Composite structure residual life prediction method based on domain adaptation under digital-analog combined driving

Abstract

The application provides a composite structure residual life prediction method and system, solves the problem of low prediction accuracy caused by the difference between simulation and real data distribution in digital-analog driving, and proposes a digital-analog fusion domain adversarial network (DMF-DANN). The feature extractor fusion channel attention (SE) module and the Inception module, the former establishes channel dependency, and the latter aggregates multi-scale features to efficiently extract multi-channel long-time sequence data features. The domain adversarial module combines the implicit adversarial of the gradient reversal layer (GRL) and the explicit measurement of the maximum mean difference (MMD) to form a double distribution alignment mechanism, forcing the model to learn domain-invariant features and achieving more robust domain adaptation. Through the design, the application can effectively transfer the mechanism knowledge in the simulation model to the real data prediction task, and significantly improves the prediction accuracy and robustness of the composite material structure residual life.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, and in particular to a method, system, and computer-readable storage medium for predicting the remaining useful life (RUL) of composite material structures using deep learning and digital-analog fusion. Background Technology

[0002] Carbon fiber reinforced polymer (CFRP) composites, due to their excellent properties such as lightweight and high strength, have been widely used in the aerospace field, including key load-bearing components of modern aircraft such as the Boeing B787 and Airbus A350. Ensuring the safety and reliability of these composite structures, especially accurately predicting their remaining service life, is of paramount importance for developing reasonable maintenance plans and preventing catastrophic accidents.

[0003] With the development of monitoring technology, data-driven methods based on deep learning have shown great potential in RUL prediction. However, these methods typically rely on two key assumptions: first, the training and testing data must follow independent and identically distributed (i.i.d.) data; and second, a large amount of monitoring data covering the entire lifecycle of the structure is required. In practical engineering, obtaining full lifecycle data for high-reliability composite material structures is costly and time-consuming.

[0004] To address the problem of data scarcity, researchers began using physical mechanism-based simulation models to generate full-lifecycle simulation data, combining this data with a small amount of real monitoring data to form a prediction method driven by joint numerical simulation. However, this approach faces a core challenge: simulation models struggle to fully simulate the complex external conditions of real service environments, leading to significant differences in the feature distributions of simulation and real data. This distributional discrepancy violates the independent and identically distributed assumption, resulting in a significant decrease in accuracy if models trained on simulation data are directly used to predict the lifespan of real structures.

[0005] Domain adaptation in transfer learning offers an effective approach to this problem. Its core idea is to learn domain-invariant features between the simulation domain (source domain) and the real domain (target domain) to reduce distribution discrepancies. For example, Domain Adversarial Networks (DANNs) and their improved models have been applied to life prediction of rotating machinery such as bearings, aligning data distributions under different operating conditions through adversarial training or metric learning. However, directly applying existing domain adaptation methods to RUL prediction of composite material structures still has limitations:

[0006] 1. Monitoring data for composite material structures are usually multi-channel long-term series data from multiple sensors. Traditional feature extraction networks are unable to effectively capture the interdependence between data channels and the long-distance dependence in the time dimension, resulting in insufficient feature extraction.

[0007] 2. Relying solely on gradient inversion layers (GRL) for implicit adversarial training or using only a single explicit metric such as maximum mean difference (MMD) may lead to training instability or failure to fully align distributions when the differences between the source and target domains are too large. This results in low efficiency in transferring physical mechanism knowledge from the source domain and ultimately low prediction accuracy.

[0008] Therefore, there is an urgent need for a new method that can efficiently extract key features from multi-channel long-time-series data and more robustly and fully align the distribution differences between simulation and real data, so as to improve the prediction accuracy of the remaining life of composite material structures under the joint simulation and numerical simulation. Summary of the Invention

[0009] The main objective of this invention is to overcome the shortcomings of the prior art and provide a method, system, and storage medium for predicting the remaining life of composite material structures based on domain adaptation under the joint driving of numerical simulation and simulation. This aims to solve the problem of low prediction accuracy caused by the large difference between the distribution of simulation data and real data and the difficulty in extracting features from multi-channel long time-series data.

