A structural damage identification method based on improved adaptive noise complete ensemble empirical mode decomposition and storage medium

Abstract

The application discloses a structural damage identification method based on improved adaptive noise complete ensemble empirical mode decomposition and a storage medium, and obtains the dynamic response of a structure through continuous vibration test; the power spectral density of each mode is calculated based on improved adaptive noise complete ensemble empirical mode decomposition and a frequency spectrum feature inspection method; the power spectral density secondary peak suppression algorithm is used to overcome the influence of the mode aliasing phenomenon in the decomposition result on the classification accuracy; the K-means++ is used for frequency domain clustering of the processed mode components, and a damage sub-signal is extracted; the instantaneous amplitude, phase and other features are calculated, a composite energy factor is constructed, and the existence and position of the structural damage are automatically judged according to the PELT algorithm. The PSD secondary peak suppression frequency domain clustering algorithm based on the improved adaptive noise complete ensemble empirical mode decomposition is proposed, the damage feature sub-signal is extracted, the robust composite energy damage identification factor is constructed, and the existence and position of the damage are visualized and automatically warned through the PELT.

Description

Technical Field

[0001] This invention relates to the field of engineering testing technology, and in particular to a structural damage identification method and storage medium based on improved adaptive noise complete integrated empirical mode decomposition. Background Technology

[0002] In existing technologies, health diagnosis of wharves is mainly based on the dynamic characteristics of the structure, such as frequency, modes, and strain energy. However, the external excitation of the structure usually contradicts the white noise excitation conditions used in previous modal parameter identification. Damage detection using modal features lacks sensitivity, which greatly reduces the effectiveness of structural damage identification. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a structural damage identification method and storage medium based on improved adaptive noise complete integrated empirical mode decomposition, which can realize the visualization of damage existence and damage location and automatic early warning processing.

[0004] To address the aforementioned technical problems, this invention provides a structural damage identification method based on improved adaptive noise complete ensemble empirical mode decomposition, comprising the following steps:

[0005] Step 1: Obtain the dynamic response of the structure through continuous vibration testing;

[0006] Step 2: Calculate the power spectral density of each mode based on the improved adaptive noise complete ensemble empirical mode decomposition and spectral feature verification method;

[0007] Step 3: Use the power spectral density subpeak suppression algorithm to overcome the impact of mode aliasing in the decomposition results on classification accuracy;

[0008] Step 4: Perform frequency domain clustering on the processed modal components using K-means++ to extract the damage sub-signals;

[0009] Step 5: Calculate instantaneous amplitude, phase and other characteristics and construct composite energy factor. Automatically determine the existence and location of structural damage based on the PELT algorithm.

[0010] In step 1, the dynamic response signals are displacement, velocity, and acceleration.

[0011] In step 2, based on the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and the spectral feature verification method, the power spectral density of each mode is calculated as follows: the vibration signal is decomposed into n IMFs and one residual term r(t) using ICEEMDAN, i.e.:

[0012]

[0013] Where x(t) represents the original signal with T time samples;

[0014] The calculation of power spectral density (PSD) treats the original signal x(t) as a sequence with finite energy, assuming N=2. M M is taken as the smallest positive integer when N≥T. The original signal is subjected to a Fast Fourier Transform (FFT) to obtain FFT[x(t)], thus obtaining the estimated value of PSD:

[0015]

[0016] Where fs represents the sampling frequency. The frequency corresponding to the k-th value of the PSD is:

[0017]

[0018] In step 3, the power spectral density subpeak suppression algorithm is used to overcome the impact of mode aliasing in the decomposition results on classification accuracy. Specifically, it is assumed that the PSD frequency range is [f1, f2]. N The core idea of ​​local linear kernel regression is to minimize the weighted least squares formula:

[0019]

[0020] Where h represents the bandwidth, K(·) represents the kernel function, and a Gaussian kernel function is used, thus obtaining As an estimator of the restricted linear estimator g(x), As an estimator of the marginal effect β(x):

[0021]

[0022] To overcome the bandwidth parameter selection problem, least squares cross-validation is used to obtain an adaptive bandwidth, the idea of ​​which is to minimize the function:

[0023]

[0024] in It is composed of g(X) k When generating the prediction for the k-th observation point, the curve SPSD after PSD smoothing is automatically obtained by omitting that point.

