A structural modal parameter intelligent screening method and system based on feature fusion
By employing a feature fusion-based intelligent screening method for structural modal parameters, utilizing the random subspace algorithm and the XGBoost model, the problems of incomplete false modality identification and reliance on manual parameter tuning are solved. This enables efficient and automated screening of modal parameters for facilities such as bridges and dams, improving identification accuracy and reliability.
Patent Information
- Application Number
- CN202610154375.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-03
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies for modal parameter identification in the engineering structural health monitoring of major infrastructure such as bridges and dams suffer from problems such as the incorporation of false modes, strong reliance on manual parameter tuning, and inaccurate identification and high operation and maintenance costs due to the single feature dimension.
A feature fusion-based intelligent screening method for structural modal parameters is adopted. Initial modal parameters are identified through a random subspace algorithm, a multi-dimensional feature fusion vector is constructed and input into a pre-trained XGBoost model, the true and false probabilities of the modal are output, a probability threshold is set to screen the true modal and eliminate the false modal, and the process is repeated in subsequent data blocks.
It has achieved the complete elimination of spurious modalities, reduced operation and maintenance costs, improved the accuracy of modal parameters and the level of automation in health monitoring, and enhanced the reliability of structural damage early warning for critical infrastructure.
Smart Images
Figure CN122262991A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of engineering structure health monitoring and machine learning, and in particular to an intelligent screening method and system for structural modal parameters based on feature fusion. Background Technology
[0002] In the field of structural health monitoring for major infrastructure projects such as bridges and dams, modal parameters (frequency, damping ratio, and mode shape) are core indicators for assessing structural safety, and their identification accuracy directly determines the reliability of the health monitoring system. Currently, mainstream technologies in the industry rely on algorithms such as the Stochastic Subspace Method (SSI) to extract modal parameters, and use clustering algorithms such as K-means and DBSCAN combined with manual parameter tuning to screen for true modes.
[0003] Existing technologies suffer from the following drawbacks: 1. Traditional clustering algorithms are sensitive to multi-source noise such as environmental noise and sensor interference, failing to effectively distinguish between structural physical modes and noise-induced spurious modes. This results in a large amount of invalid data being mixed into the modal parameters, affecting the accuracy of structural safety assessments. 2. Key parameters such as cluster radius and density threshold require repeated trial-and-error adjustments based on expert experience, which is not only inefficient but also makes it difficult for the system to achieve automated continuous operation, significantly increasing maintenance costs. 3. Existing methods focus only on the single feature of frequency stability, neglecting the influence of damping ratio consistency, mode shape correlation, and environmental factors (such as temperature and water level) on modal identification, leading to insufficient criteria for distinguishing genuine and spurious modes.
[0004] Existing machine learning solutions have limitations: although some studies have attempted to introduce machine learning models such as SVM, they are limited by using only a single feature of frequency variance and have low computational efficiency, which cannot meet the engineering needs of real-time monitoring of major infrastructure. Summary of the Invention
[0005] The main objective of this invention is to provide an intelligent screening method for structural modal parameters based on feature fusion, which solves the technical problems of incomplete removal of false modalities, strong reliance on manual parameter tuning, and insufficient judgment criteria due to single feature dimensions.
[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an intelligent screening method for structural modal parameters based on feature fusion, comprising the following steps: S1. The structural response signal is divided into blocks using a preset time window to obtain several data blocks arranged in chronological order. S2. Modal parameter identification is performed on the data blocks using a random subspace algorithm to obtain an initial set of modal parameters including frequency, damping ratio, and mode shape vector; S3. Construct a multi-dimensional feature fusion vector based on the initial modal parameter set and the structural response signal, input the multi-dimensional feature fusion vector into the pre-trained classification model, output the modal authenticity probability, and set a probability threshold to filter out real modalities and eliminate false modalities. S4. Based on the stable modal parameters obtained from the screening, repeat the process from S2 to S3 on the subsequent data blocks to output the final modal parameters.
