturbation under the premise of ensuring that the local linear approximation is valid.

[0146] The objective function domain and the correction parameter domain correspond to the structural response space (output side) and input parameter space (input side) of the objective function construction subsystem, respectively. The data in the objective function domain comes from the sensitivity principal component screening results, that is, the physical features most sensitive to the structural response. Specifically, it includes three types of features: mode shape vector features, natural frequency features, and nodal time history response features. The data in the correction parameter domain comes from the input definition of the finite element model (i.e., finite element model) during the construction process. It is an adjustable quantity used to control the structural features. In cable net structures, the correction parameters include radial cable length, nodal mass, and tension stroke.

[0147] In the objective function domain, a finite-amplitude perturbation based on the maximum gradient of the sensitivity tensor is introduced by adding a local perturbation function that follows a zero-mean Gaussian distribution to each feature sampling point; in the correction parameter domain, a finite-amplitude perturbation based on the step size of the sensitivity tensor is introduced by applying a fixed proportion of finite difference perturbation to the parameter set.

[0148] The consistency constraint establishment module is used to define feature domain mapping functions and parameter domain mapping functions based on a dual-domain perturbation dataset, so as to characterize the correspondence between input perturbation and output features.

[0149] The feature domain mapping function represents the mapping from the input feature tensor (objective function features, such as mode shape, frequency, displacement response) to its output feature space in the deep network. It is used to perform deep characterization of the input features and reflects the impact of input perturbations (structural response changes) on the model output features. The parameter domain mapping function represents the mapping from the input modified parameter set (such as cable length, nodal mass, etc.) to the model parameter prediction space. It is used to invert the feature characterization results to the modified parameter space and reflects the impact of model input parameter perturbations on the output estimation.

[0150] The process of defining the mapping function is as follows: input the first perturbation sample into the feature mapping path of the network to obtain the perturbation feature output. By inputting the original samples into the same mapping path, the corresponding benchmark output is obtained. Map the input parameters of the second perturbation sample to the path to obtain the perturbation output. By inputting their original parameters into the same path, the corresponding baseline output can be obtained. Compare the output differences before and after the perturbation, and establish a consistency constraint function;

[0151] The mapping path refers to the feature transfer route of data from input to output in the neural network structure. Since this system focuses on the bidirectional correlation between structural response features (output side) and structural parameters (input side), two independent but co-trainable mapping paths are established:

[0152] The feature mapping path is used to describe the flow and representation of features in the network starting from the feature domain (such as objective function, mode shape, frequency, nodal displacement), and reflects the variation law of structural response features (such as mode shape vector, nodal displacement, etc.) under different disturbance conditions. In other words, it is equivalent to simulating the disturbance propagation at the "structural performance" level.

[0153] The parameter mapping path is used to describe the inversion and influence of parameters on the model output, starting from the parameter domain (such as cable length, elastic modulus, mass density, and boundary stiffness), reflecting the inverse mapping relationship from "observed features" to "structural parameters". In other words, it is the main channel for achieving parameter inversion and correction in the structural model correction stage.

[0154] These two paths constitute a bidirectional mapping mechanism of "input space ↔ output space" in the network. By simultaneously inputting perturbation samples and original samples into these two paths and comparing the output results of the two, a consistency constraint relationship can be constructed to measure whether the model's output results remain consistent when facing input perturbations. If the model can maintain stable output in both paths in response to perturbation inputs, it indicates that the model has good robustness and generalization ability.

[0155] The residual values ​​are calculated using the first and second perturbation samples. The residual values ​​are used as perturbation feedback signals. The difference between the output results before and after the perturbation is compared to construct a consistency constraint function. The consistency constraint function is used to realize the consistency constraint relationship between the objective function domain and the correction parameter domain, so as to ensure that the mapping output remains consistent before and after the input perturbation, and to promote the network to maintain mapping stability under perturbation input.

[0156] To ensure the consistency of the input before and after the perturbation, a consistency constraint function is constructed:

[0157] Where L is the two-domain consistency constraint loss, used to measure the difference in output before and after the perturbation. The q-th feature is the original input (mode shape, frequency, or displacement) at time t. Let q be the perturbation input for the q-th feature at time t. This represents the j-th correction parameter. For the perturbation sample of the j-th correction parameter, f( ) is the feature domain mapping function, g( Let q be the feature number, J be the total number of correction parameters, and A be the total number of features. and These represent the characteristic outputs with and without perturbations, respectively. and These represent the parameter outputs (reference outputs) under perturbation and undisturbed conditions, respectively.

