Ceramic-based composite material layering defect identification method based on deep learning
By using a hybrid architecture of mask-guided SAGAN and enhanced U-Net network, the problems of small sample adaptability and weak defect identification in the layered defect identification of ceramic matrix composites are solved, achieving high-precision and efficient industrial inspection results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2025-12-11
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies for identifying delamination defects in ceramic matrix composites suffer from poor adaptability to small samples, insufficient sensitivity to weak or microscale defects, and difficulty in meeting the real-time industrial detection requirements due to the inference efficiency of the models.
We employ a hybrid architecture design that combines mask-guided self-attention generative adversarial network (SAGAN) with an enhanced U-Net network. By constructing high-quality dedicated datasets and employing a three-stage collaborative training strategy, along with data augmentation and multi-scale feature fusion, we improve the accuracy and efficiency of defect identification.
It significantly improves the accuracy and efficiency of identifying delamination defects in ceramic matrix composites, reduces the risk of false positives and false negatives, enhances the adaptability and robustness of the model under different process conditions, and meets the high reliability and high efficiency requirements of industrial non-destructive testing.
Smart Images
Figure CN121904435A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ceramic matrix composites technology, and more specifically to a method for identifying delamination defects in ceramic matrix composites based on deep learning. Background Technology
[0002] Ceramic matrix composites, due to their advantages such as low density, high temperature resistance, high strength, and high designability, are widely used in high-end fields such as aerospace, energy, and military industries, and are indispensable, especially in hot-end components of aero-engines and thermal protection systems of hypersonic vehicles. However, their manufacturing process is complex, involving multiple stages such as fiber layup, matrix densification, and processing and assembly, which can easily introduce defects such as delamination. Delamination often originates from uneven layup tension, thermal expansion mismatch, densification shrinkage, or processing impact, which can significantly weaken interlaminar shear strength and overall load-bearing capacity, and induce crack initiation and propagation under extreme conditions such as thermo-mechanical coupling, seriously threatening structural integrity and lifespan.
[0003] Currently, the identification of layered defects mainly relies on manual interpretation of non-destructive testing images such as ultrasonic C-scans, X-ray CT, or infrared thermography. This approach suffers from low efficiency, high subjectivity, and insufficient sensitivity to minute or deeply buried defects, and it is difficult to achieve online monitoring during the manufacturing process. Therefore, developing intelligent and high-precision layer identification methods is crucial for not only accurately assessing component performance and revealing damage evolution mechanisms, but also for optimizing layup design, interface coatings, and densification processes. This represents a key breakthrough in improving the manufacturing quality and service reliability of ceramic matrix composites.
[0004] Self-Attention Generative Adversarial Networks (SAGANs) represent a substantial improvement over existing image generation techniques. The effective integration of self-attention technology enables the network to realistically capture and correlate long-range spatial information for better synthesis of new images, while maintaining computational efficiency. By combining WGAN-GP with autoencoder pre-training to suppress mode collapse in both the discriminator and generator networks, SAGANs improve training stability. SAGANs are often used to supplement the training dataset of U-Net networks by synthesizing high-quality, high-fidelity labeled data, particularly in the field of medical imaging, to address the scarcity and high cost of medical data annotation, thus helping U-Net networks better perform core tasks such as image segmentation and modality transformation.
[0005] However, the application of SAGAN and U-Net networks to delamination defects in ceramic matrix conformal materials still faces the following challenges:
[0006] On the one hand, SAGAN is prone to instability during training and is at risk of mode collapse, resulting in insufficient diversity of generated samples. Especially under small sample conditions, it is difficult to effectively capture and reproduce the fine features of tiny layered defects. The generated defect images deviate significantly from real defects in terms of morphology, texture distribution and spatial location, which limits its practicality in high-precision defect simulation and data augmentation.
[0007] On the other hand, the U-Net network is limited by the scarcity of defect samples in ceramic matrix composites, making it prone to overfitting and exhibiting weak generalization ability across production lines or process scenarios. At the same time, its inherent encoder-decoder structure is prone to losing the spatial coordinate information of sub-millimeter-level defects during deep feature extraction, resulting in a decrease in defect localization accuracy. Furthermore, the traditional skip connection mechanism fails to achieve effective fusion of multi-scale features, causing the feature responses of defects of different scales, such as micron-level pores and millimeter-level stratification, to interfere with each other, significantly increasing the false detection rate.
