Composite material performance degradation high-fidelity modeling method based on deep learning

Abstract

A composite material performance degradation high-fidelity modeling method based on deep learning relates to the field of composite materials, and comprises the following steps: defining a degradation index based on a degradation mechanism, and carrying out feature importance analysis on degradation data to obtain preprocessed degradation data; the method comprises the following steps of: establishing a LightGBM-LSTM (Lightweight Gradient LightGBM-Long Short Term Memory) model fusing a Lightweight Gradient LightGBM model and an LSTM network; training the LightGBM-LSTM model by using degradation data, and optimizing model hyper-parameters by using a genetic algorithm to obtain an optimal model; constructing a normal degradation data set and an unseen degradation data set; and applying the trained high-fidelity model to nonlinear composite material performance degradation data prediction. According to the method, feature items highly related to performance degradation can be screened out, and the influence of redundant features on model prediction accuracy is reduced; the feature extraction and generalization ability of high-dimensional, complex and nonlinear degradation data is remarkably improved, and the method is suitable for composite material performance degradation prediction scenes under variable structure parameters.

Description

Technical Field

[0001] This invention relates to the field of composite material performance degradation assessment technology, and more specifically, to a high-fidelity modeling method for composite material performance degradation based on deep learning. Background Technology

[0002] Composite materials, as important structural materials, have been widely used in numerous engineering fields such as aerospace, automotive, and wind power due to their superior mechanical properties. However, during their service life, composite materials are inevitably subjected to various external loads and environmental factors, leading to gradual damage at their internal structures or interfaces. As service time increases, their performance gradually degrades, potentially causing structural failure and resulting in safety accidents involving personnel and equipment. Therefore, high-fidelity degradation modeling and accurate prediction of the performance degradation process of composite materials are of great significance for assessing the health status of composite materials and enabling predictive maintenance.

[0003] Due to the complex and variable structural properties of composite materials, their degradation mechanism is influenced by a combination of factors such as fiber type, laminate thickness, and layup sequence. The degradation process typically exhibits high-dimensionality, nonlinearity, and dynamic evolution. Generally, the degradation process can be viewed as a gradual accumulation of damage rather than a sudden failure. Currently, accelerated degradation experiments are often used to obtain real dynamic response data under different operating conditions to reveal the degradation laws and processes of composite materials. However, this method is time-consuming, costly, and resource-intensive, making it difficult to cover composite materials with all structural parameters.

[0004] In practical applications, composite material performance prediction methods are mainly divided into two categories: physics-driven and data-driven. Physics-driven methods rely on theoretical mechanics models and, through analytical or finite element simulations, can quantitatively describe the performance evolution of composite materials. Currently, various degradation models, including classical mechanics, semi-empirical, and empirical models, have been proposed for predicting composite material performance degradation. Finite element simulation, as a primary auxiliary tool, can effectively reduce modeling and experimental complexity, improve adaptability to different composite material structures, and enhance performance prediction accuracy. However, the reliability and accuracy of methods based on classical physical models and finite element simulations are highly dependent on the setting of parameters such as material constitutive relations, boundary conditions, and failure criteria. These methods require extensive professional knowledge and prior conditions, and the simulation process involves a large amount of computation, making it difficult to meet the demand for rapid and accurate performance prediction under varying structural parameters.

[0005] To address the limitations of physics-driven methods, data-driven methods, as an end-to-end feature representation approach, have seen rapid development in various fields in recent years due to their powerful nonlinear learning capabilities. Data-driven methods do not require complex physical modeling processes or prior knowledge; they can automatically establish deep-level mapping relationships between inputs and outputs based on the analysis of existing data. Currently, data-driven methods have shown broad prospects in areas such as material performance prediction and remaining life assessment. Nevertheless, when applied to predicting the performance degradation of composite laminates, data-driven methods still face challenges such as scarce degradation data, strong structural heterogeneity, and high-dimensional nonlinearity of degradation data, leading to insufficient model prediction accuracy and poor generalization, especially for nonlinear degradation data of unknown structures. Furthermore, the selection of hyperparameters for prediction models often relies on manual experience, making parameter tuning cumbersome and difficult to achieve global optimization, thus affecting the model's prediction performance. Therefore, establishing an efficient, accurate, and highly generalizable high-fidelity prediction method for composite material performance degradation is a pressing technical challenge in this field. Summary of the Invention

[0006] The purpose of this invention is to solve the problem of high-fidelity prediction of high-dimensional nonlinear performance degradation data of composite materials mentioned in the background art, and to propose a high-fidelity modeling method for composite material performance degradation based on deep learning.