[0025] The `find_peaks` function in the SciPy library is used to find the peak points of SPSD. Based on the concept of topographic prominence, local maxima of a set of data are detected. The prominence is set to 0.2 × max(SPSD). Only when the prominence of a peak is higher than 0.2 × max(SPSD) will it be identified, and then the next step of secondary peak suppression will be performed.

[0026] Assume that a total of p peaks of SPSD are found, and the coordinates of the peak points are as follows:

[0027] (H1,SPSD(H1)),...,(H p SPSD(H) p )),H1<H p

[0028] Among them (H) m SPSD(H) m ), 1≤m≤p, representing the coordinates of the main peak, and the rest are the coordinates of the secondary peaks; similarly, after taking the negative value of SPSD, peak identification is performed to obtain at least p-1 and at most p+1 peaks and valleys of SPSD, assuming the coordinates of these peaks and valleys are:

[0029] (L1,SPSD(L1)),(L2,SPSD(L2)),(L3,SPSD(L3)),...

[0030] Based on the peaks and valleys, the frequency ranges of each primary and secondary peak under different conditions are determined. After determining the frequency range corresponding to each secondary peak, the secondary peaks in the PSD are weighted. The weighting coefficient is the exponential form of the ratio of each secondary peak to the primary peak. The calculation formula is as follows:

[0031]

[0032] Where f is the frequency range of the secondary peak j, and c is a constant representing the degree of weighting of the secondary peak, which is taken as c = 4.

[0033] In step 4, the processed modal components are clustered in the frequency domain using K-means++ to extract the damage sub-signals. Specifically, after obtaining the PSD after secondary peak suppression, the i-th IMF is calculated at frequency f. k CSD at the location:

[0034]

[0035] Among them, F i * The WPSD corresponding to the i-th IMF;

[0036] Each CSD is treated as a data object, and K-means++ is used to cluster objects with common data characteristics, thereby achieving automatic and robust reconstruction of IMFs. Euclidean distance is chosen as the synchronization metric, and the CSDs are standardized before clustering, using the following formula:

[0037]

[0038] The silhouette coefficient is chosen as a reference for the optimal number of clusters. Assuming that sample i is assigned to cluster A, its silhouette coefficient and the silhouette coefficient of this clustering are defined as follows:

[0039]

[0040]

[0041] Where a(i) represents the average distance between sample i and other samples in the same cluster, and b(i) represents the minimum average distance between sample i and samples in other clusters. The silhouette coefficient SC ranges from [-1, 1]. The larger the value, the more reasonable and effective the clustering result is when the number of clusters corresponding to that value is selected. The silhouette coefficient curve can be used to determine the selection of the number of clusters.

[0042] In step 5, instantaneous amplitude, phase, and other characteristics are calculated, and a composite energy factor is constructed. The PELT algorithm is used to automatically determine the existence and location of structural damage. Specifically, the formula corresponding to the Hilbert transform is:

[0043]

[0044] Where c(t) refers to the signal to be transformed; after the Hilbert transform, the analytic signal is obtained. The instantaneous amplitude and phase can be obtained from the analytical signal:

[0045]

[0046]

[0047] When the structure of a system changes or its state becomes abnormal, the system's energy also changes accordingly. Therefore, energy can usually reflect structural damage well. The formulas for calculating instantaneous energy and the energy of a certain signal are as follows:

[0048]

[0049]

[0050] Damage or anomalies in a structural system typically lead to phase changes; furthermore, analyzing the phase slope indirectly reveals changes in the information contained in the signal. Therefore, after obtaining the energy and phase characteristics, damage indices are constructed based on energy and phase slope, respectively.

[0051]

[0052]

[0053] Where E h (t), E d (t) represents the energy obtained under healthy and injured conditions, respectively. PS h PS d These represent the slopes of the phases under healthy and damaged conditions, respectively, and are obtained by fitting a linear regression equation based on the least squares principle.