[0010] The above-mentioned technical objective of the present invention is achieved through the following technical solution: The present invention proposes a domain adversarial network (DMF-DANN) model.

[0011] On the one hand, addressing the challenge of feature extraction from multi-channel long-term data, this invention designs a feature extractor that combines channel attention and multi-scale fusion. First, a channel attention module (SE Block) is used to assign adaptive attention weights to each sensor data channel, enabling the model to focus on key channels containing more damage information and explicitly model the interdependencies between channels. Second, the Inception module uses convolutional kernels and pooling operations of different scales in parallel to extract and aggregate multi-scale features of the data, thereby enhancing the model's ability to capture long-distance dependencies over time and achieving more comprehensive feature extraction.

[0012] On the other hand, to address the issue of insufficient alignment between the data distributions of the source and target domains, this invention constructs a dual distribution alignment mechanism in the domain adversarial module. This mechanism combines the implicit metric of the Gradient Reversal Layer (GRL) with the explicit metric of Maximum Mean Difference (MMD). GRL, through adversarial training, prompts the feature extractor to learn domain-independent features, while MMD, as a non-parametric statistical test, directly calculates and minimizes the feature distribution distance between the two domains in the reproducing kernel Hilbert space. This dual distribution alignment mechanism can more stably and fully align the feature distributions of simulated and real data, improving the model's ability to extract domain-invariant features and its robustness to domain adaptation.

[0013] Compared with the prior art, the beneficial effects of the present invention are:

[0014] 1. Significantly Improved Prediction Accuracy: Through an improved feature extractor and a dual distribution alignment mechanism, the DMF-DANN model proposed in this invention reduces the mean absolute error (MAE) by 44.7% and the root mean square error (RMSE) by 47.1% compared to the baseline DANN model, while also improving the coefficient of determination (R²). 2 It increased by 11.5%.

[0015] 2. Strong feature extraction capability: Combining the SE module and the Inception module, it effectively solves the problem of capturing the dependencies between channels and on the time scale in multi-channel long-term series data. Ablation experiments have demonstrated that each module has a significant positive contribution to the model performance.

[0016] 3. High robustness of domain adaptation: The dual distribution alignment mechanism combining GRL and MMD can more stably and effectively reduce the distribution difference between simulation data and real data compared with a single alignment method, and efficiently transfer the mechanism knowledge in the simulation model to the life prediction task under real working conditions.

[0017] 4. The superiority of the model has been verified: When processing time series data of different lengths, the performance of the model of this invention is consistently better than that of a variety of existing domain adaptive transfer learning models, demonstrating its superiority in processing complex data. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0019] Figure 1 This is a schematic diagram of the structure of Domain Adversarial Networks (DANNs) in the prior art;

[0020] Figure 2 This is a schematic diagram of the structure of the DMF-DANN model proposed in this embodiment of the invention;

[0021] Figure 3 This is a schematic diagram of the channel attention (SE) module used in an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of the Inception module used in this embodiment of the invention;

[0023] Figure 5 This is an overall flowchart of the prediction of the remaining life of composite material structures in the embodiments of the present invention;

[0024] Figure 6 These are X-ray images of the composite material samples used in the verification experiments of this invention and the development of delamination damage during the experiments;

[0025] Figure 7 This is a schematic diagram of a finite element simulation model of different levels of layered damage established based on the actual damage trend in the verification experiment of this invention;

[0026] Figure 8 These are the prediction results and fitting curves of the remaining lifetime of composite material structures using different transfer learning models in the ablation experiment of this invention.

[0027] Figure 9 This is a bar chart showing the results of the DMF-DANN ablation experiment in the ablation experiment of this invention;

[0028] Figure 10 This is a graph showing the evaluation metrics of the domain adaptive transfer learning model in the comparative experiment of this invention.