[0008] In summary, SAGAN and U-Net face common challenges in the defect identification of ceramic matrix composites: poor adaptability to small sample data, insufficient sensitivity to weak or microscale defects, and model inference efficiency that is difficult to meet the needs of real-time industrial detection. Summary of the Invention
[0009] To address the key issues of the standard U-Net network in detecting delamination defects in ceramic matrix composites, such as overfitting due to small samples, loss of spatial information of sub-millimeter-level defects, and multi-scale feature confusion caused by data scarcity, this invention aims to provide a deep learning-based method for identifying delamination defects in ceramic matrix composites. This method utilizes a hybrid network architecture design that combines the advantages of high-quality dedicated datasets for construction and fusion, as well as generation and segmentation, to support the construction of a highly reliable, high-precision, and efficient intelligent non-destructive testing system for ceramic matrix composites.
[0010] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:
[0011] A deep learning-based method for identifying delamination defects in ceramic matrix composites, the method comprising the following steps:
[0012] Step 1: High-resolution X-ray computed tomography (CT) scans were performed on ceramic matrix composite samples with typical ply structures to obtain voxel-level three-dimensional volume data. Three-dimensional reconstruction was completed using filtered back projection or iterative reconstruction algorithms to generate a complete volume dataset containing information on the internal microstructure of the material. Two-dimensional XCT slice images, including cross-sectional, longitudinal, and arbitrarily angled slice images, were extracted as real samples to comprehensively characterize the geometric morphology of the interlayer interface and the spatial distribution characteristics of potential defects.
[0013] Step 2: Perform data augmentation on the XCT 2D slice images extracted in Step 1 to obtain augmented samples;
[0014] Step 3: Using the spatial distribution features of potential defects extracted in Step 1 and the enhanced samples obtained in Step 2 as prior conditions, a mask-guided conditional generation framework is constructed. A SAGAN model with an integrated self-attention mechanism is adopted to model long-distance spatial dependence in the generator and discriminator, reproducing the extensibility and diffusion characteristics of layered defects. Using real XCT two-dimensional slice images and their corresponding defect masks as supervision signals, the SAGAN model is guided to generate high-fidelity virtual samples with sheet-like or diffuse layered structures only in the defect area, so that their texture, shape and spatial distribution are consistent with the real defects.
[0015] Step 4: Collect and organize real samples, augmented samples, and high-fidelity virtual samples for training. Label the layered defect regions in each image to generate corresponding polygon annotation files and form an image-label pairing dataset.
[0016] Step 5: Construct an enhanced U-Net model, which includes an encoder, a decoder, and a classification head; take XCT two-dimensional slice images as input and pixel-level defect segmentation results as output; combine real samples, enhanced samples, and high-fidelity virtual samples to construct an image-label pairing dataset to train the enhanced U-Net model, so as to improve the enhanced U-Net model's context-aware ability for lamellar and diffuse layered defects;
[0017] The encoder employs a pre-trained VGG network, extracting features from low to high layers of the XCT 2D slice image through convolution and max pooling operations. Low-layer feature maps highlight the microstructural details of the material, while mid-to-high-layer feature maps, with increased depth, focus on capturing the overall structural characteristics of the composite material and its potential layering interfaces, clearly displaying layered defects against complex backgrounds. The decoder of the enhanced U-Net model receives a deep feature map with a resolution of 1 / 32 of the original input from the encoder, then gradually restores the spatial resolution through upsampling and skip connections, fusing detailed information from the encoder's corresponding layers to ultimately generate a high-resolution feature map to reconstruct the morphology of layered defects. A channel-spatial dual attention gating mechanism optimizes skip connections, enhancing the U-Net model's ability to focus on key features, highlighting information sensitive to layered defects, and distinguishing between real defects and process clutter interference structures. The classification head captures contextual information, achieving pixel-level classification through 1×1 convolutional layers to generate corresponding defect region segmentation maps.
[0018] Step 6: Use the trained enhanced U-Net model to identify layered defect regions in the XCT slice images of the ceramic matrix composite material to be tested.