[0007] In a preferred embodiment, the step of identifying modal parameters of the data blocks using a random subspace algorithm includes: S21. A covariance-driven random subspace algorithm is used to identify each data block to obtain the modal parameters of each data block, wherein the modal parameters include the frequency, the damping ratio and the mode shape vector. S22. Before performing block processing on the structural response signal, the structural response signal is preprocessed. The preprocessing includes drift elimination, filtering, and resampling. The preprocessed structural response signal is obtained through the preprocessing and is used as the input for the block processing.
[0008] In a preferred embodiment, before dividing the structural response signal into blocks, step S1 further includes preprocessing the original structural response signal, wherein the preprocessing method includes drift elimination, filtering, and resampling.
[0009] In a preferred embodiment, constructing a multi-dimensional feature fusion vector based on the initial modal parameter set and the structural response signal includes: S31. Extract frequency stability features, damping ratio consistency features, and mode shape correlation features from the initial modal parameter set, and extract environmental features from the structural response signal; S32. The frequency stability feature, the damping ratio consistency feature, the mode shape correlation feature and the environmental feature are fused into the multi-dimensional feature fusion vector, wherein the multi-dimensional feature fusion vector is a multi-dimensional vector; S33. Input the multi-dimensional feature fusion vector into the pre-trained classification model, and process the multi-dimensional feature fusion vector through the pre-trained classification model to obtain the modality authenticity probability.
[0010] In the preferred embodiment, the extraction of frequency stability features, damping ratio consistency features, and mode shape correlation features from the initial modal parameter set includes: S311. Calculate the short-term frequency variance, medium-term frequency variance, and long-term frequency variance for the frequency to obtain the frequency stability characteristics; S312. Calculate the adjacent damping difference rate and cumulative damping offset for the damping ratio to obtain the consistency characteristics of the damping ratio; S313. Calculate the correlation between the current mode and the previous mode, the correlation between the current mode and the filtered true mode, the correlation between the current mode and the adjacent lower-order mode, the correlation between the current mode and the adjacent higher-order mode, and the correlation between the current mode and the dominant mode for the mode vector to obtain the mode correlation characteristics. S314. Obtain real-time water level and real-time temperature from the structural response signal to obtain the environmental characteristics, and use the frequency stability characteristics, the damping ratio consistency characteristics, the mode shape correlation characteristics, and the environmental characteristics as components of the multi-dimensional feature fusion vector.
[0011] In the preferred scheme, the constructed multi-dimensional feature fusion vector is a 12-dimensional feature vector, specifically including: Three-dimensional frequency stability characteristics: short-term frequency variance, medium-term frequency variance, long-term frequency variance, and typical vibration period of the covered structure; Damping ratio consistency characteristics in 2D: adjacent damping difference rate, cumulative damping offset; The modal correlation characteristics are 5-dimensional: the correlation between the current time and the previous time, the correlation between the current time and the filtered true modes, the correlation between the current time and the adjacent lower-order modes, the correlation between the current time and the adjacent higher-order modes, and the correlation between the current time and the dominant mode. Two-dimensional environmental characteristics: real-time water level and real-time temperature.
[0012] In the preferred embodiment, inputting the multi-dimensional feature fusion vector into the pre-trained classification model includes: S34. The XGBoost model is used as the pre-trained classification model, the multi-dimensional feature fusion vector is used as the input of the XGBoost model, and the modality true / false probability is output as a continuous value through the XGBoost model. S35. Compare the modality authenticity probability with the probability threshold. If the modality authenticity probability is greater than or equal to the probability threshold, it is determined to be the true modality; otherwise, it is determined to be the false modality and is removed. S36. The initial modal parameter set is processed through the decision mechanism of the XGBoost model to obtain the filtered modal parameters, and the filtered modal parameters are used to determine the stable modal parameters.
[0013] In the preferred embodiment, the use of the XGBoost model as the pre-trained classification model includes: S341. Train the XGBoost model based on labeled samples from a historical monitoring dataset, wherein the labeled samples contain real modality labels and fake modality labels; S342. Optimize the hyperparameters of the XGBoost model using cross-validation, including learning rate, maximum tree depth, and subsample ratio, and control the number of iterations using early stopping. S343. Solidify the trained XGBoost model and embed it into an edge computing device, and perform inference of the multi-dimensional feature fusion vector through the edge computing device to obtain the modality true / false probability; S344. Support the incremental learning mechanism of the XGBoost model, and update the weights of the XGBoost model using new labeled data when the structural response signal changes.