[0007] The technical solution adopted by this invention to solve its technical problem is: A deep learning-based high-fidelity modeling method for composite material performance degradation includes the following steps: S1. Data preparation steps: Define degradation indices based on degradation mechanisms, perform feature importance analysis on degradation data, and obtain preprocessed degradation data. S2. Model building steps: Build a LightGBM-LSTM model that integrates the Lightweight Gradient Boosting Machine (LightGBM) model and the LSTM network. S3. Model training and optimization steps: Train the LightGBM-LSTM model using degraded data, and optimize the model hyperparameters using a genetic algorithm to obtain the optimal model; S4. Dataset construction steps: Construct normal degradation dataset and no degradation dataset; S5. Prediction Step: The trained high-fidelity model is used to predict the performance degradation data of nonlinear composite materials.

[0008] Furthermore, the data preparation steps specifically include: We collected progressive degradation data of composite laminates from different researchers, including experimental data and finite element simulation data; Based on the degradation mechanism of composite laminates, potential structural mechanical parameters related to the degradation process are defined as input features; Multiple nonlinear feature selection methods are employed to quantitatively determine the importance of different feature parameters. Based on the feature importance metric, feature parameters that are highly correlated with the degradation index are selected as the input to the model.

[0009] Furthermore, the nonlinear feature selection method includes at least one of LASSO regression, ridge regression, decision tree, random forest, and gradient boosting.

[0010] Furthermore, in the model building step, the LightGBM-LSTM model uses a histogram algorithm and a leaf-shaped splitting strategy in its LightGBM part to segment and fit the data through a tree structure. The LSTM part captures long-term dependencies of different temporal features by introducing a long-term memory mechanism.

[0011] Furthermore, the model training and optimization steps specifically include: A genetic algorithm was used to jointly optimize the key hyperparameters of the LightGBM model, including the number of leaves, the maximum depth of the tree, the learning rate, and the number of decision trees. The feature vectors output by the trained LightGBM model are used as input to the LSTM network, and a genetic algorithm is used to optimize the hidden layer dimension and learning rate of the LSTM model.

[0012] Furthermore, in the dataset construction step, the normal degradation dataset includes a training dataset, a validation dataset, and a test dataset. The test dataset includes multiple sets of experimental data and limited metadata from different data sources. The unseen degradation dataset includes training datasets, validation datasets, and test datasets selected from different data sources, and the test dataset is partially or completely different from the test dataset in the normal degradation dataset.

[0013] Furthermore, in the prediction step, the trained high-fidelity model is tested on normal degradation datasets and unseen degradation datasets to verify the model's prediction performance and generalization ability.

[0014] Furthermore, it also includes an evaluation step, which uses at least one of the following indicators—root mean square error, coefficient of determination, mean square error, and mean absolute percentage error—to quantitatively evaluate the predictive performance of the model.

[0015] Furthermore, in the evaluation step, if the RMSE values ​​on both the normal degradation dataset and the unseen degradation dataset are less than a preset threshold, and R... 2 If all values ​​are greater than the preset threshold, the model is determined to have nonlinear prediction performance and generalization ability.

[0016] Furthermore, it includes a comparison step, which compares the proposed LightGBM-LSTM model with other classic models, including at least one of the LSTM model, GRU model, LightGBM model alone, and LightGBM-GRU model.

[0017] Compared with existing technologies, the beneficial effects of this invention are: this invention can screen out feature terms that are highly correlated with performance degradation, reducing the impact of redundant features on the model's prediction accuracy; the proposed fusion deep learning model combines the advantages of tree models and recurrent neural networks, significantly improving the feature extraction and generalization ability for high-dimensional, complex, and nonlinear degradation data; the introduction of genetic algorithms for hyperparameter optimization of the model can effectively save time and manpower costs in model optimization, while improving prediction accuracy; this method is applicable to composite material performance degradation prediction scenarios with variable structural parameters, and also has good prediction effects for unseen degradation data, exhibiting better nonlinear data prediction capabilities compared to traditional degradation models. Attached Figure Description

