Real-time damage prediction method for ceramic matrix composites based on deep learning

Abstract

The present invention belongs to the technical field of material structure health monitoring, and specifically provides a method for predicting the real-time damage degree of ceramic matrix composite materials based on deep learning, including the following steps: establishing an initial damage data set; building an initial damage model, and driving it in combination with physical constraints; establishing a process damage data set for ceramic matrix composite materials; establishing a noise filtering model, and performing noise-free and normalized processing; building a process damage model, and driving it in combination with the damage mechanism; coupling initial damage and process damage to establish a real-time damage degree mapping model, and using the least squares method and nonlinear optimization to fit parameters; applying the mapping model to predict the real-time damage degree. The present invention is simple and easy to operate, and has high efficiency. By integrating initial defect characteristics, service process damage and physical parameters into the machine learning model, the constructed physical information-deep learning model obtains better physical consistency, powerful real-time health monitoring capabilities and higher prediction robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of material structure health monitoring, and specifically provides a real-time damage degree prediction method for ceramic-based composite materials based on deep learning. Background Art

[0002] Faced with the increasing requirements for high temperature and lightweight of aircraft structures, ceramic matrix composites have attracted much attention in the aerospace field due to their excellent properties of low density, high specific strength and strong temperature resistance. However, their inherent dispersion and brittleness severely limit their large-scale application in structural components, such as radomes, turbine blades and thermal protection structures. Therefore, the intrinsic characteristics of dispersion and brittleness of ceramic matrix composites should be studied in depth to promote their large-scale application.

[0003] The mechanical properties of composite materials depend on their component properties, fiber volume fraction, weaving structure and loading conditions. Even under the same weaving structure, raw materials and process conditions, ceramic matrix composites still show significant performance dispersion in the same service environment. In order to explore the influencing factors of dispersion, people have studied the performance dispersion of ceramic matrix composites from different perspectives. By constructing a representative volume element model and combining the Weibull distribution, neural network and Monte Carlo simulation methods, the researchers explored the influence of component performance distribution and manufacturing defect uncertainty on the performance dispersion of ceramic matrix composites. These studies mainly explore the influence of changes in component properties, weaving structure and micro defects on mechanical properties, and lack research on the quantitative evaluation of the damage state of specific ceramic matrix composite structures. In addition, a large number of studies that explore the influence of material variability through defect changes focus on simulating materials using representative volume elements with simple shape defects. This has achieved certain results, but the ongoing challenge is to introduce real complex defects and quantify their influence on the damage state of ceramic matrix composites.

[0004] The reinforcement and toughening effect of fibers improves the brittleness of ceramic matrix composites to a certain extent, but there is still the problem of no obvious characteristics before damage, which brings challenges to real-time monitoring. In order to ensure the real-time integrity of ceramic matrix composite structures, many scholars have developed continuous state monitoring technology, especially technology based on acoustic emission. Acoustic emission signals contain information on the initiation and evolution of damage. Accurate identification and separation of different damage sources based on acoustic emission signal analysis has become a development trend in composite damage pattern recognition. Existing studies usually select amplitude and frequency as clustering indicators of acoustic emission signals to identify the damage mechanism of ceramic matrix composites in real time. This shows that in the field of damage diagnosis, significant research progress has been made in identifying the damage mode of composite materials using acoustic emission signal characteristics. However, research on quantitative assessment of composite damage based on deep learning methods is still very limited, which makes real-time health status assessment face bottlenecks.

[0005] Brittleness and dispersion are important factors that hinder the effective monitoring of the damage degree of ceramic matrix composites, so it is necessary to further explore how to conduct real-time quantitative evaluation of damage during service. Deep learning has made great progress in the field of performance prediction of ceramic matrix composites, and has applications in strength and life prediction, but there are few reports on the prediction of real-time damage degree of ceramic matrix composites; on the other hand, multi-source fusion methods for multiple monitoring methods have not yet been established. The powerful data fusion capability of deep learning models provides new ideas for dynamically extracting and quantifying the damage accumulation and evolution process of materials under different load conditions.