[0029] Figure 11 This is a comparative experiment of the present invention, showing the prediction results and fitting curves of the remaining life of composite material structures using adaptive transfer learning models in various domains. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0031] This invention aims to provide a method and system for predicting the remaining life of composite material structures based on domain adaptation under the joint driving of numerical simulation and model. Its core is to construct a novel domain adversarial network (DMF-DANN) model to efficiently and accurately solve the problem of low prediction accuracy caused by the huge difference between the distribution of simulation data and real data, as well as the difficulty in feature extraction of multi-channel long time series data in the existing technology.

[0032] Overall architecture of the DMF-DANN model

[0033] Reference Figure 2 The DMF-DANN model proposed in this invention mainly consists of three core parts:

[0034] 1. Feature Extractor G f A specially designed network for automatically extracting high-dimensional feature representations from input source domain data (simulation data) and target domain data (real monitoring data).

[0035] 2. Predictor G y A regression network that receives source domain features from a feature extractor and predicts their corresponding remaining lifetime (RUL).

[0036] 3. Domain Adversarial Module: Contains a domain discriminator G d A maximum mean difference (MMD) calculation module is included to align the feature distributions of the source and target domains.

[0037] The specific implementation of each part will be explained in detail below.

[0038] 1. Improved Feature Extractor G f

[0039] To address the multi-channel, long-time characteristics of composite material structure monitoring data, this invention designs a feature extractor that combines channel attention and multi-scale feature fusion to replace the traditional CNN network.

[0040] 1.1 Channel Attention (SE) Module

[0041] Reference Figure 3 To enable the model to adaptively focus on sensor channels containing more damage information, this invention first introduces a channel attention (SE) module. Its workflow is as follows:

[0042] Compression (Squeeze): Adaptive average pooling is performed on the input multi-channel data (e.g., a 36x1000 array representing 36 sensor channels and 1000 time steps) to compress the one-dimensional signal data of each channel into a real number, resulting in a channel descriptor vector Z representing the global information of each channel. This step aggregates the global receptive field information of each channel.

[0043] Excitation: To capture the non-linear dependencies between channels, the channel descriptor Z is fed into a bottleneck structure consisting of two 1x1 convolutional layers for learning. The first convolutional layer is used for dimensionality reduction, and the second convolutional layer restores the dimensionality. A sigmoid activation function is then used to generate an attention weight for each channel, ranging from 0 to 1.

[0044] Reweighting: The learned attention weights are multiplied back onto the original input multichannel signal data channel by channel, thereby relabeling the original features, enhancing the feature response of key channels, and suppressing the features of secondary channels.

[0045] 1.2 Inception Module

[0046] Reference Figure 4 The feature data processed by the SE module is then fed into the Inception module to extract and aggregate multi-scale temporal features. The Inception module consists of four parallel branches:

[0047] Branch 1: A 1x1 convolutional layer, used to reduce computation and enhance non-linear representation.

[0048] Branch 2: A 1x1 convolutional layer followed by a 3x3 convolutional layer is used to capture local features of the medium-scale receptive field.

[0049] Branch 3: A 1x1 convolutional layer followed by a 5x5 convolutional layer is used to capture a larger receptive field feature.

[0050] Branch 4: A 3x3 max pooling layer followed by a 1x1 convolutional layer is used to preserve important features and perform channel adjustment through downsampling.

[0051] Finally, the feature data output from these four branches are concatenated along the channel dimension to form a fused feature containing rich multi-scale information. This structure can improve the network's ability to express complex time-series patterns without significantly increasing network depth and the number of parameters.

[0052] 2. A dual distribution alignment mechanism based on GRL and MMD

[0053] To achieve a more stable and comprehensive alignment of the feature distributions of the source and target domains, this invention employs a dual distribution alignment mechanism that combines a gradient inversion layer (GRL) and maximum mean difference (MMD) in the domain adversarial module.