[0018] Figure 1 This is the overall flowchart of the present invention; Figure 2 The illustration shows the feature importance analysis results of five different feature selection methods of the present invention. Figure 3 This is a network structure diagram of the present invention; Figure 4 The illustration shows the prediction results of the proposed network on a normal degradation dataset. Figure 5 This diagram illustrates the difference between the predicted and actual values ​​for all training and testing data in this invention. Figure 6 The illustration shows the prediction results of the proposed network on a dataset without degraded data. Figure 7 This diagram illustrates the difference between the predicted and actual values ​​for all training and testing data in this invention. Figure 8 This diagram illustrates the difference in prediction results between the proposed deep learning network model and other network models. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention will be further described with reference to the accompanying drawings and embodiments: A high-fidelity modeling method for composite material performance degradation based on deep learning is characterized by the following steps: S1. Data preparation steps: Define degradation indices based on degradation mechanisms, perform feature importance analysis on degradation data, and obtain preprocessed degradation data. Furthermore, in step S1, the data preparation step specifically includes: S11. Collect progressive degradation data of composite laminates from different researchers, including experimental data and finite element simulation data; To better evaluate the nonlinear prediction performance of the proposed framework in experiments, 35 sets of data from different researchers were collected from 14 published works.[0] n / 90 m Progressive degradation data for orthogonal composite laminates. The degradation data primarily involve glass fiber reinforced polymer (GFPR) and carbon fiber reinforced polymer (CFPR) laminates with different stacking configurations and matrix crack densities. Specifically, 23 sets of degradation data were collected from experiments and 12 sets from finite element simulations. Each set represents a progressive degradation process for a composite laminate with a given structure and parameters. In this case, the structure and material parameters of the composite laminates are fixed; the variations in degradation data are due to changes in matrix crack density under load. Each set of degradation data contains a different number of samples. In this work, for the entire data source, 35 sets of degradation data contain a total of 356 samples. Of these, 269 samples are experimental results and 87 samples are finite element simulation results. For each set of data, even from the same study, the structure and distribution of the data may differ due to variations in material, layer thickness, crack density, and external testing conditions. Since composite laminates are typically elasto-plastic materials, the degradation process mainly consists of two distinct stages: a linear response stage before damage begins (such as matrix cracks) and a nonlinear response stage from the onset of damage to eventual failure.

[0020] S12. Based on the degradation mechanism of composite laminates, define the potential structural and mechanical parameters related to the degradation process as input features; To establish a degradation model describing the degradation index and input characteristics, based on the degradation mechanism of composite laminates, potential structural mechanical parameters related to the degradation process can be obtained as input characteristics. Specifically, for single-layer composite structures, stiffness can directly reflect the degradation performance of the structure and can be defined using the material's mechanical parameters. (1) in, and These are the axial modulus and transverse modulus of the structure, respectively. , It is Poisson's ratio. , It is the shear modulus. From the stiffness matrix in equation (1), it can be seen that the mechanical parameters related to the degradation process of composite materials include... , , and , can be used as input features for the model.

[0021] Since multilayer composite laminates are composed of stacked single-layer composite structures with different fiber angles, variations in material properties are also related to layer design information, such as layer angles and layer thicknesses. In this case, for [0 n / 90 m ] s Composite laminate, 0° layup thickness is The thickness of the 90° layup is These can serve as two additional input features related to the degradation index. Furthermore, when composite laminates are subjected to external loads, internal cracks appear and gradually propagate, affecting the composite's properties and even leading to complex damage patterns and structural failure. Therefore, the performance degradation of composite laminates is highly correlated with the number of structural cracks. To better clarify the impact of cracks, crack density... It can be defined as the number of cracks per unit length of structure.

[0022] Young's modulus, a classic mechanical parameter, is frequently used to evaluate the elastic deformation process of composite materials under load. Modulus degradation can be viewed as a decline in material performance. Typically, the modulus of a composite material in a certain direction can be defined as the ratio of stress to strain. When the degradation process of a composite material is in the linear response stage, the corresponding modulus is close to a constant value. As damage occurs and evolves, the degradation process of the composite material enters the nonlinear response stage, and the corresponding modulus gradually decreases. In fact, during this process, the composite structure may lose its ability to resist the corresponding load, and its performance gradually degrades; this can be called a progressive failure process in the nonlinear evolution mode. Specifically, in the nonlinear response stage of composite materials, potential internal damage, such as matrix cracking, delamination, and fiber breakage, can lead to changes in the material's constitutive relationship. In this case, the degradation modulus caused by damage can be described as the equivalent elastic modulus. Therefore, the equivalent elastic modulus can be used to define the performance degradation index of composite laminates, described as... and The ratio between .