[0006] In addition, in order to improve the interpretability and physical consistency of deep learning models, deep learning models that are bidirectionally driven by physical information and data information also need to be vigorously developed. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a real-time damage degree prediction method for ceramic-based composite materials based on deep learning.

[0008] The present invention is implemented by providing a real-time damage degree prediction method for ceramic matrix composite materials based on deep learning, comprising the following steps:

[0009] S1: X-ray computed tomography is used to reconstruct the structure of ceramic matrix composites. The initial defects of ceramic matrix composites are segmented based on image segmentation methods to identify the spatial characteristics of defects. The extracted defective images are used to establish the initial damage data set.

[0010] S2: Construct an initial damage model based on a U-shaped neural network, drive it in combination with physical constraints, input the initial damage data set into the initial damage model, output the initial damage factor, and predict the initial damage degree of the material;

[0011] S3: Based on the acoustic emission method, the damage signal of the ceramic matrix composite material in the service process is obtained, the time domain characteristic parameters and frequency domain characteristic parameters of the damage signal are extracted, the evolution of the defect characteristics over time is characterized, the defect time characteristics are obtained, and the process damage data set is constructed;

[0012] S4: Establish a noise filtering model to de-noise and normalize the process damage data set;

[0013] S5: Using the improved K-means unsupervised learning algorithm, the damage pattern is extracted from the process damage data set processed by S4, and then the artificial neural network is used to establish a nonlinear mapping between the damage pattern and the degree of material process damage, and a process damage model is constructed. The output of the process damage model is the process damage factor, and the process damage degree of the ceramic matrix composite material is obtained;

[0014] S6: Based on the damage accumulation theory, the initial damage model and the process damage model are coupled to construct a mapping model between the defect spatial characteristics, defect temporal characteristics and the real-time damage degree of ceramic matrix composites, and the least squares method and nonlinear optimization are used to fit the model parameters to obtain the optimal mapping model;

[0015] S7: Use the trained optimal mapping model to predict the real-time damage degree of ceramic matrix composites.

[0016] Preferably, step S1 specifically includes the following steps:

[0017] S1.1: Obtain high-resolution three-dimensional images of the ceramic matrix composite material by X-ray computed tomography to form a reconstructed data set of the internal structure of the ceramic matrix composite material;

[0018] S1.2: Based on the gray value difference of the three-dimensional image, the interactive thresholding method is used to perform preliminary segmentation of the initial defects in the ceramic matrix composite material to determine the initial damage area;

[0019] S1.3: For areas that are difficult to accurately segment due to grayscale value overlap, the Interactive Top-Hat method is used to enhance the contrast of tiny structural features through morphological operations to further extract initial defect features;

[0020] S1.4: Combining the results of the Interactive Thresholding and Interactive Top-Hat methods, the segmentation results of the 3D image are optimized by algorithm fusion to ensure that all initial defect spatial features in the ceramic matrix composite are accurately identified;

[0021] S1.5: The 3D image segmentation results optimized by S1.4 are scaled using the nearest neighbor interpolation method and transformed to a uniform standard size to match the scanning sizes of different ceramic matrix composite samples. After the size adjustment is completed, the initial damage data set is established based on the defective section of the sample gauge section to ensure that the data between different samples are comparable.

[0022] Further preferably, in step S1.4, the initial defect space characteristics include shape, size, and position parameters of microcracks, pores, and matrix voids.

[0023] Further preferably, in step S2, the physical constraints include porosity, large defects adjacent to the surface, and uniformity of defect distribution. The initial damage model adjusts the learning emphasis of these three physical constraints through a loss function, optimizes the training of the U-type neural network, and obtains the optimal initial damage model.

[0024] Further preferably, in step S3, the time domain characteristic parameters of the damage signal include amplitude, duration, rise time, ringing count, rise count, energy, effective value voltage and average signal level; the frequency domain characteristic parameters are extracted by Fourier transform, including centroid frequency and peak frequency.