[0014] In the preferred embodiment, the step of repeatedly executing steps S2 to S3 on subsequent data blocks based on the stable modal parameters obtained through screening includes: S41. Using the stable modal parameters as a benchmark, the random subspace algorithm is applied to the subsequent data blocks to identify the modal parameters, thereby obtaining a new set of initial modal parameters. S42. Construct a new multi-dimensional feature fusion vector based on the new initial modality parameter set, and input it into the pre-trained classification model to obtain the new modality true / false probability; S43. Based on the new modality true / false probability and the probability threshold, filter out new false modalities, and update the stable modality parameters; S44. Repeat the process until all data blocks are processed, and output the final modal parameters as the overall result.
[0015] Secondly, the present invention provides an intelligent screening system for structural modal parameters based on feature fusion, applicable to the aforementioned intelligent screening method for structural modal parameters based on feature fusion, comprising: The data segmentation module is used to segment the structural response signal into blocks using a set time window, resulting in several data blocks arranged in chronological order. The parameter identification module is used to identify modal parameters in the data blocks using a random subspace algorithm, and obtain an initial set of modal parameters including frequency, damping ratio and mode shape vector. The feature fusion module is used to construct a multi-dimensional feature fusion vector based on the initial modal parameter set and the structural response signal, input the multi-dimensional feature fusion vector into the pre-trained classification model, output the modal authenticity probability, and set a probability threshold to filter the real modal and directly eliminate the false modal. The modal parameter module is used to repeatedly execute the process from the parameter identification module to the feature fusion module on subsequent data blocks based on the stable modal parameters obtained through screening, and output the final modal parameters.
[0016] This invention provides an intelligent screening method for structural modal parameters based on feature fusion, comprising the following steps: S1, segmenting the structural response signal into blocks using a preset time window; S2, identifying modal parameters in the data blocks using SSI-Cov to obtain an initial set of modal parameters including frequency, damping ratio, and mode shape vector; S3, constructing a multi-dimensional feature fusion vector based on the initial set of modal parameters and the structural response signal, inputting the multi-dimensional feature fusion vector into a pre-trained classification model, outputting the modal authenticity probability, and performing screening; S4, repeating the process for subsequent data blocks based on the screened stable modal parameters, and outputting the final modal parameters. This invention improves the automation level of health monitoring; increases the false modal rejection rate, solving the problem of incomplete false modal rejection; reduces operation and maintenance costs, improves error identification accuracy, and enhances the reliability of early warning for damage to arch dam structures. Attached Figure Description
[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the multidimensional feature construction of the present invention; Figure 3 This is a schematic diagram of the XGBoost model inference principle of this invention; Figure 4 This is a comparison image before and after the spurious modality screening of this invention. Detailed Implementation
[0018] Example 1 like Figure 1-4 As shown, an intelligent screening method for structural modal parameters based on feature fusion includes the following steps: S1. The structural response signal is segmented using a preset time window to obtain several data blocks arranged in chronological order.
[0019] S2. Modal parameter identification is performed on the data blocks using the random subspace algorithm (SSI-Cov) to obtain an initial set of modal parameters containing frequency, damping ratio and mode shape vector.
[0020] S3. Construct a multi-dimensional feature fusion vector based on the initial modal parameter set and structural response signal. Input the multi-dimensional feature fusion vector into the pre-trained classification model, output the modal true and false probabilities, and set a probability threshold to filter true modalities and eliminate false modalities.
[0021] S4. Based on the stable modal parameters obtained from the screening, repeat the process from S2 to S3 on the subsequent data blocks to output the final modal parameters.
