A Smart Correction Method for Structural Finite Element Models Based on Multi-Level Classification

Abstract

This invention provides an intelligent correction method for structural finite element models based on multi-level classification, belonging to the field of intelligent structural engineering technology. The method includes the following steps: constructing a dynamic response dataset of the structure under various preset damage conditions through finite element simulation; pre-dividing the structure into N damage grouping regions based on structural mechanics principles, and assigning a location label to each region; constructing and training a multi-level classification recognition model based on the dynamic response dataset; acquiring the response data of the structure to be evaluated, inputting it into the trained multi-level classification recognition model, and obtaining output results regarding the necessity of correction, the correction location label, and the correction degree level; and automatically correcting the stiffness parameters of the corresponding regions in the finite element model according to the output results. This invention achieves intelligent damage localization and parameter correction, avoiding low computational efficiency and local optima problems, making it more efficient and intelligent.

Description

Technical Field

[0001] This invention relates to the field of intelligent structural engineering technology, and in particular to an intelligent correction method for structural finite element models based on multi-level classification. Background Technology

[0002] Traditional finite element model correction methods for structures have significant limitations when dealing with complex spatial structures. These methods primarily rely on global optimization algorithms for parameter search, such as genetic algorithms and particle swarm optimization. While these methods can theoretically find the global optimum, they suffer from low computational efficiency and are prone to getting trapped in local optima when dealing with large-scale spatial truss structures. This is due to the large number of design variables (typically including stiffness parameters of dozens to hundreds of components) and the extremely high dimensionality of the optimization space. Furthermore, traditional methods lack structured damage localization strategies, often relying on component-by-component inspection for damage identification. This approach is ineffective in handling complex conditions with multiple damages coexisting, especially when damage exists in multiple regions of the structure simultaneously, making it difficult to accurately identify the spatial distribution patterns of the damage.

[0003] More importantly, existing technologies neglect the mechanical interrelationships between components in space truss structures. In practical engineering, the components of space truss structures often exhibit distinct grouping characteristics based on their position and function within the structural system: the main column system bearing the primary vertical loads, the horizontal support system responsible for transmitting horizontal forces, and the diagonal bracing system ensuring overall stability, etc. These component groups with different functions often show correlation when damage occurs; that is, components within the same functional group are prone to coordinated damage due to similar stress patterns, similar construction processes, and consistent usage environments. However, traditional correction methods treat all components as independent design variables, failing to fully utilize this structural characteristic, resulting in a lack of physical meaning and low efficiency in the correction process.

[0004] Furthermore, existing methods are highly subjective in determining the degree of correction, often relying on engineers' experience or pre-set correction ranges, and lack systematic quantitative criteria. This not only affects the accuracy of the correction results but also makes the correction process difficult to standardize and automate. Especially in emergency detection and rapid assessment scenarios, traditional methods cannot meet the dual requirements of real-time performance and accuracy.

[0005] Therefore, there is an urgent need for an intelligent correction method for structural finite element models based on multi-level classification to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent correction method for structural finite element models based on multi-level classification, to solve the technical problems of low correction efficiency, inaccurate damage localization, and strong subjectivity in judging the degree of correction in existing technologies. The various technical effects of the preferred solutions among the many technical solutions provided by this invention are detailed below.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an intelligent correction method for structural finite element models based on multi-level classification, comprising the following steps: A dynamic response dataset of a structure under various preset damage conditions was constructed using finite element simulation. Based on the principles of structural mechanics, the structure is pre-divided into N damage grouping regions, and each region is assigned a location label, where N is an integer greater than 1; Based on the dynamic response dataset, a multi-level classification and recognition model is constructed and trained, wherein the multi-level classification and recognition model includes at least: The first-level classifier is used to determine whether the structural finite element model needs to be corrected. The second-level classifier is used to identify the location labels of the damaged grouping regions that need to be corrected. The output of this classifier is the multi-label classification result corresponding to the N regions. The third-level classifier is used to determine the degree of correction for the identified damage grouping regions; The response data of the structure to be evaluated is obtained and input into the trained multi-level classification and recognition model to obtain output results on the necessity of correction, the correction position label and the degree of correction. Based on the output results, the stiffness parameters of the corresponding region in the finite element model are automatically corrected.

