Intelligent structural damage identification method based on LSTM-Transform fusion model

Abstract

The invention discloses an intelligent structural damage identification method based on an LSTM-Transform fusion model, and relates to the technical field of structural damage identification. Comprising the following steps: S1, multi-source data acquisition: aiming at a key stress part, a node concentration area, a vertical stiffness mutation area and a dynamic coupling part of the lattice type cylinder structure, deploying a strain gauge, an accelerometer and a visual acquisition device, synchronously acquiring a structure vibration signal, stress-strain data, a surface visual image and environmental parameters, the collection range covers a transparent area, a node core area and a mast connection part, and a multi-dimensional original data set is formed; s2, data preprocessing; S3, fusion model construction and training; S4, structure damage identification; and S5, result feedback and application, wherein a damage identification result is transmitted to a structure health management platform.

Description

Technical Field

[0001] This invention relates to the field of structural damage identification technology, and in particular to a method for intelligent structural damage identification based on an LSTM-Transformer fusion model. Background Technology

[0002] Lattice-type cylindrical structures are the core structural form of tall, iconic buildings, widely used in engineering fields such as television towers and communication towers. These structures often employ personalized design concepts and generally possess unique characteristics such as geometric asymmetry, abrupt changes in vertical stiffness, multiple open sections, and dynamic coupling between the top mast and the main body. The structures primarily exhibit horizontal vibration accompanied by significant torsional vibration effects, resulting in complex dynamic responses and pronounced nonlinear characteristics.

[0003] Currently, there are several technical challenges in the field of damage identification and seismic toughness assessment for this type of structure. First, the complexity of the structure makes it difficult to capture dynamic responses. Conventional linear analysis methods struggle to accurately characterize the nonlinear responses caused by geometric asymmetry and abrupt stiffness changes. They also lack the ability to identify the dynamic coupling effect between the mast and the main body, as well as damage related to torsional vibration, thus failing to comprehensively reflect the true stress state of the structure. Second, research on the coupling effects of multiple disasters has limitations. Existing technologies mostly focus on single seismic actions, failing to fully consider the synergistic effects of far-field long-period earthquakes, near-field pulse earthquakes, wind loads, and temperature deformation, making it difficult to adapt to the actual monitoring needs of complex service environments. Third, there are discrepancies between simplified models and actual engineering conditions. Existing studies are mostly based on idealized rigid or hinged models, failing to fully incorporate actual working condition factors such as material nonlinearity, welding residual stress, and node construction errors, leading to deviations between damage identification results and the true structural state. Fourth, damage identification methods have inherent limitations. Traditional technologies rely on single sensor data, and models often focus on a single dimension of temporal or spatial characteristics, lacking deep fusion of multi-source data. Furthermore, they do not incorporate core structural seismic resistance methods such as modal analysis and elastoplastic time history analysis, making it difficult to meet engineering requirements in terms of generalization ability and identification accuracy. Fifth, there is insufficient coordination between monitoring and simulation. The lack of dynamic linkage between physical structures and virtual models makes it impossible to predict damage evolution trends and achieve intelligent full-cycle management, thus hindering support for structural seismic optimization design and disaster emergency response.

[0004] Existing technologies have failed to form a complete technical system adapted to the specific characteristics of lattice tube structures, making it difficult to meet the actual needs of accurate identification of structural damage and seismic toughness assessment under complex working conditions. There is an urgent need for an intelligent identification method that deeply integrates core technologies of structural seismic analysis, adapts to multiple disaster scenarios, and takes into account both virtual and real collaboration. Summary of the Invention

[0005] This invention provides a method for intelligent structural damage identification based on an LSTM-Transformer fusion model. The technical solution is as follows: The intelligent structural damage identification method based on the LSTM-Transformer fusion model includes the following steps: S1 Multi-Source Data Acquisition: For key stress-bearing parts, concentrated node areas, vertical stiffness abrupt change areas, and dynamic coupling parts of the lattice tube structure, strain gauges, accelerometers, and visual acquisition devices are deployed to simultaneously acquire structural vibration signals, stress and strain data, surface visual images, and environmental parameters. The acquisition range covers the open area, the core node area, and the mast connection parts to form a multi-dimensional raw dataset. S2 data preprocessing: Noise suppression, data alignment and feature enhancement operations are performed sequentially on the multi-source raw data to eliminate deviations caused by equipment vibration, environmental interference and asynchronous acquisition. The time domain peak value, frequency domain main frequency, extreme value features of stress and strain data and surface texture and geometric deformation features of visual images are extracted to form a standardized dataset. S3 Fusion Model Construction and Training: An LSTM-Transformer fusion model is constructed. The Transformer module extracts the spatial correlation features of the data, and the LSTM module captures the temporal dynamic features of the data. The model is trained using labeled structural damage sample data and modal analysis results completed by the multiple Ritz vector method. This enables the model to learn the coupling features of horizontal vibration and torsional vibration of the structure, and optimizes the model parameters to improve the accuracy of damage feature recognition. S4 Structural Damage Identification: A standardized dataset is input into the trained model, which outputs damage feature vectors for each monitored area. A damage assessment system is constructed by utilizing the stiffness matrix and damping coefficient from the structural mechanical property parameters, the material strength parameters and structural geometric dimensions from the design parameters, and the peak displacement, peak acceleration, and stress distribution data from the seismic response data generated by elastoplastic time history analysis. S5 Results Feedback and Application: The damage identification results are transmitted to the structural health management platform and simultaneously pushed to the digital twin model, providing data support for structural maintenance and reinforcement, construction process adjustment and disaster emergency response.

[0006] Beneficial effects Precisely adapting to the specific characteristics of the structure and overcoming core technical pain points: The results of multi-Ritz vector modal analysis are deeply integrated into the training of the LSTM-Transformer fusion model to accurately capture the coupling characteristics of horizontal and torsional vibrations of the lattice cylindrical structure. It specifically solves the problem of damage identification in areas of abrupt change in vertical stiffness, open areas and dynamic coupling parts, significantly improving the accuracy of damage feature capture under complex structural morphology, which is different from the generalized technical logic of general structural damage identification.

[0007] Enhance multi-hazard coupling adaptation capability to cover complex service scenarios: Through seismic motion spectrum adaptation design, the model feature extraction weights are dynamically adjusted to address the differences in spectral characteristics between far-field long-period earthquakes and near-field pulse earthquakes, combined with the coupling effect of wind load and temperature deformation. This strengthens the mapping relationship between long-period vibration signals, stress mutation signals and damage characteristics, filling the gap in single-condition adaptation of existing technologies and ensuring the reliability and comprehensiveness of damage identification in complex disaster scenarios.