[0022] like Figure 1As shown, in this embodiment, the structural response signal is segmented using a preset time window. After identifying modal parameters in the data segments using SSI-Cov, an initial set of modal parameters is obtained. Then, a multi-dimensional feature fusion vector is constructed by combining the structural response signal. The multi-dimensional feature fusion vector is input into a pre-trained classification model to output the true and false probabilities of the modalities, and a probability threshold is set for filtering. Based on the stable modal parameters obtained by filtering, the corresponding process is repeated for subsequent data segments to output the final modal parameters. By deeply combining multi-dimensional feature fusion with intelligent algorithms, the technical problems of incomplete removal of false modalities, strong reliance on manual parameter tuning, and insufficient real-time performance are solved. This achieves fully automated and highly efficient screening for health monitoring of major infrastructure, improves the integrity and reliability of the initial modal parameters, and enhances the accuracy of the modal parameters.
[0023] This embodiment is applied to the structural health monitoring system of an arch dam in Southwest China. The vibration response signal sampling frequency of the arch dam monitoring system is 100Hz, and the monitoring period covers different seasons throughout the year (including flood season and dry season) and extreme temperature weather. It is necessary to accurately identify modal parameters (frequency, damping ratio, vibration mode).
[0024] Step S1, signal preprocessing and segmentation, includes: Raw signal acquisition: Structural response signals are acquired by accelerometers deployed on the surface of the arch dam. The signals include vibration acceleration data and synchronously acquired environmental data (water level, temperature).
[0025] In the preferred embodiment, before segmenting the structural response signal in step S1, the original structural response signal is preprocessed. The preprocessing methods include drift elimination, filtering, and resampling, specifically: A linear trend removal method is used to perform a preprocessing procedure on the original structural response signal, including drift elimination, 50Hz low-pass filtering to remove power frequency interference, and resampling while maintaining a 100Hz sampling frequency, to obtain a high-quality structural response signal.
[0026] Data segmentation: The preprocessed structural response signal is segmented using a set time window. The time window length is set to 20 minutes, and the overlap rate of adjacent segments is 15%. Finally, 1752 data segments are obtained in chronological order (covering the entire year monitoring period).
[0027] In this embodiment, preprocessing such as drift elimination, filtering, and resampling is performed before signal segmentation, which effectively removes trend interference, power frequency noise, and data redundancy in the original signal, reduces noise interference, and improves data accuracy.
[0028] Step S2: Initial modal parameter identification.
[0029] In the preferred scheme, modal parameter identification is performed on data blocks using SSI-Cov, including: S21. A covariance-driven random subspace algorithm is used to identify each data block and obtain the modal parameters of each data block, including frequency, damping ratio and mode shape vector.
[0030] S22. Before performing block processing on the structural response signal, the structural response signal is preprocessed. The preprocessing includes drift elimination, filtering, and resampling. The preprocessed structural response signal is obtained through preprocessing and is used as the input for block processing.
[0031] In this embodiment, a single data block contains 120,000 data points (20 min × 60 s × 100 Hz). The covariance matrix is constructed from the data blocks to determine a reasonable range of system order (10-50).
[0032] The system matrix is extracted by subspace decomposition, and the initial modal parameter set of each data block is calculated: the set contains the frequency, damping ratio and mode shape vector of each mode; the dimension of the mode shape vector is consistent with the number of sensors, which is 8-dimensional in this embodiment.
[0033] This embodiment uses the SSI-Cov algorithm to extract initial modal parameters, which enhances the anti-interference ability of vibration signals and improves the integrity and reliability of initial modal parameters.
[0034] Step S3, multi-dimensional feature fusion vector construction, specific extraction process and parameter features, such as Figure 2 As shown in Table 1.
[0035] Table 1 Multidimensional Feature Fusion Vector Data Table
[0036] like Figure 2 As shown, the transformation process from the output of the SSI-Cov algorithm to the generation of feature vectors is presented: the input modal parameters (frequency, damping ratio, mode shape vector) and the original vibration signal are input on the left, and after passing through four feature analysis modules (frequency stability analysis to calculate the variance of three time scales, damping ratio consistency analysis to extract adjacent difference rate and cumulative offset, mode shape correlation analysis to generate five sets of MAC values, and environmental features), they are finally gathered into a multi-dimensional feature vector.
[0037] In the preferred embodiment, a multi-dimensional feature fusion vector is constructed based on the initial modal parameter set and the structural response signal, including: S31. Extract frequency stability features, damping ratio consistency features, and mode shape correlation features from the initial modal parameter set, and extract environmental features from the structural response signal.