[0008] Furthermore, the number of the N damage group regions is six; the multi-label classification result output by the second-level classifier is a six-dimensional binary vector, where each dimension corresponds to the damage state of a damage group region.

[0009] Furthermore, the training process of the multi-level classification and recognition model includes: Using a predefined label mapping table, each of the preset damage conditions is mapped to a six-dimensional main label vector and a six-dimensional severity label vector. The main label vector is used to indicate whether damage exists in the six regions, and the severity label vector is used to indicate the degree of damage in the regions where damage exists.

[0010] Furthermore, the judgment result of the first-level classifier is a binary classification result; the third-level classifier outputs a classification result representing the degree of correction for each region identified as needing correction, wherein the degree of correction level includes at least a first level corresponding to the first stiffness reduction rate and a second level corresponding to the second stiffness reduction rate.

[0011] Furthermore, the first stiffness reduction rate is 10%, and the second stiffness reduction rate is 20%.

[0012] Furthermore, the multi-level classification and recognition model is an end-to-end deep learning network, which includes a shared feature extractor and three dedicated classification heads corresponding to the first-level classifier, the second-level classifier, and the third-level classifier, respectively.

[0013] Furthermore, the loss function of the deep learning network is a multi-task loss function, which is a weighted combination of the first-level classification loss, the second-level multi-label classification loss, and the third-level classification or regression loss.

[0014] Furthermore, the process of constructing the dynamic response dataset of the structure under various preset damage conditions includes: Damage state is simulated by reducing the stiffness of specific components or regions in the finite element model, and dynamic response data of the structure are collected under various excitations.

[0015] Furthermore, the automatic correction of the stiffness parameters of the corresponding region in the finite element model specifically includes: The region to be corrected is determined based on the location label output by the second-level classifier, and the stiffness parameters of the components within the region to be corrected are reduced accordingly based on the correction level output by the third-level classifier.

[0016] Furthermore, after performing the correction, an iterative verification step is also included, which specifically includes: Compare the corrected model simulation response with the measured or target response; If the accuracy requirement is not met, the response data of the corrected finite element model is reacquired, and the judgment and parameter correction process of the multi-level classification and recognition model is executed again until the simulation response meets the convergence condition.

[0017] This invention provides an intelligent correction method for structural finite element models based on multi-level classification, comprising: constructing a dynamic response dataset of the structure under various preset damage conditions through finite element simulation; pre-dividing the structure into N damage grouping regions based on structural mechanics principles, and assigning a location label to each region; constructing and training a multi-level classification recognition model based on the dynamic response dataset; acquiring the response data of the structure to be evaluated, inputting it into the trained multi-level classification recognition model, and obtaining output results regarding the necessity of correction, the correction location label, and the correction degree level; and automatically correcting the stiffness parameters of the corresponding regions in the finite element model according to the output results. The method establishes a correspondence between damage conditions and labels through a predefined mapping table, avoiding complex clustering analysis and improving the stability and reliability of the system; the region division strategy based on structural mechanics principles fully utilizes expert knowledge and engineering experience, making damage location more accurate and with clear physical meaning; the dual-label mechanism can simultaneously handle damage identification and degree assessment, providing more comprehensive information support for model correction; the standardized data processing flow and automated label generation significantly improve the engineering practicality of the system; and the unified processing capability for multiple damage conditions enables the system to cope with complex real-world engineering scenarios, exhibiting good adaptability and scalability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the intelligent correction method for structural finite element models based on multi-level classification, as described in this invention. Figure 2 This is a flowchart illustrating the specific steps of the intelligent correction method for structural finite element models based on multi-level classification according to the present invention. Figure 3 It is a point map of structural feature extraction from finite element simulation; Figure 4 This is a diagram showing the damage location classification and member numbering; Figure 5 This is a diagram of a three-level classification and recognition network architecture; Figure 6 This is a performance curve of the three-level classifier training process. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] like Figure 1 As shown, this invention provides an intelligent correction method for structural finite element models based on multi-level classification, comprising the following steps: S1, constructing a dynamic response dataset of the structure under various preset damage conditions through finite element simulation; S2, based on the principles of structural mechanics, pre-dividing the structure into N damage group regions and assigning a location label to each region, where N is an integer greater than 1; S3, constructing and training a multi-level classification recognition model based on the dynamic response dataset, wherein: a first-level classifier is used to determine whether the structural finite element model needs correction; a second-level classifier is used to identify the location labels of the damage group regions that need correction, and the output of this classifier is the multi-label classification result corresponding to the N regions; a third-level classifier is used to determine the correction degree level of the identified damage group regions; S4, acquiring the response data of the structure to be evaluated and inputting it into the trained multi-level classification recognition model to obtain output results regarding the necessity of correction, the correction location label, and the correction degree level; S5, automatically correcting the stiffness parameters of the corresponding regions in the finite element model according to the output results.

