Machine learning adaptive model order determination method based on physical information

Abstract

The invention relates to a machine learning adaptive model order determination method based on physical information, which is suitable for real-time modal analysis of an ocean engineering structure. Therefore, the invention provides a self-adaptive order determination method fusing a complex exponential reconstruction algorithm and a light-weight gradient elevator model (LightGBM). The method comprises the following steps: firstly, performing prior order estimation on a structural vibration response signal by using complex exponential reconstruction, and extracting a corresponding optimal order; performing fast Fourier transform on the response signal, extracting a frequency domain amplitude feature, and pairing the frequency domain amplitude feature with the estimation order to construct a training sample; a nonlinear mapping relation between frequency domain features and orders is learned through a LightGBM model, and a fixed-order model which can be used for real-time prediction is obtained through training. On the basis of keeping the anti-noise advantage of a traditional method, a data driving mechanism of machine learning is introduced, empirical threshold setting and a tedious feature discrimination process are avoided, and efficient, accurate and online adaptive prediction of the structural model order is realized.

Description

TECHNICAL FIELD

[0001] The application relates to a physical information-based machine learning adaptive model order determination method, aiming to solve the problems of noise sensitivity failure and artificial experience dependence of traditional model order determination algorithms, so as to meet the needs of real-time modal analysis of ocean engineering structures. The application relates to the field of dynamic characteristic analysis of ocean engineering structures. BACKGROUND

[0002] Model order determination is a key link in dynamic characteristic analysis of engineering structures. In the structure modal parameter identification method based on the state space model, such as the autoregressive moving average time series method (ARMA), the stochastic subspace identification method (SSI) and the characteristic system realization algorithm (ERA), the model order is a necessary parameter for modal parameter calculation, and directly determines the accuracy and stability of the identification result. The limitations of traditional model order determination methods mainly come from the following three aspects: first, the true order of the structure response under environmental load excitation is unknown; second, the random noise in the measured data is mixed with the structure inherent information, introducing false modes of noise, which affects the stability of the order determination criterion; third, the high-energy harmonic noise or non-stationary random component in the measured data affects the singular value results of the system matrix, which easily causes order determination failure. Therefore, under the conditions of unknown true order and random noise that cannot be perfectly removed from the response, the essence of the model order determination problem can be understood as how to determine the appropriate system order range under the condition of limited noise interference, so as to ensure the accuracy and stability of the system parameter identification.

[0003] Traditional model order determination algorithms are mainly divided into two categories: time domain and frequency domain. Frequency domain algorithms use frequency response function (FRF) or equivalent impulse response function (IRF) as algorithm input, and determine the model order through peak value detection, curve fitting and other methods; time domain algorithms use time domain response data as algorithm input, and construct a system state space model through state space method, subspace method and system realization, and then determine the model order through singular value decomposition, stability diagram, information criterion and correlation analysis. However, the response data of ocean engineering structures under natural environmental excitation conditions have the characteristics of high complexity and strong time variation, and the traditional model order determination methods based on empirical threshold or stability criterion are usually sensitive to noise, strong in experience and low in calculation efficiency, which are difficult to meet the real-time order estimation requirements of ocean engineering structure responses under complex marine environment. Therefore, how to realize the fast estimation of model order under noise interference is a key technical problem faced by the field of adaptive monitoring and analysis of ocean engineering structures. SUMMARY

[0004] The technical problems to be solved by the present application are that the traditional model order determination method is sensitive to noise, lacks self-adaptive ability, and has low calculation efficiency, and the single threshold condition and complex calculation process cannot meet the demand of real-time estimation of the order of the structure dynamic response data model under the condition of variable load.

[0005] The technical scheme adopted by the present application is:

[0006] A machine learning adaptive model order determination method based on physical information, which combines a complex exponential reconstruction order determination algorithm based on a physical model constraint and a nonlinear prediction model based on a data-driven LightGBM model, realizes physical-data joint modeling and adaptive prediction of the order of a structure dynamic system for the dynamic characteristics of a structure vibration response signal under unknown order and noise interference conditions, and includes the following steps:

[0007] Step S1: using a complex exponential reconstruction method to estimate the order of a structure vibration response signal to obtain the corresponding optimal order result for subsequent training sample label generation;

[0008] Step S2: performing fast Fourier transform on the structure vibration response signal to convert it from time domain form to frequency domain form; extracting frequency domain amplitude information to form a frequency domain physical feature vector, and pairing the feature vector with the optimal order result obtained in step S1 to construct a training sample set for LightGBM model training;

[0009] Step S3: based on the training samples constructed in step S2, using the frequency domain feature vector as input and the optimal order obtained in step S1 as label, using the LightGBM model for supervised learning training to establish a nonlinear mapping relationship between the frequency domain features and the order, and training a LightGBM model for order prediction;

[0010] Step S4: extracting features from the new structure vibration response signal after frequency domain conversion and inputting them into the trained LightGBM model to output continuous order prediction results or order interval classification results, realizing online adaptive order determination of the structure model.

