Composite material defect modeling method based on physical modeling generation and meta-learning migration

Abstract

The invention discloses a composite material defect modeling method based on physical modeling generation and meta-learning migration, which comprises the following steps: S10, starting from a forming mechanism of a fatigue crack of a composite material, establishing a morphological function of a crack morphological simulation model based on a mechanical principle for simulating a spatial evolution process of a defect; s20, embedding the morphological function as a regular term into a generator loss function of a generative adversarial network, and generating a high-simulation pseudo-defect image; s30, introducing a transfer learning mechanism, constructing a model-irrelevant meta learning training strategy, and realizing rapid learning of the model on a new task through a two-layer nested optimization strategy; s40, introducing a domain adaptive residual module in the model training process to enhance the feature alignment capability of a cross-batch image domain; and S50, training the model by using a real sample and a synthetic defect sample in a mixed manner.

Description

Technical Field

[0001] The present invention belongs to the technical field of material defect detection, and specifically relates to a composite material defect modeling method based on physical modeling generation and meta-learning transfer. Background Art

[0002] In the aerospace industry, with the widespread application of composite materials in high-performance structural parts, materials such as carbon fiber-reinforced plastics (CFRP), glass fiber-reinforced polymers (GFRP), and metal matrix composites (MMC) have become the main materials for critical components such as aircraft skins, frames, and spars. Despite their significant advantages, such as high strength, low specific gravity, and corrosion resistance, composite materials exhibit complex internal defects, making early detection difficult. Furthermore, inspection images often suffer from high sample annotation costs and extensive manual intervention. For example, typical composite material defects include delamination, voids, cracks, and inclusions. These defects are highly random and rare in actual production, making traditional defect detection models based on large-sample supervised learning difficult to deploy. This is especially true when the number of samples is less than a few dozen. The trained models suffer from severe overfitting and poor generalization. Summary of the Invention

[0003] In view of the above problems, the present invention provides a composite material defect modeling method based on physical modeling generation and meta-learning transfer. By introducing a generation mechanism constrained by physical rules, high-quality synthetic defect image data is constructed, and combined with a fast model adaptation mechanism based on the meta-learning framework, the model can achieve rapid convergence and high-precision detection in new defect scenarios, which is particularly suitable for the typical small sample conditions in aerospace composite material detection.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions: A composite material defect modeling method based on physical modeling generation and meta-learning transfer includes the following steps: S10, starting from the formation mechanism of fatigue cracks in composite materials, a morphological function of the crack morphology simulation model based on mechanical principles is established to simulate the spatial evolution process of defects; S20, embedding the morphological function as a regular term into the generator loss function of the generative adversarial network to generate a highly realistic pseudo defect image; S30 introduces a transfer learning mechanism and builds a model-independent meta-learning training strategy, which enables the model to quickly learn new tasks through a two-layer nested optimization strategy. S40, introduces a domain-adaptive residual module during model training to enhance the ability to align features across image batches; S50, using real samples and synthetic defect samples to mix and train the model.

[0005] In a possible implementation, the morphology function of the crack morphology simulation model is: ; in A is the crack starting point, is the crack starting point, Control the crack diffusion range, 、 Control the direction and starting phase of the main crack texture.

[0006] In one possible implementation, the generator loss function is: ; in: ; The physical rule constraint loss function is used to guide the generator to produce defect images that are more consistent with the physical morphology laws; Input latent variables to the generator From its distribution The loss term corresponding to the sample obtained by sampling is the mathematical expectation; Represents the latent space Generated defect image; represents the Laplacian edge feature of the image, To control weight; The square of the L² norm, which measures the Euclidean distance between the edge maps of two images; is the morphological function of the crack morphology simulation model.