[0054] Based on DI (E) With DI (PS) Fusion to construct composite energy factors:

[0055]

[0056] The PELT algorithm works by minimizing the following cost function:

[0057]

[0058] in, It is a local cost function used to evaluate the cost from time point τ. j To τ j+1 The goodness of fit of the time series segments; k(·,·) is the RBF kernel function; β is the penalty term used to control the total number of change points and prevent overfitting, here β=1; M is the total number of detected change points, which can automatically determine the existence and location of structural damage.

[0059] Correspondingly, a structural damage identification storage medium based on improved adaptive noise complete integrated empirical mode decomposition is provided, on which a computer program is stored. When executed by a processor, the program implements a distribution network estimation fusion method for asynchronous intermittent measurements.

[0060] The beneficial effects of this invention are as follows: This invention can solve two problems in structural damage identification under complex excitation: aliasing of multiple types of signals and robust damage identification and automatic early warning; for the aliasing effect of multiple types of signals, a PSD sub-peak suppression frequency domain clustering algorithm based on ICEEMDAN is proposed to extract damage feature sub-signals; a robust composite energy damage identification factor is constructed, and the PELT algorithm is used to distinguish the existence and location of damage; it can realize the visualization of the existence and location of damage and automatic early warning processing. Attached Figure Description

[0061] Figure 1 This is an overhead view of the experimental model of the high-pile wharf of the present invention.

[0062] Figure 2 This is a side view of the experimental model of the high-pile wharf of the present invention.

[0063] Figure 3 This is a schematic diagram of the acceleration response of each node under the 10% damage condition of the present invention.

[0064] Figure 4(a) shows the IMF1-IMF4 and corresponding PSD diagrams of node 5 obtained by ICEEMDAN under the 10% damage condition of the present invention.

[0065] Figure 4(b) shows the IMF5-IMF8 and corresponding PSD diagrams of node 5 obtained by ICEEMDAN under the 10% damage condition of the present invention.

[0066] Figure 4(c) shows the IMF9-IMF12 and corresponding PSD diagrams of node 5 under the 10% damage condition of this invention, obtained by ICEEMDAN.

[0067] Figure 5 This is a diagram showing the IMFs classification results obtained by the PSD secondary peak suppression frequency domain clustering algorithm of this invention.

[0068] Figure 6 This is a schematic diagram showing the reconstruction results of different types of signals in this invention.

[0069] Figure 7 This is a graph of the composite energy factor of the present invention.

[0070] Figure 8 This is a diagram showing the automatic discrimination results of the PELT algorithm of this invention.

[0071] Figure 9 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0072] like Figure 9 As shown, a structural damage identification method based on improved adaptive noise complete ensemble empirical mode decomposition includes the following steps:

[0073] Step 1: Obtain the dynamic response of the structure through continuous vibration testing. The dynamic response signals are displacement, velocity, and acceleration.

[0074] Step 2: Based on the improved adaptive noise complete ensemble empirical mode decomposition and spectral feature verification method, calculate the power spectral density of each mode, including:

[0075] Research indicates that the dynamic response of structures in practical engineering is often non-stationary and nonlinear, and is composed of multiple types of superimposed signals. Direct use of these signals would result in poor damage identification. Therefore, a method of decomposition, classification, and reconstruction is needed to automatically extract damage-sensitive signals. Empirical Mode Decomposition (EMD) can decompose a signal into several Intrinsic Mode Functions (IMFs), while Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) significantly improves the decomposition effect by introducing the concepts of noise and local mean. The vibration signal is decomposed into n IMFs and one residual term r(t) using ICEEMDAN, i.e.:

[0076]

[0077] Where x(t) represents the original signal with a time sample number of T.

[0078] The calculation of power spectral density (PSD) treats the original signal x(t) as a sequence with finite energy. Assume N = 2. M M is taken as the smallest positive integer when N≥T. The original signal is subjected to a Fast Fourier Transform (FFT) to obtain FFT[x(t)], thus obtaining the estimated value of PSD:

[0079]

[0080] Where fs represents the sampling frequency. The frequency corresponding to the k-th value of the PSD is:

[0081]

[0082] Step 3: Employ a power spectral density secondary peak suppression algorithm to overcome the impact of mode aliasing in the decomposition results on classification accuracy, including:

[0083] To aid classification, the PSD can be integrated to obtain the Cumulative Spectral Distribution (CSD) at the frequency, thus capturing its spectral characteristics. However, the multi-peak phenomenon in the PSD can cause the CSD result to exhibit a stepped shape, elongating its overall form and affecting classification accuracy. Therefore, a PSD sub-peak suppression algorithm is needed to overcome the influence of mode aliasing.