[0054] Gradient Reversal Layer (GRL): As an implicit distribution alignment method, GRL is connected to the feature extractor G... f and the domain discriminator G d Between. During model forward propagation, GRL is an identity transformation; during back propagation, it incorporates the transformations from the neighborhood discriminator G. d The gradient is multiplied by a negative constant before being passed to the feature extractor G. f This adversarial training method forces G f Learned to "deceive" Gd The characteristic of this is that its source domain cannot be easily determined.

[0055] Maximum mean difference (MMD): As an explicit measure of distributional difference, MMD is used to directly measure the difference after G... f The distribution distance between the extracted source domain features and target domain features. Its squared form is expressed as:

[0056]

[0057] In the formula D s D t Let be the set of data samples from the source domain simulation data and the target domain real data, denoted as , x s x t For simulation data D s and real data D t Sample instances in the sample distribution set, n s n t The sample size for simulation data and real data. This is a mapping function that maps samples to the Reproducing Kernel Hilbert Space (RKHS). Using kernel tricks, this computation can be transformed into the computation of a kernel function. This invention employs a Gaussian kernel function:

[0058]

[0059] In the formula, k(·) is the Gaussian kernel function, <·, ·> denote the inner product operation, and σ is the width of the kernel function. By minimizing the MMD loss, the distributions of the two domains can be directly brought closer together in the feature space.

[0060] This dual-distribution alignment mechanism, which combines implicit adversarial approaches with explicit metrics, is more robust than a single mechanism and can more effectively promote the learning of domain-invariant features.

[0061] 3. Model Optimization Objective and Loss Function

[0062] Training the DMF-DANN model involves three optimization objectives, corresponding to three loss functions:

[0063] (1) Regression loss L for predicting labels rec This loss is calculated only on labeled source domain data and is used to optimize the feature extractor G. f and predictor G y The mean squared error (MSE) function is used:

[0064]

[0065] In the formula n sThe number of samples in the source domain. The remaining lifetime is the true value of the source domain simulation data. The remaining lifetime prediction value is the source domain simulation data.

[0066] (2) Classification loss of domain labels This loss is used to train the domain discriminator G. d and with G f To conduct adversarial analysis, using the negative log-likelihood function as the loss function, the classification loss of the neighborhood discriminator can be expressed as:

[0067]

[0068]

[0069] In the formula d i For the domain label: d i = 0 indicates that the sample comes from the source domain simulation data, d i = 1 indicates that the sample comes from real data in the target domain, W d and b d These are the neighborhood discriminators G d The weight and bias matrices in the diagram are: sigm is the sigmoid activation function, and n is the number of training samples in a training batch, i.e., n = n s + n t Training G d At that time, the goal is to minimize Training G using GRL f At that time, the goal is to maximize .

[0070] (3) Domain distribution measurement difference based on MMD: In order to achieve feature-level adaptation, the source domain simulation data D s and the target domain real data D t Features extracted by the feature extractor and Mapping to RKHS, the distribution distance is shortened by minimizing the MMD between the two. A nonlinear mapping in inner product form using a Gaussian kernel function is employed; the expression for MMD is:

[0071]

[0072] Therefore, by weighting and combining these three loss functions, we can obtain the overall loss function L:

[0073]

[0074] In the formula, α and β are hyperparameters that adjust the intensity of domain adversarial learning and domain adaptive learning.