[0025] Further preferably, in step S4, the Mahalanobis distance method is combined with the characteristic parameters of the noise sample, and a noise filtering model is established by an analytical method of setting a threshold, which specifically includes the following steps:

[0026] The mathematical definition of Mahalanobis distance is as follows:

[0027] (1)

[0028] in, D M ( n ) is the Mahalanobis distance, n is the feature vector of the data sample, μ is the mean vector of the data distribution, S is the covariance matrix of the data;

[0029] Define the distribution characteristics of the noise samples:

[0030] (2)

[0031] (3)

[0032] in, N is the number of noise samples, It is i The feature vector of the samples, and is the mean vector and co-defense difference matrix of the noise samples;

[0033] The threshold is defined using the Mahalanobis distance formula:

[0034] (4)

[0035] in α is the confidence coefficient, usually 1;

[0036] The Mahalanobis distance between the process damage data set recorded during the loading phase and the center of the noise filter model is calculated to determine whether it is noise:

[0037] (5)

[0038] If the above conditions are met, the sample is regarded as noise, otherwise it is regarded as valid data;

[0039] For the non-noise signal filtered by the noise filter model, the Z-score standardization method is used for normalization. The specific formula is:

[0040] (6)

[0041] in, X and Z represents the filtered signal samples before and after normalization, μ X and σ are the sample mean and standard deviation respectively.

[0042] Further preferably, step S5 specifically comprises the following steps:

[0043] S5.1: Determine the number of damage patterns and the optimal number of clusters for the process damage dataset using the elbow method and the silhouette coefficient method H , the elbow method calculates the number of each possible cluster k The corresponding total internal variance, select the point where the variance decreases significantly k As the optimal number of clusters, the specific calculation method is as follows:

[0044] (7)

[0045] in, is the total internal variance, k is the number of clusters considered, It is i The sample set in the cluster, m is a single sample in the cluster, It is i The centroid of the clusters;

[0046] The silhouette coefficient method measures the compactness of a data point within its cluster and its separation from other clusters. The clustering effect is evaluated by the size of the silhouette coefficient. The specific calculation formula is as follows:

[0047] (8)

[0048] Among them, among them, It is a sample i The average distance to other samples in its cluster, It is a sample i The average distance to all samples in the nearest cluster, G is the number of clusters;

[0049] S5.2: Optimize the K-means unsupervised learning algorithm by selecting new centroids based on the probability of the distance between the sample and the selected centroid, optimize the initial conditions of clustering and the global optimality of the results, and use the automatic variable weighting method to optimize the clustering feature weights. The weight adjustment uses the following formula:

[0050] (9)

[0051] in, It is i The weight of the feature, It is i The feature in j The variance in the clusters;

[0052] S5.3: Using the results in steps S5.1 and S5.2, an artificial neural network is used to learn the nonlinear relationship between different damage modes and the degree of material process damage. The input layer receives the clustering results obtained from the improved K-means unsupervised learning algorithm, and the output layer is used to predict the process damage factor of the material. A process damage model is established, and the process damage factor is used to represent the degree of process damage of different damage modes to the ceramic matrix composite material.

[0053] S5.4: Physical constraint optimization model, combined with Pearson correlation analysis, establishes a loss function to ensure the positive correlation between the process damage factor and the damage signal amplitude. Through iterative training, the network weights and biases are systematically adjusted to obtain the optimal process damage model.

[0054] Further preferably, in step S6, based on the damage accumulation theory, linear superposition is performed to couple the initial damage model and the process damage model, specifically:

[0055] (10)

[0056] in, x is the initial damage factor, y is the process damage factor, a and b are the fitting parameters;

[0057] The least square method and nonlinear optimization are used to fit the mapping model parameters and obtain the optimal mapping model.