[0011] The complex exponential reconstruction order determination algorithm and the LightGBM nonlinear prediction model principles are as follows:

[0012] 1. Complex exponential reconstruction order determination algorithm:

[0013] Due to the influence of the environment and the measuring equipment, the actual measurement signal usually contains noise of different degrees and components. Therefore, the discrete-time signal of the system response can be generally represented as:

[0014]

[0015] where x k , represent the system response component and the noise component, respectively. According to the above formula, for any finite, discrete-time system dynamic response signal s k , it can be expressed in the form of a complex exponential sum, that is:

[0016]

[0017] According to the above formula, the complex exponential order estimation method is to approximate the number of components of the signal by using the expression form of the complex exponential sum of the signal, wherein the number of terms of the complex exponential sum is the model order. In the complex exponential reconstruction order estimation process, the error between the reconstructed signal and the original signal will decrease with the increase of the order, and when the order is too large, the error between the reconstructed signal and the original signal will no longer decrease, therefore, the order corresponding to the minimum reconstruction error is the model order estimated by the algorithm.

[0018] According to the above relationship between the order and the signal reconstruction error, the present application estimates the order of the structural response signal in the structural vibration response data based on the complex exponential reconstruction method, and takes the order when the error between the reconstructed response signal and the original measured signal is less than the set threshold value as the estimated order of the complex exponential reconstruction order estimation, and the order estimation expression is as follows:

[0019]

[0020] wherein O is the model order estimated by the complex exponential algorithm, is the complex exponential reconstruction signal, J is the complex exponential decomposition order, E' is the root mean square error (RMSE) between the reconstructed signal and the original signal, and ε is the reconstruction error threshold. Since the complex exponential reconstruction process takes the noise component in the signal into account and measures the interference degree of the noise on the model order through the reconstruction error, the problem of noise order estimation failure existing in the traditional singular value order estimation method can be effectively alleviated.

[0021] 2. LightGBM nonlinear prediction model

[0022] LightGBM is a high-efficiency gradient boosting decision tree model, which adopts histogram algorithm, leaf-wise algorithm, gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB) to optimize the memory occupation and consumption of traditional decision tree algorithm, and supports efficient parallel training and distributed computing, so as to have strong data processing capacity and online computing capacity. The efficient computing advantage and nonlinear data feature mapping capacity of LightGBM algorithm provide a fast and reliable analysis and calculation model for model order reduction algorithm based on physical information machine learning.

[0023] Further, step S1 comprises the following steps:

[0024] Step S101: The structural vibration response signal is constructed into a Hankel matrix form for representing its time sequence characteristics;

[0025] Step S102: The Hankel matrix is singular value decomposed to obtain a system matrix A, and the system matrix is decomposed to obtain complex exponential parameters reflecting complex modal characteristics, and the decomposition order is traversed in ascending order;

[0026] Step S103: The signal is reconstructed based on the extracted complex exponential parameters, and the corresponding reconstruction error is calculated;

[0027] Step S104: Under the condition that the error meets the predefined tolerance limit, the order corresponding to the minimum reconstruction error is selected as the optimal order, and the traversal calculation is stopped.

[0028] Further, step S2 comprises the following steps:

[0029] Step S201: The structural vibration response signal is subjected to fast Fourier transform to obtain its amplitude information in the frequency domain to form a frequency domain physical feature vector;

[0030] Step S202: The frequency domain feature vector is corresponded to the optimal order result obtained in step S1 to generate a training sample set required for supervised learning.

[0031] Further, step S3 comprises the following steps:

[0032] Step S301: The frequency domain feature vector in the training sample set is taken as input, and the optimal order obtained in step S1 is taken as a supervised label;

[0033] Step S302: performing supervised learning based on a LightGBM model to construct a nonlinear mapping relationship between input features and order labels; wherein the LightGBM model can use an L2 loss function or other loss function suitable for regression tasks; during model training, a gradient histogram-based optimization strategy is used for feature gain sorting, and the model complexity is controlled by setting hyperparameters such as maximum tree depth and leaf node number.