[0158] The first term represents the perturbation consistency constraint in the objective function domain, and the second term represents the perturbation consistency constraint in the modified parameter domain. By comparing the differences before and after the perturbation, the calculation process of the dual-domain consistency constraint is defined to ensure that the dataset can cover perturbation information during the training phase. The above process is based on the calculation results of the sensitivity tensor, ensuring that the perturbation range is consistent with the sensitivity direction, thereby introducing perturbation samples with physical origins into the dataset.

[0159] First summation term To ensure that within the feature domain: the squared difference of the output change before and after the input perturbation is minimized, thus ensuring feature consistency. The second summation term... To ensure parameter consistency, the output prediction changes as little as possible before and after parameter perturbations. These two factors combined constitute a dual-domain consistency constraint, forcing the network to maintain stable mappings in both the feature and parameter domains. The purpose of constructing the consistency constraint function is to minimize L, allowing the model to output approximately consistent feature responses under small perturbation inputs, thus achieving robust modeling of structural state fluctuations. This is equivalent to establishing a "perturbation-invariant mapping" in the finite element sensitivity direction, ensuring that the model output does not change drastically with minor environmental perturbations.

[0160] The residual value is used to represent the offset of each feature (i.e., mode shape, frequency or displacement) at the corresponding sampling point due to the disturbance. It should be noted that the residual value is not just a record of the difference, but is fed back into the training process to guide the network on how to correct the offset caused by the disturbance.

[0161] The adversarial training and residual correction module is used to input the dual-domain perturbation dataset into the training network and set up discriminator units to distinguish between original samples (unperturbed samples) and perturbed samples (data perturbed by the objective function domain or corrected parameter domain). By minimizing the discrimination error, the parameters of the feature domain mapping function and the parameter domain mapping function are updated, ensuring that the model maintains consistent output under perturbed input. During training iterations, the residuals are fed back to the network's backpropagation path, allowing the network to converge to the perturbation-invariant mapping after multiple iterations, dynamically correcting the parameter weights of the feature domain mapping function and the parameter domain mapping function. When the consistency loss function and the discriminator error converge simultaneously, parameter updates are stopped, resulting in the trained network model optimized by dual-domain perturbation and consistency constraints. The simultaneous convergence of the consistency loss function and the discriminator error means that the network has achieved output stability and adversarial discriminant balance under perturbed input, i.e., the network has completed the learning process of the perturbation-invariant mapping.

[0162] The training network refers to the joint training system of the entire adversarial consistency framework. It consists of a feature domain mapping function, a parameter domain mapping function, a discriminator unit, and a consistency constraint function. The discriminator unit tries its best to distinguish between perturbed samples and original samples. The feature domain mapping function and the parameter domain mapping function (main mapping network) try their best to generate samples that the discriminator unit cannot distinguish. When the training reaches a certain stage, the discriminator unit can no longer distinguish between perturbed samples and original samples (discrimination accuracy ≈ 50%), indicating that the output of the main mapping network can no longer be detected under perturbation, that is, adversarial equilibrium convergence is achieved.

[0163] The discriminator unit is a neural network used to identify the source of a sample. Its function is to determine whether the input data sample is the original sample (unperturbed sample) or the perturbed sample. Training minimizes the discriminator error (or maximizes the confusion probability) so that the network continuously optimizes its parameters, so that the output is still close to the original distribution when subjected to perturbed input. This is the core logic of adversarial consistency training.

[0164] In this embodiment of the invention, the feature extraction subsystem is further equipped with a perturbation data generation module and a consistency constraint establishment module, and combined with adversarial training and residual correction modules to form a dual-domain perturbation and consistency constraint optimization mechanism, which is used to enhance the stability and generalization performance of the model, so that the model can still maintain the reliability of the prediction results when there is input noise or parameter uncertainty.