[0007] In one possible implementation, a two-layer nested optimization strategy is used to enable the model to quickly learn new tasks, including inner loop updates for rapid adaptation within each subtask. Then, we use the initial parameters θ to perform one or more gradient updates to obtain the task-specific parameters : ; Where θ is the shared initialization parameter; α is the inner loop learning rate; It's a task The loss function of the upper support set is usually cross entropy loss; To test the model; Then optimize the update effect under multi-task in the outer loop: use the updated parameters In the validation set Calculate the loss and optimize θ: ; Where β is the outer loop learning rate; The final model optimization goal, the training goal is to minimize the total expectation of all task verification losses: ; In one possible implementation, the domain-adaptive residual module includes a global channel attention mechanism and a domain-aligned batch normalization mechanism, and its output feature representation is: ; in are trainable weights, is a nonlinear activation function, The BatchNorm parameters are selected based on the source domain or target domain.

[0008] The adoption of the present invention has at least the following beneficial effects: physical rule modeling is introduced into the GAN generator structure, an explainable defect image generation mechanism is constructed, and it is ensured that the data enhancement stage not only generates diverse pseudo samples, but also guarantees high consistency with the real defect morphology. At the same time, by introducing a meta-learning strategy based on gradient nested optimization, the model's rapid adaptability under small sample conditions is achieved, and the migration domain alignment residual block is used to enhance the field generalization ability. This technical solution has extremely high promotion value in the field of aerospace composite materials, and can be expanded to defect recognition scenarios such as railway weld detection, bridge crack detection, and high-speed photolithography wafer detection, to build a universal small sample intelligent detection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flowchart of a composite material defect modeling method based on physical modeling generation and meta-learning transfer according to an embodiment of the present invention; Figure 2 2 is a flow chart of the inner loop and outer loop parameter update mechanism in the MAML optimization structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0010] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0011] See also Figure 1 , which shows a composite material defect modeling method based on physical modeling generation and meta-learning transfer according to an embodiment of the present invention, comprising the following steps: S10, starting from the formation mechanism of fatigue cracks in composite materials, a morphological function of the crack morphology simulation model based on mechanical principles is established to simulate the spatial evolution process of defects; S20, embeds the morphological function as a regular term into the generator loss function of the Generative Adversarial Networks (GAN) to generate highly realistic pseudo-defect images; S30 introduces a transfer learning mechanism and builds a model-independent meta-learning training strategy, which enables the model to quickly learn new tasks through a two-layer nested optimization strategy. S40, introduces a domain-adaptive residual module during model training to enhance the ability to align features across image batches; S50, using real samples and synthetic defect samples to mix and train the model.

[0012] As a specific application example, in S10, according to the Paris formula in material mechanics, the fatigue crack growth behavior can be approximately described as: ; in is the crack length, is the number of loading cycles, is the stress intensity factor range, 、 is a constant related to the material. The Paris formula is used to describe the rate at which cracks grow under fatigue loads as the number of cyclic loading increases. In the stable growth stage (usually before the crack enters the rapid growth or critical fracture stage), the range of the crack growth rate and the stress intensity factor is It is a power function relationship. When it is smaller, the crack growth rate is slower; as As the crack increases, it will accelerate the expansion.

[0013] In aerospace, cracks in carbon fiber laminates often extend radially and asymmetrically, accompanied by a microcrack network structure. The above process is mapped into a generative model for image defects, and the morphological function of the crack morphology simulation model is designed through morphological modeling. The morphological function of the crack morphology simulation model is: ; in A is the crack starting point, is the crack starting point, Control the crack diffusion range, 、 Controls the orientation and starting phase of the primary crack texture. Specifically, the morphology function simulates the spatial evolution of cracks in composite materials, while the parameter A controls the point at which the crack initiates—that is, the crack's initial position in image space. This parameter helps locate the initial defect when generating the crack morphology.

[0014] Furthermore, this function is introduced into the structural design of the GAN generator as a morphological control term, so that the morphological distribution of the generated samples conforms to the spatial evolution logic of the cracks in the real material. Based on the original GAN ​​framework, a physical rule constraint loss function is introduced, and the generator loss function is: ; in: ; The physical rule constraint loss function is used to guide the generator to produce defect images that are more consistent with the physical morphology laws; Input latent variables to the generator From its distribution The loss term corresponding to the sample obtained by sampling is the mathematical expectation; Represents the latent space Generated defect image; represents the Laplacian edge feature of the image, To control weight; The square of the L² norm, which measures the Euclidean distance between the edge maps of two images; is the morphological function of the crack morphology simulation model. The introduction of this loss function makes the generator more inclined to produce highly realistic defect images with "sharp crack edges + gradual texture mutations + multi-directional branching structures" during training.