[0075] The overall optimization objective of the composite structure remaining life prediction model based on digital model transfer can be expressed as:

[0076]

[0077] In the formula , , It is G f G y G d The optimal parameters, when the hyperparameter γ is the learning rate, are as follows: The order of updating the main parameters of the model is as follows:

[0078]

[0079]

[0080]

[0081] 4. Overall forecasting process

[0082] Reference Figure 5 The complete process of the method for predicting the remaining life of composite material structures proposed in this invention is as follows:

[0083] (1) Data Acquisition and Construction: Monitoring signals throughout the entire lifecycle are acquired using sensors (such as PZT sensors) on the surface of the composite material structure to construct a real dataset of the target domain. Simultaneously, based on the X-ray images of the structure (such as...),... Figure 6 Based on the actual layered damage development trend revealed by ), a corresponding finite element simulation model (such as...) is established. Figure 7 Numerical analysis is performed to obtain simulation signals throughout the entire lifecycle and to construct a source domain simulation dataset.

[0084] (2) Model construction and initialization: Construct the DMF-DANN model as described above, and set hyperparameters such as the number of iterations, learning rate, and batch size.

[0085] (3) Model Training: Input the source domain simulation data and the target domain real data training set into the model in batches. Feature Extractor G f Extracting domain-invariant features, which are then input into the predictor G. y and the domain discriminator G d In the predictor G y Calculate the predicted regression loss L rec In the adversarial training module, the domain classification loss is calculated. and MMD loss .

[0086] (4) Parameter optimization: The model parameters are optimized based on the overall loss function through backpropagation and gradient descent. , , .

[0087] (5) Iteration complete: Repeat steps 3 and 4 until the set upper limit of the number of iterations is reached to complete the model training.

[0088] (6) Lifetime prediction and verification: Input the test set of real data in the target domain into the trained model, output the remaining life prediction results of the composite material structure, and perform performance evaluation.

[0089] 5. Verification of Implementation Examples

[0090] To further illustrate the technical effects of the present invention, a specific embodiment is provided below.

[0091] 5.1 Experimental Data and Environment

[0092] Dataset: This embodiment uses a publicly available accelerated life test dataset for composite material structures. This dataset records the full life-cycle monitoring data of dog-bone-shaped composite material samples made of Torayca unidirectional prepreg under cyclic loading. A PZT sensor array is arranged on the surface of the structure. Figure 6 The sample and X-ray images of layered damage under different loading cycles are shown.

[0093] Data processing: Extract full lifecycle monitoring data as the real data for the target domain. Based on... Figure 6 The revealed damage development trend has led to the establishment of a system such as Figure 7 The finite element simulation models with different levels of layered damage are shown, and the simulation data are collected as the source domain simulation dataset.

[0094] The formula for calculating the remaining useful life (RUL) of a composite material structure is as follows:

[0095]

[0096] In the formula n max is the maximum number of load cycles, and n is the current number of load cycles.

[0097] Experimental environment: Intel i7-147000KF, NVIDIA Geforce RTX 4070 Ti SUPER graphics monitor, PyTorch 2.0.1 deep learning environment, Windows 11 operating system.

[0098] Hyperparameter settings: learning rate is set to 0.0001, batch size is 32, and training epochs are 100.

[0099] Evaluation metrics: Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used. 2 () is used as an evaluation metric for model performance. The calculation formula is as follows:

[0100]

[0101]

[0102]

[0103] In the formula, n is the number of samples in the target domain test set, and y i The remaining lifetime values ​​are the true values ​​for the test set samples. The remaining lifetime prediction value for the test set samples. This represents the average of the true remaining lifetime values ​​for the test set samples.

[0104] 5.2 Experimental Results and Analysis

[0105] 5.2.1 Ablation Test

[0106] To verify the effectiveness of each improved component in the proposed DMF-DANN method, ablation experiments were conducted. The ablation experiment model is described in Table 1.

[0107] Table 1 Ablation Experimental Model

[0108]

[0109] In the table, the baseline DANN model does not contain any improved modules; the DANN-SE model adds only the Inception and MMD modules to the DANN model, i.e., the ablation SE module; the DANN-In model adds only the SE and MMD modules to the DANN model, i.e., the ablation Inception module; the DANN-MMD model adds only the SE and Inception modules to the DANN model, i.e., the ablation MMD module; the complete DMF-DANN model contains all improved modules.