[0022] like Figures 1-6 As shown, this embodiment takes a four-story spatial steel truss structure with typical engineering representativeness and complex mechanical behavior characteristics as the application object, and elaborates on the implementation process of the method of the present invention.

[0023] like Figure 2 As shown, this embodiment mainly consists of four interconnected core modules: data preprocessing module 1, mapping table label generation module 2, three-level classification and recognition module 3, and intelligent correction execution module 4. The execution process follows the logical order of data acquisition and preprocessing, label mapping generation, deep learning training, and intelligent correction iteration.

[0024] Step S1 specifically involves constructing a structural response dataset. The aim is to build a comprehensive and diverse damage condition dataset for subsequent model training and validation. Step S1, which involves constructing a dynamic response dataset of the structure under various preset damage conditions, includes: simulating the damage state by reducing the stiffness of specific components or regions in the finite element model, and collecting dynamic response data of the structure under various excitations.

[0025] The core function of the data preprocessing module 1 in this embodiment is to efficiently extract structural dynamic response features from a large amount of simulation data and to normalize them to improve the accuracy of subsequent analysis. Firstly, as... Figure 3 and Figure 4 As shown, a precise four-story spatial truss baseline finite element model was established using professional finite element analysis software, such as ANSYS APDL. This included constructing a multi-story spatial structure model and setting appropriate inter-story spacing; using structural components of different specifications, including main load-bearing components and connecting components; setting corresponding material mechanics parameters; and applying appropriate boundary constraints and loading methods to the model. Based on this, 39 different types of typical damage conditions were systematically designed. These damage conditions cover various damage modes commonly encountered in engineering practice, including single-component damage, multi-component combined damage, and damage of varying severity. Numerical simulation of the damage was achieved using the stiffness degradation method, specifically setting two damage levels: a 10% stiffness reduction for minor damage and a 20% stiffness reduction for severe damage.

[0026] During the dynamic analysis, a broadband random excitation signal is applied to the top of the structure, with the excitation frequency range covering the main modal frequencies of the structure. The system collects three types of response data from each key node, including displacement time history response, acceleration time history response, and strain response of key sections. The simulation analysis of each damage condition generates a standard CSV format data file for easy batch processing later.

[0027] The data processing strategy employs a row-by-row expansion method, treating each row of data in each CSV file as an independent training sample. This significantly increases the size of the training dataset and improves the model's generalization ability.

[0028] In the specific implementation process, the random seed is first fixed at 42 to ensure the reproducibility of the experimental results. Then, the program automatically traverses the "Simulation" folder under the specified directory, identifies and reads all CSV format data files. For each file, the efficient reading function of the pandas data processing library is used to parse the data matrix row by row. Each row of data constitutes a complete feature vector, the dimension of which is determined by the number of monitoring points and the response type.