[0038] S32. The frequency stability characteristics, damping ratio consistency characteristics, mode shape correlation characteristics and environmental characteristics are fused into a multi-dimensional feature fusion vector, wherein the multi-dimensional feature fusion vector is a multi-dimensional vector.
[0039] S33. Input the multi-dimensional feature fusion vector into the pre-trained classification model, and obtain the modality true / false probability by processing the multi-dimensional feature fusion vector through the pre-trained classification model.
[0040] In the preferred scheme, frequency stability features, damping ratio consistency features, and mode shape correlation features are extracted from the initial modal parameter set, including: S311. Calculate the short-term frequency variance, medium-term frequency variance, and long-term frequency variance for the frequency to obtain the frequency stability characteristics.
[0041] S312. Calculate the adjacent damping difference rate and cumulative damping offset for the damping ratio to obtain the damping ratio consistency characteristics.
[0042] S313. Calculate the correlation between the current mode and the previous mode, the correlation between the current mode and the filtered true mode, the correlation between the current mode and the adjacent lower-order mode, the correlation between the current mode and the adjacent higher-order mode, and the correlation between the current mode and the dominant mode for the mode vector, and obtain the mode correlation characteristics. S314. Obtain real-time water level and real-time temperature from the structural response signal to obtain environmental characteristics, and use frequency stability characteristics, damping ratio consistency characteristics, mode shape correlation characteristics and environmental characteristics as components of the multi-dimensional feature fusion vector.
[0043] The multi-dimensional feature fusion vector is then input into the pre-trained classification model for training and modality selection.
[0044] In the preferred scheme, the XGBoost model is used as the pre-trained classification model for training and modality selection, including: S34. Using the multi-dimensional feature fusion vector as input to the XGBoost model, and outputting the modal truth probabilities as continuous values through the XGBoost model, specifically including: S341. Train the XGBoost model based on labeled samples from historical monitoring datasets. The labeled samples contain real modality labels and fake modality labels.
[0045] S342. Cross-validation is used to optimize the hyperparameters of the XGBoost model, including the learning rate, maximum tree depth, and subsample ratio, and early stopping is used to control the number of iterations.
[0046] S343. Solidify the trained XGBoost model and embed it into an edge computing device. Perform inference of multi-dimensional feature fusion vectors through the edge computing device to obtain the modality true / false probability.
[0047] S344. Supports the incremental learning mechanism of the XGBoost model, which updates the weights of the XGBoost model using newly labeled data when the structural response signal changes.
[0048] S35. Compare the modality true / false probability with the probability threshold. If the modality true / false probability is greater than or equal to the probability threshold, it is determined to be a true modality; otherwise, it is determined to be a false modality and is removed.
[0049] S36. The initial modal parameter set is processed through the decision mechanism of the XGBoost model to obtain the filtered modal parameters, and the filtered modal parameters are used to determine the stable modal parameters.
[0050] like Figure 3 The diagram illustrates the core inference process of the XGBoost model: First, a 12-dimensional feature fusion vector is input; then, adaptive splitting (Tree Splits) is performed in parallel across multiple decision trees to generate nonlinear feature combinations based on the gain maximization principle; the transformed higher-order features are input into a linear classifier, which outputs continuous probability values of 0-1 via the sigmoid function; finally, a probability threshold ≥ 0.9 is used as the criterion to automatically filter modal authenticity. In this embodiment, the specific implementation includes: 1) Model training data preparation: Select 1,000 data blocks from the monitoring data of the arch dam in the previous year, manually label the real modes (label 1) and false modes (label 0) in them, and form a training dataset containing 8,000 modal samples.
[0051] 2) Hyperparameter optimization: Five-fold cross-validation is used to optimize the model hyperparameters, with a learning rate of 0.1, a maximum tree depth of 5, and a subsample ratio of 0.8. The number of iterations is controlled by early stopping. Training is stopped when the validation set loss does not decrease for 10 consecutive rounds to prevent the model from overfitting.
