CFRP wing skin damage positioning system and method based on multi-modal signal processing and Bayesian optimization DSCN

Abstract

The invention relates to the technical field of material nondestructive testing, in particular to a CFRP wing skin damage positioning system and method based on multi-modal signal processing and Bayesian optimization DSCN, and the method comprises the steps: carrying out permutation entropy analysis and Higuchi fractal dimension analysis on a signal data set received by a sensor; performing time-frequency feature extraction on the sensor receiving signal data set deviating from the reference; performing damage feature screening on the signal data in the effective time period; training the DSCN structure to obtain an initial CFRP wing skin damage positioning prediction model; performing hyper-parameter optimization on the initial CFRP wing skin damage positioning prediction model according to a Bayesian algorithm; and performing damage positioning prediction on a to-be-detected sample according to the final CFRP wing skin damage positioning prediction model to obtain a corresponding CFRP wing skin damage positioning result. According to the method, effective damage features can be accurately and efficiently extracted from complex signals, and the accuracy and efficiency of CFRP wing skin damage positioning are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of material nondestructive testing, and in particular to a CFRP wing skin damage location system and method based on multimodal signal processing and Bayesian optimization (DSCN). Background Art

[0002] The wing skin is the outer structural component of a drone's wing, directly covering the wing frame to create a complete aerodynamic shape. The primary function of a drone's wing skin is to maintain the drone's aerodynamic shape, transmit aerodynamic loads, and protect internal equipment. Therefore, the integrity of the wing skin directly impacts the drone's flight performance and safety. Carbon fiber reinforced plastic (CFRP) has become a mainstream material in drone structures due to its light weight, high strength, excellent fatigue resistance, and customizable anisotropic properties. Globally, CFRP accounts for 60%-80% of drone structures, with wing skin being a major application. Wing skins made of CFRP, with their smooth surface, precise shape, and excellent symmetry, reduce air resistance, increase flight speed and efficiency, and provide sufficient lift and stability for the wing.

[0003] However, the brittle matrix and layered structure of CFRP make it sensitive to impact loads and prone to internal damage. During long-term service, cyclic loads will cause the material to accumulate microscopic damage and initiate fatigue cracks. If not detected in time, the cracks will expand into macroscopic damage, threatening flight safety. Accurately locating crack damage inside the wing skin is the key to preventing catastrophic failure of CFRP structures. Traditional non-destructive testing technology has problems such as complex operation, low efficiency, detection blind spots and radiation risks. In contrast, Lamb wave monitoring technology has become an ideal solution for internal damage detection of CFRP wing skins due to its millisecond response speed, long propagation distance, and high sensitivity to hidden damage.

[0004] However, Lamb wave technology faces three major technical bottlenecks in locating damage on curved CFRP wing skins: Multi-sensor signal modal aliasing and noise interference, making it difficult for existing sensor screening methods to distinguish damage signals from noise due to their reliance on a single metric; the difficulty in extracting complex signal features, which hinders traditional time-frequency analysis methods; and the redundancy of high-dimensional time series signals, which leads traditional linear dimensionality reduction methods to suffer from nonlinear failure and local feature loss. Furthermore, existing damage localization models suffer from limitations such as insufficient representation capabilities of shallow networks and inefficient optimization of deep learning models. Hyperparameter optimization also faces challenges, with traditional methods facing problems such as the curse of dimensionality and a lack of intelligent guidance.

[0005] Chinese Patent Publication No. CN116642410A discloses a non-contact CFRP structural damage monitoring system and method. The non-contact CFRP structural damage monitoring system includes a vector network analyzer (VNA), a focusing antenna, and a computer. The VNA is connected to the focusing antenna via a coaxial cable and is responsible for transmitting, receiving, processing, and analyzing microwave signals. The focusing antenna focuses and amplifies the pulse signal generated by the VNA, irradiating it onto the center of the CFRP plate and receiving the reflected signal. The capacitance of the CFRP plate is derived from the reflected microwave signal, thereby characterizing the strain in the deformed area. This solution makes it difficult to accurately and efficiently extract effective damage features from complex signals for CFRP wing skin damage localization, resulting in reduced damage localization accuracy and efficiency. Summary of the Invention

[0006] To this end, the present invention provides a CFRP wing skin damage localization system and method based on multimodal signal processing and Bayesian optimization DSCN, so as to overcome the problem in the prior art that it is difficult to accurately and efficiently extract effective damage features from complex signals in CFRP wing skin damage localization, resulting in reduced accuracy and efficiency of damage localization.

[0007] To achieve the above objectives, the present invention provides a CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN, comprising the following steps: S1. In a numerical simulation experiment of a CFRP wing skin, a damage-free model of the wing skin is established, and damage locations of the CFRP wing skin are set at multiple spatial coordinates of the damage-free model of the wing skin. Lamb wave time-domain signal data corresponding to the damage locations of the CFRP wing skin are obtained, and a corresponding sensor received signal data set is obtained based on the Lamb wave time-domain signal data; S2. Perform permutation entropy analysis and Higuchi fractal dimension analysis on the sensor receiving signal dataset based on a preset dual-index sensor signal screening strategy to obtain a sensor receiving signal dataset that deviates from the baseline; S3, extracting time-frequency features from the sensor received signal data set that deviates from the reference to obtain signal data for a valid time period; S4. Based on the functional principal component analysis strategy, damage characteristics are screened for the signal data in the effective time period to obtain damage characteristics screening results; S5. Train the DSCN structure based on the damage feature screening results to obtain the initial CFRP wing skin damage location prediction model; S6. Optimizing the hyperparameters of the initial CFRP wing skin damage location prediction model according to the Bayesian algorithm to obtain the final CFRP wing skin damage location prediction model; S7. Perform damage location prediction on the test sample according to the final CFRP wing skin damage location prediction model to obtain the corresponding CFRP wing skin damage location result.

[0008] Compared with the prior art, the present invention has the following advantages: 1. By using permutation entropy analysis and Higuchi fractal dimension analysis to screen the sensor received signal data set, the permutation entropy can quantify the randomness of the signal and reflect the dynamic complexity of the signal. The Higuchi fractal dimension can describe the self-similarity and nonlinear characteristics of the signal. Through the dual-index strategy combining permutation entropy and Higuchi fractal dimension, the sensor received signal that deviates from the baseline can be identified from multiple dimensions, and irrelevant signals caused by environmental noise, sensor errors, etc. can be effectively filtered out. Compared with the existing method of indirectly characterizing strain only through capacitance changes, this application can screen the original signal received by the sensor, more accurately retain the effective information related to the damage, and avoid the interference of invalid features in complex signals.

[0009] 2. By extracting the time-frequency features of the sensor received signal data set that deviates from the baseline, the effective time period signal data can be obtained. The time-frequency analysis can expand the signal in two dimensions of time and frequency, and clearly present the changes in the frequency components of the signal at different times. When damage occurs to the CFRP wing skin, the frequency characteristics of the related signals of the CFRP wing skin will change within a specific time period. Through time-frequency feature extraction, the frequency change information closely related to the damage can be captured. Compared with the existing technology that only analyzes the signal from a single dimension, the present application can obtain the damage signal characteristics more comprehensively and accurately, thereby improving the efficiency of extracting effective damage features from the sensor received signal, thereby improving the overall efficiency and accuracy of damage location.