[0084] Given that the PSD corresponding to each IMF component in the decomposition results typically contains multiple "spurs," direct identification would yield multiple extreme points, making it impossible to determine the true peak location. Therefore, we consider smoothing the PSD first. To reduce parameter selection issues, local linear kernel regression is chosen for smoothing. Assume the PSD frequency range is [f1, f2]. N The core idea of ​​local linear kernel regression is to minimize the weighted least squares formula:

[0085]

[0086] Where h represents bandwidth, and K(·) represents kernel function, the Gaussian kernel function is used here. Thus, we obtain... As an estimator of the restricted linear estimator g(x), As an estimator of the marginal effect β(x):

[0087]

[0088] To overcome the bandwidth parameter selection problem, least squares cross-validation is used to obtain an adaptive bandwidth, the idea of ​​which is to minimize the function:

[0089]

[0090] in It is composed of g(X) k When generating the prediction for the k-th observation point, this point is omitted. This automatically yields the PSD-smoothed SPSD curve.

[0091] This invention uses the `find_peaks` function from the SciPy library to find peak points of SPSD and detects local maxima in a set of data based on the concept of topographic prominence. A prominence of 0.2 × max(SPSD) is set; a peak is only identified when its prominence is higher than 0.2 × max(SPSD), thus allowing for the next step of secondary peak suppression.

[0092] Assume a total of p peaks are found in SPSD, and the coordinates of the peak points are as follows:

[0093] (H1,SPSD(H1)),...,(H p SPSD(H) p )),H1<H p

[0094] Among them (H) m SPSD(H) m), 1≤m≤p, represents the coordinates of the main peak, and the rest are the coordinates of the secondary peaks. Similarly, by taking the negative value of SPSD and identifying the peaks, we can obtain at least p-1 and at most p+1 peaks and valleys of SPSD. Assume the coordinates of these peaks and valleys are:

[0095] (L1,SPSD(L1)),(L2,SPSD(L2)),(L3,SPSD(L3)),...

[0096] This allows us to determine the frequency ranges of the main and secondary peaks under different conditions based on the peaks and valleys. After determining the frequency ranges corresponding to each secondary peak, we weight each secondary peak in the PSD, with the weighting coefficient being the exponential form of the ratio of each secondary peak to the main peak. The calculation formula is as follows:

[0097]

[0098] Where f is the frequency range of the secondary peak j, and c is a constant representing the degree of weighting of the secondary peak. In this invention, c = 4 is used, and all algorithms are performed in the Python 3.9.9 environment, using scipy library version 1.7.3.

[0099] Step 4: Perform frequency domain clustering on the processed modal components using K-means++ to extract the damage sub-signals, including:

[0100] After obtaining the PSD after secondary peak suppression, calculate the i-th IMF at frequency f. k CSD at the location:

[0101]

[0102] Among them, F i * Let be the WPSD corresponding to the i-th IMF.

[0103] K-means++ is an improvement on the K-means clustering algorithm. It selects centroids by introducing random probability, avoiding the local optima problem that can arise from random centroid initialization in the traditional K-means algorithm. Each CSD is treated as a data object, and K-means++ is used to cluster objects with common data characteristics, thus achieving automatic and robust reconstruction of IMFs. To obtain better clustering results, Euclidean distance is chosen as the synchronization metric, and the CSDs are standardized before clustering, using the following formula:

[0104]

[0105] Furthermore, the silhouette coefficient is chosen as a reference for the optimal number of clusters. Assuming sample i is assigned to cluster A, its silhouette coefficient and the silhouette coefficient of this clustering are defined as follows:

[0106]

[0107]

[0108] Where a(i) represents the average distance between sample i and other samples in the same cluster, and b(i) represents the minimum average distance between sample i and samples in other clusters. The silhouette coefficient SC ranges from [-1, 1]. The larger the value, the more reasonable and effective the clustering result is when the number of clusters corresponding to that value is selected. The silhouette coefficient curve can be used to determine the selection of the number of clusters.