[0165] The perturbation data generation module introduces controllable perturbations into two distinct spaces: the objective function domain (the structural response output space) and the modified parameter domain (the model input parameter space). The former expands the feature layer by adding small perturbations following a zero-mean Gaussian distribution to each sampling point of the "modified unified feature representation," thus testing the model's robustness to output perturbations. The latter applies a fixed proportion (e.g., 5%) of finite-difference perturbation to the modified parameter set to verify the model's response stability under slight fluctuations in input parameters. For example, when the cable length changes by 1% due to construction errors, the system can generate a set of perturbation samples to simulate this change, thereby testing the network's self-correction capability under small perturbation conditions.

[0166] The consistency constraint establishment module defines feature domain mapping functions and parameter domain mapping functions based on the aforementioned dual-domain perturbation dataset. The former captures feature change trends in the high-dimensional feature space, while the latter inverts these features to the corrected parameter space. By calculating the output difference before and after the perturbation (i.e., residual values), the system constructs a consistency constraint function to ensure the model maintains consistent output before and after the perturbation input. For example, if the perturbation causes a change in the intrinsic frequency but should not affect the overall trend of node displacement, the consistency constraint will prompt the network to automatically correct the mapping parameters, causing the output to return to the expected distribution. The adversarial training and residual correction module further introduces a discriminator unit during training to distinguish between original samples and perturbation samples. By minimizing the discrimination error, the mapping function parameters are continuously updated, enabling the network to learn a "perturbation-invariant mapping." For example, when a Gaussian perturbation causes a shift in the response of a node, the residual feedback will correct the weights along the backpropagation path until the model output returns to stability.

[0167] In summary, this invention, by jointly introducing perturbations in the objective function domain and the correction parameter domain, and applying consistency constraints and adversarial learning mechanisms, enables the model to possess noise resistance, bias resistance, and self-correction capabilities. Compared with existing fixed-structure networks, it significantly improves computational robustness and prediction accuracy under conditions of construction errors, measurement noise, and parameter fluctuations.

[0168] In a preferred embodiment of the present invention, the prediction training subsystem includes:

[0169] The measurement module is used to monitor multi-source sensor data during the construction of the cable net structure, including displacement data (using displacement sensors to monitor the spatial displacement of nodes), cable force data (using cable force meters to monitor the actual tension of the radial cables), and acceleration data (using accelerometers to monitor the vibration response of the structure, used to invert natural frequencies and modes). The displacement data is mapped to the corresponding node degrees of freedom in the finite element model of the construction process, serving as the node displacement boundary conditions. The cable force data is converted into equivalent tension loads and applied to the endpoints of the corresponding tension elements. The acceleration data undergoes frequency domain analysis to obtain natural frequency parameters and relative displacement modes at the sensor placement locations, which are compared with the finite element calculation results. Through the input of multi-source sensor data, the finite element model of the construction process reflects the actual construction conditions at the boundary condition level. The relative displacement modes at the sensor placement locations represent the mode shape vector components of each mode at the sensor node.

[0170] The measured processing module is used to combine the measured natural frequency parameters, the relative displacement pattern at the sensor placement position, and the displacement data to generate the measured results. Based on the measured results, the measured objective function matrix is ​​obtained.

[0171] The modified parameter prediction module is used to calculate the modified parameter prediction values ​​through forward propagation based on the trained network model optimized by dual-domain perturbation and consistency constraints. That is, the response data is directly input into the network to obtain the corresponding modified parameter prediction values ​​(such as cable length, nodal mass, boundary stiffness, etc.), realizing the inversion process from structural performance to parameter space.

[0172] Corrected parameter predictions are used to complete the mapping from the objective function to the parameters;

[0173] The prediction training subsystem also includes:

[0174] The error module is used to input the predicted values ​​of the corrected parameters into the finite element model of the construction process, recalculate the objective function matrix, and calculate the prediction error of the corrected parameters by comparing it with the measured objective function matrix. Specifically: ,in, The prediction error of the j-th correction parameter, For the j-th correction parameter prediction value, The output of the objective function is calculated based on the predicted values ​​of the corrected parameters. The actual observed values ​​are used; the difference between the predicted parameters and the measured conditions is obtained through error calculation, providing a numerical basis for parameter correction.