[0010] 3. By using the Bayesian algorithm to optimize the hyperparameters of the initial CFRP wing skin damage location prediction model, the optimal parameter combination is automatically searched, avoiding the blindness of manual parameter adjustment. Compared with the single feature characterization method in the existing technology that only relies on microwave reflection signals to deduce capacitance, this application can extract key damage features from multi-dimensional time-frequency features through a functional principal component analysis strategy, and ensure the optimal configuration of model parameters through the Bayesian algorithm. Through the synergistic effect of the functional principal component analysis strategy and the Bayesian algorithm, the final CFRP wing skin damage location prediction model can more accurately capture the mapping relationship between damage features and positions, thereby improving the accuracy of damage location.

[0011] Furthermore, the S1 comprises the following steps: S11. Establishing an undamaged model of the wing skin based on finite element simulation software, and determining multiple internal crack damage locations in the undamaged model of the wing skin based on preset crack damage rules, thereby forming a CFRP wing skin simulation structure including an undamaged state and different damaged states, wherein the preset crack damage rules include presetting internal crack damage locations in high stress areas of the CFRP wing skin and setting them according to typical damage modes of the CFRP wing skin; S12. Simulate the excitation and propagation process of Lamb waves in the CFRP wing skin simulation structure to obtain the original time domain signal data set; S13. Locate each preset sensor receiving point from the original time domain signal dataset, and analyze the fluctuation response characteristics of each sensor receiving point in the time domain dimension to obtain a sensor receiving signal dataset, wherein the sensor receiving signal dataset includes an undamaged state dataset and a damaged state dataset.

[0012] In this solution, finite element simulation software refers to an engineering simulation finite element software that includes a unit library and a material model library and can handle complex boundary and load conditions. The wing skin damage-free model refers to a model built based on the finite element simulation software to simulate the structural morphology and mechanical properties of the wing skin when it has no internal cracks or other damage. The CFRP wing skin simulation structure refers to a model based on the wing skin damage-free model that determines the formation of multiple internal crack damage locations according to preset crack damage rules, including structures with no damage and in different damage states.

[0013] By using finite element simulation software to construct a simulation structure of a CFRP wing skin containing both undamaged and different damaged states, the excitation and propagation of Lamb waves can be simulated to obtain the original time domain signal data set. The sensor receiving point is located from this original time domain signal data set and the wave response characteristics are analyzed to obtain a sensor receiving signal data set containing data in both undamaged and damaged states, which helps to accurately analyze the damage characteristics.

[0014] Furthermore, the step S2 includes the following steps: S21. Perform permutation entropy analysis on the undamaged state data set and the damaged state data set according to the permutation entropy analysis algorithm in the preset dual-index sensor signal screening strategy, and obtain the permutation entropy value of each sensor in the undamaged state and the damaged state of the CFRP wing skin. , wherein the mathematical expression of the permutation entropy value is: , represents the probability of the occurrence of the i-th arrangement pattern, represents the embedding dimension, Represents the embedding dimension The factorial of S22. performing Higuchi fractal dimension analysis on the undamaged state data set and the damaged state data set according to the Higuchi fractal dimension algorithm in the preset dual-index sensor signal screening strategy, respectively, to obtain Higuchi fractal dimension values ​​of each sensor in the undamaged state and the damaged state of the CFRP wing skin; S23, integrating the permutation entropy value and the Higuchi fractal dimension value to obtain a dual-index feature matrix; S24. Setting a permutation entropy change benchmark and a Higuchi fractal dimension change benchmark for each sensor based on the undamaged state data set, and determining a corresponding permutation entropy change rate and a Higuchi fractal dimension change rate based on the permutation entropy value and Higuchi fractal dimension value of each sensor in the damaged state of the CFRP wing skin; S25. Jointly analyzing the dual-index feature matrix according to the permutation entropy change benchmark and the Higuchi fractal dimension change benchmark to obtain a list of candidate sensitive sensors that meet preset sensitive sensor conditions, wherein the preset sensitive sensor conditions include a permutation entropy change rate exceeding the permutation entropy change benchmark, or a Higuchi fractal dimension change rate exceeding the Higuchi fractal dimension change benchmark; S26: Add the sensor receiving signal corresponding to each candidate sensitive sensor in the candidate sensitive sensor list to the sensor receiving signal data set that deviates from the baseline.

[0015] In this scheme, the preset dual-index sensor signal screening strategy refers to a method of screening sensor signals by comprehensively utilizing permutation entropy analysis and Higuchi fractal dimension analysis. By performing these two analyses on the undamaged and damaged state data sets respectively, integrating the results and combining the change benchmark to screen out sensitive sensors, the permutation entropy change benchmark refers to the change reference standard of the permutation entropy value of each sensor set based on the undamaged state data set, which is used to measure whether the change of the permutation entropy value of the sensor in the damaged state exceeds the normal range, and the Higuchi fractal dimension change benchmark refers to the change reference standard of the Higuchi fractal dimension value of each sensor set based on the undamaged state data set, which is used to judge whether the sensor is damaged. Whether the change in the Higuchi fractal dimension value under the damaged state is abnormal; the permutation entropy change rate refers to the degree of change in the permutation entropy value of each sensor in the damaged state of the CFRP wing skin relative to the undamaged state; the Higuchi fractal dimension change rate refers to the degree of change in the Higuchi fractal dimension value of each sensor in the damaged state of the CFRP wing skin relative to the undamaged state; the candidate sensitive sensor list refers to the sensor list screened out based on the comparison of the permutation entropy and Higuchi fractal dimension change rate with the corresponding change benchmark; the sensor reception signal data set that deviates from the benchmark refers to the data set formed by adding the sensor reception signals corresponding to the candidate sensitive sensor list.

[0016] By presetting a dual-indicator sensor signal screening strategy, the permutation entropy and Higuchi fractal dimension analyses are performed on the damage-free data set and the damaged state data set, respectively, to obtain a dual-indicator feature matrix. Combined with the change benchmark set in the damage-free state, the change rate of each sensor is determined, and then a list of candidate sensitive sensors is screened out and a sensor receiving signal dataset that deviates from the benchmark is formed. This can accurately identify sensors that are sensitive to damage, remove irrelevant signal interference, and highlight damage characteristics.

[0017] Furthermore, the step S3 includes the following steps: S31, perform time-frequency localization analysis on the sensor receiving signal data set that deviates from the reference by continuous wavelet transform to obtain a time-frequency feature data set, wherein the energy distribution of the time-frequency plane in the time-frequency localization analysis process is The mathematical expression is: ,in, represents the scale parameter, represents the translation parameter, represents the wavelet basis function, represents the time domain signal, represents the time variable; S32, input the time-frequency feature data set into the Markov transition field to perform state transfer analysis, and construct a state transfer probability matrix, where the state transfer probability The mathematical expression is: , represents the number of transitions from state i to state j, represents the number of transitions from state i to state k, represents the smoothing factor, k represents the number of states; S33. Obtain signal data for a valid time period by analyzing the signal evolution law of the state transition probability matrix and screening the characteristic time period.

[0018] In this scheme, the time-frequency feature dataset refers to the data set obtained by using continuous wavelet transform to perform time-frequency localization analysis on the sensor received signal dataset that deviates from the baseline. The state transition probability matrix refers to the matrix constructed after inputting the time-frequency feature dataset into the Markov transition field for state transition analysis. The signal evolution law of the state transition probability matrix refers to analyzing the transition trends and patterns of the signal between different states based on the state transition probability matrix, thereby revealing the change law of the signal over time. The effective time period signal data refers to the signal data within a specific time period containing key damage information that is screened out by analyzing the signal evolution law of the state transition probability matrix.