[0034] Step S303: updating model parameters in multiple rounds of iterative training by minimizing the loss function between the prediction result and the supervised label, and training a LightGBM model for order prediction.

[0035] Further, the output form of the LightGBM model in step S4 is a classification result, which is used to determine whether the structural model order belongs to a preset discrete order interval.

[0036] Further, the LightGBM model has both regression output path and classification output path, which are respectively used to output continuous order value and order interval classification result, supporting multi-task structural modal identification.

[0037] Further, the classification output path includes a softmax layer for normalizing the original classification scores output by the model to generate the probability distribution of the interval to which the structural model order belongs.

[0038] Further, step S4 includes the following steps:

[0039] Step S401: performing fast Fourier transform on the new structural vibration response signal to extract its frequency domain feature vector;

[0040] Step S402: inputting the frequency domain feature vector into the trained LightGBM model;

[0041] Step S403: outputting the predicted order result of the structural model, which is used for subsequent modal parameter identification or state space modeling.

[0042] Further, the method is suitable for engineering structural modal identification tasks under environmental excitation conditions, including but not limited to offshore platforms, wind towers, bridges, offshore wind power foundations and other complex structural systems.

[0043] The beneficial effects of the present application are:

[0044] The present application combines the complex exponential reconstruction order algorithm based on physical model constraints with the LightGBM model based on data-driven, fully utilizing the robustness of the complex exponential method in noisy environments and the efficient training and fast prediction ability of the LightGBM in small sample scenarios.

[0045] Compared with traditional order determination methods that rely on empirical thresholds, stable graph analysis, or singular value criteria, this invention eliminates the need for manual threshold setting and large-scale feature judgment, significantly improving the automation and computational efficiency of order prediction.

[0046] The nonlinear prediction model constructed in this invention can quickly output the optimal order prediction result after the structural response data is acquired. It provides an effective method with real-time performance, noise resistance and high generalization ability for structural modal analysis in complex environments. It is particularly suitable for online modal identification tasks of engineering systems such as marine platforms, wind turbine towers and bridge structures. Attached Figure Description

[0047] Figure 1 A framework for model order determination based on LightGBM physical information;

[0048] Figure 2 Flowchart of the model order determination method based on physical information LightGBM;

[0049] Figure 3 Time-frequency diagram of harmonic signal (Num=8);

[0050] Figure 4 Time-frequency comparison of noisy and clean signals (SNR = -10dB);

[0051] Figure 5 LightGBM classification model training and testing results (SNR = 10dB);

[0052] Figure 6 Training and testing results of the LightGBM classification model. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, this invention provides a machine learning adaptive model order determination method based on physical information, whose overall framework integrates physical modeling and machine learning methods. This method uses a complex exponential reconstruction algorithm as the means of constraining physical information and a Lightweight Gradient Boosting Machine (LightGBM) as the data-driven prediction model to achieve nonlinear mapping learning and adaptive prediction from structural response signals to model order.

[0055] Specifically, first, the extreme value residue decomposition (ESPRIT) and complex exponential reconstruction method are used to process the structural vibration response signal, and a mapping data set between the original signal and the optimal order is constructed; in order to improve the recognizability and operation efficiency of the model input features, the fast Fourier transform (FFT) is used to convert the structural response signal from the time domain to the frequency domain, and the frequency domain amplitude information is extracted as the physical feature vector input into the LightGBM model.

[0056] The calculation process of each step of the above technical solution is as follows:

[0057] Step S101: The structural vibration response signal is constructed into a Hankel matrix form for representing its time sequence characteristics. Specifically:

[0058] The structural response signal is written in the form of a Hankel matrix, that is:

[0059]

[0060] In formula 4, ξ and η are the number of rows and columns of the matrix H(k) respectively, and in general cases, the H(k) square matrix is set.