[0109] Step 5: Calculate instantaneous amplitude, phase, and other characteristics, and construct a composite energy factor. Determine whether damage exists based on whether the damage perception index exceeds a threshold, including:

[0110] The Hilbert Transform (HT) technique is suitable for processing vibration response data of buildings under complex excitations and can provide some unique information. The formula for HT is:

[0111]

[0112] Here, c(t) refers to the signal to be transformed. After passing through HT, the analytic signal can be obtained. The instantaneous amplitude and phase can be obtained from the analytical signal:

[0113]

[0114]

[0115] When the structure of a system changes or its state becomes abnormal, the system's energy also changes accordingly. Therefore, energy can usually reflect structural damage quite well. The formulas for calculating instantaneous energy and the energy of a specific signal are:

[0116]

[0117]

[0118] Phase is one of the important characteristics of a signal, and it is highly sensitive to minute changes in the signal and relative temporal variations. Damage or anomalies in a structural system usually lead to phase changes; furthermore, by analyzing the phase slope, changes in the information contained in the signal can be indirectly understood. Therefore, after obtaining the energy and phase characteristics, damage indices are constructed based on energy and phase slope, respectively:

[0119]

[0120]

[0121] Where E h (t), E d (t) represents the energy obtained under healthy and injured conditions, respectively. PS h PS d The slopes of the phases under healthy and damaged conditions are respectively obtained by fitting a linear regression equation based on the least squares principle.

[0122] Due to factors such as non-stationary dynamic response, low signal-to-noise ratio, and multiple types of aliasing under complex excitation, DI should be comprehensively considered. (E) With DI (PS) By fusing energy and phase information to construct a more sensitive combined damage index, the sensitivity and robustness of structural damage identification can be improved, achieving accurate identification of the existence, location, and extent of structural damage. This is based on DI (Digital Indicator) technology. (E) With DI (PS) Fusion to construct composite energy factors:

[0123]

[0124] Pruned Exact Linear Time (PELT) is an efficient algorithm for detecting structural change points in time series data. This algorithm combines dynamic programming and pruning techniques to ensure the optimal set of breakpoints is found in linear time. The core of the PELT algorithm is to minimize the following cost function:

[0125]

[0126] in, It is a local cost function used to evaluate the cost from time point τ. j To τ j+1 The goodness of fit of the time series segments. Here k(·,·) is the RBF kernel function. β is a penalty term used to control the total number of change points and prevent overfitting; here, β = 1. M is the total number of detected change points. This method can automatically determine the existence and location of structural damage.

[0127] like Figures 1 to 7 As shown, a structural damage identification method and apparatus based on improved adaptive noise complete integrated empirical mode decomposition is proposed for identifying pile foundation damage of high-pile wharf under wave excitation.

[0128] In this invention, we first constructed an experimental model of a high-pile wharf, which was placed in a specially designed water tank capable of simulating the effects of wind, waves, and water flow. Detailed construction of this model can be found in [link to relevant documentation]. Figure 1 and Figure 2Specifically, accelerometers were installed at 0.1m intervals along the front of the model, from top to bottom, on the second pile, for a total of 13 sensors. The primary function of these sensors is to capture vibrations perpendicular to the wharf's leading edge. Different degrees of pile damage were set based on the percentage decrease in stiffness. The damage width and experimental conditions were as follows: 4mm for 5% damage; 9mm for 10% damage; and 24mm for 30% damage.

[0129] The experiment used a regular wave for excitation, with a wave period of 1 second, a water depth of 1 meter, and a wave height of 0.1 meters. The purpose of the experiment was to collect and analyze the acceleration response data of the pile before and after damage occurred.

[0130] The damage identification method specifically includes the following steps:

[0131] Step 1: Obtain the acceleration response of the current state and the intact state of the structure.

[0132] The acquisition equipment uses the DH5920 dynamic signal acquisition and analysis system to achieve multi-channel parallel synchronous acquisition. The sampling frequency of a single channel is 1000Hz. The vibration pickup uses a YD-186 piezoelectric accelerometer. The DH5920 vibration pickup is used to acquire the acceleration signals of 13 nodes of the pile foundation model under wave excitation. Figure 3 As shown.