[0058] Compared with the prior art, the advantages of the present invention are:

[0059] The present invention proposes a real-time damage degree prediction method for ceramic matrix composite materials based on deep learning, which integrates high-resolution three-dimensional image data obtained by X-CT technology and damage signals collected by acoustic emission technology, and realizes the accurate extraction of initial defect characteristics of materials and damage behavior modeling during service. By combining U-type neural network, Mahalanobis distance noise filtering and improved K-means clustering algorithm, this method can realize real-time and high-precision damage degree prediction under various damage modes, complex load conditions and different defect distributions. Compared with the traditional defect feature-performance correlation model, the present invention has the ability to quantify the impact of damage on real-time damage degree in real time and dynamically, and significantly improves the monitoring accuracy of material health status. The model has strong robustness and can cope with various abnormal situations in actual working conditions, providing reliable technical support for the field of health monitoring. The model is suitable for real-time damage degree prediction of ceramic matrix composite materials under various complex conditions and has broad application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A flow chart of the method provided by the present invention;

[0061] Figure 2 It is a schematic diagram of the initial defect segmentation effect of the sample in step S1 of the present invention;

[0062] Figure 3 It is a framework diagram of the initial damage model in step S2 of the present invention;

[0063] Figure 4 It is a denoising effect diagram of the noise filtering model in step S4 of the present invention;

[0064] Figure 5 Schematic diagram of damage pattern extracted from acoustic emission signal in step S5 of the present invention;

[0065] Figure 6 This is a diagram showing the mapping model fitting effect in step S6 of the present invention;

[0066] Figure 7 This is a schematic diagram of predicting the fracture point of a new sample based on the real-time damage degree mapping model of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the following is a diagram of SiO 2f / SiO 2 Taking the real-time damage evaluation in room temperature and high temperature tensile tests of ceramic matrix composite materials as an example, the specific embodiments of the present invention are further described in detail. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0068] The present invention provides a real-time damage prediction method for ceramic matrix composite materials based on deep learning, see Figure 1 , is the SiO provided by the embodiment of the present invention 2f / SiO 2 A simplified flow chart of the real-time damage assessment method in room temperature and high temperature tensile tests of ceramic matrix composites. The specific steps are:

[0069] S1: X-ray computed tomography is used to reconstruct the structure of ceramic matrix composites. The initial defects of ceramic matrix composites are accurately segmented based on image segmentation methods to identify the spatial characteristics of defects. The extracted defective images are used to establish the initial damage data set.

[0070] S2: Construct an initial damage model based on a U-shaped neural network, drive it in combination with physical constraints, input the initial damage data set into the initial damage model, output the initial damage factor, and predict the initial damage degree of the material;

[0071] S3: Based on the acoustic emission method, the damage signal of the ceramic matrix composite material in the service process is obtained, the time domain characteristic parameters and frequency domain characteristic parameters of the damage signal are extracted, the evolution of the defect characteristics over time is characterized, the defect time characteristics are obtained, and the process damage data set is constructed;

[0072] S4: Establish a noise filtering model to de-noise and normalize the process damage data set;

[0073] S5: Using the improved K-means unsupervised learning algorithm, the damage pattern is extracted from the process damage data set processed by S4, and then the artificial neural network is used to establish a nonlinear mapping between the damage pattern and the degree of material process damage, and a process damage model is constructed. The output of the process damage model is the process damage factor, and the process damage degree of the ceramic matrix composite material is obtained;

[0074] S6: Based on the damage accumulation theory, the initial damage model and the process damage model are coupled to construct a mapping model between the defect spatial characteristics, defect temporal characteristics and the real-time damage degree of ceramic matrix composites, and the least squares method and nonlinear optimization are used to fit the model parameters to obtain the optimal mapping model;

[0075] S7: Use the trained optimal mapping model to predict the real-time damage degree of ceramic matrix composites.

[0076] In a specific embodiment, the initial defect and damage signals in steps S1 and S3 are from samples at room temperature and high temperature.

[0077] In a specific embodiment, steps S2-S7 are implemented by programming in Python.