[0110] The remaining lifetime prediction results of each model are as follows: Figure 8As shown, the closer the predicted result is to the actual remaining lifetime curve (RUL), the higher the prediction accuracy. The DANN model has the largest overall bias in predicting the lifetime of composite structures throughout their entire life cycle. The DANN-SE model after ablation of the SE module has a large bias in predicting the mid-term lifetime. The DANN-In model after ablation of the Inception module and the DANN-MMD model after ablation of the MMD module have large biases in predicting the initial and final lifetimes. The DMF-DANN model has the smallest overall prediction bias. Linear fitting analysis was performed on the prediction results of each model, and the fitting curves of the remaining lifetime prediction results of each model are shown in the figure. Figure 8 As shown in (b), the DMF-DANN model's prediction curve is closest to the actual lifetime curve RUL, indicating the highest prediction accuracy. Comparative analysis through ablation experiments demonstrates that the DMF-DANN model can effectively transfer mechanistic knowledge from simulation models to the target domain's data-driven remaining lifetime prediction task, validating the effectiveness of the DMF-DANN model.

[0111] Results: The evaluation indicators for the prediction results of each method are shown in Table 2 and Figure 9 As shown, the ablation experiment evaluation indicators MAE, RMSE, and R... 2 The results comparison shows that the DMF-DANN method effectively combines the SE module, Inception module, and MMD module. Compared with the DANN method, its evaluation metrics MAE are reduced by 44.7%, RMSE by 47.1%, and R... 2 The accuracy of the model's remaining lifetime prediction was significantly improved by 11.5%. Compared to the DANN-SE, DANN-In, and DANN-MMD methods, the DMF-DANN method reduced the MAE by 33.9%, 17.0%, and 17.6%, respectively; and the RMSE by 37.7%, 26.4%, and 24.14%, respectively. 2 The improvements were 6.8%, 3.5%, and 3.1%, respectively. Ablation experiments verified that the SE, Inception, and MMD modules all had a positive effect on model performance, with the SE module contributing more significantly than the Inception and MMD modules. This is because the SE module assigns different attention weights to each sensor path data channel, enabling the model to focus on channel data containing more damage features, thereby better capturing features in time-series data that are highly relevant to the prediction task.

[0112] Conclusion: The ablation experiments strongly demonstrate that the combination of the SE module, Inception module, and MMD module is effective and necessary for improving the model's prediction accuracy.

[0113] Table 2 Results of DMF-DANN ablation experiments

[0114]

[0115] 5.2.2 Comparative Experiment

[0116] To verify the superiority of the proposed DMF-DANN method, multiple comparative experiments were conducted, comparing it with several other domain adaptive transfer learning models: Wasserstein Domain Adversarial Network (WDANN), LSTM Domain Adversarial Network (LSTM-DANN), Deep Convolutional Transfer Learning Network (DCTLN), CNNLSTM Domain Adversarial Network (CNNLSTM-DANN), Subspace Domain Adversarial Discriminant Network (SDADN), and LSTM Deep Neural Network (LSTM-DNN). Long-term series data can contain more features about the prediction task, improving the model's prediction accuracy; however, long-term series data places higher demands on the model's feature extraction capabilities. Experiments were conducted using time-series data of lengths of 100, 500, and 1000, respectively, and MAE, RMSE, and R... 2 The performance of the model was evaluated using multiple metrics, and the evaluation results of the adaptive transfer learning model for each domain are shown in Tables 3-5.

[0117] Result: As Figure 10 As shown, with the increase in the length of the time series data, the MAE and RMSE of each model all decreased, and R... 2 All indicators showed improvement, with the DMF-DANN model achieving the best results across all evaluation metrics. Comparative experimental analysis demonstrates that the DMF-DANN model exhibits higher prediction accuracy for the remaining life of composite material structures in long-term data series compared to other models, validating the superiority of the DMF-DANN model.