[0175] The iterative update module is used to iteratively update the correction parameters based on the changing trend of the prediction error. The purpose is to gradually reduce the prediction error by adjusting the correction parameters; its update relationship is as follows: ,in, For the j-th correction parameter in the current iteration step, These are the corrected parameters for the next iteration.

[0176] To predict the sensitivity of the error to the correction parameter, the partial derivative of the error with respect to the correction parameter is expressed as a finite difference, calculated using this method. It reflects the degree to which the error is sensitive to this parameter. A positive partial derivative indicates that increasing the parameter will increase the error; a negative partial derivative indicates that increasing the parameter will decrease the error.

[0177] The step size coefficient is used to update the parameter update speed control coefficient, which determines the magnitude of each parameter adjustment. The larger the value, the faster the parameter update, but it may cause oscillation; if the value is too small, the convergence speed will be too slow. The value is usually 0.01 to 0.1, which is determined by experience or trial calculation.

[0178] This indicates the direction and magnitude of parameter adjustment needed in each iteration. The presence of a negative sign indicates that the parameter moves in the direction of decreasing error. t is the iteration number, representing the current iteration round. It starts at t=0 and gradually increases until the convergence condition is met (i.e., the error change is below a preset threshold).

[0179] When the change in prediction error is lower than the preset threshold after several consecutive iterations, the current set of corrected parameters is determined to be in a convergent state. The corrected parameters obtained at this time are the final stable set of parameters. The converged set of corrected parameters is then re-inputted into the finite element model of the construction process to generate a construction process model that is consistent with the actual structural state, thereby realizing automatic correction and dynamic convergence of model parameters.

[0180] Through multiple rounds of iterative updates, the deviation between the prediction error and the measured error is gradually reduced, making the set of corrected parameters tend to stabilize.

[0181] The role of the prediction training subsystem is to establish a closed-loop mechanism for model prediction and experimental correction. Through the training network, it achieves iterative updates of parameter inversion, error calculation and sensitivity guidance, so that the parameters of the finite element model gradually approach the true state in the process of cyclic adjustment.

[0182] In this embodiment of the invention, the prediction training subsystem establishes a closed-loop mechanism of model prediction, actual measurement feedback, and parameter correction, enabling the finite element model to adaptively correct structural parameters and gradually approximate the actual construction state. The subsystem consists of an actual measurement module, an actual measurement processing module, a parameter correction prediction module, an error module, and an iterative update module. Each module forms a progressive logical chain, realizing the entire process from multi-source data acquisition to dynamic parameter convergence.

[0183] The measurement module collects multi-source sensor data from the field, including three core types of information: displacement, cable force, and acceleration. Displacement data corresponds to the degrees of freedom of the finite element model nodes and is used as displacement boundary conditions. Cable force data is converted into equivalent tension loads and applied to the endpoints of tension elements. Acceleration data is analyzed in the frequency domain to obtain natural frequency parameters and relative displacement modes. Through this process, the finite element model of the construction process reflects the actual working conditions at the boundary level. For example, during cable net tensioning, if the acceleration sensor detects frequency drift, it reflects local tension changes, providing a basis for subsequent model correction. The measurement processing module integrates the above data into a unified measured objective function matrix. This matrix integrates frequency, mode shape, and node displacement information for comparison with finite element calculation results. By constructing this matrix, the model can directly compare the differences between measured and simulated data in a multi-dimensional feature space, avoiding misjudgments based on single indicators.

[0184] The corrected parameter prediction module, based on a trained network model optimized with dual-domain perturbation and consistency constraints, calculates the predicted corrected parameters through forward propagation, achieving an inversion from structural response to parameter space. For example, when the measured frequency decreases while the mode shape change is slight, the network can automatically determine that the problem originates from "cable length increase" rather than "bracket stiffness decrease." The error module then substitutes the predicted parameters back into the finite element model, recalculates the objective function matrix, and compares it with the measured matrix to obtain the prediction error. This error is used to quantify the deviation between the prediction and the actual state, providing a precise direction for correction.

[0185] The iterative update module automatically adjusts the correction parameters based on the trend of prediction error changes, achieving dynamic model convergence. When the error falls below a preset threshold, the obtained correction parameter set reaches a stable state, and the model output is highly consistent with the measured results. For example, after multiple rounds of updates, the estimated values ​​of the cable length and node mass gradually stabilize near the true values. In summary, this invention, through multi-source measurements and a self-iterative learning mechanism, enables the model to achieve parameter self-correction and error self-repair, further improving the modeling accuracy and dynamic consistency of the cable net structure construction process compared to existing static calibration or manual adjustment methods.