[0029] In this embodiment, the data preprocessing module realizes the automatic conversion from simulation data to label vectors. By reading the structural response data in CSV format, it looks up the corresponding damage label information in a predefined mapping table according to the file name. Each row of data in each CSV file is used as an independent sample and trained with the same label vector, which significantly expands the scale and diversity of the training dataset.

[0030] Step S2 specifically involves damage grouping and intelligent label generation. This is used to divide the complex structure into regions and establish an automated label mapping system. Specifically, in step S2, the number of N damage grouping regions is six. The division of these six regions is based on structural mechanics principles and engineering experience. By analyzing the stress characteristics, spatial relationships, and damage correlations of the components in the space truss structure, the structure is divided into six regions with clear physical meaning according to function and stress pattern. The components within each region have similar stress characteristics and damage sensitivities, and the number of components in each region is relatively balanced, avoiding excessive concentration or sparseness of components in certain regions, thus effectively reflecting the damage distribution pattern of the structure.

[0031] The multi-label classification system in this embodiment adopts a dual-label mechanism of main label and additional label. The main label vector represents the presence of damage in six regions, and the additional label vector represents the specific severity of damage. This hierarchical label design can simultaneously handle the two key tasks of damage localization and severity assessment, thereby improving the accuracy and practicality of classification.

[0032] The damage condition label generation adopts a predefined mapping table approach. By leveraging expert knowledge and engineering experience, a mapping relationship is established between filenames and damage labels. The system automatically finds the corresponding damage region combinations and severity information based on the simulation data filenames, generating multi-level label vectors, including a 6-dimensional main label vector representing the damage state of each region and a 6-dimensional additional label vector representing the severity level of the damage. The multi-label classification result output by the second-level classifier is a six-dimensional binary vector, where each dimension corresponds to the damage state of a damage group region. In this embodiment, the mapping table label generation module 2 achieves a precise correspondence between filenames and damage labels through a pre-constructed intelligent mapping table, effectively avoiding the instability and computational complexity problems existing in traditional clustering analysis methods.

[0033] In the specific implementation process, the system pre-establishes two core mapping data tables: a main label key-value mapping table and a severity label mapping table. The main label key-value mapping table uses a region number combination encoding method to represent the spatial distribution pattern of damage. For example, "1-3" represents simultaneous damage in regions 1 and 3, and "2-4-5" represents a combined damage pattern in regions 2, 4, and 5. The severity label mapping table uses a 6-dimensional vector to describe the degree of damage in each region. Each element in the vector corresponds to a region, where a value of 0 indicates that the region is intact, a value of 1 indicates that the region has mild damage with a 10% stiffness reduction, and a value of 2 indicates that the region has severe damage with a 20% stiffness reduction.

[0034] The specific algorithm for automatic label generation includes: First, the system searches for the corresponding region code in the main label key-value mapping table based on the input file name. If no match is found, the file is automatically skipped and the abnormal information is recorded. After a successful search, the region code is segmented using the separator "-" to obtain a specific sequence of damage region numbers. Based on this sequence, a 6-dimensional binary main label vector is constructed, with the position corresponding to the damage region assigned a value of 1, and the remaining positions kept as 0. Simultaneously, the corresponding 6-dimensional severity vector is queried in the severity mapping table. Finally, the main label vector, severity vector, and corresponding feature data are associated and bound to form a complete labeled training sample. This label generation mechanism based on a predefined mapping table ensures the consistency and accuracy of label allocation and can flexibly handle complex multi-region damage conditions.

[0035] See Figure 4 As shown in Table 1, to construct the supervised learning objective for the multi-label classification task, 39 typical damage scenarios were designed based on the structural component numbers and their physical meaning in the spatial layout, and these scenarios were systematically classified and labeled. Each damage scenario consists of one or more component numbers, representing damage events occurring at the corresponding locations.