[0019] Furthermore, in step 2, the data augmentation operation on the XCT two-dimensional slice image extracted in step 1 includes:
[0020] Rotation and translation operations are performed on XCT 2D slice images to simulate the observation effect under different viewpoints; the brightness and contrast of XCT 2D slice images are adjusted, and color changes are added to simulate the effect of different light sources or imaging parameters on XCT 2D slice images; noise is added to XCT slice images and sharpening is performed to simulate imaging noise in the real world and enhance feature edges.
[0021] Furthermore, in step 3, the network weights of the SAGAN model are constrained by spectral normalization, and mode collapse is suppressed by combining WGAN-GP and autoencoder pre-training.
[0022] Step 4 further includes:
[0023] We collected and organized real, augmented, and high-fidelity virtual samples for training, ensuring uniformity in image format, resolution, and naming conventions. Using the LabelMe annotation tool, we marked the layered defect regions in each image and performed spatial registration by inferring the actual fracture location from the fracture morphology of the failed specimen and the DIC strain field, generating corresponding polygon annotation files. We then performed quality checks on all annotation results, removing blurry, missing, or mislabeled samples to form an image-label pairing dataset.
[0024] Furthermore, in step 5, the training process of the enhanced U-Net model includes:
[0025] First, the model is warmed up with unenhanced real samples and a warmup learning rate to establish basic feature representations. Then, high-fidelity virtual samples and enhanced data are gradually introduced to alleviate inter-domain distribution bias through the learning process. Finally, cosine annealing or step decay learning rate is combined to jointly optimize Dice Loss and boundary-sensitive Focal Loss, thereby improving the enhanced U-Net model's sensitivity to minor defects, including low contrast and blurred edges, segmentation accuracy, and industrial generalization ability.
[0026] Furthermore, in step 5, the image-label pairing dataset is divided into training set, validation set and test set according to the proportion, and the hierarchical defect recognition model is trained, validated and tested; wherein, the various types of hierarchical defect samples are evenly distributed in each subset.
[0027] Furthermore, in step 5, after the enhanced U-Net model is trained, the model performance is comprehensively evaluated using a test set that has not participated in any training or parameter tuning process; specifically, the accuracy of the enhanced U-Net model in identifying layered defect regions is measured by calculating the Dice loss coefficient and the MIoU quantitative index.
[0028] Furthermore, in step 5, during the training process, a configurable Xavier, Kaiming, or orthogonal initialization method is used, and the batch normalization layer is explicitly parameterized.
[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0030] First, the deep learning-based layered defect identification method for ceramic matrix composites of the present invention introduces a SAGAN generation mechanism guided by a real layered defect mask. The extracted prior information on the spatial distribution of defects is used as a conditional input, which effectively constrains the spatial semantic consistency of the generation process. This not only significantly improves the realism and diversity of the synthesized virtual defect samples in terms of morphology, texture and spatial location, but also enhances the intra-domain consistency between the generated data and the actual XCT image. Thus, while alleviating the problem of scarce labeled samples in the field of ceramic matrix composites, it effectively narrows the distribution gap between the generated domain and the real domain, providing high-quality and high-fidelity training support for subsequent segmentation models.
[0031] Secondly, the deep learning-based method for identifying layered defects in ceramic matrix composites proposed in this invention utilizes a pre-trained VGG network as the encoder backbone. It optimizes skip connections using a channel-space dual attention gating mechanism and integrates a multi-scale dilated convolution module at the decoder end, enhancing the recovery of shallow spatial details while preserving deep semantic information. This design is specifically tailored to the characteristics of ceramic matrix composites, such as fiber interweaving, dense matrix, and complex background textures. It effectively distinguishes non-critical structures like micron-level pores and process clutter from target layered defects, alleviating multi-scale feature confusion and significantly improving the perception and localization capabilities for low-contrast, blurred-edge sub-millimeter-level sheet-like or diffuse layered defects. Thanks to the high-fidelity, pixel-level labeled virtual samples generated by mask-guided SAGAN—whose morphology and distribution strictly follow the physical laws of real defects in ceramic matrix composites—the model can fully learn the structural priors and boundary characteristics of defects during training. This effectively compensates for the insufficient representation caused by the scarcity of actual labeled data, further improving segmentation accuracy and cross-sample generalization performance, and reducing the risk of false detections and false negatives in industrial inspection.