[0052] 3) Model deployment: After the trained XGBoost model is solidified, it is embedded into the edge computing device (deployed at the arch dam monitoring station). The time for a single inference is ≤8ms, which meets the real-time monitoring requirements.
[0053] 4) Modal authenticity determination: The 12-dimensional feature fusion vector is input into the pre-trained XGBoost model, and the output modal authenticity probability value (0-1 interval) is used. The probability threshold is set to 0.9. When the output probability is ≥0.9, it is determined to be a real modal. Otherwise, the false modal is directly eliminated.
[0054] In this embodiment, when the arch dam experiences significant structural changes, such as a sudden rise in water level exceeding 10m during the flood season or a temperature change exceeding 15℃, new monitoring data is collected and 50 modal samples are manually labeled. The XGBoost model weights are dynamically updated using an incremental learning mechanism, which improves the model's adaptability to structural changes.
[0055] In this embodiment, a multi-dimensional feature joint discrimination mechanism is adopted. Compared with the traditional single frequency feature scheme, it achieves a comprehensive characterization of modal characteristics and improves the accuracy of false mode identification: the frequency stability feature covers short, medium and long periods, improving the discrimination stability of modes at different time scales; the damping ratio consistency feature solves the technical problem of ignoring damping characteristics in traditional methods through adjacent difference rate and cumulative offset; the mode shape correlation feature verifies the physical authenticity of the modes from multi-dimensional correlation, improving the authenticity of the data; the environmental features incorporate real-time water level and temperature, improving the adaptability of modal identification to complex environments.
[0056] In the preferred scheme, based on the stable modal parameters obtained through screening, the subsequent data is divided into blocks, and processes S2 to S3 are repeated, including: S41. Using the stable modal parameters as a benchmark, apply the random subspace algorithm to the subsequent data blocks to identify modal parameters and obtain a new set of initial modal parameters. S42. Construct a new multi-dimensional feature fusion vector based on the new initial modality parameter set, and input it into the pre-trained classification model to obtain the new modality true / false probability; S43. Based on the new modality true / false probabilities and probability thresholds, filter out new false modalities and update the stable modality parameters; S44. Repeat the process until all data blocks are processed, and output the final modal parameters as the overall result.
[0057] In this embodiment, step S4 involves iterative optimization and final parameter output, specifically as follows: 1) Determination of reference mode: The stable mode parameters obtained after the above steps for the first data block are used as the reference mode for subsequent data block processing.
[0058] 2) Iterative processing: Repeat steps (ii) to (iv) for the remaining 1751 data blocks. After processing each block, update the stable mode parameter benchmark with the currently selected real mode.
[0059] 3) Final output: After all data is processed in blocks, the final parameters of the first three dominant modes of the arch dam are output, including the average frequency, damping ratio and mode shape vector of each mode, where the frequency fluctuation range is ≤ ±0.05Hz and the damping ratio error is ≤5%.
[0060] like Figure 4As shown in Figure (a), the original natural frequency identification results output by the traditional SSI-Cov algorithm are as follows: a large number of false modal points (0.5–12Hz band) seriously interfere with the real physical modes, resulting in drastic frequency fluctuations (such as the abnormal fluctuation of the first mode in the 1.8–2.4Hz band). Figure (b) shows the effect of the same data after intelligent filtering in this embodiment: by using the XGBoost model that integrates multi-dimensional features (frequency stability, damping ratio consistency, mode shape correlation and environment) to make judgments with a probability threshold ≥0.9, noisy modal points are eliminated, and the fluctuation of the real modal frequency is significantly compressed; the first three real modes of the structure form a clear and stable blue cluster distribution, which significantly improves the reliability of health monitoring data of major infrastructure.
[0061] This embodiment improves the false mode rejection rate through the above steps, solves the problem of incomplete false mode rejection, reduces operation and maintenance costs, improves error identification accuracy, and enhances the reliability of early warning of damage to arch dam structures.