[0008] Improve the fit between the model and the actual project and reduce idealization deviation: Incorporate nonlinear compensation and construction error correction design for structural materials, combine measured data of material mechanical properties with actual working conditions during the construction stage, and calibrate model parameters through multi-source data to effectively reduce the deviation between the ideal model and the actual structure, making the damage identification results more consistent with the actual project and improving the practicality of the technical solution.

[0009] Optimize damage identification accuracy and generalization ability: By leveraging the synergistic advantages of the LSTM-Transformer fusion model, taking into account both the spatial correlation and temporal dynamic features of structural damage data, and using the metaheuristic NRBO optimization algorithm to automatically adjust hyperparameters, the fusion effect of multi-source features such as vibration, stress and strain, and visual images is enhanced, overcoming the limitations of traditional single models in feature capture, and significantly improving the identification accuracy and generalization ability for different types of damage.

[0010] Achieving seamless integration of virtual and physical systems and intelligent full-cycle management: Through the dynamic linkage of digital twins and damage identification, the virtual model reconstructed based on point cloud scanning and BIM technology can not only receive identification results to deduce damage development trends, but also optimize model parameters and monitoring schemes in reverse. Combined with full-process parameter adaptive control and failure mode early warning functions, a complete technical chain is constructed, encompassing data acquisition, identification and analysis, simulation and deduction, parameter optimization, and early warning and response. This upgrades the system from passive monitoring to proactive prediction, providing full-cycle data support for structural maintenance and reinforcement, construction process adjustments, and seismic optimization. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0012] Figure 1 A process flow diagram provided for an embodiment of this application. Detailed Implementation

[0013] The technical solution provided in this application will now be described with reference to the accompanying drawings.

[0014] To facilitate understanding of the embodiments of this application, the following points will be explained first: First, in this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates an "or" relationship between the preceding and following related objects, but does not exclude the possibility of indicating an "and" relationship. The specific meaning can be understood in the context. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one of a, b, or c" can represent: a, b, c; a and b; a and c; b and c; or a and b and c. Here, a, b, and c can be single or multiple.

[0015] Second, in this application, the use of prefixes such as "first," "second," etc., is merely for the purpose of distinguishing and describing different things belonging to the same name category, and does not constrain the order, size, or quantity of things. For example, "first message" and "second message" are simply different messages, and there is no temporal, size, or priority relationship between them.

[0016] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0017] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0018] like Figure 1 As shown, the intelligent structural damage identification method based on the LSTM-Transformer fusion model includes the following steps: S1 Multi-Source Data Acquisition: For key stress-bearing parts, concentrated node areas, vertical stiffness abrupt change areas, and dynamic coupling parts of the lattice tube structure, strain gauges, accelerometers, and visual acquisition devices are deployed to simultaneously acquire structural vibration signals, stress and strain data, surface visual images, and environmental parameters. The acquisition range covers the open area, the core node area, and the mast connection parts to form a multi-dimensional raw dataset. S2 data preprocessing: Noise suppression, data alignment and feature enhancement operations are performed sequentially on the multi-source raw data to eliminate deviations caused by equipment vibration, environmental interference and asynchronous acquisition. The time domain peak value, frequency domain main frequency, extreme value features of stress and strain data and surface texture and geometric deformation features of visual images are extracted to form a standardized dataset. S3 Fusion Model Construction and Training: An LSTM-Transformer fusion model is constructed. The Transformer module extracts the spatial correlation features of the data, and the LSTM module captures the temporal dynamic features of the data. The model is trained using labeled structural damage sample data and modal analysis results completed by the multiple Ritz vector method. This enables the model to learn the coupling features of horizontal vibration and torsional vibration of the structure, and optimizes the model parameters to improve the accuracy of damage feature recognition. S4 Structural Damage Identification: A standardized dataset is input into the trained model, which outputs damage feature vectors for each monitored area. A damage assessment system is constructed by utilizing the stiffness matrix and damping coefficient from the structural mechanical property parameters, the material strength parameters and structural geometric dimensions from the design parameters, and the peak displacement, peak acceleration, and stress distribution data from the seismic response data generated by elastoplastic time history analysis. S5 Results Feedback and Application: The damage identification results are transmitted to the structural health management platform and simultaneously pushed to the digital twin model, providing data support for structural maintenance and reinforcement, construction process adjustment and disaster emergency response.

[0019] As an optional embodiment, the model training process in step S3 incorporates a structural material nonlinear compensation step: The specific implementation is as follows: Obtain the measured stress and strain data from the multi-source data collected in step S1, and combine it with the natural frequency and mode shape data obtained from the structural dynamic characteristic analysis to establish the material parameter deviation correction equation: δP=k×(σmeasured-σpredicted), where δP is the material parameter correction amount, k is the correction coefficient, σmeasured is the stress value in the measured stress and strain data, σpredicted is the stress prediction value based on the initial sample parameters, and k is determined by fitting the structural material constitutive relation test data; The influence of node construction deviation, welding residual stress and material mechanical property fluctuation formed during the construction phase is incorporated to construct a sample parameter calibration model: Pcalibration = Pinitial + δP + δP1 + δP2 + δP3, where Pcalibration is the calibrated sample parameter, Pinitial is the initial sample parameter, δP1 is the parameter correction amount corresponding to node construction deviation, δP2 is the parameter correction amount corresponding to welding residual stress, and δP3 is the parameter correction amount corresponding to material mechanical property fluctuation. δP1 is calculated by the difference between the measured stress and strain data at the node and the design stress data; δP2 is calculated based on the correlation model between welding process parameters and residual stress; and δP3 is determined by the difference between the material mechanical property test data and the standard parameters. Substituting the P calibration into the model training samples completes the reverse calibration of the sample parameters. This compensation step directly correlates the multi-source data and structural modal analysis results collected in step S1 through the above equation and model.

[0020] As an optional embodiment, the multi-source data preprocessing in step S2 includes a feature fusion step: Extract the time-domain peak value and frequency-domain dominant frequency of the vibration signal, the extreme value features of the stress-strain data, and the structural surface texture features and geometric deformation features of the visual image to construct a multi-dimensional feature set; The modal response spectrum method was used to calculate the contribution of each mode. The modal contribution was determined by the weighted sum of the modal participation coefficient and the modal energy proportion. Based on the accuracy characteristics and response speed of each type of data, and combined with the modal contribution, the weight is assigned as follows: Wi=W0i×(α×Ci+β), where Wi is the final weight of the i-th type of data, W0i is the initial weight of the i-th type of data, Ci is the modal contribution of the monitoring part corresponding to the i-th type of data, and α and β are weight adjustment coefficients, which are calibrated through multi-source data fusion experiments. The multi-dimensional feature sets are weighted and fused according to the above weights to generate a fused feature vector, thus completing the feature fusion operation.