[0023] (2) in, It is the equivalent elastic modulus. and These are the elastic modulus when undamaged and the elastic modulus during degradation, respectively.

[0024] Therefore, this study initially selected seven different parameters related to the degradation process for composite laminates with different structural parameters, including... , , , , , , , as input feature parameters of the degradation model, are named F0, F1, F2, F3, F4, F5 and F6 respectively. Meanwhile, [0 n / 90 m ] s The degradation index of composite laminates, i.e. As the output feature of the degradation model, it is named F7.

[0025] S13. Employ various nonlinear feature selection methods to quantitatively determine the importance of different feature parameters, and select feature parameters that are highly correlated with degradation indices as inputs to the model based on the feature importance metric.

[0026] Because there is a high-dimensional nonlinear relationship between the performance degradation and structural properties of composite materials, the degradation mechanism is comprehensively influenced by many factors such as fiber type, layer thickness, layup sequence, crack density, and mechanical parameters. Data-driven models aim to establish a mapping relationship between input feature parameters and composite material degradation indices. However, since different input feature parameters contribute differently to the performance degradation process of composite materials, theoretically, input feature parameters with smaller contributions can be called redundant features. These not only fail to improve the performance of data-driven models but also reduce the model's nonlinear feature learning ability and computational efficiency. Therefore, a nonlinear feature selection process can be employed before inputting features into the model to reduce the impact of redundant features and improve the model's predictive performance.

[0027] To further select the feature parameters most relevant to the degradation index, the nonlinear feature selection method includes at least one of LASSO regression, ridge regression, decision tree, random forest, and gradient boosting. The importance / contribution of different feature parameters is quantitatively determined, and feature parameters with good correlation to the degradation index are selected as input to the model based on the feature importance metric. Redundant feature parameters with low importance are removed from all feature parameters. Specifically, LASSO and ridge regression are two similar linear regression models used for feature selection; they can be combined with multinomial features to capture nonlinear relationships for handling nonlinear high-dimensional data. Decision trees select the optimal features and split points to minimize impurities, support nonlinear decision boundaries, and capture complex interactions through multi-level splitting. Random forests can reduce overfitting by randomly selecting features and retain strong predictive features, making them suitable for handling high-dimensional data. Gradient boosting can automatically discover higher-order feature interactions by progressively optimizing approximations of complex functions, resulting in high prediction accuracy. Notably, employing five different nonlinear feature selection methods avoids the limitations and randomness of a single method, further improving the reliability of the model's prediction results.

[0028] like Figure 2 The image shows the feature importance analysis results for five different feature selection methods. Due to the inherent high-dimensional nonlinear relationship of composite laminate properties, five nonlinear feature selection methods were used to analyze feature importance. These five methods are Lasso regression, ridge regression, decision tree, random forest, and gradient boosting. Figure 2 It is evident that feature F6 received the highest feature importance score, indicating that feature F6 has the best correlation with the degradation index. Conversely, feature F2 received the lowest feature importance score, indicating that feature F2 has the worst correlation with the degradation index. In this case, feature F2 can be considered a redundant feature in the performance degradation prediction process. Therefore, to eliminate the influence of redundant features, after feature selection on the degradation data, the input features of the degradation model can be defined as F0, F1, F3, F4, F5, and F6, with the output feature label F7.

[0029] S2. Model building steps: Build a high-fidelity composite material performance degradation prediction model (LightGBM-LSTM) that integrates Lightweight Gradient Boosting Machine (LightGBM) and Long Short-Term Memory Network (LSTM). Furthermore, in the model building step, the LightGBM-LSTM model employs a histogram algorithm and a leaf-like splitting strategy in its LightGBM part to segment and fit the data through a tree structure; the LSTM part captures the long-term dependencies of different temporal features by introducing a long-term memory mechanism.

[0030] LightGBM is an improved decision tree model that employs a histogram algorithm and a leaf-based splitting strategy. It uses a tree structure to segment and fit data, making it suitable for handling complex nonlinear data. The histogram algorithm discretizes features into different integers using a piecewise function, constructs a histogram, and aggregates statistics during traversal. Subsequent histogram analysis determines the optimal split point. LightGBM uses a leaf-by-leaf splitting strategy instead of a level-by-level splitting strategy, selecting only the node with the largest splitting gain for branching. This avoids the cost of nodes with smaller gains, thus achieving higher accuracy with lower computational cost. Therefore, the model's performance is related to the number of leaves selected, the maximum depth of the tree, and the number of decision trees. Its loss function can be expressed as: (3) in, This represents the difference between the predicted value and the actual value. This represents a regularization term used to prevent the model from overfitting.