[0062] Example 2 To further illustrate with reference to Embodiment 1, a structural modal parameter intelligent screening system based on feature fusion is applicable to a structural modal parameter intelligent screening method based on feature fusion in the embodiment, comprising: The data segmentation module is used to segment the structural response signal into blocks using a set time window, resulting in several data blocks arranged in chronological order. The parameter identification module is used to identify modal parameters by dividing the data into blocks using a random subspace algorithm, and obtain an initial set of modal parameters including frequency, damping ratio and mode shape vector. The feature fusion module is used to construct a multi-dimensional feature fusion vector based on the initial modal parameter set and structural response signal. The multi-dimensional feature fusion vector is input into the pre-trained classification model, outputs the true and false probabilities of the modal, and sets a probability threshold to filter the true modal and directly eliminate false modalities. The modal parameter module is used to repeatedly execute the S2 to S3 process on subsequent data blocks based on the stable modal parameters obtained by filtering, and output the final modal parameters.
[0063] This embodiment provides the working process, working details and technical effects of the intelligent screening method for structural modal parameters based on feature fusion, which can be found in Embodiment 1 and will not be repeated here.
[0064] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.
Claims
1. A method for intelligent screening of structural modal parameters based on feature fusion, characterized in that, Includes the following steps: S1. The structural response signal is divided into blocks using a preset time window to obtain several data blocks arranged in chronological order. S2. Modal parameter identification is performed on the data blocks using a random subspace algorithm to obtain an initial set of modal parameters including frequency, damping ratio, and mode shape vector; S3. Construct a multi-dimensional feature fusion vector based on the initial modal parameter set and the structural response signal, input the multi-dimensional feature fusion vector into the pre-trained classification model, output the modal authenticity probability, and set a probability threshold to filter out real modalities and eliminate false modalities. S4. Based on the stable modal parameters obtained from the screening, repeat the process from S2 to S3 on the subsequent data blocks to output the final modal parameters.
2. The intelligent screening method for structural modal parameters based on feature fusion according to claim 1, characterized in that, The modal parameter identification of the data blocks using the random subspace algorithm includes: S21. A covariance-driven random subspace algorithm is used to identify each data block to obtain the modal parameters of each data block, wherein the modal parameters include the frequency, the damping ratio and the mode shape vector. S22. Before performing block processing on the structural response signal, the structural response signal is preprocessed. The preprocessing includes drift elimination, filtering, and resampling. The preprocessed structural response signal is obtained through the preprocessing and is used as the input for the block processing.
3. The intelligent screening method for structural modal parameters based on feature fusion according to claim 2, characterized in that, In step S1, before dividing the structural response signal into blocks, the original structural response signal is preprocessed. The preprocessing methods include drift elimination, filtering, and resampling.
4. The intelligent screening method for structural modal parameters based on feature fusion according to claim 1, characterized in that, The construction of a multi-dimensional feature fusion vector based on the initial modal parameter set and the structural response signal includes: S31. Extract frequency stability features, damping ratio consistency features, and mode shape correlation features from the initial modal parameter set, and extract environmental features from the structural response signal; S32. The frequency stability feature, the damping ratio consistency feature, the mode shape correlation feature and the environmental feature are fused into the multi-dimensional feature fusion vector, wherein the multi-dimensional feature fusion vector is a multi-dimensional vector; S33. Input the multi-dimensional feature fusion vector into the pre-trained classification model, and process the multi-dimensional feature fusion vector through the pre-trained classification model to obtain the modality authenticity probability.
5. The intelligent screening method for structural modal parameters based on feature fusion according to claim 4, characterized in that, The extraction of frequency stability features, damping ratio consistency features, and mode shape correlation features from the initial modal parameter set includes: S311. Calculate the short-term frequency variance, medium-term frequency variance, and long-term frequency variance for the frequency to obtain the frequency stability characteristics; S312. Calculate the adjacent damping difference rate and cumulative damping offset for the damping ratio to obtain the consistency characteristics of the damping ratio; S313. Calculate the correlation between the current mode and the previous mode, the correlation between the current mode and the filtered true mode, the correlation between the current mode and the adjacent lower-order mode, the correlation between the current mode and the adjacent higher-order mode, and the correlation between the current mode and the dominant mode for the mode vector to obtain the mode correlation characteristics. S314. Obtain real-time water level and real-time temperature from the structural response signal to obtain the environmental characteristics, and use the frequency stability characteristics, the damping ratio consistency characteristics, the mode shape correlation characteristics, and the environmental characteristics as components of the multi-dimensional feature fusion vector.
