A method for inverting delamination damage profile of composite laminates based on lamb wave data

Abstract

This invention relates to a method for inverting the delamination damage profile of composite laminate structures based on Lamb wave data, belonging to the field of composite structure damage detection technology. The propagation behavior of Lamb waves in CFRP composite laminates with delamination damage was studied using numerical methods. The influence of delamination damage on Lamb wave characteristics was analyzed using wavelet packet decomposition and reconstruction methods and constructed eigenvalues. Through eigenvalue analysis and selection, effective features sensitive to the size and location of delamination damage were obtained. An XGBoost machine learning model was established based on these effective features. A Lamb wave feature sample set of laminates with random delamination damage was constructed using numerical methods, and the XGBoost model was trained to predict the size and location of delamination damage.

Description

Technical Field

[0001] This invention relates to a method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data, belonging to the field of composite structure damage detection technology. Background Technology

[0002] In existing technologies, carbon fiber reinforced resin (CFRP) composites possess excellent specific strength, specific stiffness, and good designability, making them a key material in aerospace, military equipment, medical, and automotive industries. However, due to the influence of the CFRP composite structure molding process, delamination damage is prone to occur under alternating loads and high-speed impacts. Delamination damage is somewhat concealed and difficult to detect through visual inspection on a checklist. Furthermore, delamination damage is characterized by rapid propagation and the potential for sudden structural failure. If it is not detected in time during the structure's service life, it may lead to catastrophic structural damage, seriously threatening the service safety and lifespan of composite structures.

[0003] Current methods for detecting delamination damage in carbon fiber reinforced composite structures mainly focus on predicting the extent and location of the damage. This means they can only capture the approximate area of ​​damage or assess its severity, lacking quantitative methods for retrieving delamination damage. Furthermore, the few methods used to predict the location or shape of delamination damage are affected by detection conditions, and in-situ detection is relatively difficult.

[0004] The existing technology still has the following main drawbacks:

[0005] The difficulty in quantitatively inverting delamination damage stems primarily from the complexity of the internal structure of composite materials and the high degree of uncertainty in damage morphology. Delamination typically occurs between layers, and its propagation path is influenced by a combination of fiber orientation, resin properties, and interface conditions, resulting in nonlinear and multimodal characteristics in damage shape, size, and depth. While existing detection methods (such as ultrasound, X-rays, and acoustic emission) can identify the presence of damage and roughly assess its severity, they struggle to accurately reconstruct the three-dimensional geometric parameters of delamination (such as edge morphology, thickness distribution, and degree of interface peeling). Furthermore, signal response is often affected by material anisotropy, noise interference, and the coupling effects of multiple damages, making it difficult to establish a high-fidelity quantitative damage model from limited data, which lacks sufficient theoretical support and computational stability, thus hindering the establishment of a precise quantitative evaluation system.

[0006] In-situ detection of delamination damage faces numerous practical constraints, particularly the contradiction between the complexity of the structural service environment and the feasibility of detection methods. Most high-precision detection technologies (such as industrial CT and transmission ultrasound) rely on off-site or contact measurements, making them difficult to apply directly in assembled components or confined spaces. While methods suitable for in-situ monitoring, such as acoustic emission and fiber optic sensing, can achieve dynamic response acquisition, they are susceptible to mechanical vibration, temperature fluctuations, and electromagnetic interference, resulting in low signal-to-noise ratios. Furthermore, composite material structures are often in enclosed or load-bearing states, limiting sensor placement space and signal accessibility, leading to incomplete damage information. These factors combined make in-situ detection not only costly and cumbersome, but also difficult to balance real-time performance, accuracy, and reliability, hindering accurate on-site diagnosis and long-term health management of delamination damage.

[0007] Therefore, there is an urgent need for a method that can achieve in-situ detection and quantitative inversion of delamination damage in order to accurately assess the damage state of carbon fiber reinforced resin matrix composite structures and provide support for the safety and functional reliability evaluation of composite structures with long service life. Summary of the Invention

[0008] The core technical problem to be solved by this invention is to overcome the shortcomings of the prior art in terms of inaccurate quantitative parameters for delamination damage and difficulty in in-situ detection, and to provide a method for in-situ quantitative inversion of delamination damage in carbon fiber reinforced resin matrix composite laminated structures.