[0061] It has been proved that H(k) can be expressed in the form of matrix multiplication:

[0062]

[0063] In formula 5, is an observation matrix, is a control matrix, A∈R J×J is a system matrix, where r is the rank of the matrix H(k), J<ξ and J<η. Therefore, when k=0, then:

[0064]

[0065] Step S102: Singular value decomposition is performed on the Hankel matrix to obtain the system matrix A, and then the system matrix is decomposed to obtain the complex exponential parameters reflecting the complex modal characteristics, and the decomposition order is traversed in ascending order. Specifically:

[0066] The singular value decomposition is performed on the matrix H(0):

[0067]

[0068] In formula 7, U∈R ξ×J and V∈R η×L are orthogonal matrices, S∈R J×J is a singular value diagonal matrix, represents matrix transposition. Then, a possible balance mode is taken for formula 5 and formula 6, and the observation matrix and the control matrix are represented as: When k = 1, the balanced expression of the observation matrix and the control matrix is substituted into formula (3.16) to obtain:

[0069]

[0070] The left side of the above formula is multiplied by and the right side is multiplied by to obtain the system matrix A:

[0071]

[0072] The eigenvalue z n of the matrix A is n = 1, 2, … J, and the extreme value λ n of the system is ln(z n ) / Δt. Substitute the extreme value into formula 2, and then obtain the residue r n of the system by linear least squares method.

[0073] Step S103: reconstruct the signal based on the extracted complex exponential parameters and substitute it into formula 2, and calculate the corresponding reconstruction error.

[0074] Step S104: under the condition that the error satisfies the pre-defined tolerance limit as shown in formula 3, select the order corresponding to the minimum reconstruction error as the optimal order, and stop the traversal calculation.

[0075] Step S201: perform fast Fourier transform on the structural vibration response signal to obtain its amplitude information in the frequency domain to form a frequency domain physical feature vector;

[0076] Step S202: correspond the frequency domain feature vector to the optimal order result obtained in step S1 to generate a training sample set required for supervised learning.

[0077] Step S301: take the frequency domain feature vector in the training sample set as input, and the optimal order obtained in step S1 as a supervised label;

[0078] Step S302: perform supervised learning based on the LightGBM model to construct a nonlinear mapping relationship between the input features and the order label;

[0079] Step S303: update the model parameters in multiple rounds of iterative training by minimizing the loss function between the prediction result and the supervised label, and train to obtain a LightGBM model for order prediction.

[0080] Step S401: perform fast Fourier transform on the new structural vibration response signal to extract its frequency domain feature vector;

[0081] Step S402: input the frequency domain feature vector into the trained LightGBM model;

[0082] Step S403: output the predicted order result of the structural model.

[0083] As Figure 2 shown, the model order determination process includes the following three parts:

[0084] 1. Prior data construction: perform complex exponential modeling and error evaluation on the structural vibration response data to obtain the corresponding optimal order, and construct the data set of frequency domain features and order labels;

[0085] 2. Model training stage: divide the data set into a training set and a test set, input the training set into the LightGBM model for supervised learning, the goal is to minimize the error between the model output and the supervised label, update the parameters through multiple iterations, and establish a nonlinear mapping relationship between the frequency domain features and the order;

[0086] 3. Model verification and application: use the test set data to evaluate the accuracy, stability and generalization ability of the trained model, and the model can be used for real-time online order prediction tasks of structural response data after passing the verification.

[0087] In addition, Figure 1 The model structure shown in the above formula can be selected according to application requirements, and the regression output path, the classification output path or the multi-task output structure combining the two can be selected, wherein the classification path can be normalized by a softmax layer to realize the order interval judgment function.

[0088] To verify the effectiveness and robustness of the physical information model order determination method proposed in the present application compared with a plurality of typical order determination algorithms, including frequency spectrum peak value counting, singular value threshold, Akaike information criterion (AIC), simple object access protocol (SAMOS) and confidence criterion, a harmonic signal model is used for test analysis. In the test, 4 groups of harmonic signals with different component numbers (Num=2, 4, 6, 8) are set for comparison test, and the components are increased by 2 each time according to the order number in Table 1. When setting the frequency of the harmonic signal, a plurality of different frequency components and adjacent frequency components are set, such as 0.05 Hz and 0.1 Hz, 1 Hz and 1.01 Hz, so as to improve the diversity of the harmonic signal components and ensure the reliability of the robustness verification result of the algorithm. The length n of the discrete time undamped harmonic signal is defined as follows:

[0089]

[0090] Wherein the sampling time interval Δt=0.1s, the parameter values of the signal x are shown in Table 1, and the time and frequency diagram of the harmonic signal is shown in Figure 3 .