[0019] Through continuous wavelet transform, time-frequency analysis of signals that deviate from the baseline is performed to obtain a time-frequency feature dataset, which comprehensively presents the time-frequency characteristics of the signal. The time-frequency feature dataset is input into the Markov transition field to construct a state transition probability matrix, and the signal evolution law is analyzed. The trend of signal state changes can be accurately grasped. Based on this, signal data in the effective time period is screened out, irrelevant and redundant information is removed, and key features related to damage are highlighted.

[0020] Further, the S4 includes the following steps: S41. Standardizing the effective time period signal data based on a functional principal component analysis strategy to obtain standardized effective time period signal data, and constructing a mean function based on the standardized effective time period signal data; S42, performing centralization processing on the normalized effective time period signal data according to the mean function to obtain the centralized effective time period signal data; S43, performing basis function expansion on the effective time period signal data after centralization processing according to the B-spline basis function to obtain a basis function matrix, and constructing a projection matrix according to the basis function matrix; S44, projecting the signal data of the valid time period after the centralization processing according to the projection matrix to obtain projected signal data; S45, performing covariance calculation on the projected signal data to obtain a covariance matrix, and processing the covariance matrix according to an L2 regularization method to obtain a regularized covariance matrix; S46. Perform Cholesky decomposition on the regularized covariance matrix to obtain an intermediate result including basis coefficients, and perform damage feature screening on the projected signal data based on the intermediate result including basis coefficients to obtain a damage feature screening result.

[0021] In this solution, the functional principal component analysis strategy refers to a method for extracting main features from data, reducing data dimensions, and highlighting key information in the data. The B-spline basis function refers to a basis function that can flexibly fit various complex function shapes by expanding functional data. The projection matrix refers to a matrix constructed from the basis function matrix and used to project the signal data of the valid time period after centralization into a new space. The L2 regularization method refers to the addition of a small regularization parameter to the diagonal of the covariance matrix in covariance matrix processing. The Cholesky decomposition refers to a method for decomposing a symmetric positive definite matrix into a lower triangular matrix and its transpose product.

[0022] Through a functional principal component analysis strategy, the signal data of the effective time period is first standardized and centered, and then the projection matrix is ​​constructed using B-spline basis function expansion for projection. The covariance matrix is ​​then processed using the L2 regularization method, and finally Cholesky decomposition is performed. This can effectively extract the main features of the signal data, reduce the data dimension, and improve the efficiency and accuracy of data processing. The stability of the covariance matrix is ​​enhanced through regularization processing, which can more accurately reflect the damage characteristics of the CFRP wing skin.

[0023] Furthermore, the step S5 includes the following steps: S51. Initializing the DSCN structure according to the progressive architecture construction method, wherein the network initialization includes setting the DSCN structure to a single hidden layer and setting the hidden layer to a single node, and all hidden layer nodes are fully connected to the output layer of the DSCN structure; S52. Constructing training samples based on the damage feature screening results, and training the DSCN structure based on the training samples. Dynamically adjusting node parameters through a supervised learning mechanism to obtain optimized node parameters and corresponding model performance indicators. S53, based on the optimized node parameters and the corresponding model performance indicators, gradually increase the number of hidden layers and nodes along the depth direction and the width direction to obtain an expanded DSCN structure; S54. Perform output weight analysis on the expanded DSCN structure, calculate the connection weights between all hidden layer nodes and the output layer using the least squares method, and establish an initial CFRP wing skin damage location prediction model.

[0024] In this solution, the DSCN (Deep Stochastic Configuration Networks) structure refers to a deep neural network model based on random configuration that combines the hierarchical feature extraction capabilities of deep learning with random configuration methods. The progressive architecture construction method refers to a method for gradually building and optimizing the neural network structure. The supervised learning mechanism refers to the process of adjusting network node parameters through the results of known damage feature screening when training the DSCN structure. The network calculates the error between the predicted output and the actual label based on the input training samples and corresponding labels, and then dynamically adjusts the node parameters through the backpropagation algorithm to minimize the error.

[0025] By adopting a progressive architecture construction method, the DSCN structure is initialized, and the number of hidden layers and nodes is gradually increased. The network complexity can be flexibly adjusted according to the training samples and model performance indicators. Through the supervised learning mechanism, the damage feature screening results are used as training samples, and the node parameters are dynamically adjusted, so that the DSCN structure can better learn the characteristics and patterns in the data. Finally, the output weights are determined by the least squares method, and an initial CFRP wing skin damage location prediction model is established, which improves the accuracy and adaptability of the model.

[0026] Further, the S6 includes the following steps: S61. Determine the input feature dimension and hyperparameter set of the initial CFRP wing skin damage location prediction model based on the Bayesian algorithm; S62. Determine the test error between the predicted coordinates and the true coordinates of the test set samples based on the input feature dimension and the hyperparameter set, and construct a target optimization function with the goal of minimizing the test error; S63. Performing Bayesian probability modeling analysis on the target optimization function through prior distribution and mixed kernel function to construct a Gaussian process model; S64. Constructing a time-weighted expectation improved acquisition function based on the predicted mean, standard deviation, and current optimal observation value of the Gaussian process model, and selecting a hyperparameter set for the next evaluation point through the time-weighted expectation improved acquisition function; S65. Update the Gaussian process model based on the hyperparameter set of the next evaluation point, combined with the model prediction value and the actual test error, until the preset convergence condition is met, and output the optimal hyperparameter combination; S66. Input the optimal hyperparameter combination into the initial CFRP wing skin damage location prediction model to perform hyperparameter optimization to obtain the final CFRP wing skin damage location prediction model.

[0027] In this scheme, the input feature dimension refers to the number of features used to input the initial CFRP wing skin damage location prediction model. The hyperparameter set refers to the parameters that need to be manually set before model training, such as the learning rate and regularization coefficient. Different hyperparameter combinations will affect the training effect and prediction performance of the model. The objective optimization function refers to a function constructed with the goal of minimizing the test error between the predicted coordinates and the true coordinates of the test set samples. It is used to evaluate the performance of the model under different input feature dimensions and hyperparameter combinations. The Gaussian process model refers to a model that provides a probabilistic prediction of the objective function obtained by performing Bayesian probability modeling analysis on the objective optimization function. The time-weighted expectation improvement acquisition function refers to a hyperparameter set constructed based on the predicted mean, standard deviation, and current optimal observation value of the Gaussian process model and is used to select the next evaluation point. The preset convergence condition refers to the standard used to determine whether the model optimization process has ended, such as reaching the maximum number of iterations or the change in the objective function value being less than a threshold. When this condition is met, the optimal hyperparameter combination is output.

[0028] By using the Bayesian algorithm to determine the input feature dimension and hyperparameter set, a basis for model optimization is provided. By constructing a target optimization function to evaluate the model performance, and then using Bayesian probability modeling to obtain a Gaussian process model, the target function can be accurately predicted. The time-weighted expectation improvement acquisition function is used to effectively select the next evaluation point hyperparameter set, accelerate the optimization process, and update the Gaussian process model according to the next evaluation point until the preset convergence conditions are met to output the optimal hyperparameter combination, which is input into the initial CFRP wing skin damage location prediction model to obtain the final CFRP wing skin damage location prediction model. The model parameters can be automatically optimized, the accuracy and stability of the CFRP wing skin damage location prediction model can be improved, the positioning error can be reduced, and the damage location effect can be improved.