[0009] A method for detecting delamination damage profiles in carbon fiber reinforced resin (CFRP) composite laminates based on Lamb wave properties was developed. The propagation behavior of Lamb waves in CFRP composite laminates with delamination damage was investigated using numerical methods. The influence of delamination damage on Lamb wave characteristics was analyzed using wavelet packet decomposition and reconstruction methods and constructed eigenvalues. Through eigenvalue analysis and selection, effective features sensitive to the size and location of delamination damage were obtained. An XGBoost machine learning model was established based on these effective features. A Lamb wave feature sample set of laminates with random delamination damage was constructed using numerical methods, and the XGBoost model was trained to predict the size and location of delamination damage.

[0010] The technical solution of this invention is: A method for inverting delamination damage profiles in composite laminate structures based on Lamb wave data, the method comprising the following steps: Step 1: Establish a numerical simulation model of the composite laminate. Step 2: Implant layered damage of the same size but different center positions into the numerical simulation model established in Step 1, calculate the Lamb signal under different layered damage states, and filter effective feature values ​​based on all the obtained Lamb signals to obtain the center position sensitive effective feature value filtering results. The center position changes according to a fixed step size gradient; The numerical simulation model established in step one is implanted with layered damage of different sizes but with the same center position. Lamb signals under different layered damage states are calculated. Effective feature values ​​are screened based on all the obtained Lamb signals to obtain the size-sensitive effective feature value screening results. The dimensions change according to a fixed step size gradient; Step 3: Implant random shape layered damage into the numerical simulation model established in Step 1, calculate the Lamb signal under different layered damage states, perform wavelet packet decomposition and reconstruction on the Lamb signal to obtain the signal features in the time domain, frequency domain, and time-frequency domain, and then calculate the signal features. According to the center position sensitive effective feature value screening result, extract effective feature value A as sample set A, and according to the size sensitive effective feature value screening result, extract effective feature value B as sample set B. Step 4: Build the XGBoost machine learning model; Step 5: Use the sample set A obtained in Step 3 to train the XGBoost machine learning model established in Step 4, and obtain the trained XGBoost machine learning model A. The XGBoost machine learning model established in step four is trained using the sample set B obtained in step three to obtain the trained XGBoost machine learning model B. Step 6: Quantitatively invert the delamination damage of the composite laminate under test using the trained XGBoost machine learning model A and XGBoost machine learning model B to obtain the delamination damage profile.

[0011] In step one, when establishing the numerical simulation model of the composite laminate, it is required that the calculated Lamb waveform is the same as the experimentally tested Lamb waveform under non-destructive conditions.

[0012] In step two, the method for obtaining the center position sensitive effective feature value screening results is as follows: Step 211: Perform wavelet packet decomposition and reconstruction on the Lamb signal to obtain the signal features in the time domain, frequency domain, and time-frequency domain. Step 212: Calculate the feature values ​​of the signal features obtained in step 211 to obtain several feature values; Step 213: Filter the feature values ​​obtained in step 212 based on whether there is at most one inflection point in the change of the center position. If there is at most one inflection point, the feature value is used as the result of the center position sensitive effective feature value filtering; if there are more than two inflection points, the feature value is removed. In step two, the method for obtaining the size-sensitive effective feature value screening results is as follows: Step 221: Perform wavelet packet decomposition and reconstruction on the Lamb signal to obtain the signal features in the time domain, frequency domain, and time-frequency domain. Step 222: Calculate the feature values ​​of the signal features obtained in step 221 to obtain several feature values; Step 223: Filter the feature values ​​obtained in step 222 based on whether there is at most one inflection point in the size change. If there is at most one inflection point, the feature value is used as the result of the size-sensitive valid feature value screening; if there are more than two inflection points, the feature value is removed. In step five, RobustScaler standardization, PowerTransformer transformation, and dual feature selection (F regression + mutual information regression) feature engineering are used during training to further filter and standardize the pre-screened dataset; In the first phase of training, 5-fold cross-validation combined with root mean square error (RMSE) is used to determine the range of hyperparameter selection. After determining the range of hyperparameter selection, incremental training combined with an early stopping mechanism is used to further train the training and validation sets. During training, if the training objective is reached or the early stopping mechanism is triggered, the initial training is considered complete, and the current model parameters are saved. Then, the validation set error is analyzed. RMSE represents the average deviation between predicted and true values, R² represents the model's ability to explain changes in the target variable, MAPE is used to evaluate outlier detection, and the maximum error is used to evaluate the model's worst predictive ability. If the error meets the standard, the training is considered complete, and the current model parameters are saved as the final model parameters. If the error is too large, error analysis is performed. First, the training level is analyzed to see if it is sufficient. If the number of iterations during training is less than 50, the early stopping mechanism is modified. Second, the number of features selected through feature engineering is checked. If the number of features does not exceed 80% of the total number of features, feature expansion is performed. Finally, the model structure is checked. If the model structure is adjusted more than 10 times... If the validation criteria are not met after each error analysis, the entire training process will be terminated. After each error analysis, the model will be retrained and the current model parameters will be updated for model validation. In step three, when implanting random-shaped layered damage, the size of the layered damage refers to the length of the damage along the detection path, while the location of the layered damage refers to the distance between the damage center and the excitation sensor. Sample set A contains 300 samples, with lesion sizes ranging from 5 mm to 60 mm and lesion locations ranging from 55 mm to 95 mm. The lesion locations are the training labels for sample set A. Sample set B contains 300 samples with lesion sizes ranging from 5 mm to 60 mm and lesion locations ranging from 55 mm to 95 mm. The lesion size is the training label for sample set B.