[0036] To enhance the comprehensiveness and diversity of model training, damage scenarios are divided into three categories: single damage, double damage, and triple damage, corresponding to scenarios where 1, 2, and 3 components in the structure are damaged, respectively. To achieve multi-label classification, it is necessary to systematically categorize the various damage locations in the complex structure and standardize the label naming rules.

[0037] In this embodiment, the spatial arrangement and component numbering of the structure are arranged counterclockwise, starting from x-1 and sequentially numbered to x-4. Based on the basic principles of structural mechanics and engineering practice, the four-story spatial truss structure is divided into six damage grouping areas. Each grouping area corresponds to a structural label category, numbered from label 1 to label 6. Each label category represents a set of components that may be damaged. Once damage appears at any location within that set of components, it is determined that the label category exists.

[0038] The specific grouping scheme is shown in Table 1: Table 1 Correspondence Table of Damage Group Numbers The regional division fully considers key factors such as the main stress modes of components, their impact on the overall structural performance, and their damage sensitivity. This scientific zoning strategy has the following significant advantages: the stress properties and damage response characteristics of components within the same region are similar, which is beneficial for the identification and analysis of damage modes; different regions have clear functional divisions and hierarchical structures, facilitating accurate damage location and impact assessment; the regional division considers possible damage propagation paths, which helps predict and control damage development trends; and the number of components in each region is relatively balanced, avoiding the problem of excessive concentration or sparseness of components in certain regions. Through this zoning method based on mechanical principles, the complex problem of multi-component damage identification is transformed into a classification problem of six regions, significantly reducing the computational complexity and solution difficulty.

[0039] Step S3 specifically involves building and training an end-to-end deep learning model, including a shared feature extractor and three dedicated classification heads corresponding to the first-level classifier, the second-level classifier, and the third-level classifier, respectively.

[0040] The training process of the multi-level classification recognition model includes: mapping each preset damage condition to a six-dimensional main label vector and a six-dimensional severity label vector through a predefined label mapping table; wherein, the main label vector is used to indicate whether there is damage in the six regions, and the severity label vector is used to indicate the degree of damage in the regions where damage exists.

[0041] The first-level classifier determines the model by binary classification. Specifically, it assesses the deviation between the structural response data and the baseline model to determine if the current model needs correction, outputting a "yes" or "no" binary classification result. The second-level classifier identifies the specific correction location for models deemed to require correction, classifying the structure into six regions (labels 1 to 6) for multi-label classification. The third-level classifier outputs a classification result representing the degree of correction for each identified region. The correction degree level includes at least a first level corresponding to the first stiffness reduction rate and a second level corresponding to the second stiffness reduction rate. The first stiffness reduction rate is 10%, and the second stiffness reduction rate is 20%. Specifically, the loss function of the deep learning network is a multi-task loss function, composed of a weighted combination of the first-level classification loss, the second-level multi-label classification loss, and the third-level classification or regression loss.

[0042] In this embodiment, the first-level classification criterion is to analyze the structural response feature vector through a deep learning classifier and directly output a binary classification result indicating whether correction is needed. The classifier learns damage feature patterns through training and automatically determines the structural state. The second-level multi-label classification is implemented by extracting the structural response feature vector; establishing a multi-label classification model, where each label corresponds to the correction requirement of a structural region; when any component in a certain damage condition belongs to a certain label group, that label is activated; and outputting a 6-dimensional binary label vector reflecting the spatial distribution pattern of structural damage. The third-level correction degree classification method includes dividing the damage severity into two categories: mild damage and severe damage; learning the mapping relationship between damage features and severity levels through training a classifier; mild correction corresponds to a 10% stiffness adjustment, suitable for cases with small response deviations; severe correction corresponds to a 20% stiffness adjustment, suitable for cases with large response deviations; and learning the mapping relationship between the degree of deviation and the correction level through training a classifier.

[0043] In the specific implementation process, the network model is first constructed and trained, such as... Figure 5 and Figure 6 As shown, the three-level classification and recognition model in this embodiment adopts an end-to-end deep learning network architecture, which mainly consists of a shared feature extractor, a dual attention mechanism module, and three specialized classification heads. Deep feature extraction and adaptive enhancement are achieved through residual connections and attention mechanisms. The shared feature extractor employs a cascaded design of residual convolutional blocks to progressively map the input structural response features to the high-level semantic feature space.