[0032] Third, the deep learning-based method for identifying layered defects in ceramic matrix composites employed in this invention utilizes a three-stage collaborative training strategy. This strategy starts with pure real samples, gradually fusing enhanced data and high-fidelity virtual samples, and combines warmup learning rate preheating, course learning scheduling, and cosine annealing or step-wise learning rate decay mechanisms. This achieves a smooth transition from basic feature learning to generalization in complex scenarios. This strategy significantly enhances the model's training stability under small sample conditions, improves its consistent performance under different production lines, process parameters, or imaging conditions, and overall enhances the system's robustness, adaptability, and practicality in complex industrial environments, better meeting the comprehensive requirements of high reliability, high efficiency, and high precision for non-destructive testing of ceramic matrix composites. Attached Figure Description
[0033] Figure 1 CT image of internal delamination defects in a typical L-shaped ceramic matrix composite component; Figure 2 Augment the graph for the data; Figure 3 A schematic diagram illustrating the process of generating samples for SAGAN; Figure 4 This is a schematic diagram of the VGG backbone network; Figure 5 For training parameter curves; Figure 6 This is a diagram illustrating the effect of layered defect identification. Figure 7 This is a schematic diagram of the training process of the enhanced U-Net model of the present invention. Detailed Implementation
[0034] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0035] This invention discloses a method for identifying delamination defects in ceramic matrix composites based on deep learning, the method comprising the following steps:
[0036] Step 1: High-resolution X-ray computed tomography (CT) scans were performed on ceramic matrix composite samples with typical ply structures to obtain voxel-level three-dimensional volume data. Three-dimensional reconstruction was completed using filtered back projection or iterative reconstruction algorithms to generate a complete volume dataset containing information on the internal microstructure of the material. Two-dimensional XCT slice images, including cross-sectional, longitudinal, and arbitrarily angled slice images, were extracted as real samples to comprehensively characterize the geometric morphology of the interlayer interface and the spatial distribution characteristics of potential defects.
[0037] Step 2: Perform data augmentation on the XCT 2D slice images extracted in Step 1 to obtain augmented samples;
[0038] Step 3: Using the spatial distribution features of potential defects extracted in Step 1 and the enhanced samples obtained in Step 2 as prior conditions, a mask-guided conditional generation framework is constructed. A SAGAN model with an integrated self-attention mechanism is adopted to model long-distance spatial dependence in the generator and discriminator, reproducing the extensibility and diffusion characteristics of layered defects. Using real XCT two-dimensional slice images and their corresponding defect masks as supervision signals, the SAGAN model is guided to generate high-fidelity virtual samples with sheet-like or diffuse layered structures only in the defect area, so that their texture, shape and spatial distribution are consistent with the real defects.
[0039] Step 4: Collect and organize real samples, augmented samples, and high-fidelity virtual samples for training. Label the layered defect regions in each image to generate corresponding polygon annotation files and form an image-label pairing dataset.
[0040] Step 5: Construct an enhanced U-Net model, which includes an encoder, a decoder, and a classification head; take XCT two-dimensional slice images as input and pixel-level defect segmentation results as output; combine real samples, enhanced samples, and high-fidelity virtual samples to construct an image-label pairing dataset to train the enhanced U-Net model, so as to improve the enhanced U-Net model's context-aware ability for lamellar and diffuse layered defects;
[0041] The encoder employs a pre-trained VGG network, extracting features from low to high layers of the XCT 2D slice image through convolution and max pooling operations. Low-layer feature maps highlight the microstructural details of the material, while mid-to-high-layer feature maps, with increased depth, focus on capturing the overall structural characteristics of the composite material and its potential layering interfaces, clearly displaying layered defects against complex backgrounds. The decoder of the enhanced U-Net model receives a deep feature map with a resolution of 1 / 32 of the original input from the encoder, then gradually restores the spatial resolution through upsampling and skip connections, fusing detailed information from the encoder's corresponding layers to ultimately generate a high-resolution feature map to reconstruct the morphology of layered defects. A channel-spatial dual attention gating mechanism optimizes skip connections, enhancing the U-Net model's ability to focus on key features, highlighting information sensitive to layered defects, and distinguishing between real defects and process clutter interference structures. The classification head captures contextual information, achieving pixel-level classification through 1×1 convolutional layers to generate corresponding defect region segmentation maps.
[0042] Step 6: Use the trained enhanced U-Net model to identify layered defect regions in the XCT slice images of the ceramic matrix composite material to be tested.