[0031] LSTM is a classic recurrent neural network designed to address the vanishing and exploding gradient problems during model training. Unlike other neural networks, LSTM, due to its structural characteristics, is better suited for handling problems with complex multi-input features. For the performance degradation prediction of composite laminates in this work, the model input consists of multi-dimensional feature parameters, including material elastic parameters, layer thickness, crack density, etc. Furthermore, by introducing a long-term memory mechanism, LSTM can effectively capture long-term dependencies between different time features, such as periodicity and trends, thereby significantly improving the model's ability to represent nonlinear features of time-series data. Its loss function can be expressed as... (4) in, The total number of samples input to the LSTM model. It is to input the actual value. This is the predicted output value.

[0032] Therefore, combining the advantages of LightGBM and LSTM in processing time series, a high-fidelity composite material performance degradation prediction model, LightGBM-LSTM, was established to provide high-fidelity modeling of nonlinear degradation data of composite materials.

[0033] like Figure 3As shown, the proposed LightGBM-LSTM model architecture is illustrated. The proposed framework combines the advantages of LightGBM and LSTM, and utilizes the GA algorithm to optimize the hyperparameters. Initially, the framework is trained using experimental and finite element analysis (FEA) datasets to determine the optimal model parameters. After the training phase, the well-trained deep learning framework is tested on different degenerate datasets to verify its power and generalization ability in handling nonlinear degenerate data. Finally, comparative experiments further elucidate the superior nonlinear prediction performance of the proposed method, thus validating the effectiveness of the proposed framework.

[0034] S3. Model training and optimization steps: Train the LightGBM-LSTM model using degraded data, and optimize the model hyperparameters using a genetic algorithm (GA) to obtain the optimal model; Furthermore, in step S3, the model training and optimization steps specifically include: A genetic algorithm is used to jointly optimize the key hyperparameters of the LightGBM model, including the number of leaves, the maximum depth of the tree, the learning rate, and the number of decision trees. By minimizing the relevant loss function, the robustness and generalization ability of the model are improved, thereby obtaining the optimal parameter configuration of the LightGBM model and outputting the feature vector.

[0035] The feature vectors output by the trained LightGBM model are used as input to the LSTM network, and a genetic algorithm is used to optimize the hidden layer dimension and learning rate of the LSTM model.

[0036] By minimizing the LSTM loss function, the final network prediction output is obtained, achieving high-fidelity nonlinear modeling of the degradation process of composite laminates. The above training process is implemented on both standard degradation datasets and unseen datasets, effectively ensuring the prediction accuracy of the constructed deep learning model in high-dimensional, nonlinear degradation scenarios.

[0037] The trained and optimized LightGBM-LSTM-GA deep learning model was applied to different degradation datasets. The degradation metrics output by the model were used to accurately predict performance degradation trends and to quantitatively evaluate model performance. Thus, the network model training and hyperparameter optimization process was completed using the preprocessed dataset, resulting in the optimal predictive model parameters.

[0038] S4. Dataset construction steps: Construct normal degradation dataset and no degradation dataset; To verify the effectiveness and generalization of the proposed framework in predicting nonlinear degradation data, two different degradation datasets were constructed: a normal degradation dataset and an unseen degradation dataset. First, a normal degradation dataset was constructed based on all degradation data to verify the model's predictive performance on complex nonlinear degradation data of composite laminates with variable structural parameters. In this case, the normal degradation dataset includes a training dataset, a validation dataset, and a test dataset. The test dataset includes multiple sets of experimental data from different data sources and limited metadata. Specifically, four different test datasets were constructed from four sets of degradation data from different data sources, including three sets of experimental data and one set of limited metadata, named T1, T2, T3, and T4, as shown in Table 1. Then, the remaining degradation data were randomly divided into a training dataset and a validation dataset. The training dataset included 25 sets of data, and the validation dataset included 6 sets of data. Six relevant features F0, F1, F3, F4, F5, and F6 were used as input to the network, and the degradation index F7 was used as the network's output label.