[0015] Furthermore, after the construction of the generated data enhancement module is completed, another challenge is how to effectively integrate the generated data into the small sample detection model to avoid overfitting. To this end, a transfer learning mechanism is introduced, and a meta-learning training strategy based on model-independent meta-learning (MAML) is constructed. The core idea is to train a general model initialization parameter through gradient nested optimization on multiple task distributions, so that it can quickly adapt to new tasks with only a small number of samples. . In this framework, the defect detection model is regarded as a function learner that can be quickly generalized through a few samples, and the model is quickly adapted to new tasks through two-layer nested optimization. Specifically, its structure includes two main processes: "task batch sampling", "inner loop update (Inner Loop)" and "outer loop optimization (Outer Loop)", and cooperates with task grouping, shared model, gradient conduction and other mechanisms to achieve efficient learning. For a few-sample task ,The update of model parameters can be divided into two parts: “inner loop update” and “outer loop optimization”.

[0016] Task sampling involves selecting from a set of tasks Randomly sample several subtasks: ; Each subtask Contains support set and validation set (query set) .

[0017] Inner loop updates are performed to quickly adapt to each subtask within the task Then, we use the initial parameters θ to perform one or more gradient updates to obtain the task-specific parameters : ; Where θ is the shared initialization parameter; α is the inner loop learning rate; It's a task The loss function on the support set is often the cross entropy loss. To test the model.

[0018] Then optimize the update effect under multi-task in the outer loop: use the updated parameters In the validation set Calculate the loss and optimize θ: ; Where β is the outer loop learning rate.

[0019] The final model optimization goal, the training goal is to minimize the total expectation of all task verification losses: ; Figure 2 The arrows clearly distinguish the inner loop local update path and the outer loop global update path of the model parameters. The model starts from θ and generates multiple task-specific models through inner loop gradient descent. , and then use the feedback from the validation set to adjust θ to form an initialization that can generalize to multiple tasks. The entire process can be completed iteratively using optimizers such as SGD or Adam.

[0020] This approach enables the model to quickly adapt to the distribution of new samples during the training phase. Combined with the highly realistic samples generated by the aforementioned physical modeling GAN, only 3 to 5 annotated images are needed during the testing phase to achieve high-precision classification and positioning in new material batches or new defect types.

[0021] Furthermore, the Domain Adaptive Residual Block (DARB) structure is embedded in the feature extraction backbone network to enhance the domain alignment capability during transfer learning. The DARB block includes a global channel attention mechanism and a domain-specific batch normalization mechanism. Its output features can be expressed as: ; in are trainable weights, is a nonlinear activation function, is the BatchNorm parameter selected based on the source or target domain. Through this structure, the model can adaptively model the sample distribution drift between composite batches caused by factors such as manufacturing process and imaging resolution.

[0022] To validate the effectiveness of the method described in this embodiment, experiments were conducted using multiple datasets of real-world aerospace composite material defect images, including samples of various types of cracks and delamination. Five images of each defect type were randomly selected as training samples, with the remaining images used for testing. Comparisons were made using traditional CNNs, a Fine-Tune pre-trained ResNet model, a ProtoNet prototype network, and a Meta-Learning method without GAN enhancement. The experimental results show that the method described in this embodiment outperforms the baseline model across all metrics, achieving an approximately 15% improvement in accuracy and a greater than 30% reduction in false positives. Robust performance was particularly evident under extremely small sample sizes (e.g., only one or two images per class).