[0043] The following section uses a common L-shaped CMC structure as a specific example to illustrate the training process of the enhanced U-Net model and the hierarchical defect identification process of this invention. (See also...) Figure 7 The model training and recognition process of this invention includes:
[0044] Step 1: Perform high-resolution XCT scanning on ceramic matrix composite samples with typical ply structures to obtain voxel-level three-dimensional volume data; complete three-dimensional reconstruction through filtered back projection or iterative reconstruction algorithms to generate a volume dataset containing information on the internal microstructure of the material; on this basis, extract cross-sectional (XY), longitudinal (XZ, YZ) and oblique slice images at arbitrary angles to comprehensively characterize the geometry of interlayer interfaces and the distribution of potential defects. Figure 1 CT image of internal delamination defects in a typical L-shaped ceramic matrix composite component;
[0045] Step 2: Rotate and translate the XCT 2D slice image to simulate the observation effect from different perspectives; adjust the brightness and contrast of the XCT 2D slice image and add color changes to simulate the effect of different light sources or imaging parameters on the XCT 2D slice image; add noise and perform sharpening processing on the XCT slice image to simulate imaging noise in the real world and enhance feature edges. Figure 2 To augment the data.
[0046] Step 3: Using the spatial distribution features of defects extracted in Step 1 and the data-enhanced samples obtained in Step 2 as prior conditions, a mask-guided conditional generation framework is constructed. A SAGAN model with an integrated self-attention mechanism is employed to model long-range spatial dependencies in both the generator and discriminator, accurately reproducing the extensibility and diffusion characteristics of layered defects. Spectral normalization constrains the network weights, and WGAN-GP and autoencoder pre-training suppress mode collapse. Using real images and their corresponding defect masks as supervisory signals, the model is guided to generate high-fidelity virtual samples only in defect regions, ensuring that their texture, morphology, and spatial distribution are consistent with real defects. This effectively expands the U-Net training data and alleviates the overfitting problem caused by small sample sizes. Figure 3 A schematic diagram of the process of generating samples for SAGAN.
[0047] Step 4: Collect and organize real, augmented, and virtual samples for training, ensuring uniformity in image format, resolution, and naming conventions; use the LabelMe annotation tool to mark the layered defect regions in each image, and perform spatial registration by combining the actual fracture location obtained by back-calculating the fracture morphology of the failed specimen with the DIC strain field to generate corresponding polygon annotation files; perform quality checks on all annotation results, removing blurry, missing, or mislabeled samples to form an image-label pairing dataset.
[0048] Step 5: Construct an enhanced U-Net model for delamination defect detection in ceramic matrix composites, using XCT 2D slice images as input and pixel-level defect masks as output. To address the issues of the standard U-Net in this scenario, such as loss of spatial information of sub-millimeter-level defects, multi-scale feature confusion, and overfitting due to small sample sizes, this step resolves these problems through collaborative optimization of the network architecture and data strategies.
[0049] In terms of network structure, the following is adopted: Figure 4 The VGG network shown serves as the encoder backbone, incorporating a hybrid architecture design that integrates the advantages of generation and segmentation. A channel-spatial dual attention gating mechanism optimizes skip connections, and a multi-scale dilated convolution module is integrated at the decoder end. This enhances the recovery of shallow spatial details while preserving deep semantic information, resulting in an enhanced U-Net model for detecting layered defects in ceramic matrix composites. Specifically, in the enhanced U-Net model, the focus when processing XCT 2D images for layered defect detection in ceramic matrix composites is to effectively extract and characterize the unique layered defect features of this type of material. Due to the characteristics of ceramic matrix composites—densely interwoven fibers, a dense matrix, and complex background textures—accurately locating sub-millimeter-level lamellar or diffuse layered defects is quite challenging. This process first utilizes a pre-trained VGG network as the encoder, extracting features from low to high layers of the image through convolution and max pooling operations. Low-layer feature maps highlight the microstructural details of the material, which is crucial for identifying subtle defects; while as the layers deepen, mid-to-high-layer feature maps focus more on capturing the overall structural characteristics of the composite material and its potential layering interfaces, clearly displaying layered defects against a complex background.