[0021] As an optional embodiment, the structural damage identification in step S4 is integrated into the seismic motion spectrum adaptation step: In advance, the spectral characteristic parameters of far-field long-period earthquakes, near-field pulse earthquakes and wind loads and temperature deformation coupling scenarios are extracted using ground motion spectrum analysis tools. These parameters include spectral peak value, dominant period, and spectral intensity. Combined with the dynamic response test data of the lattice tube structure, a mapping library of different scenarios and feature extraction rules is established. For long-period earthquake scenarios in the far field, a correlation matrix M between long-period vibration signals and damage characteristics is constructed. The elements of the correlation matrix are determined by training with structural damage test data under long-period earthquake loading. Matrix multiplication is used to multiply the feature vector of the long-period vibration signal with M to strengthen the mapping relationship between the two. For near-field pulsed seismic scenarios, a feature weight adjustment module is set up. Based on the peak ground acceleration and pulse duration of near-field pulsed seismic events, the weight coefficient ω of the stress mutation signal is calculated through a linear weighting algorithm. The weight coefficient ω is directly multiplied by the stress mutation feature vector to adjust the feature weight. Vibration signals from the dynamic coupling parts are collected, and the phase difference Δφ and amplitude ratio λ of the vibration signals are calculated. A deviation correction model is established: Fcorrected = Foriginal - α × Δφ - β × λ, where Fcorrected is the corrected damage feature vector, Foriginal is the uncorrected damage feature vector, and α and β are correction coefficients. The model is calibrated using damage simulation test data of the dynamic coupling parts and corrects the damage identification deviation caused by the dynamic coupling effect between the mast and the main body.

[0022] As an optional embodiment, the metaheuristic NRBO optimization algorithm is introduced during the model training process in step S3 to automatically adjust the hyperparameters of the LSTM-Transformer fusion model: The specific implementation is as follows: The hyperparameters that need to be adjusted include the number of attention heads in the Transformer module, the number of hidden layer neurons in the LSTM module, and the model learning rate. The damage probability density of each region of the structure is calculated by the seismic vulnerability curve. Regions with damage probability density higher than the preset density threshold are identified as regions with concentrated damage probability. The number of hidden layer neurons in the LSTM module corresponding to this region is adjusted according to the following formula: Nnew=Nold×(1+γ×ρ), where Nnew is the number of neurons after adjustment, Nold is the number of neurons before adjustment, γ is the number adjustment coefficient, and ρ is the ratio of damage probability density to threshold. The time-series variation variance of structural damage data is calculated. Data with variance higher than a preset variance threshold is identified as damage data with drastic time-series variation. The learning rate of the model corresponding to this type of data is adjusted according to the following formula: ηnew=ηold×(1-δ×σ), where ηnew is the adjusted learning rate and ηold is the number of attention heads in the Transformer module. Based on the spatial distribution characteristics of structural damage data, the data is divided into cluster regions by a spatial clustering algorithm. The number of attention heads is adjusted according to the spatial density of the cluster regions. The higher the spatial density, the more attention heads are required. The hyperparameter tuning process is performed iteratively using the NRBO optimization algorithm until the model recognition accuracy meets the preset requirements.

[0023] As an optional embodiment, the noise suppression in step S2 employs an adaptive filtering step: To address visual image noise caused by dust interference: Morphological opening operations are used to remove floating dust noise. A Bayesian image restoration algorithm is constructed based on the gray-level distribution law of the surface material to repair pixel information in the feature point region. A gray-level gradient threshold T is set for the feature point. When the gray-level gradient of the feature point is lower than T, the repair parameter is adjusted according to the following formula: θnew=θold×(1+ε×(Tg) / T), where θnew is the repair parameter after adjustment, θold is the repair parameter before adjustment, ε is the parameter adjustment coefficient, and g is the actual gray-level gradient of the feature point. To address environmental vibration interference in vibration signals: extract the structure's natural frequency f0, construct an adaptive notch filter algorithm with a center frequency of f0, and determine the filter bandwidth according to the following formula: B=κ×f0, where B is the filter bandwidth and κ is the bandwidth coefficient. Through environmental vibration frequency analysis and calibration, components in the vibration signal with frequencies exceeding the range of [f0-B / 2, f0+B / 2] are removed. To address temperature drift interference in stress-strain data: a linear compensation model for ambient temperature and stress data is established: σcorrection = σoriginal - kt × (t - t0), where σcorrection is the corrected stress data, σoriginal is the original stress data, kt is the temperature compensation coefficient, t is the measured ambient temperature, t0 is the standard ambient temperature, and kt is determined by fitting temperature-stress test data. All filtering parameters are dynamically adjusted based on a parameter mapping model constructed from real-time collected environmental parameters and structural dynamic response data.

[0024] As an optional embodiment, the result feedback of step S5 is integrated into the digital twin collaboration step: The digital twin model is reconstructed based on point cloud scanning data and BIM technology, and includes the geometric parameters, material properties, construction stage information, historical damage data and elastic-plastic time history analysis results of the lattice tube structure. After receiving the damage identification results, the digital twin model calls the stress concentration coefficient K of the damaged part and the structural material strength parameter [σ] to construct a damage development trend prediction model: d(t)=d0×exp(K×t / [σ]), where d(t) is the degree of damage at time t, d0 is the initial degree of damage, and t is the time variable. The virtual simulation prediction of the damage development trend is completed through this model. The digital twin model has a built-in model parameter sensitivity analysis module. It uses the Sobol sensitivity analysis algorithm to calculate the influence coefficient S of the number of attention heads in the Transformer module and the temporal feature capture window of the LSTM module on the damage recognition accuracy. The influence coefficient S is determined by the ratio of the parameter change to the recognition accuracy change. Set an accuracy threshold T, calculate the deviation Δ between the damage identification result and the simulation result of the digital twin model, and when Δ>T, sort the parameters based on the magnitude of the influence coefficient S and select the parameter with the largest S value as the priority adjustment object; Regarding the number of attention heads in the Transformer module: Based on the spatial distribution density ρ of the structural damage region, construct an integer programming objective function: min|ρ-ω×n|, where n is the number of attention heads and ω is the density-number correlation coefficient. Solve this function to output the suggested value for adjusting the number of attention heads. For the LSTM module's temporal feature capture window: Based on the temporal variation frequency f of the damage data, a sliding window optimization model is constructed: Lopt=round(1 / f×τ), where Lopt is the optimal window length and τ is the time constant. Through temporal data feature analysis and calibration, the model outputs the suggested values ​​for adjusting the temporal feature capture window. The digital twin model directly transmits the above adjustment suggestions to the model training unit in step S3 to guide the optimization of model training parameters; the damage identification accuracy data in step S4 is fed back to the digital twin model in real time, and the simulation parameters of the digital twin model are adjusted by correcting the boundary condition parameters in the simulation model.