[0013] Beneficial effects 1. Numerical Simulation Method for Lamb Propagation in CFRP Composite Laminates with Delamination Damage: This method establishes a numerical model of a CFRP laminate containing random delamination damage to systematically simulate the propagation behavior of Lamb waves in the damaged structure. Key points include: employing an explicit dynamic analysis method, setting the mesh size and time step to satisfy the CFL stability condition; extracting signal characteristics caused by delamination damage by comparing waveform changes in healthy and damaged states; and introducing sinusoidal excitation modulated by a Hanning window. The focus is on the numerical model establishment process, boundary condition settings, excitation signal design, and damage implantation simulation methods, providing a reproducible and high-precision computational foundation for subsequent damage characteristic analysis.

[0014] 2. Lamb signal analysis method based on wavelet packet decomposition and reconstruction combined with eigenvalue construction: This method utilizes wavelet packet decomposition and reconstruction techniques to decompose the Lamb wave signal into multiple scales according to frequency bands, extracting time-domain, frequency-domain, and time-frequency-domain eigenvalues ​​at different frequency bands. Key points include: obtaining signal components from multiple sub-frequency bands through multi-level wavelet packet decomposition; constructing 28 types of feature parameters, including standard deviation, peak value, kurtosis, and energy percentage; analyzing the correlation between eigenvalues ​​and damage size and location, and screening out sensitive features. Its focus is on the signal processing flow of wavelet packet decomposition and reconstruction, the definition of multi-dimensional eigenvalues, and the feature screening mechanism based on damage sensitivity, providing highly discriminative feature inputs for quantitative damage inversion.

[0015] 3. XGBoost Machine Learning Model Based on Pre-selected Features: Building upon feature selection, an XGBoost machine learning model is established for the inversion of layered damage size and location. Key points include: constructing a pre-selected feature set through manual intervention and feature engineering (RobustScaler normalization, PowerTransformer transformation, and dual feature selection using F-regression and mutual information regression); optimizing hyperparameters using cross-validation and early stopping mechanisms to improve model generalization ability; training inversion models for damage size and location separately, and evaluating model performance using metrics such as root mean square error and R². The focus is on the XGBoost model construction process based on pre-selected features, feature engineering methods, model training and tuning strategies, and the dual-model structure design suitable for layered damage inversion.

[0016] 4. An automated method for establishing numerical models of CFRP laminates with random delamination damage: This method automates the modeling process for randomly introducing delamination damage of different shapes and locations along the detection path in CFRP laminates. Key points include: parametric control of damage size (5 60mm) and center position (55 The damage parameters are randomly distributed (95 mm); a damage dataset of 300 samples is constructed to ensure the diversity of damage morphology and spatial distribution; the model size, material parameters, boundary conditions, and sensor layout are kept consistent to study the damage impact as a single variable. The focus of this study is on the algorithm for generating random damage parameters, the program implementation for automated modeling, and the standardized sample set construction method for machine learning, providing a large-scale, high-quality data foundation for model training.

[0017] The method of the present invention monitors the Lamb wave signal using a PZT piezoelectric element and inverts the layered damage, thereby achieving in-situ and quantitative layered damage detection.

[0018] Analysis of existing technologies: Existing Lamb wave detection methods for delamination damage in composite materials are mostly focused on qualitative or coarse localization of damage, such as using single features like signal amplitude attenuation and time delay to determine the presence of damage or to assess the approximate area. Although these methods can achieve in-situ monitoring, they often struggle to quantitatively invert the size, shape, and precise location of damage, limiting their application in high-precision condition sensing such as digital twins.