[0039] Since degraded data from the same data source exhibit similar constitutive relationships and degradation trends, a dedicated unseen degradation dataset can be constructed to better validate the generalization and robustness of the proposed framework. Unlike the normal degradation dataset, the unseen degradation dataset is built based on partially degraded data. The unseen degradation dataset contains training, validation, and test datasets selected from different data sources. Specifically, to increase the difference between the training and test datasets, the test dataset differs partially or entirely from the test dataset in the normal degradation dataset; the training and test datasets are selected from different data sources. Specifically, the test dataset for the unseen degradation dataset is the same as the test dataset for the normal degradation dataset in Table 1. The training and validation datasets can be constructed from other remaining data sources, obtained by removing the entire data source corresponding to the four test datasets in Table 1 from the total data. It is worth noting that the constructed training dataset contains only 18 data sets, and the validation dataset contains only 5 data sets.

[0040]

[0041] S5. Prediction Step: The trained high-fidelity model is used to predict the performance degradation data of nonlinear composite materials. In this prediction step, the trained high-fidelity model is tested on both normal degradation datasets and unseen degradation datasets to verify the model's predictive performance and generalization ability.

[0042] The well-trained model was tested on a normal degradation dataset, including three sets of experimental data and a set of limited metadata. Figure 4As shown, real-world degradation data and corresponding prediction results are presented. To evaluate the performance of the proposed framework on normal datasets, LightGBM-LSTM-GA was trained and validated using normal datasets. The well-trained model was then tested on four different test datasets, including three sets of experimental data and a set of limited metadata. Figure 4 It is evident that the degradation process of composite laminates exhibits significant nonlinear characteristics, increasing the difficulty of performance degradation prediction. LightGBM-LSTM-GA can accurately predict the entire performance degradation process of composite laminates under different datasets, demonstrating its superior nonlinear prediction capabilities. In particular, compared with... Figure 4 Compared with the experimental data results in (a)-(c), Figure 4 The results in (d) with limited metadata show better predictive performance. Therefore, it can be concluded that the proposed framework demonstrates superior performance in predicting the nonlinear degradation of composite laminates with limited data, providing a new approach for establishing an accurate and reliable mapping between composite structures and composite degradation properties.

[0043] Next, to better evaluate the overall predictive performance of the model, we can obtain the differences between the predicted and actual values ​​of all samples in the training and test datasets, such as... Figure 5 As shown, to better evaluate the overall predictive performance of the model, the difference between the predicted and actual values ​​for all samples in both the training and test datasets can be obtained. Ideally, the predicted and actual values ​​should be identical. However, in reality, zero difference between predicted and actual values ​​is very difficult due to prediction errors and model uncertainties. Generally, the model's predictive performance is better when the predicted values ​​are close to the actual values. Figure 5 It can be clearly observed that the predicted values ​​on both the training and testing data are concentrated around the actual values. Especially for the testing data, only a few outliers deviate from the actual values; most predicted values ​​are close to the actual values, such as... Figure 5 As shown in (b). Therefore, the results demonstrate that the proposed framework exhibits superior performance in nonlinear degradation prediction.

[0044] In addition, to quantitatively evaluate the predictive performance of the model, an evaluation step is included, using root mean square error (RMSE) and coefficient of determination (R²). 2 The predictive performance of a model is quantitatively evaluated using at least one of the following metrics: mean squared error (MSE) and mean absolute percentage error (MAPE).

[0045] The definition is as follows: (5) (6) (7) (8) First, for the normal degradation dataset, four different metrics were tested on four test datasets, as shown in Table 2. It can be seen that the RMSE is less than 0.012 for each test dataset, indicating that R... 2 The accuracy reaches over 95%, MSE is close to 0, and MAPE is below 5%. Quantitative results show that the proposed framework achieves excellent nonlinear prediction performance on various test datasets. In particular, the prediction results on the T4 (simulation data) dataset are superior to those on other test datasets (experimental data). Figure 4 The results are the same.

[0046]

[0047] Subsequently, to evaluate the generalization ability of the proposed framework, predictions were made using a dataset without known degradation, such as... Figure 6 As shown, the actual degradation data and prediction results on unseen datasets are presented. First, to evaluate the generalization ability of the proposed framework to unseen data, LightGBM-LSTM-GA was trained using an unseen dataset. Then, the well-trained model was tested on four different test datasets, including three sets of experimental data and one set of limited metadata. It can be clearly observed that LightGBM-LSTM-GA performs well on unseen datasets from different data sources, even when degradation occurs. Figure 6 Even slight fluctuations in (c) can yield good predictive performance. In this case, LightGBM-LSTM-GA not only demonstrates superior generalization ability in nonlinear degradation prediction but also effectively captures the trend and magnitude of composite material property degradation. In particular, compared with... Figure 6 Compared with the experimental data results in (a)-(c), Figure 6 Better prediction performance can be obtained on the dataset in (d), which may be due to the data consistency in the FEA dataset. Therefore, it can be concluded that the proposed framework exhibits significant generalization ability in the prediction of nonlinear degradation of composite laminates, and provides a new and feasible approach for the accurate prediction of complex nonlinear degradation processes of composite materials in practical applications, even for limited unseen data.