[0044] Specifically, the shared feature extractor in this embodiment adopts a residual convolutional block cascade design. By introducing a skip connection mechanism, the residual convolutional blocks effectively solve the gradient vanishing and gradient exploding problems during deep network training. Each residual block contains two one-dimensional convolutional layers, configured with batch normalization layers and the GELU activation function, achieving effective feature transfer and preservation through residual connections. The network as a whole employs a layer-by-layer feature dimensionality reduction strategy, progressively mapping the original high-dimensional input features to hierarchical feature spaces of 256, 128, and 64 dimensions, achieving an effective transformation from low-level physical features to high-level semantic features.

[0045] The dual attention mechanism module combines the CBAM and SE attention modules. The CBAM module, through the organic cascading of channel and spatial attention mechanisms, adaptively learns and adjusts the importance weights of different feature channels, highlighting key feature information directly related to damage. The SE module, through a combination of global average pooling and fully connected layers, deeply learns the interdependencies between feature channels, further enhancing the network's feature representation and discrimination capabilities. The synergistic effect of the dual attention mechanism significantly improves the network's ability to identify weak damage features in structural response data, enhancing the overall classification accuracy and robustness.

[0046] Three specialized classification heads are assigned to different levels of classification and recognition tasks. The damaged / indestructible classification head employs a classic binary classification network design, using a combination of fully connected layers and a softmax activation function to output a probability assessment of whether the overall structure is damaged. The main label classification head adopts a multi-label classification design, outputting a 6-dimensional probability vector through a sigmoid activation function. Each element in the vector represents the probability assessment of damage in the corresponding region, enabling simultaneous identification and judgment of damage states in multiple regions. The additional label classification head uses a multi-class regression design, directly outputting a 6-dimensional predicted damage severity value, with each element quantifying the specific damage severity level of the corresponding region.

[0047] Then, a loss function for multi-objective collaborative optimization is designed. The multi-stage loss function adopts an intelligent weighted fusion strategy, adaptively combining and optimizing the loss functions of the three different levels of classification tasks. The first stage of lossy and lossless detection uses the standard cross-entropy loss function: in Labels representing the actual damage state, This represents the damage probability predicted by the model.

[0048] The second stage, main label classification, addresses the class imbalance problem commonly found in multi-label classification by employing an improved focus loss function: in, and These are adjustable focusing parameters for positive and negative samples, respectively. By dynamically adjusting the loss weights for easy and difficult samples, the model's ability to identify and learn samples from minority classes is significantly improved.

[0049] The third stage of additional label prediction uses the cross-entropy loss function. By preprocessing the severity label by subtracting 1 and limiting the minimum value to 0, the classification of the severity level of injury is optimized.

[0050] The total loss function of the system is defined as: in, , , The weight parameters are adjustable, and the optimal parameter combination is determined through extensive experimental verification and grid search methods. This multi-stage joint loss function design can simultaneously optimize the three core tasks of damage detection, precise localization, and severity assessment, achieving true end-to-end joint training and optimization.

[0051] Finally, data partitioning and model training optimization were performed. The dataset was strictly partitioned using a fixed random seed and standardized proportions to ensure the reproducibility of experimental results. Specifically, the complete dataset was directly divided into training, validation, and test sets in a 7:2:1 ratio, with the training set comprising 70%, the validation set 20%, and the test set 10%, forming a standard three-part data distribution. This ensured the stability of model training and the reproducibility of results. Batch processing and data shuffling techniques were also employed to improve training efficiency.

[0052] Data loading utilizes PyTorch's efficient DataLoader mechanism, with a batch size set to 128 to ensure training efficiency while avoiding memory overflow. Random shuffling is enabled on the training set to increase sample randomness and diversity, while shuffling is disabled on the validation and test sets to maintain consistency and stability in the evaluation process.