[0019] Advantages of this invention: This invention achieves quantitative inversion of layered damage contours (size and location) through multi-feature fusion and machine learning models, possessing in-situ, rapid, and high-precision detection capabilities. This method not only meets the real-time and accurate perception requirements of digital twin systems for the damage state of physical models, but also supports dynamic updates of damage models and prediction of residual strength, providing an effective tool for intelligent structural health management.

[0020] The method of this invention simulates the propagation behavior of Lamb waves in composite laminates using a validated numerical method, thereby improving the applicability and cost of delamination damage detection methods.

[0021] Analysis of existing technologies: Existing research largely relies on experimental methods to obtain damage signals, which are costly, time-consuming, and make it difficult to systematically study the influence of damage parameters (such as size and location) on the signals. In addition, some numerical simulation studies lack sufficient experimental verification.

[0022] Advantages of this invention: This invention establishes an experimentally validated numerical simulation method that can effectively simulate the propagation behavior of Lamb waves in laminated composite structures with delamination damage. This method significantly reduces research costs and time, providing a reliable and reproducible computational foundation for systematically analyzing the relationship between damage characteristics and signal response, and constructing large-scale, high-quality sample sets.

[0023] The method of this invention proposes a Lamb data analysis method that combines wavelet packet decomposition and reconstruction with eigenvalue construction, which broadens the feature space of Lamb signals (time domain, frequency domain, and time-frequency domain) and filters out effective features.

[0024] Analysis of existing technologies: Existing Lamb wave signal analysis mostly uses single-dimensional features in the time or frequency domain, such as signal amplitude, arrival time, and spectral peak value. The feature dimensions are limited, the ability to distinguish complex damage morphologies is insufficient, and it is easily affected by noise and material anisotropy.

[0025] Advantages of this invention: This invention decomposes the signal into multiple sub-frequency bands using wavelet packet decomposition and reconstruction techniques, and systematically constructs 28 multi-dimensional features in the time, frequency, and time domains, greatly expanding the feature space. By analyzing the correlation and inheritance evolution laws between features and damage parameters, features sensitive to damage size and location are effectively screened, laying a crucial foundation for subsequent quantitative inversion.

[0026] The method of this invention establishes an automated method for building numerical calculation models of CFRP composite materials with random shape delamination damage, thereby improving the efficiency of sample set acquisition.

[0027] Analysis of existing technologies: The application of existing machine learning methods in damage detection is often limited by the acquisition of high-quality, large-scale sample sets. Traditional methods generate samples through manual modeling or single damage patterns, which is inefficient and lacks sample diversity, making it difficult to fully cover the randomness and complexity of actual damage and affecting the model's generalization ability.

[0028] Advantages of this invention: This invention develops a parameterized and automated numerical modeling process that can quickly generate random, layered damage samples with different sizes, locations, and shapes. This method can construct large-scale, diverse, and standardized sample sets in a short time, greatly improving data acquisition efficiency and providing sufficient and high-quality data for machine learning model training, effectively enhancing the robustness and generalization performance of the models.

[0029] The method of this invention establishes a hierarchical damage inversion XGBoost machine learning model based on pre-screened features, which has good prediction accuracy in the case of small samples.

[0030] Analysis of existing technologies: Existing machine learning-based damage inversion methods often use raw or simply processed features directly for training, resulting in high feature redundancy and noise, leading to low model training efficiency and easy overfitting, especially when the sample size is limited, resulting in insufficient prediction accuracy and stability.

[0031] Advantages of this invention: Before model training, this invention performs in-depth processing on the pre-selected sensitive feature set through feature engineering (standardization, transformation, and dual feature selection), effectively reducing data noise and redundancy. The constructed XGBoost model is optimized by combining cross-validation and early stopping mechanisms, achieving accurate inversion of damage size and location even with a limited sample size (300 groups). Attached Figure Description

[0032] Figure 1 The geometric model and boundary conditions of the laminate; Figure 2 This is a schematic diagram of the wavelet packet decomposition and reconstruction process; Figure 3 This is the effective characteristic value of the damage size; Figure 4 This represents the effective feature value of the damage location; Figure 5 This is a sample set for a numerical model of layered damage with random shapes. Figure 6 The XGBoost machine learning model was constructed; Figure 7 For the training process and validation results; Figure 8 The results of the damage contour inversion are compared with the actual contour. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] Example A method for inverting delamination damage profiles in composite laminate structures based on Lamb wave data, the method comprising the following steps: Step 1: Based on the orthogonal fabric structure characteristics of carbon fiber reinforced resin matrix composites (CFRP), a geometric model of a laminate with dimensions of 450×450×0.9 mm was established, with a layup pattern of [0 / 45]. s The model includes excitation and receiving points, and applies a sinusoidal excitation load modulated with a 150 kHz, 5-cycle Hanning window. Numerical simulation is performed using explicit dynamic analysis, with an upper limit of 0.4 mm for element size and 1 × 10⁻⁶ for time step. -8 The boundary conditions include displacement constraints in the thickness direction and triaxial constraints at the corners. The model is as follows: Figure 1 As shown, the material parameters are shown in Table 1.