[0029] Further, the S7 includes the following steps: S71, performing multimodal signal acquisition on the CFRP wing skin to be tested to obtain a sample to be tested, and preprocessing the sample to be tested to obtain a preprocessed sample to be tested; S72. Input the preprocessed test sample into the final CFRP wing skin damage location prediction model to perform damage location prediction, obtain a spatial comparison map containing the real three-dimensional coordinates of the damage and the predicted three-dimensional coordinates, and perform spatial constraint analysis on the spatial comparison map to obtain the corresponding CFRP wing skin damage location result.

[0030] In this scheme, the test samples are obtained and preprocessed through multimodal signal acquisition, which can comprehensively and accurately obtain the status information of the CFRP wing skin to be tested. The preprocessed test samples are input into the final CFRP wing skin damage location prediction model for damage location prediction. A spatial comparison map containing the real three-dimensional coordinates of the damage and the predicted three-dimensional coordinates is obtained, which can intuitively present the damage distribution. The spatial comparison map is then subjected to spatial constraint analysis to obtain the damage location result, thereby improving the accuracy of damage location.

[0031] On the other hand, the present invention also provides a CFRP wing skin damage localization system based on multimodal signal processing and Bayesian optimization DSCN, comprising: a sensor received signal data set acquisition module, configured to establish a wing skin undamaged model in a numerical simulation experiment of the CFRP wing skin, set a CFRP wing skin damage location at multiple spatial coordinates of the wing skin undamaged model, acquire corresponding Lamb wave time domain signal data based on the CFRP wing skin damage location, and obtain a corresponding sensor received signal data set based on the Lamb wave time domain signal data; A sensor signal screening module is used to perform permutation entropy analysis and Higuchi fractal dimension analysis on the sensor receiving signal data set according to a preset dual-index sensor signal screening strategy to obtain the sensor receiving signal data set that deviates from the benchmark; The time-frequency feature extraction module is used to extract the time-frequency features of the sensor received signal data set that deviates from the reference and obtain the signal data of the valid time period; The damage feature screening module is used to screen the damage features of the signal data in the effective time period according to the functional principal component analysis strategy to obtain the damage feature screening results; The model training module is used to train the DSCN structure based on the damage feature screening results to obtain the initial CFRP wing skin damage location prediction model; A model optimization module is used to optimize the hyperparameters of the initial CFRP wing skin damage location prediction model based on the Bayesian algorithm to obtain the final CFRP wing skin damage location prediction model; The damage location prediction module is used to perform damage location prediction on the test sample according to the final CFRP wing skin damage location prediction model to obtain the corresponding CFRP wing skin damage location result. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of a process for locating CFRP wing skin damage based on multimodal signal processing and Bayesian optimization DSCN according to an embodiment of the present invention; Figure 2Schematic diagram of the structure of a CFRP wing skin damage localization system based on multimodal signal processing and Bayesian optimization DSCN according to an embodiment of the present invention; Figure 3 Graph showing performance differences among different prediction models in an embodiment of the present invention; Figure 4 This is a diagram showing the error distribution of different prediction models on different test samples in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following is further described in detail through specific implementation methods: See also Figure 1 , which is a flow chart of a CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to an embodiment of the present invention, including the following steps: S1. In a numerical simulation experiment of a CFRP wing skin, a damage-free model of the wing skin is established, and damage locations of the CFRP wing skin are set at multiple spatial coordinates of the damage-free model of the wing skin. Lamb wave time-domain signal data corresponding to the damage locations of the CFRP wing skin are obtained, and a corresponding sensor received signal data set is obtained based on the Lamb wave time-domain signal data; S2. Perform permutation entropy analysis and Higuchi fractal dimension analysis on the sensor receiving signal dataset based on a preset dual-index sensor signal screening strategy to obtain a sensor receiving signal dataset that deviates from the baseline; S3, extracting time-frequency features from the sensor received signal data set that deviates from the reference to obtain signal data for a valid time period; S4. Based on the functional principal component analysis strategy, damage characteristics are screened for the signal data in the effective time period to obtain damage characteristics screening results; S5. Train the DSCN structure based on the damage feature screening results to obtain the initial CFRP wing skin damage location prediction model; S6. Optimizing the hyperparameters of the initial CFRP wing skin damage location prediction model according to the Bayesian algorithm to obtain the final CFRP wing skin damage location prediction model; S7. Perform damage location prediction on the test sample according to the final CFRP wing skin damage location prediction model to obtain the corresponding CFRP wing skin damage location result.

[0034] Specifically, S1 includes the following steps: S11. Establishing an undamaged model of the wing skin based on finite element simulation software, and determining multiple internal crack damage locations in the undamaged model of the wing skin based on preset crack damage rules, thereby forming a CFRP wing skin simulation structure including an undamaged state and different damaged states. The preset crack damage rules include presetting internal crack damage locations in high stress areas of the CFRP wing skin and setting them according to typical damage modes of the CFRP wing skin. S12. Simulate the excitation and propagation process of Lamb waves in the CFRP wing skin simulation structure to obtain the original time domain signal data set; S13. Locate each preset sensor receiving point from the original time domain signal dataset, and analyze the fluctuation response characteristics of each sensor receiving point in the time domain dimension to obtain a sensor receiving signal dataset. The sensor receiving signal dataset includes an undamaged state dataset and a damaged state dataset.

[0035] In this example, the excitation and propagation of Lamb waves in a simulated CFRP wing skin structure are simulated using finite element simulation software. First, a complete geometric model of the wing skin is constructed, including an undamaged state and multiple pre-set internal crack damage locations. Cracks can be simulated by reducing local material stiffness or introducing cohesive elements to ensure that their geometric characteristics (e.g., length and angle) conform to the actual damage scenario. Subsequently, orthotropic material properties, including elastic modulus, Poisson's ratio, and shear modulus, are assigned to the CFRP laminate, and the fiber orientation of each ply is defined. An appropriate excitation signal (e.g., a Hanning window-modulated sine wave) is selected, and a displacement or force load is applied to selected locations on the wing skin surface to simulate the excitation of a piezoelectric sensor. The excitation direction must match the target Lamb wave mode. The solution is performed using an explicit dynamic analysis step, with a suitable mesh size (to meet wavelength resolution) and time step (to ensure stability). The displacement or stress field output during wave propagation is recorded. After the simulation is complete, the time-domain displacement / acceleration data for each pre-set sensor receiving point is extracted to form the raw time-domain signal dataset.

[0036] To analyze the time-domain fluctuation response characteristics of the sensor receiving point, the sensor locations are first located within the original dataset, ensuring that the crack region is avoided to capture the directly propagating wave signal. The time-domain signal is extracted using the History Output function of the finite element simulation software, and time alignment and filtering are performed to remove noise and retain the valid frequency band. Subsequently, the amplitude variation, wave packet arrival time (ToF), and waveform distortion of the time-domain signal are analyzed, and the signal energy (such as the RMS value) is calculated or cross-correlation analysis is performed to compare the difference between the damaged state and the undamaged baseline signal. Finally, the analyzed features are classified and stored according to sensor location and damage condition, and a sensor receiving signal dataset containing the undamaged state and different damage states is constructed. Throughout the entire process, attention must be paid to the dispersion effect of CFRP and the crack modeling accuracy, and computational efficiency must be optimized to ensure simulation feasibility.