[0025] As an optional embodiment, the damage determination in step S4 is based on the structural mechanical properties and seismic response characteristics: By combining structural design safety standards, material strength parameters, structural service life and seismic vulnerability analysis results, a damage characteristic threshold library is constructed, which contains characteristic thresholds for different structural parts and different service environments. The difference ΔV between the damage feature vector output by the fusion model and the feature threshold corresponding to the structural elastic limit state is calculated. Combined with the structural dynamic response change rate γ, where γ is the ratio of the difference between dynamic response data at adjacent time points to the initial dynamic response data, the degree of damage is determined according to the following rules: When ΔV<0 and γ<0, it is determined to be without damage; When 0≤ΔV<ΔV1 and 0≤γ<γ1, it is judged as a minor injury; When ΔV1≤ΔV<ΔV2 and γ1≤γ<γ2, it is judged as moderate injury; When ΔV≥ΔV2 and γ≥γ2, it is determined to be a severe injury; Where ΔV1 and ΔV2 are the segmented values ​​of the characteristic threshold, and γ1 and γ2 are the segmented values ​​of the dynamic response change rate, all of which are determined based on structural mechanics test data; The feature threshold is dynamically adjusted according to the following formula: Tadjust=Tbase×(1+μ×C+ν×E), where Tadjust is the adjusted threshold, Tbase is the baseline threshold, C is the structure type correction coefficient, E is the service environment correction coefficient, and μ and ν are adjustment coefficients. Threshold correction for regions with significant torsional vibration effects: Calculate the coupling coefficient ζ = A × f between the vibration amplitude A and frequency f in this region, and correct the threshold according to Ttwist = Tadjust × (1 + ξ × ζ), where ξ is the torsional correction coefficient; Threshold correction for dynamic coupling parts: Calculate the coupling strength τ of the part (τ is the vibration energy transfer efficiency between the dynamic coupling part and the main structure), and correct the threshold according to Tcouple=Tadjust×(1+ψ×τ), where ψ is the coupling correction coefficient; All correction factors were calibrated using structural dynamic response test data.

[0026] As an optional embodiment, a full-process parameter adaptive adjustment step is also included: State monitoring modules are respectively set in the data acquisition device, the data processing unit and the model operation platform. The state monitoring module collects the device operation parameters in real time, including the sensor sampling frequency, the device power consumption, =, the algorithm operation parameters, including the filtering parameters, the feature extraction threshold, =, and the model operation status data, including the iteration number, the loss function value; The control unit receives the feedback on the data preprocessing effect: calculate the signal-to-noise ratio SNR of the preprocessed data. When SNR < SNR threshold, adjust the data preprocessing algorithm parameters according to the following formula: Pproc = Pproc0 × (1 + λ × (SNR threshold - SNR) / SNR threshold), where Pproc is the adjusted algorithm parameter, Pproc0 is the initial algorithm parameter, and λ is the preprocessing parameter adjustment coefficient; The control unit receives the feedback on the model recognition accuracy: calculate the mean square error MSE between the damage recognition result and the digital twin simulation result. When MSE > MSE threshold, adjust the model hyperparameters according to the hyperparameter adjustment rule of claim 5; The control unit receives the feedback on the digital twin deduction result: extract the damage evolution rate v. When v > v threshold, increase the sensor sampling frequency according to the following formula: fsensor = fsensor0 × (1 + φ × v / v threshold), where fsensor is the adjusted sampling frequency, fsensor0 is the initial sampling frequency, and φ is the frequency adjustment coefficient; Calculate the damage recognition error e, where e is the difference between the recognition result and the actual damage state. When e > e threshold, and e threshold is determined based on the allowable range of structural elastic deformation, start the model incremental training process: add the latest damage data and the structural dynamic response change data to the training sample set according to the time series, keep the sample set size constant, use the sliding window mechanism to剔除 the earliest sample data, and retrain the model.

[0027] As an optional embodiment, the structural damage recognition in step S4 incorporates the failure mode association step: Combined with the structural seismic vulnerability curve data, the modal decomposition response spectrum analysis results and the potential failure mode recognition data, establish the association mapping between damage and structural failure modes, bind the damage location to the structural weak parts, and associate the damage degree with failure modes such as plastic hinge formation and joint buckling; When it is recognized that the damage feature vector exceeds the elastic limit threshold, predict the possible structural failure path caused by the damage through the fusion model; Simultaneously generate targeted emergency disposal suggestions and reinforcement plans. This step directly calls the digital twin model deduction data of step S7, and the seismic ground motion spectrum adaptation result of step S4 provides the working condition adaptation basis for the early warning plan.

[0028] Embodiment: The following explanation, based on a case study of a large-scale lattice-type cylindrical TV tower project, follows the standard format with examples, comparative examples, and results analysis. All operations and data are derived from on-site engineering verification and strictly adhere to the core technologies of the "seismic toughness of large-scale spatial lattice-type cylindrical structures" topic, ensuring the reproducibility of the technical solutions.

[0029] I. Basic Experimental Conditions (a) Experimental subjects A lattice-type cylindrical TV tower was selected. The main body is an all-steel structure with a total height of over 300m. It contains multiple open sections and exhibits abrupt changes in vertical stiffness. A mast antenna is installed at the top, which has a dynamic coupling effect with the main structure. The structure is mainly subjected to horizontal vibration, accompanied by significant torsional vibration. The ratio of the third torsional period to the first translational period meets the requirements of the current seismic design code for buildings. It is a typical lattice-type cylindrical structure with unique characteristics of geometric asymmetry and complex dynamic response.

[0030] (II) Experimental Equipment and Software category Specific configuration Multi-source acquisition equipment Strain gauges (monitor structural stress and strain), triaxial accelerometers (capture X / Y / Z vibration signals), industrial vision cameras (acquire images of structural surfaces), and environmental monitoring equipment (record environmental parameters such as temperature, humidity, and wind load). Auxiliary analysis equipment Total station (for reference calibration), spectrum analyzer (for processing vibration signals), industrial control computer (for data processing and model training). Core Software and Algorithms The system includes a deep learning framework-based LSTM-Transformer fusion model, a metaheuristic NRBO optimization algorithm, finite element analysis software (for modal analysis and elastoplastic time history analysis), a BIM+point cloud scanning digital twin platform, a modal decomposition response spectrum analysis module, and seismic vulnerability analysis tools. II. Implementation Examples Example 1: Damage Recognition Based on Modal Analysis Fusion Model Training Multi-source data acquisition: Acquisition equipment is deployed layer by layer along key stress-bearing parts of the TV tower, concentrated node areas, regions of abrupt changes in vertical stiffness, and dynamic coupling points, comprehensively covering the open areas, core node areas, and mast connection points. Strain gauges, accelerometers, industrial vision cameras, and environmental monitoring equipment are simultaneously activated to continuously acquire structural vibration signals, stress-strain data, surface visual images, and environmental parameters, forming a multi-dimensional raw dataset. During acquisition, the results of structural dynamic characteristic analysis are referenced to ensure complete data coverage of the coupling characteristics of horizontal and torsional vibrations, avoiding the loss of key response data.