[0035] Table 1 Material Parameter Table

[0036] Step 2: By introducing delamination damage between the third and fourth layers of the laminate, the effects of damage size and location variations on Lamb wave propagation characteristics were investigated. For location sensitivity analysis, the damage size was kept consistent (circular damage with a diameter of 20 mm), and the damage center position (from 30 mm to 70 mm from the excitation sensor) was changed along the detection path with a fixed step size gradient, and the Lamb wave signal in each state was calculated. For size sensitivity analysis, the damage center position was kept constant, and the damage size (from 20 mm to 50 mm) was changed along the detection path with a fixed step size gradient, and the corresponding Lamb wave response was obtained. Wavelet packet decomposition and reconstruction methods were used ( Figure 2 Combining eigenvalue calculations (Table 2), a set of effective eigenvalues ​​for the damage location (such as C) is extracted and screened from the time domain, frequency domain, and time-frequency domain. t8 Such as exhibiting significant correlation within a specific frequency band, such as Figure 3 As shown), and the set of effective feature values ​​for damage size (such as C). t5 Such as exhibiting regular changes within a specific frequency band, such as Figure 4 (As shown).

[0037] Table 2 Eigenvalue Table

[0038] Step 3: Introduce layered damage with random shapes and locations into the numerical model to generate a damage dataset containing 300 samples, such as... Figure 5 As shown. The damage size ranges from 5 to 60 mm, and the location ranges from 55 to 95 mm. Wavelet packet decomposition and reconstruction are performed on the Lamb wave signal under each damage state to obtain time-domain, frequency-domain, and time-frequency-domain signals in 15 frequency bands, and then various feature values ​​are calculated. Based on the location-sensitive features selected in step two, sample set A is formed, and size-sensitive features form sample set B. Each sample uses a feature value vector as input and the corresponding damage size or location as a label.

[0039] Step 4: Construct an XGBoost machine learning model for the inversion prediction of hierarchical damage size and location. The model framework includes three main parts: data preprocessing, feature engineering, and hyperparameter optimization, such as... Figure 6 As shown, the preprocessing stage performs RobustScaler standardization and PowerTransformer transformation on the features; feature engineering uses F-regression and mutual information regression for dual feature selection to eliminate redundant features; hyperparameter optimization is dynamically adjusted through five-fold cross-validation and early stopping mechanism, with the goal of minimizing the root mean square error (RMSE).

[0040] Step 5: Train the damage location inversion model using sample set A (XGBoost-A) and the damage size inversion model using sample set B (XGBoost-B). During training, the model rapidly reduces error in the initial iterations and stabilizes after approximately 20–30 iterations. The validation set error of the location inversion model is generally below 10%, and the error of the size inversion model is mostly below 20%. A few small-sized damages have slightly higher errors, but the absolute deviation is small. Figure 7 As shown. After training, the model parameters are saved, and the model performance is evaluated using error metrics (RMSE, R², MAPE) to ensure that it meets the inversion accuracy requirements.

[0041] Step Six: After wavelet packet decomposition and reconstruction of the Lamb wave detection signal of the CFRP laminate to be tested (inducing delamination damage by implanting a polytetrafluoroethylene film), the effective feature values ​​selected in Step Two are extracted and input into the trained XGBoost-A and XGBoost-B models respectively to predict the location and size of the delamination damage center along the detection path. By fusing the multi-path detection results, the two-dimensional contour of the delamination damage is reconstructed, as shown below. Figure 8 As shown, this method exhibits high accuracy (positional error 3.52%, dimensional error 6.87%), enabling rapid, in-situ, and quantitative inversion of delamination damage in composite materials.