[0037] Specifically, S2 includes the following steps: S21. Perform permutation entropy analysis on the undamaged state data set and the damaged state data set according to the permutation entropy analysis algorithm in the preset dual-index sensor signal screening strategy, and obtain the permutation entropy value of each sensor in the undamaged state and the damaged state of the CFRP wing skin. , where the mathematical expression of the permutation entropy is: , represents the probability of the occurrence of the i-th arrangement pattern, represents the embedding dimension, Represents the embedding dimension The factorial of S22. performing Higuchi fractal dimension analysis on the undamaged state data set and the damaged state data set according to the Higuchi fractal dimension algorithm in the preset dual-index sensor signal screening strategy, respectively, to obtain Higuchi fractal dimension values ​​of each sensor in the undamaged state and the damaged state of the CFRP wing skin; S23, integrating the permutation entropy value and the Higuchi fractal dimension value to obtain a dual-index feature matrix; S24. Setting a permutation entropy change benchmark and a Higuchi fractal dimension change benchmark for each sensor based on the undamaged state data set, and determining a corresponding permutation entropy change rate and a Higuchi fractal dimension change rate based on the permutation entropy value and Higuchi fractal dimension value of each sensor in the damaged state of the CFRP wing skin; S25. Jointly analyzing the dual-index feature matrix according to the permutation entropy change benchmark and the Higuchi fractal dimension change benchmark to obtain a list of candidate sensitive sensors that meet preset sensitive sensor conditions, wherein the preset sensitive sensor conditions include a permutation entropy change rate exceeding the permutation entropy change benchmark, or a Higuchi fractal dimension change rate exceeding the Higuchi fractal dimension change benchmark; S26: Add the sensor receiving signal corresponding to each candidate sensitive sensor in the candidate sensitive sensor list to the sensor receiving signal data set that deviates from the baseline.

[0038] In this embodiment, the calculation steps of the Higuchi fractal dimension value are as follows: For the time series corresponding to the non-damaged state data set and the damaged state data set Perform segment processing and set the maximum number of segments , divide the time series into part.

[0039] Calculate the curve length of each segment , the formula is: ;in, is the number of segments, and N is the number of time series corresponding to the non-damaged state data set and the damaged state data set. and Perform linear fitting, and the slope is the Higuchi fractal dimension value.

[0040] Specifically, S3 includes the following steps: S31, perform time-frequency localization analysis on the sensor receiving signal data set that deviates from the reference by continuous wavelet transform to obtain a time-frequency feature data set, wherein the energy distribution of the time-frequency plane in the time-frequency localization analysis process is The mathematical expression is: ,in, represents the scale parameter, represents the translation parameter, represents the wavelet basis function, represents the time domain signal, represents the time variable; S32, input the time-frequency feature data set into the Markov transition field to perform state transfer analysis, and construct a state transfer probability matrix, where the state transfer probability The mathematical expression is: , represents the number of transitions from state i to state j, represents the number of transitions from state i to state k, represents the smoothing factor, k represents the number of states; S33. Obtain signal data for a valid time period by analyzing the signal evolution law of the state transition probability matrix and screening the characteristic time period.

[0041] Specifically, S4 includes the following steps: S41. Standardizing the effective time period signal data based on a functional principal component analysis strategy to obtain standardized effective time period signal data, and constructing a mean function based on the standardized effective time period signal data; S42, performing centralization processing on the normalized effective time period signal data according to the mean function to obtain the centralized effective time period signal data; S43, performing basis function expansion on the effective time period signal data after centralization processing according to the B-spline basis function to obtain a basis function matrix, and constructing a projection matrix according to the basis function matrix; S44, projecting the signal data of the valid time period after the centralization processing according to the projection matrix to obtain projected signal data; S45, performing covariance calculation on the projected signal data to obtain a covariance matrix, and processing the covariance matrix according to an L2 regularization method to obtain a regularized covariance matrix; S46. Perform Cholesky decomposition on the regularized covariance matrix to obtain an intermediate result including basis coefficients, and perform damage feature screening on the projected signal data based on the intermediate result including basis coefficients to obtain a damage feature screening result.

[0042] In this embodiment, the mathematical basis of the functional principal component analysis strategy is the spectral decomposition of the covariance operator. Define the effective time period signal data as a function is a square-integrable random process, and the noise term Independent and identically distributed, expressed by basis function expansion: ,in, is the mean function, For the principal component basis functions (orthogonal), For the The signal data of the valid time period is in The scores on the principal components, K is the number of principal components.

[0043] Covariance operator Defined as: ,in, is the mathematical expectation operator, is a random process At the point A random variable at is a random process At the point The mean value at is a random process At the point A random variable at is a random process At the point The mean at , spectral decomposition is performed by Mercer theorem: ,in, is the eigenvalue, represents the covariance operator The k-th eigenfunction of , the principal component score is calculated by numerical integration: ,in, is the time interval, is the observation value of the i-th sample at time t, and the score variance is the eigenvalue .

[0044] In this embodiment, the B-spline basis function is used. It has local support and smoothness, can better capture local mutations caused by damage, and is suitable for processing non-periodic and smoothly changing functional data. The mathematical expression of the B-spline basis function is: ,in, is the B-spline basis function, is the B-spline basis function The coefficient of , combined with L2 regularization, can eliminate noise interference and enhance the robustness of the data.

[0045] The objective function F of L2 regularization is: ,in For parameter set Optimize, is the original data matrix, is the basis function matrix, is the principal component score matrix, is the square of the Frobenius norm, is the coefficient of regularization strength. Through Cholesky decomposition, the covariance matrix C is decomposed into the product of the lower triangular matrix L and its transpose, and combined with numerical integration to efficiently calculate the signal energy distribution and statistical characteristics, further improving the computational efficiency. The mathematical expression of the covariance matrix C is: .

[0046] Assuming that the cumulative variance explanation rate of the first two principal components reaches 95%, it means that they have captured most of the information in the data, and the remaining principal components are usually considered to be noise or redundant information. In order to comprehensively consider the information of the two principal components, the Euclidean distance is used to calculate the comprehensive score, which reflects the first The comprehensive change intensity of the signal data in the two principal component directions in the effective time period , ,in, and Respectively The score of the signal data in a valid time period on the first principal component and the score on the second principal component.

[0047] Based on the absolute value of the comprehensive score, 5% of the significant time points are screened as the data set for subsequent experiments. The specific screening logic is as follows: Given a time series, the joint score is recorded as ,in Indicates the The scores of each time point are first calculated as the threshold value of the 95th percentile of the joint score (i.e., 95% quantile). , ,in, is the quantile function, then the index set of significant time points It can be expressed as: ,in, Contains all satisfaction Time point index The indices of significant time points can be obtained, i.e., those time points where the joint score is greater than the 95th percentile. .

[0048] Specifically, S5 includes the following steps: S51. Initializing the DSCN structure according to the progressive architecture construction method, wherein the network initialization includes setting the DSCN structure to a single hidden layer and setting the hidden layer to a single node, and all hidden layer nodes are fully connected to the output layer of the DSCN structure; S52. Constructing training samples based on the damage feature screening results, and training the DSCN structure based on the training samples. Dynamically adjusting node parameters through a supervised learning mechanism to obtain optimized node parameters and corresponding model performance indicators. S53, based on the optimized node parameters and the corresponding model performance indicators, gradually increase the number of hidden layers and nodes along the depth direction and the width direction to obtain an expanded DSCN structure; S54. Perform output weight analysis on the expanded DSCN structure, calculate the connection weights between all hidden layer nodes and the output layer using the least squares method, and establish an initial CFRP wing skin damage location prediction model.