[0048] Next, to better evaluate the overall predictive performance of the model, we can obtain the differences between the predicted and actual values ​​of all samples in the training and test datasets, such as... Figure 7As shown. Ideally, the predicted value should be the same as the actual value. However, in reality, due to prediction errors and model uncertainties, achieving zero difference between the predicted and actual values ​​is very difficult. Generally, the model's predictive performance is better when the predicted value is close to the actual value. Figure 7 As shown, to better evaluate the model's predictive performance on unseen data, the differences between the predicted and actual values ​​for all samples in both the training and test datasets can be obtained. It is clear that the predicted values ​​are clustered around the actual values. Especially for the test data, only a few predicted values ​​differ from the actual values; most predicted values ​​are close to the actual values, such as... Figure 7 As shown in (b), this further demonstrates that the proposed framework has superior performance in predicting nonlinear degradation in unseen data.

[0049] Furthermore, to quantitatively evaluate the model's predictive performance, four different metrics were tested on four test datasets, as shown in Table 3. It can be seen that the RMSE value for all test datasets was consistently below 0.027, while the R... 2 In each case, the MSE value is greater than 0.91. The MSE value is close to zero, and the MAPE is below 7.3%. These quantitative results validate the effectiveness of the proposed framework in predicting nonlinear degradation on unseen datasets. Notably, the model maintains robust performance not only on the T1-T3 (experimental data) dataset but also on the T4 (simulation data) dataset, where the T4 dataset yields particularly favorable results. Overall, these results reveal the generalization ability of the proposed framework, which will promote the application of data-driven methods in predicting the degradation of composite material properties in real-world scenarios.

[0050]

[0051] To evaluate the effectiveness and advantages of the proposed method, a comparison step is also included, in which the proposed LightGBM-LSTM model is compared with other classic models on different test datasets T2 and T4. The other classic models include at least one of LSTM model, GRU model, LightGBM model alone, and LightGBM-GRU model, to verify the effectiveness and superiority of the proposed model.

[0052] like Figure 8 As shown, the comparison results of different models are presented. To evaluate the effectiveness of the proposed method, four different data-driven models—LSTM, GRU, LightGBM, and LightGBM-GRU—were specifically compared on different test datasets T2 and T4. Figure 8(a) It can be clearly observed that the proposed LightGBM-LSTM-GA framework achieves the best prediction performance on the experimental dataset T2, even with a few non-optimal prediction points. In this case, the predictions of LightGBM-LSTM-GA are closer to the actual values ​​than those of the other four classic models. Similarly, on the finite element simulation dataset T4, the predictions of the LightGBM-LSTM-GA network are also closer to the actual values ​​than those of other models.

[0053] Furthermore, to quantitatively compare the prediction performance of the proposed method with other methods, four different metrics were evaluated on the degraded datasets T2 and T4, as shown in Tables 4 and 5. It can be seen that the prediction accuracy of LightGBM-LSTM-GA is significantly better than other models. Specifically, on the experimental test dataset T2, for RMSE, LightGBM-LSTM-GA showed the smallest reduction, decreasing by 4.90%, 6.74%, 19.77%, and 12.18% compared to LSTM, GRU, LightGBM, and LightGBM-GRU models, respectively. In R... 2 In terms of performance, LightGBM-LSTM outperforms LSTM, GRU, LightGBM, and LightGBM-GRU models by 0.5%, 0.72%, 2.70%, and 1.43%, respectively. Regarding MSE, LightGBM-LSTM-GA is lower than LSTM, GRU, LightGBM, and LightGBM-GRU models by 9.59%, 13.16%, 35.92%, and 23.26%, respectively. For MAPE, LightGBM-LSTM-GA is lower than LSTM, GRU, LightGBM, and LightGBM-GRU models by 2.03%, 2.09%, 80.00%, and 3.39%, respectively. Furthermore, similar results are obtained on the finite element simulation dataset T4. Specifically, compared to the LSTM model, the proposed framework R... 2 The index rose 3.35%, while RMSE, MSE, and MAPE fell by 35.58%, 58.46%, and 0.01%, respectively. Compared with the GRU model, R... 2 R increased by 2.88%, while RMSE, MSE, and MAPE decreased by 32.75%, 55.00%, and 0.01%, respectively. Compared with the LightGBM model, R... 2 Rise by 2.32%, while RMSE, MSE, and MAPE fall by 29.07%, 49.69%, and 16.33%, respectively. Compared to the LightGBM-GRU model, R... 2The index rose by 2.72%, while RMSE, MSE, and MAPE fell by 31.87%, 53.45%, and 0.020%, respectively. This comparison further demonstrates that the proposed LightGBM-LSTM model significantly outperforms other classic models in predicting the nonlinear degradation of composite laminates.