[0091] Table 1 Harmonic signal parameters

[0092]

[0093] Nine order estimation methods including spectral counting method were used to estimate the order of harmonic signals, and the results are shown in Table 2. From the results in the table, it can be seen that the singular value order estimation, confidence criterion order estimation and reconstruction order estimation algorithms accurately identify the order of the four groups of harmonic signals. Among them, the three reconstruction algorithms use the time domain correlation coefficient of the reconstructed signal and the original signal as the index, and the correlation coefficient threshold of the singular value and complex exponential reconstruction is 0.99, and the correlation coefficient threshold of the Fourier reconstruction is 0.87. The spectral counting method is affected by the spectral leakage of Fourier transform and cannot effectively identify the adjacent frequency components, so the order estimation result is smaller. The AIC criterion method also has poor identification effect on frequency components with close energy and frequency. The order estimation effect of the empirical mode decomposition (EMD) method is the worst. Since EMD uses data extrema as the scale for signal decomposition, the intrinsic mode function (IMF) obtained is not in the form of a single frequency, and it is more difficult to distinguish data with close frequency and energy.

[0094] Table 2 Comparison of order estimation results of different harmonic number data

[0095]

[0096] * The accurate identification results are in bold.

[0097] To further verify the robustness of the above algorithms in order estimation under real measurement signal conditions, four groups (SNR = 10 dB, 5 dB, -5 dB, -10 dB) of random noise with different signal-to-noise ratios were added to the harmonic signal with frequency number Num = 8. The time and frequency comparison of the noisy signal and the original harmonic signal is shown in Figure 4 The order estimation results of each order estimation method are shown in Table 3.

[0098] Table 3 Comparison of order estimation results of different signal-to-noise ratio data

[0099]

[0100] * The accurate identification results are in bold.

[0101] From the results of the above table, for the order estimation of harmonic signals under Gaussian noise interference, the spectral counting method has a large deviation from the true order due to the wideband interference of Gaussian noise in the frequency domain. The singular value truncation method can accurately identify the true harmonic order in the signal for high signal-to-noise ratio signals, but for negative signal-to-noise ratio signals, the order identification result is less than the true order, which is because the energy of the noise component in the signal is greater than that of the harmonic component, and the mutation inflection point of the singular value curve is not obvious. Similarly, the Fourier reconstruction and singular value reconstruction order determination methods are affected by the same reason, resulting in inaccurate order determination results. The AIC criterion method is affected by noise, leading to an increase in the overall complexity of the model. However, due to the masking of weak energy harmonics by noise and the overfitting constraint penalty in the criterion, the model only fits the high-energy components in the signal, which can easily cause the order identification result to be less than the true signal order. The EMD method identifies the accurate order results for the three groups of data in the test, indicating that the EMD order determination method has strong anti-noise robustness. The confidence criterion method has similar results as the singular value truncation method, and the difference between the two methods is that the noise interference causes the modal difference discrimination rule to be inconsistent with the assumed rule, resulting in a smaller order determination result. The complex exponential reconstruction method identifies the accurate order results for the four kinds of signal-to-noise ratio data, showing good anti-noise robustness.

[0102] To further verify the effectiveness of the order estimation algorithm based on the LightGBM model, the data set is constructed with the signal parameters in Table 1, and the parameters are shown in Table 4. The number of training samples and test samples of the four kinds of sample data is 9:1. The LightGBM model can solve classification or regression problems according to the output target requirements. According to the characteristics of the data samples, the classification model is selected for order estimation in the test, and the classification and regression model parameter configurations are shown in Table 5. At the same time, in order to increase the complexity of the analysis data, random noise with a signal-to-noise ratio of 10 is added to the sample data.

[0103] Table 4 Data set parameters

[0104]

[0105] Table 5 LightGBM classification and regression model parameters

[0106]

[0107] Since the true order result of the simulation signal is known, the sample data is directly labeled accordingly. The training data is randomly sorted using a random seed to improve the robustness of the model. After training, the test set data is input into the trained model to verify the accuracy of the model order prediction, and the prediction results are shown in Figure 5 .

[0108] Figure 5The order estimation results of the LightGBM classification model-based order estimation algorithm for the sample set with a signal-to-noise ratio of 10 dB. The total number of samples is 220, of which the number of training samples is 200 and the number of test samples is 20. The accuracy of the model test results is 100%. As can be seen from the figure, the accuracy of the model test results is 100%. This test result is basically predictable. Since the simulation signal is a simple harmonic signal, the complexity of the data is relatively low, and the similarity between the training set and the test set is high, so the accuracy of the test result is also high. To further verify the anti-noise ability of the prediction model, two groups of noise signals with signal-to-noise ratios of SNR=5 dB and SNR=0 dB are tested, and the test results are shown in Figure 6