[0049] In this embodiment, the algorithm description of the DSCN structure can be summarized as follows: The input features of dimensions are expressed as , and has The corresponding labels of the dimensions are Here, Indicates the number of training points. Assume that in the hidden layer of DSCN Generated in nodes, the current output can be defined as . Therefore, the residual vector can be calculated using the equation : , among which, when hour, is the input feature of the training data, and when hour, , Represents the hidden layer Output.

[0050] From and Initialize candidate weights and biases from a uniform distribution on And maximize the objective function The input weight and deviation As a node Parameters, where: ; ;in, Representation node The output, , .

[0051] Output weight vector It can be calculated based on the least squares method as the equation: ;in, , , represents the Moore-Penrose inverse of H.

[0052] Specifically, S6 includes the following steps: S61. Determine the input feature dimension and hyperparameter set of the initial CFRP wing skin damage location prediction model based on the Bayesian algorithm; S62. Determine the test error between the predicted coordinates and the true coordinates of the test set samples based on the input feature dimension and the hyperparameter set, and construct a target optimization function with the goal of minimizing the test error; S63. Performing Bayesian probability modeling analysis on the target optimization function through prior distribution and mixed kernel function to construct a Gaussian process model; S64. Constructing a time-weighted expectation improved acquisition function based on the predicted mean, standard deviation, and current optimal observation value of the Gaussian process model, and selecting a hyperparameter set for the next evaluation point through the time-weighted expectation improved acquisition function; S65. Update the Gaussian process model based on the hyperparameter set of the next evaluation point, combined with the model prediction value and the actual test error, until the preset convergence condition is met, and output the optimal hyperparameter combination; S66. Input the optimal hyperparameter combination into the initial CFRP wing skin damage location prediction model to perform hyperparameter optimization to obtain the final CFRP wing skin damage location prediction model.

[0053] In this implementation, the initial CFRP wing skin damage location prediction model is assumed to be ,in: is the input feature, is the parameter combination to be optimized, and its space is defined as: ; The optimization goal is: ;in, is the optimal parameter combination, is the test error, where The mathematical expression is: ;in, Is DSCN to Sample No. The predicted value of the dimensional coordinate, is the corresponding real coordinate value.

[0054] Gaussian processes model the target function, providing predictions and uncertainty estimates for any parameter point. Characteristics of Gaussian processes: They are nonparametric models that adapt to arbitrarily complex function forms; their output is a probability distribution with a mean and variance; and they capture correlations between parameters using a covariance kernel.

[0055] The mathematical expression of the prior distribution is: ;in, is the objective function, is a Gaussian process, is the mean function, is the covariance function, is the observation noise variance, is the identity matrix.

[0056] The hybrid kernel function is designed as: ; Among them, the continuous parameter kernel : ; Integer parameter kernel : ; Discrete parameter kernel : The design of the time-weighted expectation improvement acquisition function is based on the prediction results of the Gaussian process model, defining the strategy for selecting the next evaluation point and balancing exploration and development. The mathematical expression is: ,in, , is the current optimal observation value, for The mean and standard deviation of the forecast; To estimate the training time cost.

[0057] Specifically, S7 includes the following steps: S71, performing multimodal signal acquisition on the CFRP wing skin to be tested to obtain a sample to be tested, and preprocessing the sample to be tested to obtain a preprocessed sample to be tested; S72. Input the preprocessed test sample into the final CFRP wing skin damage location prediction model to perform damage location prediction, obtain a spatial comparison map containing the real three-dimensional coordinates of the damage and the predicted three-dimensional coordinates, and perform spatial constraint analysis on the spatial comparison map to obtain the corresponding CFRP wing skin damage location result.

[0058] See also Figure 2 As shown in FIG, it is a schematic structural diagram of a CFRP wing skin damage localization system based on multimodal signal processing and Bayesian optimization DSCN according to an embodiment of the present invention, including: a sensor received signal data set acquisition module, for establishing a wing skin undamaged model in a numerical simulation experiment of the CFRP wing skin, setting a CFRP wing skin damage position at multiple spatial coordinates of the wing skin undamaged model, acquiring corresponding Lamb wave time domain signal data according to the CFRP wing skin damage position, and obtaining a corresponding sensor received signal data set according to the Lamb wave time domain signal data; A sensor signal screening module is used to perform permutation entropy analysis and Higuchi fractal dimension analysis on the sensor receiving signal data set according to a preset dual-index sensor signal screening strategy to obtain the sensor receiving signal data set that deviates from the benchmark; The time-frequency feature extraction module is used to extract the time-frequency features of the sensor received signal data set that deviates from the reference and obtain the signal data of the valid time period; The damage feature screening module is used to screen the damage features of the signal data in the effective time period according to the functional principal component analysis strategy to obtain the damage feature screening results; The model training module is used to train the DSCN structure based on the damage feature screening results to obtain the initial CFRP wing skin damage location prediction model; A model optimization module is used to optimize the hyperparameters of the initial CFRP wing skin damage location prediction model based on the Bayesian algorithm to obtain the final CFRP wing skin damage location prediction model; The damage location prediction module is used to perform damage location prediction on the test sample according to the final CFRP wing skin damage location prediction model to obtain the corresponding CFRP wing skin damage location result.

[0059] Based on the above experimental analysis: First, the finite element simulation software is used to establish a damage-free model of the wing skin. The model size is ,Width , skin thickness The curvature radii are 、 、 、 The model is constructed using a single-layer modeling method, and a six-layer structure is achieved through assembly operations, with each layer having a uniform thickness of The mesh size is set to 1 mm.

[0060] Internal crack damage was set at 45 different locations on the upper half of the wing skin surface. The coordinates of the crack damage center can be seen from Table 1. The size of the crack damage at each location is 25 mm long, 4 mm wide, and 3 layers thick (3rd, 4th, and 5th layers).

[0061] Table 1 Center coordinates of different crack damage locations

[0062] Through the simulation experiment with finite element simulation software, the Lamb wave time domain signal data of 26 different receiving sensors with no damage and 45 different damage positions can be obtained. For the Lamb wave time domain signal data obtained by the simulation experiment with finite element simulation software. If the time domain signal data of all channels are directly used as the model input, the pressure of model data analysis and processing will increase, resulting in the model being unable to effectively improve the learning ability and generalization performance. Therefore, effective preprocessing of the data before the Lamb wave data is input into the model can improve the learning efficiency of the model and reduce the burden of model fitting. The signal data preprocessing of this application is divided into: permutation entropy and Higuchi fractal dimension sensor sensitivity screening, continuous wavelet and Markov transition field time-frequency feature extraction, and functional principal component analysis damage feature compression. The multimodal signal processing experiment is designed as follows.

[0063] Permutation entropy and Higuchi fractal dimension sensor sensitivity screening: The experimental data set includes Lamb time domain data of 1 undamaged and 45 damaged locations, and each damaged location corresponds to the monitoring data of 26 receiving sensors. By setting the embedding dimension , Higuchi fractal dimension parameter and statistical significance thresholds , and calculated the PE and HFD values ​​of the undamaged and damaged datasets, respectively. To screen common sensitive sensors, the minimum number of sensitive cases was set to 35, that is, sensors that showed significant changes in at least 35 damage cases were identified as common sensitive sensors.