[0042] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data, characterized in that... The steps of this method include: Step 1: Establish a numerical simulation model of the composite laminate. Step 2: Implant layered damage of the same size but different center positions into the numerical simulation model established in Step 1, calculate the Lamb signal under different layered damage states, and filter effective feature values ​​based on all the obtained Lamb signals to obtain the center position sensitive effective feature value filtering results. The center position changes according to a fixed step size gradient; The numerical simulation model established in step one is implanted with layered damage of different sizes but with the same center position. Lamb signals under different layered damage states are calculated. Effective feature values ​​are screened based on all the obtained Lamb signals to obtain the size-sensitive effective feature value screening results. The dimensions change according to a fixed step size gradient; Step 3: Implant random shape layered damage into the numerical simulation model established in Step 1, calculate the Lamb signal under different layered damage states, perform wavelet packet decomposition and reconstruction on the Lamb signal to obtain the signal features in the time domain, frequency domain, and time-frequency domain, and then calculate the signal features. According to the center position sensitive effective feature value screening result, extract effective feature value A as sample set A, and according to the size sensitive effective feature value screening result, extract effective feature value B as sample set B. Step 4: Build the XGBoost machine learning model; Step 5: Use the sample set A obtained in Step 3 to train the XGBoost machine learning model established in Step 4, and obtain the trained XGBoost machine learning model A. The XGBoost machine learning model established in step four is trained using the sample set B obtained in step three to obtain the trained XGBoost machine learning model B. Step 6: Quantitatively invert the delamination damage of the composite laminate under test using the trained XGBoost machine learning model A and XGBoost machine learning model B to obtain the delamination damage profile.

2. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step one, when establishing the numerical simulation model of the composite laminate, it is required that the calculated Lamb waveform is the same as the experimentally tested Lamb waveform under non-destructive conditions.

3. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step two, the method for obtaining the center position sensitive effective feature value screening results is as follows: Step 211: Perform wavelet packet decomposition and reconstruction on the Lamb signal to obtain the signal features in the time domain, frequency domain, and time-frequency domain. Step 212: Calculate the feature values ​​of the signal features obtained in step 211 to obtain several feature values; Step 213: The feature values ​​obtained in step 212 are filtered based on whether there is at most one inflection point in the change of the center position. When there is at most one inflection point, the feature value is used as the result of the center position sensitive effective feature value filtering; when there are more than two inflection points, the feature value is removed.

4. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step two, the method for obtaining the size-sensitive effective feature value screening results is as follows: Step 221: Perform wavelet packet decomposition and reconstruction on the Lamb signal to obtain the signal features in the time domain, frequency domain, and time-frequency domain. Step 222: Calculate the feature values ​​of the signal features obtained in step 221 to obtain several feature values; Step 223: The feature values ​​obtained in step 222 are filtered based on whether there is at most one inflection point in the size change. If there is at most one inflection point, the feature value is used as the result of the size-sensitive valid feature value filtering; if there are more than two inflection points, the feature value is removed.

5. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step three, when implanting randomly shaped layered damage, the size of the layered damage refers to the length of the damage along the detection path, while the location of the layered damage refers to the distance between the damage center and the excitation sensor.

6. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step three, sample set A contains 300 samples with damage sizes ranging from 5 mm to 60 mm and damage locations ranging from 55 mm to 95 mm. The damage locations are the training labels of sample set A.

7. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step three, sample set B contains 300 samples with damage sizes ranging from 5 mm to 60 mm and damage locations ranging from 55 mm to 95 mm. The damage size is the training label of sample set B.

8. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step five, RobustScaler standardization, PowerTransformer transformation, and dual feature selection feature engineering are used during training to further filter and standardize the pre-screened dataset.

9. The method for inverting the delamination damage profile of composite laminated structures based on Lamb wave data according to claim 1, characterized in that: In step five, during the first stage of training, the range of hyperparameter selection is determined by combining 5-fold cross-validation with root mean square error. After determining the range of hyperparameter selection, incremental training combined with an early stopping mechanism is used to further train the training and validation sets. During training, if the training objective is reached or the early stopping mechanism is triggered, the initial training is considered complete, and the current model parameters are saved. Then, the validation set error is analyzed. If the error meets the standard, the training is considered complete, and the current model parameters are saved as the final model parameters. If the error is too large, error analysis is performed. First, the training level is analyzed to see if it is sufficient. If the number of iterations during training is less than 50, the early stopping mechanism is modified. Second, the number of features selected through feature engineering is checked. If the number of features does not exceed 80% of the total number of features, feature expansion is performed. Finally, the model structure is checked. If the model structure is adjusted more than 10 times and still does not meet the validation standard, the entire training process is terminated. After each error analysis, the model will be retrained, and the current model parameters will be updated for model validation.

Images