6. The intelligent screening method for structural modal parameters based on feature fusion according to claim 1, characterized in that, The constructed multi-dimensional feature fusion vector is a 12-dimensional feature vector, specifically including: Three-dimensional frequency stability characteristics: short-term frequency variance, medium-term frequency variance, long-term frequency variance, and typical vibration period of the covered structure; Damping ratio consistency characteristics in 2D: adjacent damping difference rate, cumulative damping offset; The modal correlation characteristics are 5-dimensional: the correlation between the current time and the previous time, the correlation between the current time and the filtered true modes, the correlation between the current time and the adjacent lower-order modes, the correlation between the current time and the adjacent higher-order modes, and the correlation between the current time and the dominant mode. Two-dimensional environmental characteristics: real-time water level and real-time temperature.
7. The intelligent screening method for structural modal parameters based on feature fusion according to claim 1, characterized in that, The step of inputting the multi-dimensional feature fusion vector into the pre-trained classification model includes: S34. The XGBoost model is used as the pre-trained classification model, the multi-dimensional feature fusion vector is used as the input of the XGBoost model, and the modality true / false probability is output as a continuous value through the XGBoost model. S35. Compare the modality authenticity probability with the probability threshold. If the modality authenticity probability is greater than or equal to the probability threshold, it is determined to be the true modality; otherwise, it is determined to be the false modality and is removed. S36. The initial modal parameter set is processed through the decision mechanism of the XGBoost model to obtain the filtered modal parameters, and the filtered modal parameters are used to determine the stable modal parameters.
8. The intelligent screening method for structural modal parameters based on feature fusion according to claim 7, characterized in that, The use of the XGBoost model as the pre-trained classification model includes: S341. Train the XGBoost model based on labeled samples from a historical monitoring dataset, wherein the labeled samples contain real modality labels and fake modality labels; S342. Optimize the hyperparameters of the XGBoost model using cross-validation, including learning rate, maximum tree depth, and subsample ratio, and control the number of iterations using early stopping. S343. Solidify the trained XGBoost model and embed it into an edge computing device, and perform inference of the multi-dimensional feature fusion vector through the edge computing device to obtain the modality true / false probability; S344. Support the incremental learning mechanism of the XGBoost model, and update the weights of the XGBoost model using new labeled data when the structural response signal changes.
9. The intelligent screening method for structural modal parameters based on feature fusion according to claim 1, characterized in that, The process of repeatedly executing steps S2 to S3 on subsequent data blocks based on the stable modal parameters obtained through screening includes: S41. Using the stable modal parameters as a benchmark, the random subspace algorithm is applied to the subsequent data blocks to identify the modal parameters, thereby obtaining a new set of initial modal parameters. S42. Construct a new multi-dimensional feature fusion vector based on the new initial modality parameter set, and input it into the pre-trained classification model to obtain the new modality true / false probability; S43. Based on the new modality true / false probability and the probability threshold, filter out new false modalities, and update the stable modality parameters; S44. Repeat the process until all data blocks are processed, and output the final modal parameters as the overall result.
10. A structural modal parameter intelligent screening system based on feature fusion, characterized in that, The method for intelligent screening of structural modal parameters based on feature fusion, applicable to any one of claims 1-9, comprises: The data segmentation module is used to segment the structural response signal into blocks using a set time window, resulting in several data blocks arranged in chronological order. The parameter identification module is used to identify modal parameters in the data blocks using a random subspace algorithm, and obtain an initial set of modal parameters including frequency, damping ratio and mode shape vector. The feature fusion module is used to construct a multi-dimensional feature fusion vector based on the initial modal parameter set and the structural response signal, input the multi-dimensional feature fusion vector into the pre-trained classification model, output the modal authenticity probability, and set a probability threshold to filter the real modal and directly eliminate the false modal. The modal parameter module is used to repeatedly execute the process from the parameter identification module to the feature fusion module on subsequent data blocks based on the stable modal parameters obtained through screening, and output the final modal parameters.