[0023] More importantly, the physical modeling enhancement mechanism proposed in the embodiment of the present invention not only makes the generated samples discriminable at the visual feature level, but also achieves alignment with the distribution of real defect samples in the high-dimensional feature distribution space. Using t-SNE visualization to perform dimensionality reduction analysis on the feature embedding space, the results show that the sample distribution clustering structure after processing using the GAN+Meta model is clearer, and the distance between classes is significantly increased. In contrast, models without generated data often exhibit intra-class mixing and blurred boundaries, validating the effectiveness of this method in constructing a discriminant-friendly feature space.

[0024] In summary, the method of the embodiment of the present invention innovatively introduces physical rule modeling into the GAN generator structure, constructs an explainable defect image generation mechanism, and ensures that the data enhancement stage not only generates diverse pseudo samples, but also ensures high consistency with the real defect morphology. At the same time, by introducing a meta-learning strategy based on gradient nested optimization, the model's rapid adaptability under small sample conditions is achieved, and the migration domain alignment residual block is used to enhance the field generalization capability. This technical solution has extremely high promotion value in the field of aerospace composite materials, and can be extended to defect recognition scenarios such as railway weld detection, bridge crack detection, and high-speed lithography wafer detection to build a universal small-sample intelligent detection system.

[0025] It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the present invention have been described in conjunction with the accompanying drawings, it should be understood by those skilled in the art that various changes in form and details may be made without departing from the spirit and scope of the present invention as defined by the appended claims.

Claims

1. A composite material defect modeling method based on physical modeling generation and meta-learning transfer, characterized in that: The following steps are involved: S10, starting from the formation mechanism of fatigue cracks in composite materials, a morphological function of the crack morphology simulation model based on mechanical principles is established to simulate the spatial evolution process of defects; S20, embedding the morphological function as a regular term into the generator loss function of the generative adversarial network to generate a highly realistic pseudo defect image; S30 introduces a transfer learning mechanism and builds a model-independent meta-learning training strategy, which enables the model to quickly learn new tasks through a two-layer nested optimization strategy. S40, introduces a domain-adaptive residual module during model training to enhance the ability to align features across image batches; S50, using real samples and synthetic defect samples to mix and train the model.

2. The composite material defect modeling method based on physical modeling generation and meta-learning transfer according to claim 1, characterized in that: The morphological function of the crack morphology simulation model is: ； in A is the crack starting point, is the crack starting point, Control the crack diffusion range, 、 Control the direction and starting phase of the main crack texture.

3. The composite material defect modeling method based on physical modeling generation and meta-learning transfer according to claim 2, characterized in that: The generator loss function is: ； in: ； The physical rule constraint loss function is used to guide the generator to produce defect images that are more consistent with the physical morphology laws; Input latent variables to the generator From its distribution The loss term corresponding to the sample obtained by sampling is the mathematical expectation; Represents the latent space Generated defect image; represents the Laplacian edge feature of the image, To control weight; The square of the L² norm, which measures the Euclidean distance between the edge maps of two images; is the morphological function of the crack morphology simulation model.

4. The composite material defect modeling method based on physical modeling generation and meta-learning transfer according to claim 3, characterized in that: The model can quickly learn new tasks through a two-layer nested optimization strategy, including inner loop updates for rapid adaptation within each subtask 𝒯 Then, we use the initial parameters θ to perform one or more gradient updates to obtain the task-specific parameters : ； Where θ is the shared initialization parameter; α is the inner loop learning rate; It's a task The loss function of the upper support set is usually cross entropy loss; To test the model; Then optimize the update effect under multi-task in the outer loop: use the updated parameters In the validation set Calculate the loss and optimize θ: ； Where β is the outer loop learning rate; The final model optimization goal, the training goal is to minimize the total expectation of all task verification losses: 。 5. The composite material defect modeling method based on physical modeling generation and meta-learning transfer according to claim 4, characterized in that: The domain-adaptive residual module includes a global channel attention mechanism and a domain-aligned batch normalization mechanism, and its output features are expressed as: ； in are trainable weights, is a nonlinear activation function, The BatchNorm parameters are selected based on the source domain or target domain.

Images