[0053] Model training optimization employs the AdamW optimization algorithm, which combines the adaptive learning rate characteristic of the Adam optimizer with the advantages of weight decay regularization. The initial learning rate is set to 4 × 10^-5, and the weight decay coefficient is set to 6 × 10^-5. The learning rate scheduling strategy uses the ReduceLROnPlateau adaptive adjustment mechanism, which automatically halves the learning rate when the validation set loss shows no significant improvement for five consecutive training epochs, thus achieving finer parameter optimization.

[0054] To effectively prevent overfitting, a multi-layered regularization strategy was employed: Dropout regularization with a dropout probability set to 0.3; and an early stopping mechanism with a patience parameter set to 30 training epochs. During training, detailed training metrics, including loss values ​​and classification accuracy at each stage, were printed every 5 epochs, enabling real-time monitoring and performance evaluation. The early stopping mechanism was automatically triggered when the validation loss showed no improvement for 30 consecutive epochs, and the model parameters at the moment of minimum validation loss were saved as the final optimal model.

[0055] Steps S4 and S5 specifically involve using a multi-label classification algorithm to analyze the structural response characteristics and implement intelligent model correction. This means applying the trained model to new data to achieve automated diagnosis and correction. Step S5, specifically the automatic correction of stiffness parameters in the corresponding region of the finite element model, includes: determining the region to be corrected based on the location labels output by the second-level classifier, and reducing the stiffness parameters of the components within the region to be corrected by a corresponding magnitude based on the correction level output by the third-level classifier. After correction in step S5, an iterative verification step is included. This step involves comparing the corrected model simulation response with the measured or target response. If the accuracy requirements are not met, steps S4 and S5 are repeated until the convergence condition is met.

[0056] In this embodiment, the intelligent correction execution module 4 automatically formulates and executes the parameter correction scheme of the finite element model based on the output results of the trained three-level classifier. For newly input structural response data, the standardized feature vector is first extracted through the data preprocessing module, and then it is input into the trained deep learning classification network for intelligent reasoning analysis.

[0057] The network inference outputs prediction results at three levels: a probability assessment of whether the overall structure is damaged or not, a 6-dimensional probability vector of regional damage, and a 6-dimensional numerical vector of the severity of regional damage. Based on a preset decision threshold (usually set to 0.5), the system converts the regional damage probability into a clear binary judgment vector, accurately determining the specific regions requiring correction.

[0058] For each area identified as needing correction, the system determines a specific correction strategy based on the predicted damage severity value: when the predicted value is closer to 1, it is determined to be minor damage, and a 10% reduction correction is applied to the elastic modulus of all related components in that area; when the predicted value is closer to 2, it is determined to be severe damage, and a 20% reduction correction is applied to the elastic modulus.

[0059] The intelligent correction execution module automatically determines and executes the correction strategy based on the output of the three-level classifier. The specific process is as follows: Input the structural response data to be corrected, and sequentially perform three-level classification judgments: correction necessity judgment, correction location identification, and correction degree determination; automatically adjust the stiffness parameters of the corresponding region in the finite element model according to the classification results.

[0060] For example, the first-level classifier determines whether the input structural response data indicates that the structure is damaged; if it is determined to be in an undamaged state, the system outputs "the structure is healthy and no correction is needed"; if it is determined to be in a damaged state, the second-level classifier automatically identifies the specific location of the damaged area, and the third-level classifier simultaneously determines the severity level of each damaged area.

[0061] The system then automatically formulates a correction plan based on the classification results. This includes determining the corresponding stiffness reduction for the identified damaged areas based on the severity prediction results (10% for minor damage and 20% for severe damage), and automatically updating the relevant parameters of the finite element model to complete the intelligent model correction process.

[0062] Through the above implementation, this intelligent correction method for structural finite element models based on multi-level classification can fully utilize the inherent characteristics of spatial truss structures, realize intelligent damage location and parameter correction, and provide an efficient, accurate and automated technical solution for structural health monitoring and damage identification.