[0064] Continuous wavelet and Markov transition field time-frequency feature extraction: The experimental data set includes the monitoring data of sensors 6, 22, 24, 25, and 26 in 45 damage locations selected by the combination of permutation entropy and Higuchi fractal dimension. The key parameters are set as follows: the sampling frequency is 100kHz, the scale range of continuous wavelet transform analysis is [1,128], the state score bin of Markov transition field is 6, the characteristic scale is [32,64], and the smoothing factor is . Functional principal component analysis damage feature compression: The experimental data is the data of the main distribution time range of the features after continuous wavelet and Markov transfer field, that is, in the time range of 3~5×10 -4 s, and the number of sampling points ranges from 3000 to 5000. There are 45 different damage locations involved, and each location has data from 5 receiving sensors. Therefore, the total data set is 2000 rows × 225 columns, that is, 2000 time points and 225 samples.

[0065] The damage localization experiment dataset is preprocessed with multimodal signal data. The total dataset contains data from 45 different damage locations, each consisting of 5 samples. This translates to 225 samples, with each sample sequence length being 100. The dataset is divided into two parts: one randomly selected damage location serves as the test set, and the other 44 locations serve as the training set.

[0066] In this embodiment, the present application is defined as a BODSCN model. To demonstrate the damage localization performance of the proposed BODSCN model, ELM (Extreme Learning Machine), RVFLN (Random Vector Functional-Link Network), and DSCN (Dynamic Stochastic Configuration Network) are selected as comparison algorithms. ELM and RVFLN use one hidden layer, while the hidden layer of DSCN is Select from each layer of nodes from Select from DSCN and BODSCN Set to 20, ; In order to avoid overfitting of random neural networks, L2 regularization is used to calculate the output weights. , BODSCN is in , , , Perform combinatorial optimization in .

[0067] This application uses positioning error ( ), 3D RMS error ( ), 3D mean absolute error ( ) and the three-dimensional coefficient of determination ( )Evaluate the performance and positioning capabilities of different models. Essentially, it is the space vector difference between the measured value and the true value. and The spatial prediction deviation of the model is quantified by the sum of squared errors and the average of absolute errors, More sensitive to larger errors, It directly reflects the average distance deviation and more intuitively reflects the overall accuracy of the model. and The smaller the value, the stronger the positioning ability of the model. Measures the model's ability to explain the joint variation of the true coordinates. The closer the value is to 1, the better the spatial fit between the predicted value and the true value.

[0068] Table 2 Damage coordinate prediction results of different models

[0069] Table 3 Comparison of evaluation index results of different models

[0070] As shown in Tables 2 and 3, the BODSCN model shows the best overall performance. The positioning error of this model for all coordinates is significantly lower than that of other models, with RMSE=4.14 and MAE=3.78 being the lowest values ​​among the four groups of methods. =0.9910 is close to the ideal value of 1, indicating that it has the best prediction accuracy and stability. In contrast, the error indicators of other models are significantly higher. These data fully demonstrate the superior performance of the BODSCN model in spatial positioning tasks.

[0071] like Figure 3As shown in the figure, the true coordinates and the predicted coordinates of the four methods clearly demonstrate the performance differences among the prediction models. The coordinate system covers the X-axis (0-140mm), Y-axis (0-10mm), and Z-axis (260-350mm) ranges. The red circles in the figure mark the true coordinate points, the yellow triangles represent the ELM predicted coordinates, the blue squares correspond to the RVFLN predicted coordinates, the cyan diamonds show the DSCN predicted coordinates, and the green stars represent the BODSCN predicted coordinates. The distribution characteristics in three-dimensional space clearly show that the RVFLN blue square predicted points are the most dispersed, with several samples exhibiting significant deviations exceeding 10mm, especially in the Z-axis. The ELM yellow triangle predicted points show improvement over the RVFLN, but still exhibit significant deviations. The DSCN cyan diamond predicted points show significantly improved prediction stability, but some areas still have errors of 3-5mm. In contrast, all BODSCN green star predicted points are tightly clustered around the true red coordinate points, maintaining minimal deviation in all three dimensions, with no significant outliers, validating the positioning accuracy advantage of BODSCN.

[0072] like Figure 4 The figure below shows the error distribution of each model on different test samples. The horizontal axis represents the sequence of coordinate points 1-16, and the vertical axis represents the positioning error value. From the overall trend, the BODSCN model marked with a green star has significantly lower positioning errors than the other three models at all test points, demonstrating better positioning accuracy. While the errors of the other models fluctuate significantly, the BODSCN model is more stable.

[0073] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme are not described in detail here. Ordinary technicians in the field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN is characterized by: The following steps are involved: S1. In a numerical simulation experiment of a CFRP wing skin, a damage-free model of the wing skin is established, and damage locations of the CFRP wing skin are set at multiple spatial coordinates of the damage-free model of the wing skin. Lamb wave time-domain signal data corresponding to the damage locations of the CFRP wing skin are obtained, and a corresponding sensor received signal data set is obtained based on the Lamb wave time-domain signal data; S2. Perform permutation entropy analysis and Higuchi fractal dimension analysis on the sensor receiving signal dataset based on a preset dual-index sensor signal screening strategy to obtain a sensor receiving signal dataset that deviates from the baseline; S3, extracting time-frequency features from the sensor received signal data set that deviates from the reference to obtain signal data for a valid time period; S4. Based on the functional principal component analysis strategy, damage characteristics are screened for the signal data in the effective time period to obtain damage characteristics screening results; S5. Train the DSCN structure based on the damage feature screening results to obtain the initial CFRP wing skin damage location prediction model; S6. Optimizing the hyperparameters of the initial CFRP wing skin damage location prediction model according to the Bayesian algorithm to obtain the final CFRP wing skin damage location prediction model; S7. Perform damage location prediction on the test sample according to the final CFRP wing skin damage location prediction model to obtain the corresponding CFRP wing skin damage location result.

2. The CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to claim 1 is characterized in that: Said S1 comprises the following steps: S11. Establishing an undamaged model of the wing skin based on finite element simulation software, and determining multiple internal crack damage locations in the undamaged model of the wing skin based on preset crack damage rules, thereby forming a CFRP wing skin simulation structure including an undamaged state and different damaged states, wherein the preset crack damage rules include presetting internal crack damage locations in high stress areas of the CFRP wing skin and setting them according to typical damage modes of the CFRP wing skin; S12. Simulate the excitation and propagation process of Lamb waves in the CFRP wing skin simulation structure to obtain the original time domain signal data set; S13. Locate each preset sensor receiving point from the original time domain signal dataset, and analyze the fluctuation response characteristics of each sensor receiving point in the time domain dimension to obtain a sensor receiving signal dataset, wherein the sensor receiving signal dataset includes an undamaged state dataset and a damaged state dataset.