[0133] Step 2: Based on the improved adaptive noise complete ensemble empirical mode decomposition and spectral feature verification method, the signal decomposition results and power spectral densities of each mode are shown in Figures 4(a)-(c). It can be seen that the IMF components are arranged from high frequency to low frequency, and mode aliasing exists in the results. For example, the PSD corresponding to IMF1-2 exhibits a double-peak phenomenon, and the frequencies of IMF2-3 are all around 100Hz. The power spectral density subpeak suppression algorithm is used to overcome the mode aliasing phenomenon in the decomposition results. K-means++ is used for frequency domain clustering of the processed mode components to extract the damage sub-signals. The classification results and reconstruction results are shown in Figures 4(a)-(c). Figure 5 , Figure 6 As shown in the figure. Figure (a) shows the local bending response, which changes significantly only when the wave arrives and contains information about local structural damage, which is the damage characteristic sub-signal needed for the study; (b) shows the global bending response, which shows a gradual decay trend in the short time following the impact and can be used to identify wave impact; (c) and (d) show the rigid body dynamic response and quasi-static response, respectively, neither of which has obvious impact-induced characteristics; (e) shows the baseline drift term.

[0134] Step 3: Calculate instantaneous amplitude, phase, and other characteristics, and construct a composite energy factor. Based on the PELT algorithm, automatically determine the existence and location of structural damage. Figure 7 , Figure 8 As shown.

[0135] Based on the damage identification results, the curves obtained under different damage levels generally show a trend of first rising and then falling, and all reach their maximum values ​​at nodes 5 or 6. Since the damage is set between nodes 5 and 6, this indicates that the composite energy factor can identify the existence and location of pile foundation damage. Furthermore, the three curves show a significant magnitude relationship, indicating that this index successfully identifies the degree of pile foundation damage. According to the PELT algorithm results, this method can achieve automatic early warning of damage existence and location.

[0136] This invention can effectively extract multiple types of signals from the dynamic response of structures under complex excitation, as well as identify the presence and location of damage, providing a new means for automatic damage early warning during structural service.

Claims

1. A structural damage identification method based on improved adaptive noise complete ensemble empirical mode decomposition, characterized in that, Comprising the following steps: Step 1, obtain the dynamic response of the structure through continuous vibration test; Step 2, calculate the power spectral density of each order mode based on the improved adaptive noise complete ensemble empirical mode decomposition and spectral feature test method; Step 3, the power spectral density sub-peak suppression algorithm is used to overcome the influence of modal aliasing in the decomposition result on the classification accuracy; assuming that the PSD frequency range is The core idea of the local linear kernel regression is to minimize the weighted least square formula: where denotes the bandwidth, denotes the kernel function, a Gaussian kernel function is used, resulting in as a local linear estimator of the estimate, as a marginal effect of the estimate: In order to overcome the problem of bandwidth parameter selection, the least square cross validation method is used to obtain the adaptive bandwidth, the idea of which is to minimize the function: wherein is derived from In generating the prediction for the first observation point, the curve SPSD is automatically obtained by omitting this point, thus smoothing the PSD. The peak points of SPSD are found using the find_peaks function in the Scipy library, which detects local maxima of a set of data according to the concept of terrain prominence. The parameters are set as follows Only when the prominence of a certain peak is higher than It will be identified and the next step of sub-peak suppression will be performed. Suppose that SPSD is found to have peaks, with peak point coordinates in order: wherein , denotes the main peak coordinate, the rest are the coordinates of the peaks; similarly, the negative of SPSD is taken to identify the peaks, obtaining at least , at most peaks and valleys of SPSD, assuming the coordinates of these peaks and valleys are: According to the peak-to-valley, the frequency range of each main peak and sub-peak under different conditions is determined, and after determining the frequency range corresponding to each sub-peak, each sub-peak in the PSD is weighted, and the weighting coefficient is the exponential form of the ratio of each sub-peak to the main peak, and the calculation formula is: in, Secondary peak The frequency range It is a constant; Step 4, K-means++ is used for frequency domain clustering of the processed modal components to extract damage sub-signals; Step 5, calculate the instantaneous amplitude, phase feature and construct the composite energy factor, and automatically judge the existence and location of structural damage according to the PELT algorithm.