[0109] Figure 6 The results shown indicate that the recognition accuracy of the model trained using sample data with a signal-to-noise ratio of SNR=10 for two low signal-to-noise ratio data with signal-to-noise ratios of SNR=5 and 0 is 100% and 72%, respectively. Figure 6 (b) The results shown indicate that the data with recognition failure are mainly concentrated in the two groups of data with orders of 12 and 8. The main reasons for this recognition error include two points: first, the sample data contains two harmonic components with similar frequencies and similar energies; second, the random noise in the data reduces the difference of the frequency spectrum of the sample data of different orders. Overall, the order determination method based on the LightGBM classification model can accurately estimate the order of low signal-to-noise ratio harmonic signals.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for part or all of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.​

Claims

1. A method for determining the order of an adaptive machine learning model based on physical information, characterized in that, The method combines a complex exponential reconstruction order determination algorithm based on physical model constraints and a nonlinear prediction model based on a data-driven LightGBM model, and realizes physical-data joint modeling and adaptive prediction of the order of a structural dynamic system under unknown order and noise interference conditions of a structural vibration response signal, including the following steps: Step S1: using a complex exponential reconstruction method to estimate the order of the structural vibration response signal, obtaining the corresponding optimal order result for subsequent training sample label generation; Step S2: performing fast Fourier transform on the structural vibration response signal to convert it from time domain form to frequency domain form; extracting the frequency domain amplitude information to form a frequency domain physical feature vector, and pairing the feature vector with the optimal order result obtained in step S1 to construct a training sample set for LightGBM model training; Step S3: based on the training samples constructed in step S2, taking the frequency domain feature vector as the input and the optimal order obtained in S1 as the label, using the LightGBM model for supervised learning training, establishing a nonlinear mapping relationship between the frequency domain features and the order, and training a LightGBM model for order prediction; Step S4: extracting the features of the new structural vibration response signal after frequency domain conversion and inputting them into the trained LightGBM model to output continuous order prediction results or order interval classification results, realizing online adaptive order determination of the structural model.

2. The adaptive model order estimation method of claim 1, wherein, Step S1 includes the following steps: Step S101: constructing the structural vibration response signal into a Hankel matrix form for representing its time sequence characteristics; Step S102: performing singular value decomposition on the Hankel matrix to obtain a system matrix A, and then decomposing the system matrix to obtain complex exponential parameters reflecting complex modal characteristics, and the decomposition order is traversed in ascending order; Step S103: reconstructing the signal based on the extracted complex exponential parameters and calculating the corresponding reconstruction error; Step S104: under the condition that the error meets the pre-defined tolerance limit, selecting the order corresponding to the minimum reconstruction error as the optimal order, and stopping the traversal calculation.

3. The method of adaptive model order selection according to claim 1, wherein, Step S2 includes the following steps: Step S201: performing fast Fourier transform on the structural vibration response signal to obtain its amplitude information in the frequency domain to form a frequency domain physical feature vector; Step S202: corresponding the frequency domain feature vector with the optimal order result obtained in step S1 to generate a training sample set required for supervised learning.

4. The method of adaptive model order selection according to claim 1, wherein, Step S3 includes the following steps: Step S301: taking the frequency domain feature vector in the training sample set as the input and the optimal order obtained in step S1 as the supervised label; Step S302: performing supervised learning based on the LightGBM model to construct a nonlinear mapping relationship between the input features and the order label; Step S303: by minimizing the loss function between the prediction result and the supervised label, updating the model parameters in multiple rounds of iterative training to train a LightGBM model for order prediction.

5. The adaptive model order estimation method of claim 1, wherein, The output form of the LightGBM model in step S4 is a classification result, which is used to determine whether the order of the structural model belongs to a preset discrete order interval.

6. The method of claim 6, wherein, The LightGBM model has both a regression output path and a classification output path, which are respectively used to output continuous order values and order interval classification results, supporting multi-task structural modal identification.

7. The method of claim 6, wherein, The classification output path includes a softmax layer, which is used to normalize the original classification scores output by the model to generate a probability distribution of the order interval to which the structural model belongs.

8. The method of adaptive model order selection according to claim 1, wherein, Step S4 includes the following steps: Step S401: performing fast Fourier transform on the new structural vibration response signal to extract a frequency domain feature vector thereof; Step S402: inputting the frequency domain feature vector into the trained LightGBM model; Step S403: outputting a predicted order result of the structural model, which is used for subsequent modal parameter identification or state space modeling.

9. The method of adaptive model order selection of claim 1, wherein, The method is suitable for engineering structural modal identification tasks under environmental excitation conditions, including complex structural systems such as offshore platforms, wind towers, bridges and offshore wind power foundations.

Images