[0050] In the decoder stage of the enhanced U-Net model, the model first obtains a deep feature map with a resolution of 1 / 32 of the original input. Then, it gradually recovers the spatial resolution through upsampling and skip connections, and fuses detailed information from the corresponding encoder layers to finally generate a high-resolution feature map of size (512, 512, 64). This process helps to accurately reconstruct the morphology of layered defects. Furthermore, the channel-spatial dual attention gating mechanism enhances the model's ability to focus on key features, effectively highlighting information sensitive to layered defects, thereby better distinguishing real defects from interference structures such as process clutter.
[0051] The classification head enables the enhanced U-Net model to capture broader contextual information while maintaining resolution. This is particularly important for identifying and segmenting diffuse or layered defects and can mitigate the multi-scale feature confusion caused by complex material structures. Pixel-level classification is achieved through 1×1 convolutional layers, generating detailed defect region segmentation maps to provide a basis for material quality assessment.
[0052] At the data level, by combining high-quality real samples with the virtual defect data generated in step 3, a more representative image-label pairing dataset is constructed. This enhances the model's contextual awareness of layered and diffuse layered defects, effectively mitigating the overfitting problem. During training, the image-label pairing dataset is first divided into training, validation, and test sets proportionally to train, validate, and test the enhanced U-Net model. The distribution of various types of layered defect samples is balanced across the subsets.
[0053] Preferably, the enhanced U-Net model employs a three-stage progressive learning approach: first, it warms up with unenhanced real samples using a warmup learning rate to establish basic feature representations; then, it gradually introduces high-fidelity virtual samples generated in step 3 and enhanced data, mitigating inter-domain distribution bias through the learning process; finally, it combines cosine annealing or a stepped decay learning rate to jointly optimize DiceLoss and boundary-sensitive Focal Loss, significantly improving the model's sensitivity to minor defects such as low contrast and blurred edges, segmentation accuracy, and industrial generalization ability. During training, configurable Xavier, Kaiming, or orthogonal initialization methods are used, and explicit parameter settings are applied to the batch normalization layers. Figure 5 The training parameter curves are shown. After the enhanced U-Net model is trained, its performance is comprehensively evaluated using a test set that has not participated in any training or parameter tuning process; specifically, the accuracy of the hierarchical defect recognition model in recognizing hierarchical defect regions is measured by calculating the Dice loss coefficient and the MIoU quantitative index.
[0054] Step 7: Use the trained layered defect recognition model to identify layered defect regions in the XCT slice image of the ceramic matrix composite material to be tested. Figure 6 This is a diagram illustrating the effect of layered defect identification in this embodiment.
[0055] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0056] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for identifying delamination defects in ceramic matrix composites based on deep learning, characterized in that, The method includes the following steps: Step 1: High-resolution X-ray computed tomography (CT) scans were performed on ceramic matrix composite samples with typical ply structures to obtain voxel-level three-dimensional volume data. Three-dimensional reconstruction was completed using filtered back projection or iterative reconstruction algorithms to generate a complete volume dataset containing information on the internal microstructure of the material. Two-dimensional XCT slice images, including cross-sectional, longitudinal, and arbitrarily angled slice images, were extracted as real samples to comprehensively characterize the geometric morphology of the interlayer interface and the spatial distribution characteristics of potential defects. Step 2: Perform data augmentation on the XCT 2D slice images extracted in Step 1 to obtain augmented samples; Step 3: Using the spatial distribution features of potential defects extracted in Step 1 and the enhanced samples obtained in Step 2 as prior conditions, a mask-guided conditional generation framework is constructed. A SAGAN model with an integrated self-attention mechanism is adopted to model long-distance spatial dependence in the generator and discriminator, reproducing the extensibility and diffusion characteristics of layered defects. Using real XCT two-dimensional slice images and their corresponding defect masks as supervision signals, the SAGAN model is guided to generate high-fidelity virtual samples with sheet-like or diffuse layered structures only in the defect area, so that their texture, shape and spatial distribution are consistent with the real defects. Step 4: Collect and organize real samples, augmented samples, and high-fidelity virtual samples for training. Label the layered defect regions in each image to generate corresponding polygon annotation files and form an image-label pairing dataset. Step 5: Construct an enhanced U-Net model, which includes an encoder, a decoder, and a classification head; take XCT two-dimensional slice images as input and pixel-level defect segmentation results as output; combine real samples, enhanced samples, and high-fidelity virtual samples to construct an image-label pairing dataset to train the enhanced U-Net model, so as to improve the enhanced U-Net model's context-aware ability for lamellar and diffuse layered defects; The encoder employs a pre-trained VGG network, extracting features from low to high layers of the XCT 2D slice image through convolution and max pooling operations. Low-layer feature maps highlight the microstructural details of the material, while mid-to-high-layer feature maps, with increased depth, focus on capturing the overall structural characteristics of the composite material and its potential layering interfaces, clearly displaying layered defects against complex backgrounds. The decoder of the enhanced U-Net model receives a deep feature map with a resolution of 1 / 32 of the original input from the encoder, then gradually restores the spatial resolution through upsampling and skip connections, fusing detailed information from the encoder's corresponding layers to ultimately generate a high-resolution feature map to reconstruct the morphology of layered defects. A channel-spatial dual attention gating mechanism optimizes skip connections, enhancing the U-Net model's ability to focus on key features, highlighting information sensitive to layered defects, and distinguishing between real defects and process clutter interference structures. The classification head captures contextual information, achieving pixel-level classification through 1×1 convolutional layers to generate corresponding defect region segmentation maps. Step 6: Use the trained enhanced U-Net model to identify layered defect regions in the XCT slice images of the ceramic matrix composite material to be tested.