[0063] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent correction of structural finite element models based on multi-level classification, characterized in that, Includes the following steps: A dynamic response dataset of a structure under various preset damage conditions was constructed using finite element simulation. Based on the principles of structural mechanics, the structure is pre-divided into N damage grouping regions, and each region is assigned a location label, where N is an integer greater than 1; Based on the dynamic response dataset, a multi-level classification and recognition model is constructed and trained, wherein the multi-level classification and recognition model includes at least: The first-level classifier is used to determine whether the structural finite element model needs to be corrected. The second-level classifier is used to identify the location labels of the damaged grouping regions that need to be corrected. The output of this classifier is the multi-label classification result corresponding to the N regions. The third-level classifier is used to determine the degree of correction for the identified damage grouping regions; The response data of the structure to be evaluated is obtained and input into the trained multi-level classification and recognition model to obtain output results on the necessity of correction, the correction position label and the degree of correction. Based on the output results, the stiffness parameters of the corresponding region in the finite element model are automatically corrected.

2. The intelligent correction method for structural finite element models based on multi-level classification according to claim 1, characterized in that: The number of the N damage group regions is six; the multi-label classification result output by the second-level classifier is a six-dimensional binary vector, where each dimension corresponds to the damage state of a damage group region.

3. The intelligent correction method for structural finite element models based on multi-level classification according to claim 2, characterized in that: The training process of the multi-level classification and recognition model includes: Using a predefined label mapping table, each of the preset damage conditions is mapped to a six-dimensional main label vector and a six-dimensional severity label vector. The main label vector is used to indicate whether damage exists in the six regions, and the severity label vector is used to indicate the degree of damage in the regions where damage exists.

4. The intelligent correction method for structural finite element models based on multi-level classification according to claim 3, characterized in that: The first-level classifier's judgment result is a binary classification result; The third-level classifier outputs a classification result representing the degree of correction for each region identified as needing correction. The degree of correction level includes at least a first level corresponding to the first stiffness reduction rate and a second level corresponding to the second stiffness reduction rate.

5. The intelligent correction method for structural finite element models based on multi-level classification according to claim 4, characterized in that: The first stiffness reduction rate is 10%, and the second stiffness reduction rate is 20%.

6. The intelligent correction method for structural finite element models based on multi-level classification according to any one of claims 1-5, characterized in that: The multi-level classification and recognition model is an end-to-end deep learning network, which includes a shared feature extractor and three dedicated classification heads corresponding to the first-level classifier, the second-level classifier and the third-level classifier, respectively.

7. The intelligent correction method for structural finite element models based on multi-level classification according to claim 6, characterized in that: The loss function of the deep learning network is a multi-task loss function, which is a weighted combination of the first-level classification loss, the second-level multi-label classification loss, and the third-level classification or regression loss.

8. The intelligent correction method for structural finite element models based on multi-level classification according to any one of claims 1-5, characterized in that, The process of constructing a dynamic response dataset of the structure under various preset damage conditions includes: Damage state is simulated by reducing the stiffness of specific components or regions in the finite element model, and dynamic response data of the structure are collected under various excitations.

9. The intelligent correction method for structural finite element models based on multi-level classification according to any one of claims 1-5, characterized in that, The automatic correction of stiffness parameters of the corresponding region in the finite element model specifically includes: The region to be corrected is determined based on the location label output by the second-level classifier, and the stiffness parameters of the components within the region to be corrected are reduced accordingly based on the correction level output by the third-level classifier.

10. The intelligent correction method for structural finite element models based on multi-level classification according to claim 9, characterized in that, After the correction is performed, an iterative verification step is also included, which specifically includes: Compare the corrected model simulation response with the measured or target response; If the accuracy requirement is not met, the response data of the corrected finite element model is reacquired, and the judgment and parameter correction process of the multi-level classification and recognition model is executed again until the simulation response meets the convergence condition.

Images