[0186] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A structural feature extraction and model correction system integrating convolution and self-attention mechanisms, characterized by: include: The finite element modeling subsystem is used to establish a finite element model of the construction process based on the initial static parameter set of the cable net structure during construction, and to obtain finite element calculation results based on the finite element model of the construction process. The finite element calculation results are used to reflect the dynamic behavior characteristics of the cable net structure. The objective function construction subsystem will extract the mode shape vectors at the local level, the natural frequencies at the global level, and the nodal displacements at the intermediate level based on the finite element calculation results, and form the objective function matrix. The principal component screening subsystem is used to obtain the sensitivity tensor based on the objective function matrix and the initial static parameter set, and then obtain the sensitivity principal components and high-sensitivity combinations after matrix decomposition. The feature extraction subsystem is used to construct a hierarchical convolutional and attention fusion feature extraction network, convert the highly sensitive combination into a modified unified feature representation, introduce finite amplitude perturbation based on the sensitivity tensor in the set objective function domain and modified parameter domain, and establish consistency constraint relationship and iterative feedback to construct a training network model optimized by dual-domain perturbation and consistency constraint. The feature extraction subsystem includes: The network module is used to construct a hierarchical convolutional and attention-fusion feature extraction network. Based on the input-output pairs constructed using a highly sensitive combination, it builds a feature mapping structure to generate a unified feature representation that can be mapped to the modified parameter prediction space. The hierarchical convolutional and attention-fusion feature extraction network includes an input construction unit, a convolutional feature extraction unit, a temporal modeling unit, an attention feature fusion unit, and a residual mapping unit. Specifically: The input building unit is used to transform the objective function into an input feature tensor and pair it with the corresponding correction parameters; The convolutional feature extraction unit and the time-series modeling unit are used to perform convolution calculations on the input feature tensor in the spatial and temporal dimensions, respectively, to extract local and dynamic features; The attention feature fusion unit is used to perform self-attention operations on the results of convolution calculations to capture the global dependencies between principal components with different sensitivities and obtain a weighted feature matrix; The residual mapping unit is used to perform an identity mapping between the weighted feature matrix and the input feature tensor, generating a unified feature representation by element-wise superposition; The prediction training subsystem will use the training network model optimized by dual-domain perturbation and consistency constraints to perform correction parameter prediction, error calculation and iterative update on multi-source sensor data, so as to obtain the converged correction parameter set and update the finite element model of the construction process.

2. The structural feature extraction and model correction system integrating convolution and self-attention mechanisms according to claim 1, characterized in that: The finite element modeling subsystem includes: The element division module is used to take the initial static parameter set of the cable net structure as input to the three-dimensional geometric model and establish a finite element model of the construction process in conjunction with the finite element analysis platform. The cable net structure includes radial cables, nodes and support ends. The radial cables are made of tension elements, the nodes are made of mass elements, and the support ends are set with variable boundary conditions. The initial static parameter set includes geometric parameters, material parameters and variable boundary conditions of the support ends. The finite element analysis module is used to call the finite element solver to perform modal and transient analyses after the model is built, so as to obtain the finite element calculation results of the cable net structure. The finite element calculation results include several natural frequencies and corresponding mode shape vectors, as well as the displacement time history response of each node at each sampling time point.

3. The structural feature extraction and model correction system integrating convolution and self-attention mechanisms according to claim 2, characterized in that: The objective function construction subsystem includes: The partitioning module is used to take the first three mode shape vectors in the finite element calculation results as the local level response; take the first five natural frequencies in the finite element calculation results as the global level response; and take the program sequence of each node displacement in the finite element calculation results as the intermediate level response. The building module is used to integrate the mode shape vectors of the local level, the natural frequencies of the global level, and the time histories of the nodal displacements of the intermediate level, and to form the objective function matrix based on the corresponding positions of the features of each level in the time and space dimensions. The objective function matrix is ​​used to describe the multidimensional coupling relationship of the objective function at each level.