[0118] The remaining life prediction results of the adaptive transfer learning model for composite material structures with a time series data of length 1000 are as follows: Figure 11 As shown in (a), the DMF-DANN prediction results have the smallest overall bias; linear fitting analysis was performed on the prediction results of each model, and the fitting curves of the remaining lifetime prediction results of each model are shown in Figure 1. Figure 11 As shown in (d), the DMF-DANN model's prediction curve is closest to the actual lifetime curve RUL, indicating the highest prediction accuracy. Comparative analysis of experiments with different domain adaptive transfer learning models demonstrates the superiority of the DMF-DANN model.

[0119] Conclusion: Comparative experiments demonstrate that the DMF-DANN model proposed in this invention has significant advantages over other domain adaptive methods, especially in handling complex long-term time-series data, where it can achieve higher prediction accuracy.

[0120] Table 3 Evaluation metrics for each model in time series data with a length of 100.

[0121]

[0122] Table 4. Evaluation metrics for various models of time series data with a length of 500.

[0123]

[0124] Table 5. Evaluation metrics for each model in time series data with a length of 1000.

[0125]

[0126] This invention addresses the problem of low prediction accuracy caused by the large distribution difference between simulation and real data in composite material remaining life prediction methods driven by joint numerical simulation. It proposes a domain adversarial model (DMF-DANN) based on numerical simulation fusion for predicting the remaining life of composite material structures. The model was validated through ablation experiments and comparative experiments with other domain adaptive transfer learning models using a publicly available composite material full life cycle dataset. The results demonstrate the effectiveness and superiority of the DMF-DANN method. The experimental conclusions are as follows:

[0127] (1) The DMF-DANN method effectively combines the SE module, Inception module, and MMD module. Compared with the DANN method, its evaluation metrics MAE are reduced by 44.7%, RMSE by 47.1%, and R 2 The accuracy of the model's remaining lifetime prediction was significantly improved by 11.5%, indicating that the DMF-DANN model can effectively extract domain-independent but task-related features from long-term series data, enabling the transfer of mechanistic knowledge from simulation models to data-driven remaining lifetime prediction tasks in the target domain.

[0128] (2) Ablation experiments verified that, compared with the DANN-SE, DANN-In, and DANN-MMD methods, the MAE of the DMF-DANN method was reduced by 33.9%, 17.0%, and 17.6%, respectively; the RMSE was reduced by 37.7%, 26.4%, and 24.14%, respectively; R 2 The improvements were 6.8%, 3.5%, and 3.1%, respectively. The results show that the SE module, Inception module, and MMD module all have a positive effect on improving model performance, with the SE module contributing the most, thus validating the effectiveness of each component of the model.

[0129] (3) Experimental verification of adaptive transfer learning models in different domains was conducted using time series data of lengths of 100, 500, and 1000. The evaluation metrics (MAE, RMSE, R) of adaptive transfer learning methods in each domain, including WDANN, LSTM-DANN, DCTLN, SDADN, CNNLSTM-DANN, and LSTM-DNN, were used. 2 The results show that the DMF-DANN model has higher prediction accuracy for the remaining life of composite material structures in long-term series data, surpassing other domain adaptive transfer learning models, thus verifying the superiority of the model.

[0130] In summary, the specific embodiments of this invention, through detailed ablation experiments and comparative experiments, fully validated the effectiveness and superiority of the proposed DMF-DANN model. The experimental results clearly demonstrate that this invention, by combining the channel attention (SE) module and the Inception module, effectively improves the feature extraction capability for multi-channel long-term data; and by combining the gradient inversion layer (GRL) and the maximum mean difference (MMD) dual distribution alignment mechanism, it significantly enhances the robustness of domain adaptation, successfully transferring the mechanistic knowledge from the simulation model to the prediction task of real data. Compared with existing technologies, this invention achieves higher accuracy and reliability in predicting the remaining life of composite material structures, successfully fulfilling the objectives of this invention.