[0031] Data preprocessing: Noise suppression, data alignment, and feature enhancement operations are performed sequentially on the raw data. For environmental interference components in the vibration signal, an adaptive notch filter algorithm is used to remove irrelevant frequency components, taking into account the inherent frequency characteristics of the structure. For temperature drift in stress-strain data, a linear compensation algorithm based on ambient temperature and stress data is used to correct the deviation. For dust interference in visual images, morphological filtering is used to remove floating dust noise, and then an image restoration algorithm corresponding to the grayscale distribution of the structural surface material is used to repair the pixel information of feature point regions. After noise suppression, the temporal peak value and frequency domain dominant frequency of the vibration signal, the extreme value features of the stress-strain data, and the surface texture and geometric deformation features of the visual image are extracted. Data alignment is then performed to unify the temporal dimension, forming a standardized dataset.

[0032] Model Construction and Training: An LSTM-Transformer fusion model was constructed. The Transformer module extracted spatial correlation features from the data, focusing on capturing the stress distribution differences in areas of concentrated nodes and abrupt stiffness changes. The LSTM module captured the temporal dynamic features of the data, accurately identifying the periodicity and abrupt changes in vibration signals. Annotated structural damage sample data was used to train the model. Modal analysis results from the multiple Ritz vector method were incorporated into the training process, enabling the model to actively learn the coupling law between horizontal and torsional vibrations of the structure, enhancing its ability to capture torsional damage features. A metaheuristic NRBO optimization algorithm was introduced to automatically adjust the number of attention heads in the Transformer module, the number of hidden layer neurons in the LSTM module, and the model learning rate. These adjustments were based on the temporal continuity, spatial distribution characteristics, and probability distribution of the seismic vulnerability curve of the structural damage data, ensuring that the model parameters were adapted to the structural dynamic response characteristics.

[0033] Damage identification: Standardized datasets are input into the trained model, which outputs damage feature vectors for each monitored area. Combining structural mechanical properties, design parameters, and seismic response results obtained from elastoplastic time history analysis, damage locations are identified through feature vector matching and comparison with dynamic response patterns. Damage types are determined based on the dimensional changes of the damage feature vectors and the mechanical properties of the structural materials. The degree of damage is quantified by combining seismic vulnerability analysis results with the structural elastic limit state parameters.

[0034] Application of Results: Damage identification results are pushed to the structural health management platform and digital twin model, providing accurate data support for the formulation of structural maintenance and reinforcement plans and the adjustment of construction techniques.

[0035] Example 2: Damage Identification Optimization Based on Multi-Hazard Coupling Adaptation Basic data acquisition and preprocessing: Following the multi-source data acquisition and preprocessing steps of Example 1, structural response data under simulated far-field long-period earthquakes and near-field pulse earthquakes were acquired. The simulation conditions were set with reference to the differences in the seismic motion spectrum characteristics to ensure that the data covered the structural dynamic response characteristics under different seismic conditions. During the preprocessing process, the filtering algorithm and feature extraction logic remained consistent, and the filtering parameters were only fine-tuned based on the signal characteristics of the newly added seismic data.

[0036] Seismic motion spectrum adaptation: For scenarios involving far-field long-period earthquakes, near-field pulse earthquakes coupled with wind loads and temperature deformation, spectral characteristic parameters under different working conditions are first extracted using seismic motion spectrum analysis tools. Combined with structural dynamic response test results, the dominant damage characteristics under each working condition are identified. Based on the analysis results, the feature extraction weights of the fusion model are adjusted: For far-field long-period earthquake scenarios, the mapping relationship between long-period vibration signals and damage characteristics is strengthened, focusing on capturing cumulative damage characteristics induced by resonance; for near-field pulse earthquake scenarios, the feature weights of stress mutation signals are strengthened to accurately identify sudden damage caused by instantaneous loads. Simultaneously, through analysis of the phase difference and amplitude ratio of vibration signals at dynamic coupling points, the damage identification bias caused by the dynamic coupling effect between the mast and the main body is corrected to ensure that damage characteristics at coupling points are not masked.

[0037] Damage identification and verification: The adapted dataset is input into the model, and the damage identification results are output. The seismic vulnerability analysis tool is called, and the probability distribution of vulnerability curves under the corresponding working conditions is combined to verify the rationality of the damage location, type and degree. For areas where the identification results are biased, the feature extraction weights and model parameters are adjusted retrospectively to form a closed loop of working condition adaptation and model optimization.

[0038] Results Feedback: The verified damage identification results are fed back to the digital twin model. The damage development trend is inferred through virtual simulation, and the model feature extraction weights and working condition adaptation logic are optimized in reverse to improve the model's adaptability in complex disaster scenarios.

[0039] Example 3: Damage Identification through Digital Twin Collaboration and End-to-End Control Basic process execution: The acquisition, preprocessing, model training and damage identification steps of Example 2 are followed to ensure that the technical logic of each step is consistent with the research method and that data transmission and processing remain continuous.

[0040] Digital Twin Collaboration: The digital twin model is reconstructed based on point cloud scan data and BIM technology, integrating structural geometric parameters, material properties, construction stage information, historical damage data, and elastoplastic time history analysis results. After receiving the damage identification results, the digital twin model simulates the stress evolution process of the damaged area under different load conditions through virtual simulation, based on the stress concentration factor of the damaged area and the structural material strength parameters. Based on the simulation results, it outputs back-end suggestions for model parameter adjustment, including local optimization of the number of attention heads in the Transformer module and adjustment of the temporal feature capture window in the LSTM module, directly guiding the optimization of model training parameters; at the same time, it optimizes the deployment location and acquisition frequency of multi-source acquisition devices according to the damage propagation path in the virtual simulation.

[0041] Full-process parameter control: Status monitoring modules are set up in the data acquisition equipment, data processing unit, and model computing platform to collect equipment operating parameters, algorithm running parameters, and model computing status data in real time. The control unit receives feedback on the data preprocessing effect. When the signal-to-noise ratio of the noise-suppressed signal does not meet the feature extraction requirements, it automatically adjusts the filtering algorithm parameters; it receives feedback on the model recognition accuracy. When the deviation between the damage recognition result and the digital twin simulation result exceeds the allowable range, it adjusts the model hyperparameters; it receives feedback on the digital twin inference result. When the predicted damage evolution rate accelerates, it increases the sensor acquisition frequency. When the damage recognition error exceeds the allowable range of structural elastic deformation, the incremental training process of the model is initiated, integrating the incremental training samples into the latest damage data and structural dynamic response change data to improve the model's adaptability to structural damage evolution.