3. The CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to claim 2 is characterized in that: The S2 comprises the following steps: S21. Perform permutation entropy analysis on the undamaged state data set and the damaged state data set according to the permutation entropy analysis algorithm in the preset dual-index sensor signal screening strategy, and obtain the permutation entropy value of each sensor in the undamaged state and the damaged state of the CFRP wing skin. , wherein the mathematical expression of the permutation entropy value is: , represents the probability of the occurrence of the i-th arrangement pattern, represents the embedding dimension, Represents the embedding dimension The factorial of S22. performing Higuchi fractal dimension analysis on the undamaged state data set and the damaged state data set according to the Higuchi fractal dimension algorithm in the preset dual-index sensor signal screening strategy, respectively, to obtain Higuchi fractal dimension values ​​of each sensor in the undamaged state and the damaged state of the CFRP wing skin; S23, integrating the permutation entropy value and the Higuchi fractal dimension value to obtain a dual-index feature matrix; S24. Setting a permutation entropy change benchmark and a Higuchi fractal dimension change benchmark for each sensor based on the undamaged state data set, and determining a corresponding permutation entropy change rate and a Higuchi fractal dimension change rate based on the permutation entropy value and Higuchi fractal dimension value of each sensor in the damaged state of the CFRP wing skin; S25. Jointly analyzing the dual-index feature matrix according to the permutation entropy change benchmark and the Higuchi fractal dimension change benchmark to obtain a list of candidate sensitive sensors that meet preset sensitive sensor conditions, wherein the preset sensitive sensor conditions include a permutation entropy change rate exceeding the permutation entropy change benchmark, or a Higuchi fractal dimension change rate exceeding the Higuchi fractal dimension change benchmark; S26: Add the sensor receiving signal corresponding to each candidate sensitive sensor in the candidate sensitive sensor list to the sensor receiving signal data set that deviates from the baseline.

4. The CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to claim 1 is characterized in that: The S3 includes the following steps: S31, perform time-frequency localization analysis on the sensor receiving signal data set that deviates from the reference by continuous wavelet transform to obtain a time-frequency feature data set, wherein the energy distribution of the time-frequency plane in the time-frequency localization analysis process is The mathematical expression is: ,in, represents the scale parameter, represents the translation parameter, represents the wavelet basis function, represents the time domain signal, represents the time variable; S32, input the time-frequency feature data set into the Markov transition field to perform state transfer analysis, and construct a state transfer probability matrix, where the state transfer probability The mathematical expression is: , represents the number of transitions from state i to state j, represents the number of transitions from state i to state k, represents the smoothing factor, k represents the number of states; S33. Obtain signal data for a valid time period by analyzing the signal evolution law of the state transition probability matrix and screening the characteristic time period.

5. The CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to claim 1 is characterized in that: The S4 comprises the following steps: S41. Standardizing the effective time period signal data based on a functional principal component analysis strategy to obtain standardized effective time period signal data, and constructing a mean function based on the standardized effective time period signal data; S42, performing centralization processing on the normalized effective time period signal data according to the mean function to obtain the centralized effective time period signal data; S43, performing basis function expansion on the effective time period signal data after centralization processing according to the B-spline basis function to obtain a basis function matrix, and constructing a projection matrix according to the basis function matrix; S44, projecting the signal data of the valid time period after the centralization processing according to the projection matrix to obtain projected signal data; S45, performing covariance calculation on the projected signal data to obtain a covariance matrix, and processing the covariance matrix according to an L2 regularization method to obtain a regularized covariance matrix; S46. Perform Cholesky decomposition on the regularized covariance matrix to obtain an intermediate result including basis coefficients, and perform damage feature screening on the projected signal data based on the intermediate result including basis coefficients to obtain a damage feature screening result.

6. The CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to claim 1 is characterized in that: The S5 comprises the following steps: S51. Initializing the DSCN structure according to the progressive architecture construction method, wherein the network initialization includes setting the DSCN structure to a single hidden layer and setting the hidden layer to a single node, and all hidden layer nodes are fully connected to the output layer of the DSCN structure; S52. Constructing training samples based on the damage feature screening results, and training the DSCN structure based on the training samples. Dynamically adjusting node parameters through a supervised learning mechanism to obtain optimized node parameters and corresponding model performance indicators. S53, based on the optimized node parameters and the corresponding model performance indicators, gradually increase the number of hidden layers and nodes along the depth direction and the width direction to obtain an expanded DSCN structure; S54. Perform output weight analysis on the expanded DSCN structure, calculate the connection weights between all hidden layer nodes and the output layer using the least squares method, and establish an initial CFRP wing skin damage location prediction model.

7. The CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to claim 1 is characterized in that: The S6 comprises the following steps: S61. Determine the input feature dimension and hyperparameter set of the initial CFRP wing skin damage location prediction model based on the Bayesian algorithm; S62. Determine the test error between the predicted coordinates and the true coordinates of the test set samples based on the input feature dimension and the hyperparameter set, and construct a target optimization function with the goal of minimizing the test error; S63. Performing Bayesian probability modeling analysis on the target optimization function through prior distribution and mixed kernel function to construct a Gaussian process model; S64. Constructing a time-weighted expectation improved acquisition function based on the predicted mean, standard deviation, and current optimal observation value of the Gaussian process model, and selecting a hyperparameter set for the next evaluation point through the time-weighted expectation improved acquisition function; S65. Update the Gaussian process model based on the hyperparameter set of the next evaluation point, combined with the model prediction value and the actual test error, until the preset convergence condition is met, and output the optimal hyperparameter combination; S66. Input the optimal hyperparameter combination into the initial CFRP wing skin damage location prediction model to perform hyperparameter optimization to obtain the final CFRP wing skin damage location prediction model.

8. The CFRP wing skin damage localization method based on multimodal signal processing and Bayesian optimization DSCN according to claim 1 is characterized in that: The S7 comprises the following steps: S71, performing multimodal signal acquisition on the CFRP wing skin to be tested to obtain a sample to be tested, and preprocessing the sample to be tested to obtain a preprocessed sample to be tested; S72. Input the preprocessed test sample into the final CFRP wing skin damage location prediction model to perform damage location prediction, obtain a spatial comparison map containing the real three-dimensional coordinates of the damage and the predicted three-dimensional coordinates, and perform spatial constraint analysis on the spatial comparison map to obtain the corresponding CFRP wing skin damage location result.

9. A CFRP wing skin damage localization system based on multimodal signal processing and Bayesian optimization DSCN is characterized by: include: a sensor received signal data set acquisition module, configured to establish a wing skin undamaged model in a numerical simulation experiment of the CFRP wing skin, set a CFRP wing skin damage location at multiple spatial coordinates of the wing skin undamaged model, acquire corresponding Lamb wave time domain signal data based on the CFRP wing skin damage location, and obtain a corresponding sensor received signal data set based on the Lamb wave time domain signal data; A sensor signal screening module is used to perform permutation entropy analysis and Higuchi fractal dimension analysis on the sensor receiving signal data set according to a preset dual-index sensor signal screening strategy to obtain the sensor receiving signal data set that deviates from the benchmark; The time-frequency feature extraction module is used to extract the time-frequency features of the sensor received signal data set that deviates from the reference and obtain the signal data of the valid time period; The damage feature screening module is used to screen the damage features of the signal data in the effective time period according to the functional principal component analysis strategy to obtain the damage feature screening results; The model training module is used to train the DSCN structure based on the damage feature screening results to obtain the initial CFRP wing skin damage location prediction model; A model optimization module is used to optimize the hyperparameters of the initial CFRP wing skin damage location prediction model based on the Bayesian algorithm to obtain the final CFRP wing skin damage location prediction model; The damage location prediction module is used to perform damage location prediction on the test sample according to the final CFRP wing skin damage location prediction model to obtain the corresponding CFRP wing skin damage location result.

Images