[0131] It should be emphasized that the above description is merely a preferred embodiment of the present invention and is 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 predicting the remaining life of composite material structures based on domain adaptation under joint numerical simulation and numerical modeling, characterized in that, The method includes the following steps: Acquire full-life-cycle simulation data of composite material structures as source domain data and real-life-cycle monitoring data of composite material structures as target domain data; A domain adversarial network model fused with digital model is constructed, the model including a feature extractor, a predictor and a domain adversarial module; The feature extractor is used to extract features from the source domain data and the target domain data to obtain their respective feature representations. The feature extractor adopts a structure that combines a channel attention module and a multi-scale feature fusion module, which assigns adaptive weights to different channels of the input data and extracts and fuses the multi-scale features of the data. The source domain data features extracted by the feature extractor are input into the predictor to obtain the remaining lifetime prediction value and calculate the prediction loss; The feature representations of the source domain data and the target domain data are input into the domain adversarial module, and domain adaptive training is performed through a dual distribution alignment mechanism. The dual distribution alignment mechanism includes an implicit metric based on a gradient reversal layer and an explicit metric based on the maximum mean difference, so as to reduce the feature distribution difference between the source domain and the target domain. The network model parameters are jointly optimized based on the predicted loss and the output of the domain adversarial module. The target domain data to be tested is input into the trained network model, and the remaining lifetime prediction results of the composite material structure are output. The channel attention module is an SE module, which adaptively learns and assigns channel weights by performing compression and activation operations on each data channel in order to establish the interdependence between data channels. The multi-scale feature fusion module is the Inception module, which performs multi-scale feature extraction and aggregation on the data by paralleling convolution kernels of different sizes and pooling operations, in order to capture the long-distance dependency characteristics of the data. The step of jointly optimizing the network model parameters is performed through an overall loss function that includes prediction regression loss, domain classification loss, and maximum mean difference loss.

2. The method according to claim 1, characterized in that, The gradient inversion layer in the domain adversarial module inverts the gradient during the model backpropagation process, enabling the feature extractor to learn domain-invariant features through adversarial training.

3. The method according to claim 1, characterized in that, The maximum mean difference in the domain adversarial module is used to explicitly measure the difference in feature distribution between the source domain data and the target domain data after processing by the feature extractor, and inter-domain distribution alignment is achieved by minimizing this difference.

4. A system for predicting the remaining life of composite material structures, characterized in that, include: The data acquisition unit is used to acquire full-life-cycle simulation data of composite material structures as source domain data and real-life-cycle monitoring data of composite material structures as target domain data. The model processing unit is configured with a domain adversarial network model that integrates data and model, the model comprising: A feature extractor, which adopts a combined structure of a channel attention module and a multi-scale feature fusion module, is used to extract features from the source domain data and the target domain data to obtain their respective feature representations; A predictor is used to output a predicted remaining lifetime value based on the source domain data features extracted by the feature extractor. A domain adversarial module includes a gradient inversion layer and a maximum mean difference calculation function. It processes the feature representations of the source and target domains through a dual distribution alignment mechanism to reduce their distribution differences. A model training unit is used to jointly optimize the network model based on the prediction results of the predictor and the processing results of the domain adversarial module; A lifetime prediction unit is used to input the test data into the optimized network model and output the remaining lifetime of the composite material structure.

5. The system according to claim 4, characterized in that, The channel attention module in the feature extractor is the SE module, and the multi-scale feature fusion module is the Inception module.

6. The system according to claim 4, characterized in that, The domain adversarial module performs adversarial training through a gradient inversion layer and uses the maximum mean difference to explicitly measure the feature distribution distance between the source and target domains, working together to enable the feature extractor to learn domain-invariant features.

7. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the remaining life of a composite material structure as described in any one of claims 1 to 3.

Images