4. The structural feature extraction and model correction system integrating convolution and self-attention mechanisms according to claim 3, characterized in that: The principal component screening subsystem includes: The sensitivity tensor module is used to take the initial static parameter set as the correction parameter set, and calculate the sensitivity tensor based on the objective function matrix and the correction parameter set using the multidimensional finite difference method. The sensitivity tensor is used to describe the joint response relationship of the objective function in the parameter space and time dimension. The decomposition module is used to input the sensitivity tensor into the matrix decomposition program, obtain the principal components of sensitivity through singular value decomposition, find the direction with the largest singular value based on the principal component results, and take the objective function and correction parameters corresponding to the direction with the largest singular value as a high-sensitivity combination, and then form a new input-output pair by the high-sensitivity combination.

5. The structural feature extraction and model correction system integrating convolution and self-attention mechanisms according to claim 4, characterized in that: The feature extraction subsystem also includes: The perturbation data generation module is used to define the objective function domain and the correction parameter domain, and introduce finite-amplitude perturbations based on the sensitivity tensor into the objective function domain and the correction parameter domain respectively. Specifically, the perturbation of the objective function domain is achieved by adding a small perturbation generated by a zero-mean Gaussian distribution to each sampling point of the unified feature representation after correction, generating the first perturbation sample, thereby expanding the perturbation of the feature layer; the perturbation of the correction parameter domain is achieved by applying a fixed percentage finite difference perturbation to the correction parameter set, generating the second perturbation sample, so as to form a dual-domain perturbation dataset of the objective function domain and the correction parameter domain. The consistency constraint establishment module is used to define the feature domain mapping function and the parameter domain mapping function based on the dual-domain perturbation dataset. It calculates the residual value using the first perturbation sample and the second perturbation sample, uses the residual value as the perturbation feedback signal, compares the difference of the output results before and after the perturbation, and constructs the consistency constraint function. The consistency constraint function is used to realize the consistency constraint relationship between the objective function domain and the modified parameter domain. The adversarial training and residual correction module is used to input the dual-domain perturbation dataset into the training network and set up a discriminator unit to distinguish between the original samples and the perturbation samples. By minimizing the discrimination error, the parameters of the feature domain mapping function and the parameter domain mapping function are updated. During the training iteration, the residual is fed back to the backpropagation path of the network to dynamically correct the parameter weights of the feature domain mapping function and the parameter domain mapping function. When the consistency loss function and the discriminator error converge simultaneously, the parameter update is stopped, and the training network model optimized by dual-domain perturbation and consistency constraints is obtained.

6. The structural feature extraction and model correction system integrating convolution and self-attention mechanisms according to claim 5, characterized in that: The prediction training subsystem includes: The measurement module is used to monitor multi-source sensor data during the construction of the cable net structure, including displacement data, cable force data, and acceleration data. The displacement data is mapped to the corresponding node degrees of freedom in the finite element model of the construction process, serving as the node displacement boundary conditions. The cable force data is converted into equivalent tension loads and applied to the endpoints of the corresponding tension elements. The acceleration data is analyzed in the frequency domain to obtain natural frequency parameters and relative displacement modes at the sensor placement locations. The relative displacement modes at the sensor placement locations represent the mode shape vector components of each mode at the node where the sensor is located. The measured processing module is used to combine the measured natural frequency parameters, the relative displacement pattern at the sensor placement position, and the displacement data to generate the measured results. Based on the measured results, the measured objective function matrix is ​​obtained. The modified parameter prediction module is used to calculate the modified parameter prediction values ​​through forward propagation based on the trained network model optimized by dual-domain perturbation and consistency constraints, thereby realizing the inversion process from structural representation to parameter space.

7. The structural feature extraction and model correction system integrating convolution and self-attention mechanisms according to claim 6, characterized in that: The prediction training subsystem also includes: The error module is used to input the predicted values ​​of the correction parameters into the finite element model of the construction process, recalculate the objective function matrix, and calculate the prediction error of the correction parameters by comparing it with the measured objective function matrix. The iterative update module is used to iteratively update the correction parameters according to the changing trend of the prediction error. When the change of the prediction error is lower than the preset threshold after several consecutive iterations, the current correction parameter set is determined to be in a convergent state, and the converged correction parameter set is re-inputted into the finite element model of the construction process.

Images