[0054]

[0055] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A high-fidelity modeling method for composite material performance degradation based on deep learning, characterized in that, Includes the following steps: S1. Data preparation steps: Define degradation indices based on degradation mechanisms, perform feature importance analysis on degradation data, and obtain preprocessed degradation data. S2. Model building steps: Build a LightGBM-LSTM model that integrates the Lightweight Gradient Boosting Machine (LightGBM) model and the LSTM network. S3. Model training and optimization steps: Train the LightGBM-LSTM model using degraded data, and optimize the model hyperparameters using a genetic algorithm to obtain the optimal model; S4. Dataset construction steps: Construct normal degradation dataset and no degradation dataset; S5. Prediction Step: The trained high-fidelity model is used to predict the performance degradation data of nonlinear composite materials.

2. The method according to claim 1, characterized in that, The data preparation steps specifically include: We collected progressive degradation data of composite laminates from different researchers, including experimental data and finite element simulation data; Based on the degradation mechanism of composite laminates, potential structural mechanical parameters related to the degradation process are defined as input features; Multiple nonlinear feature selection methods are employed to quantitatively determine the importance of different feature parameters. Based on the feature importance metric, feature parameters that are highly correlated with the degradation index are selected as the input to the model.

3. The method according to claim 2, characterized in that, The nonlinear feature selection method includes at least one of LASSO regression, ridge regression, decision tree, random forest, and gradient boosting.

4. The method according to claim 1, characterized in that, The LightGBM-LSTM model in the model building step uses a histogram algorithm and a leaf-shaped splitting strategy in its LightGBM part to segment and fit the data through a tree structure; The LSTM part captures long-term dependencies of different temporal features by introducing a long-term memory mechanism.

5. The method according to claim 1, characterized in that, The model training and optimization steps specifically include: A genetic algorithm was used to jointly optimize the key hyperparameters of the LightGBM model, including the number of leaves, the maximum depth of the tree, the learning rate, and the number of decision trees. The feature vectors output by the trained LightGBM model are used as input to the LSTM network, and a genetic algorithm is used to optimize the hidden layer dimension and learning rate of the LSTM model.

6. The method according to claim 1, characterized in that, In the dataset construction step, the normal degradation dataset includes a training dataset, a validation dataset, and a test dataset. The test dataset includes multiple sets of experimental data and limited metadata from different data sources. The unseen degradation dataset includes a training dataset, a validation dataset, and a test dataset selected from different data sources, and the test dataset is partially or completely different from the test dataset in the normal degradation dataset.

7. The method according to claim 6, characterized in that, In the prediction step, the trained high-fidelity model is tested on normal degradation datasets and undegraded datasets to verify the model's prediction performance and generalization ability.

8. The method according to any one of claims 1 to 7, characterized in that, It also includes an evaluation step, which uses at least one of the following indicators to quantitatively evaluate the predictive performance of the model: root mean square error, coefficient of determination, mean square error, and mean absolute percentage error.

9. The method according to claim 8, characterized in that, In the evaluation step, if the RMSE values ​​on both the normal degradation dataset and the undegradable dataset are less than a preset threshold, and R... 2 If all values ​​are greater than the preset threshold, the model is determined to have nonlinear prediction performance and generalization ability.

10. The method according to any one of claims 1 to 9, characterized in that, It also includes a comparison step, which compares the proposed LightGBM-LSTM model with other classic models, including at least one of the LSTM model, GRU model, LightGBM model alone, and LightGBM-GRU model.

Images