[0042] Failure Mode Early Warning: Combining structural seismic vulnerability curve data, modal decomposition response spectrum analysis results, and potential failure mode identification data, a correlation mapping between damage and structural failure modes is established—binding the damage location to weak points in the structure (nodal concentration areas, stiffness abrupt change regions), and associating the damage degree with failure modes such as plastic angle formation and nodal buckling. When the damage feature vector is identified as exceeding the elastic limit threshold, the fusion model predicts the structural failure path that the damage may trigger; based on the damage propagation rate and structural mechanical properties obtained from digital twin simulation, targeted emergency response suggestions and reinforcement schemes are generated, clarifying the reinforcement locations, reinforcement material selection, and key construction techniques.

[0043] III. Comparative Example Traditional structural damage identification schemes are adopted: stress and vibration data are collected only through strain gauges and accelerometers, and data processing uses a single threshold judgment without incorporating modal analysis, seismic motion spectrum adaptation, and digital twin collaborative logic; the model uses a single time series model without combining spatial feature extraction modules and hyperparameter optimization algorithms; the torsional vibration effect, dynamic coupling characteristics, and multi-hazard coupling effects of lattice cylindrical structures are not considered, and it is only suitable for single-condition damage identification of simple symmetrical structures, and cannot capture damage characteristics under complex dynamic responses.

[0044] IV. Experimental Results and Comparative Analysis Monitoring indicators Example 1 Example 2 Example 3 Comparative Example Accuracy of Damage Location Identification No obvious deviation No deviation No deviation Frequent deviations exist Torsional vibration-related damage identification Recognizable Accurate identification Accurate identification Difficult to identify Multi-hazard coupling adaptability Partial adaptation Fully compatible Fully compatible Only compatible with single seismic action Damage evolution trend prediction none Initially possessing Accurate prediction none Environmental interference resistance Strong powerful powerful Weak Results Analysis Example 1 effectively solves the problem of identifying torsional vibration-related damage in lattice cylindrical structures by integrating the modal analysis results obtained from the multiple Ritz vector method into the training of the LSTM-Transformer fusion model and introducing the NRBO optimization algorithm to automatically adjust the model hyperparameters. Compared with traditional methods, its damage location identification accuracy is significantly improved, its environmental interference resistance is greatly enhanced, and it can accurately capture the damage characteristics of the open area and the stiffness abrupt change area, which is consistent with the unique characteristics of the complex dynamic response of the structure.

[0045] Example 2 introduces seismic motion spectrum adaptation logic. For scenarios involving far-field long-period earthquakes, near-field pulse earthquakes coupled with wind loads and temperature deformation, it optimizes the model feature extraction weights by combining the differences between structural dynamic response laws and seismic motion spectrum characteristics, correcting identification biases caused by dynamic coupling effects. This design overcomes the shortcomings of traditional methods that only adapt to a single seismic action, achieving accurate damage identification under complex disaster conditions and ensuring that damage characteristics are not missed or misjudged under different conditions.

[0046] Example 3, through digital twin collaboration and full-process parameter control, constructs a dynamic linkage mechanism. This mechanism can not only accurately identify the location and extent of damage, but also predict the damage evolution trend based on elastoplastic time history analysis results, and generate emergency response plans by combining seismic vulnerability data. Compared with the previous two examples, its newly added full-process parameter control function can dynamically optimize the monitoring and identification logic according to the structural status, and the failure mode early warning function can avoid structural failure risks in advance, providing full-chain support for the full-cycle safety management of the structure and fully meeting the actual needs of the construction and operation and maintenance of lattice tube structures.

[0047] Due to the lack of structural-specific characteristic adaptation, multi-condition optimization, and collaborative control logic, the comparative model is unable to address the core pain points of lattice cylindrical structures such as torsional vibration and dynamic coupling. The accuracy of damage identification, adaptability to operating conditions, and anti-interference ability are all significantly insufficient, making it difficult to meet the damage identification needs of complex lattice cylindrical structures.

[0048] V. Conclusions of the Examples This invention addresses the core challenges of damage identification in lattice-type cylindrical structures by deeply integrating the LSTM-Transformer fusion model with the core analytical methods of the research topic. Example 3 demonstrates the best overall performance, enabling accurate identification of torsional vibration-related damage, complete adaptation to multi-hazard coupling scenarios, and precise prediction of damage evolution trends. It also possesses full-process dynamic control and failure mode early warning capabilities, fully meeting the safety monitoring needs of lattice-type cylindrical structures during construction and operation.

[0049] The embodiments described herein have been described in sufficient detail to enable those skilled in the art to practice the disclosed teachings. Other embodiments may be used and derived therefrom, such that structural and procedural substitutions and changes may be made without departing from the scope of this disclosure. Therefore, the detailed description should not be construed as limiting, and the scope of the various embodiments is defined only by the appended claims and the full scope of their equivalents.

Claims

1. A method for intelligent structural damage identification based on an LSTM-Transformer fusion model, characterized in that, Includes the following steps: S1 Multi-Source Data Acquisition: For key stress-bearing parts, concentrated node areas, vertical stiffness abrupt change areas, and dynamic coupling parts of the lattice tube structure, strain gauges, accelerometers, and visual acquisition devices are deployed to simultaneously acquire structural vibration signals, stress and strain data, surface visual images, and environmental parameters. The acquisition range covers the open area, the core node area, and the mast connection parts to form a multi-dimensional raw dataset. S2 data preprocessing: Noise suppression, data alignment and feature enhancement operations are performed sequentially on the multi-source raw data to eliminate deviations caused by equipment vibration, environmental interference and asynchronous acquisition. The time domain peak value, frequency domain main frequency, extreme value features of stress and strain data and surface texture and geometric deformation features of visual images are extracted to form a standardized dataset. S3 Fusion Model Construction and Training: Construct an LSTM-Transformer fusion model, extract spatial correlation features of data through the Transformer module, and capture temporal dynamic features of data using the LSTM module; The model is trained using labeled structural damage sample data and modal analysis results obtained by the multiple Ritz vector method. This allows the model to learn the coupling characteristics of horizontal and torsional vibrations of the structure, and optimize the model parameters to improve the accuracy of damage feature identification. S4 Structural Damage Identification: A standardized dataset is input into the trained model, which outputs damage feature vectors for each monitored area. A damage assessment system is constructed by utilizing the stiffness matrix and damping coefficient from the structural mechanical property parameters, the material strength parameters and structural geometric dimensions from the design parameters, and the peak displacement, peak acceleration, and stress distribution data from the seismic response data generated by elastoplastic time history analysis. S5 Results Feedback and Application: The damage identification results are transmitted to the structural health management platform and simultaneously pushed to the digital twin model, providing data support for structural maintenance and reinforcement, construction process adjustment and disaster emergency response.

2. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 1, characterized in that, Step S3 incorporates a structural material nonlinear compensation step into the model training process: The specific implementation is as follows: Obtain the measured stress and strain data from the multi-source data collected in step S1, and combine it with the natural frequency and mode shape data obtained from the structural dynamic characteristic analysis to establish the material parameter deviation correction equation: δP=k×(σmeasured-σpredicted), where δP is the material parameter correction amount, k is the correction coefficient, σmeasured is the stress value in the measured stress and strain data, σpredicted is the stress prediction value based on the initial sample parameters, and k is determined by fitting the structural material constitutive relation test data; The influence of node construction deviation, welding residual stress and material mechanical property fluctuation formed during the construction phase is incorporated to construct a sample parameter calibration model: Pcalibration = Pinitial + δP + δP1 + δP2 + δP3, where Pcalibration is the calibrated sample parameter, Pinitial is the initial sample parameter, δP1 is the parameter correction amount corresponding to node construction deviation, δP2 is the parameter correction amount corresponding to welding residual stress, and δP3 is the parameter correction amount corresponding to material mechanical property fluctuation. δP1 is calculated by the difference between the measured stress and strain data at the node and the design stress data; δP2 is calculated based on the correlation model between welding process parameters and residual stress; and δP3 is determined by the difference between the material mechanical property test data and the standard parameters. Substituting the P calibration into the model training samples completes the reverse calibration of the sample parameters. This compensation step directly correlates the multi-source data and structural modal analysis results collected in step S1 through the above equation and model.

3. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 2, characterized in that, Step S2, multi-source data preprocessing, includes a feature fusion step: Extract the time-domain peak value and frequency-domain dominant frequency of the vibration signal, the extreme value features of the stress-strain data, and the structural surface texture features and geometric deformation features of the visual image to construct a multi-dimensional feature set; The modal response spectrum method was used to calculate the contribution of each mode. The modal contribution was determined by the weighted sum of the modal participation coefficient and the modal energy proportion. Based on the accuracy characteristics and response speed of each type of data, and combined with the modal contribution, the weight is assigned as follows: Wi=W0i×(α×Ci+β), where Wi is the final weight of the i-th type of data, W0i is the initial weight of the i-th type of data, Ci is the modal contribution of the monitoring part corresponding to the i-th type of data, and α and β are weight adjustment coefficients, which are calibrated through multi-source data fusion experiments. The multi-dimensional feature sets are weighted and fused according to the above weights to generate a fused feature vector, thus completing the feature fusion operation.

4. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 3, characterized in that, Step S4, structural damage identification, is integrated into the seismic motion spectrum adaptation step: In advance, the spectral characteristic parameters of far-field long-period earthquakes, near-field pulse earthquakes and wind loads and temperature deformation coupling scenarios are extracted using ground motion spectrum analysis tools. These parameters include spectral peak value, dominant period, and spectral intensity. Combined with the dynamic response test data of the lattice tube structure, a mapping library of different scenarios and feature extraction rules is established. For long-period earthquake scenarios in the far field, a correlation matrix M between long-period vibration signals and damage characteristics is constructed. The elements of the correlation matrix are determined by training with structural damage test data under long-period earthquake loading. Matrix multiplication is used to multiply the feature vector of the long-period vibration signal with M to strengthen the mapping relationship between the two. For near-field pulsed seismic scenarios, a feature weight adjustment module is set up. Based on the peak ground acceleration and pulse duration of near-field pulsed seismic events, the weight coefficient ω of the stress mutation signal is calculated through a linear weighting algorithm. The weight coefficient ω is directly multiplied by the stress mutation feature vector to adjust the feature weight. Vibration signals from the dynamic coupling parts are collected, and the phase difference Δφ and amplitude ratio λ of the vibration signals are calculated. A deviation correction model is established: Fcorrected = Foriginal - α × Δφ - β × λ, where Fcorrected is the corrected damage feature vector, Foriginal is the uncorrected damage feature vector, and α and β are correction coefficients. The model is calibrated using damage simulation test data of the dynamic coupling parts and corrects the damage identification deviation caused by the dynamic coupling effect between the mast and the main body.

5. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 4, characterized in that, In step S3, the metaheuristic NRBO optimization algorithm is introduced during model training to automatically adjust the hyperparameters of the LSTM-Transformer fusion model. The specific implementation is as follows: The hyperparameters that need to be adjusted include the number of attention heads in the Transformer module, the number of hidden layer neurons in the LSTM module, and the model learning rate. The damage probability density of each region of the structure is calculated by the seismic vulnerability curve. Regions with damage probability density higher than the preset density threshold are identified as regions with concentrated damage probability. The number of hidden layer neurons in the LSTM module corresponding to this region is adjusted according to the following formula: Nnew=Nold×(1+γ×ρ), where Nnew is the number of neurons after adjustment, Nold is the number of neurons before adjustment, γ is the number adjustment coefficient, and ρ is the ratio of damage probability density to threshold. The time-series variation variance of structural damage data is calculated. Data with variance higher than a preset variance threshold is identified as damage data with drastic time-series variation. The learning rate of the model corresponding to this type of data is adjusted according to the following formula: ηnew=ηold×(1-δ×σ), where ηnew is the adjusted learning rate and ηold is the number of attention heads in the Transformer module. Based on the spatial distribution characteristics of structural damage data, the data is divided into cluster regions by a spatial clustering algorithm. The number of attention heads is adjusted according to the spatial density of the cluster regions. The higher the spatial density, the more attention heads are required. The hyperparameter tuning process is performed iteratively using the NRBO optimization algorithm until the model recognition accuracy meets the preset requirements.

6. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 5, characterized in that, The noise suppression in step S2 employs an adaptive filtering process: To address visual image noise caused by dust interference: Morphological opening operations are used to remove floating dust noise. A Bayesian image restoration algorithm is constructed based on the gray-level distribution law of the surface material to repair pixel information in the feature point region. A gray-level gradient threshold T is set for the feature point. When the gray-level gradient of the feature point is lower than T, the repair parameter is adjusted according to the following formula: θnew=θold×(1+ε×(Tg) / T), where θnew is the repair parameter after adjustment, θold is the repair parameter before adjustment, ε is the parameter adjustment coefficient, and g is the actual gray-level gradient of the feature point. To address environmental vibration interference in vibration signals: extract the structure's natural frequency f0, construct an adaptive notch filter algorithm with a center frequency of f0, and determine the filter bandwidth according to the following formula: B=κ×f0, where B is the filter bandwidth and κ is the bandwidth coefficient. Through environmental vibration frequency analysis and calibration, components in the vibration signal with frequencies exceeding the range of [f0-B / 2, f0+B / 2] are removed. To address temperature drift interference in stress-strain data: a linear compensation model for ambient temperature and stress data is established: σcorrection = σoriginal - kt × (t - t0), where σcorrection is the corrected stress data, σoriginal is the original stress data, kt is the temperature compensation coefficient, t is the measured ambient temperature, t0 is the standard ambient temperature, and kt is determined by fitting temperature-stress test data. All filtering parameters are dynamically adjusted based on a parameter mapping model constructed from real-time collected environmental parameters and structural dynamic response data.

7. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 6, characterized in that, The results feedback of step S5 is integrated into the digital twin collaboration steps: The digital twin model is reconstructed based on point cloud scanning data and BIM technology, and includes the geometric parameters, material properties, construction stage information, historical damage data and elastic-plastic time history analysis results of the lattice tube structure. After receiving the damage identification results, the digital twin model calls the stress concentration coefficient K of the damaged part and the structural material strength parameter [σ] to construct a damage development trend prediction model: d(t)=d0×exp(K×t / [σ]), where d(t) is the degree of damage at time t, d0 is the initial degree of damage, and t is the time variable. The virtual simulation prediction of the damage development trend is completed through this model. The digital twin model has a built-in model parameter sensitivity analysis module. It uses the Sobol sensitivity analysis algorithm to calculate the influence coefficient S of the number of attention heads in the Transformer module and the temporal feature capture window of the LSTM module on the damage recognition accuracy. The influence coefficient S is determined by the ratio of the parameter change to the recognition accuracy change. Set an accuracy threshold T, calculate the deviation Δ between the damage identification result and the simulation result of the digital twin model, and when Δ>T, sort the parameters based on the magnitude of the influence coefficient S and select the parameter with the largest S value as the priority adjustment object; Regarding the number of attention heads in the Transformer module: Based on the spatial distribution density ρ of the structural damage region, construct an integer programming objective function: min|ρ-ω×n|, where n is the number of attention heads and ω is the density-number correlation coefficient. Solve this function to output the suggested value for adjusting the number of attention heads. For the LSTM module's temporal feature capture window: Based on the temporal variation frequency f of the damage data, a sliding window optimization model is constructed: Lopt=round(1 / f×τ), where Lopt is the optimal window length and τ is the time constant. Through temporal data feature analysis and calibration, the model outputs the suggested values ​​for adjusting the temporal feature capture window. The digital twin model directly transmits the above adjustment suggestions to the model training unit in step S3 to guide the optimization of model training parameters; the damage identification accuracy data in step S4 is fed back to the digital twin model in real time, and the simulation parameters of the digital twin model are adjusted by correcting the boundary condition parameters in the simulation model.

8. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 7, characterized in that, The damage assessment in step S4 is based on the structural mechanical properties and seismic response characteristics: By combining structural design safety standards, material strength parameters, structural service life and seismic vulnerability analysis results, a damage characteristic threshold library is constructed, which contains characteristic thresholds for different structural parts and different service environments. The difference ΔV between the damage feature vector output by the fusion model and the feature threshold corresponding to the structural elastic limit state is calculated. Combined with the structural dynamic response change rate γ, where γ is the ratio of the difference between dynamic response data at adjacent time points to the initial dynamic response data, the degree of damage is determined according to the following rules: When ΔV<0 and γ<0, it is determined to be without damage; When 0≤ΔV<ΔV1 and 0≤γ<γ1, it is judged as a minor injury; When ΔV1≤ΔV<ΔV2 and γ1≤γ<γ2, it is judged as moderate injury; When ΔV≥ΔV2 and γ≥γ2, it is determined to be a severe injury; Among them, ΔV1 and ΔV2 are characteristic threshold segmentation values, and γ1 and γ2 are dynamic response change rate segmentation values, both determined based on structural mechanics test data; The characteristic threshold is dynamically adjusted according to the following formula: Tadjust = Tbase×(1 + μ×C + ν×E), where Tadjust is the adjusted threshold, Tbase is the reference threshold, C is the structural type correction coefficient, E is the service environment correction coefficient, and μ and ν are adjustment coefficients; Threshold correction for the region with significant torsional vibration effect: Calculate the coupling coefficient ζ = A×f of the vibration amplitude A and frequency f in this region, and the threshold is corrected according to Ttwist = Tadjust×(1 + ξ×ζ), where ξ is the torsional correction coefficient; Threshold correction for the dynamically coupled part: Calculate the coupling strength τ of this part (τ is the vibration energy transfer efficiency between the dynamically coupled part and the main structure), and the threshold is corrected according to Tcouple = Tadjust×(1 + ψ×τ), where ψ is the coupling correction coefficient; All correction coefficients are calibrated through structural dynamic response test data.

9. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 8, characterized in that, It also includes the step of adaptive regulation of full-process parameters: State monitoring modules are respectively set in the data acquisition device, data processing unit, and model operation platform. The state monitoring module collects the device working parameters in real time, including the sensor sampling frequency, device power consumption =, and algorithm operation parameters, including filtering parameters, feature extraction threshold =, and model operation status data, including the number of iterations and loss function values; The control unit receives the feedback on the data preprocessing effect: Calculate the signal-to-noise ratio SNR of the preprocessed data. When SNR < SNR threshold, adjust the data preprocessing algorithm parameters according to the following formula: Pproc = Pproc0×(1 + λ×(SNR threshold - SNR) / SNR threshold), where Pproc is the adjusted algorithm parameter, Pproc0 is the initial algorithm parameter, and λ is the preprocessing parameter adjustment coefficient; The control unit receives the feedback on the model recognition accuracy: Calculate the mean square error MSE between the damage recognition result and the digital twin simulation result. When MSE > MSE threshold, adjust the model hyperparameters according to the hyperparameter adjustment rule in claim 5; The control unit receives the feedback on the digital twin deduction result: Extract the damage evolution rate v. When v > v threshold, increase the sensor sampling frequency according to the following formula: fsensor = fsensor0×(1 + φ×v / v threshold), where fsensor is the adjusted sampling frequency, fsensor0 is the initial sampling frequency, and φ is the frequency adjustment coefficient; Calculate the damage recognition error e, where e is the difference between the recognition result and the actual damage state. When e > e threshold, and e threshold is determined based on the allowable range of structural elastic deformation, start the model incremental training process: Add the latest damage data and the structural dynamic response change data to the training sample set in time series, keep the scale of the sample set constant, and use the sliding window mechanism to剔除 the earliest sample data and retrain the model.

10. The intelligent structural damage recognition method based on the LSTM-Transformer fusion model as described in claim 9, characterized in that, The structural damage recognition in step S4 incorporates the failure mode association step: By combining structural seismic vulnerability curve data, modal decomposition response spectrum analysis results and potential failure mode identification data, a correlation mapping between damage and structural failure modes is established, binding the damage location to the weak parts of the structure and associating the degree of damage with failure modes such as plastic hinge formation and nodal buckling. When the damage feature vector is found to exceed the elastic limit threshold, the structural failure path that the damage may trigger is predicted by the fusion model. Simultaneously generate targeted emergency response suggestions and reinforcement plans. This step directly calls the digital twin model simulation data from step S7, and the seismic motion spectrum adaptation results from step S4 provide a basis for the working condition adaptation of the early warning plan.

Images