2. The method for structural damage identification based on improved adaptive noise complete ensemble empirical mode decomposition according to claim 1, wherein, In step 1, the dynamic response signal is displacement, velocity and acceleration.

3. The method for structural damage identification based on improved adaptive noise complete ensemble empirical mode decomposition according to claim 1, wherein, In step 2, based on the improved adaptive noise complete ensemble empirical mode decomposition and spectrum feature test method, the power spectrum density of each order mode is calculated, specifically: the vibration signal is decomposed into IMF and 1 residual term that is: wherein represents a raw signal with a number of time samples representing The calculation of power spectral density will use the original signal Consider it as a sequence with finite energy, assuming , Take The smallest positive integer at time is used to perform a Fast Fourier Transform on the original signal to obtain... Thus, the estimated value of PSD is obtained: wherein, represents the sampling frequency, the PSD value corresponding to the frequency: f = n * fs 。 4. The method for structural damage identification based on improved adaptive noise complete ensemble empirical mode decomposition according to claim 1, wherein, represents the degree of weighting to the secondary peak, taken as .

5. The method for structural damage identification based on improved adaptive noise complete ensemble empirical mode decomposition according to claim 1, wherein, In step 4, K-means++ is used for frequency domain clustering of the processed modal components to extract damage sub-signals. Specifically, after obtaining the PSD after peak suppression, the CSD of the first IMF at the frequency is calculated: ​ wherein, is the WPSD corresponding to the th IMF. Each CSD is regarded as a data object, K-means++ is used to cluster objects with common data characteristics, thereby realizing automatic and robust reconstruction of IMFs, the Euclidean distance is selected as the synchronization measure, and the CSD is standardized before clustering, and the formula is: The silhouette coefficient provides a reference for choosing the optimal number of clusters, assuming that the samples are assigned to cluster A, the silhouette coefficient of the sample and the silhouette coefficient of the cluster are defined as: wherein, representing the sample the average distance to other samples in the same cluster, representing the sample the minimum value of the average distance to other samples in other clusters; the silhouette coefficient SC has a value range of [-1, 1], and the larger the value is, the more reasonable and effective the clustering result obtained by selecting the cluster number corresponding to the value is, and the classification number selection is determined through the silhouette coefficient curve.

6. The method for structural damage identification based on improved adaptive noise complete ensemble empirical mode decomposition according to claim 1, wherein, In step 5, the instantaneous amplitude, phase feature and composite energy factor are calculated, and the existence and location of structural damage are automatically judged according to the PELT algorithm, specifically: wherein to be transformed; after Hilbert transformation, analytic signal is obtained According to the analytic signal, the instantaneous amplitude and phase can be obtained respectively When the structure of the system itself changes or the state of the system becomes an abnormal state, the energy of the system itself will also change, so energy can usually better reflect structural damage, and the calculation formula of instantaneous energy and the energy of a signal is: When the structure system is damaged or abnormal, it usually causes the change of phase; in addition, by analyzing the phase slope, the change of information contained in the signal can be indirectly understood; therefore, after obtaining the energy and phase characteristics, damage indexes are constructed based on energy and phase slope respectively: wherein respectively represent the energy obtained in the healthy, damaged condition, , respectively represent the slope of the phase in the healthy, damaged condition, obtained by fitting a linear regression equation according to the least squares principle; Based on With Fusion constructs composite energy factor: The PELT algorithm minimizes the following cost function: wherein, is a local cost function for assessing the goodness of fit of a time series column segment from a time point to ; , is an RBF kernel function; is a penalty term, is the total number of detected change points, which can automatically determine the existence of structural damage and the location of damage.

7. The method for structural damage identification based on improved adaptive noise complete ensemble empirical mode decomposition according to claim 6, wherein, to control the total number of change points to prevent overfitting.

8. A storage medium for structural damage identification based on improved adaptive noise complete ensemble empirical mode decomposition, which stores a computer program, the program being executed by a processor to realize a structural damage identification method based on improved adaptive noise complete ensemble empirical mode decomposition according to any one of claims 1 to 7.

Images