2. The method for identifying delamination defects in ceramic matrix composites based on deep learning according to claim 1, characterized in that, Step 2 involves performing data augmentation on the XCT 2D slice images extracted in Step 1, including: Rotation and translation operations are performed on XCT 2D slice images to simulate the observation effect under different viewpoints; the brightness and contrast of XCT 2D slice images are adjusted, and color changes are added to simulate the effect of different light sources or imaging parameters on XCT 2D slice images; noise is added to XCT slice images and sharpening is performed to simulate imaging noise in the real world and enhance feature edges.
3. The method for identifying delamination defects in ceramic matrix composites based on deep learning according to claim 1, characterized in that, In step 3, the network weights of the SAGAN model are constrained by spectral normalization, and mode collapse is suppressed by combining WGAN-GP and autoencoder pre-training.
4. The method for identifying delamination defects in ceramic matrix composites based on deep learning according to claim 1, characterized in that, Step 4 further includes: We collected and organized real, augmented, and high-fidelity virtual samples for training, ensuring uniformity in image format, resolution, and naming conventions. Using the LabelMe annotation tool, we marked the layered defect regions in each image and performed spatial registration by inferring the actual fracture location from the fracture morphology of the failed specimen and the DIC strain field, generating corresponding polygon annotation files. We then performed quality checks on all annotation results, removing blurry, missing, or mislabeled samples to form an image-label pairing dataset.
5. The method for identifying delamination defects in ceramic matrix composites based on deep learning according to claim 1, characterized in that, Step 5, the training process of the enhanced U-Net model includes: First, the model is warmed up with unenhanced real samples and a warmup learning rate to establish basic feature representations. Then, high-fidelity virtual samples and enhanced data are gradually introduced to alleviate inter-domain distribution bias through the learning process. Finally, cosine annealing or step decay learning rate is combined to jointly optimize Dice Loss and boundary-sensitive Focal Loss, thereby improving the enhanced U-Net model's sensitivity to minor defects, including low contrast and blurred edges, segmentation accuracy, and industrial generalization ability.
6. The method for identifying delamination defects in ceramic matrix composites based on deep learning according to claim 1, characterized in that, In step 5, the image-label pairing dataset is divided into training set, validation set and test set according to the proportion, and the hierarchical defect recognition model is trained, validated and tested; among them, the various types of hierarchical defect samples are evenly distributed in each subset.
7. The method for identifying delamination defects in ceramic matrix composites based on deep learning according to claim 5, characterized in that, In step 5, after the enhanced U-Net model is trained, the model performance is comprehensively evaluated using a test set that has not participated in any training or parameter tuning process; specifically, the accuracy of the enhanced U-Net model in identifying layered defect regions is measured by calculating the Dice loss coefficient and the MIoU quantitative index.
8. The method for identifying delamination defects in ceramic matrix composites based on deep learning according to claim 1, characterized in that, In step 5, during the training process, configurable Xavier, Kaiming, or orthogonal initialization methods are used, and explicit parameter settings are made for the batch normalization layer.