[0078] In a specific embodiment, the training set of the deep learning model described in steps S2-S7 is 2 room temperature tests and 2 high temperature tests, and the validation set is 1 high temperature test.

[0079] In order to demonstrate the application of the technical means of the present invention more clearly, the present invention has actually applied the material data. 2f / SiO 2 Taking ceramic-based composite materials as an example, the above technical means are used to realize real-time prediction of the real-time damage degree of ceramic-based composite materials.

[0080] S1. X-ray microfocus CT system (Phoenix V|tome|x M300, Baker Hughes) was used to analyze the 5 SiO 2f / SiO 2 X-CT scan of composite material sample before stretching. The fusion technology of Interactive Thresholding and Interactive Top-Hat method is used to accurately segment the initial defects of the sample. Analysis Figure 2 The porosity changes with different segmentation thresholds as shown in Figure 1. The segmentation threshold of 12760 corresponding to the platform area with low threshold sensitivity is selected to extract the initial defects. The extraction results are shown in Figure 1. Figure 2 As shown in the figure, the extracted defect image was scaled using the nearest neighbor interpolation method to transform it to a standard size of 1300 pixels × 620 pixels. Finally, 660 cross-sectional slices containing initial defect information were obtained, and the initial damage dataset for subsequent analysis was constructed.

[0081] S2. Construct an initial damage model based on a U-shaped neural network. The framework is as follows Figure 3 As shown in the figure. The model structure uses multiple layers of convolutional layers and maximum pooling layers to extract and compress image features, and each layer is followed by a ReLU activation function to enhance nonlinear processing capabilities. The high-resolution features extracted from the encoder are directly transferred to the corresponding layer of the decoder through jump connections to retain key local information, while the decoder uses upsampling and convolution operations to gradually restore the image size and integrate the transferred features to improve the image reconstruction accuracy. In addition, the features output by the decoder are further processed by a convolutional neural network, and the features are refined and integrated through a fully connected layer. Physical constraints such as porosity, large defects adjacent to the surface, and uniformity of defect distribution are introduced, and the learning of these features is weighted and optimized through a loss function. Finally, the Sigmoid activation function is used to output the average value of the initial damage factor of each specimen slice, thereby characterizing the initial damage state of the entire specimen.

[0082] S3. During the room temperature and high temperature stretching process of the sample, an 8-channel DS5-8A acoustic emission detection system was used to collect damage signals. The damage signals monitored by acoustic emission characterized the temporal evolution of defect characteristics and provided key information for the change of defects. The time domain and frequency domain characteristic parameters of the damage signal were used as the input of the model, with a total of 10 dimensional features, to establish a process damage data set. The specific input parameters are shown in Table 1.

[0083] Table 1

[0084]

[0085] S4. At the beginning and end of the room temperature and high temperature tensile tests, noise samples are recorded as noise source data. The noise filtering model is constructed by the analytical method shown in formulas (1)-(5), and the damage signal in the process damage data set is judged and filtered. The noise removal effect is shown in Figure 4 Finally, the Z-score method is used to normalize the denoised process damage data set. The specific calculation formula is shown in formula (6).

[0086] S5. By using the elbow method and the silhouette coefficient method, the optimal number of clusters for the process damage data set is determined to be 3 using formulas (7) and (8). On this basis, based on the improved K-means clustering model, the cluster feature weights are optimized according to formula (9), and the damage modes are divided into three categories: matrix cracking, fiber fracture, and interface debonding, as shown in Figure 5 At the same time, combined with Pearson correlation analysis, a loss function is constructed to ensure that the damage factor and the damage signal amplitude are positively correlated. Through multiple iterative training, the network weights and bias parameters are effectively adjusted, and finally the optimal process damage prediction model is constructed.

[0087] S6. Based on the damage accumulation theory, the linear superposition damage model shown in formula (10) is used to comprehensively apply the initial state assessment and process damage assessment. The mapping model parameters are fitted by the least squares method and nonlinear optimization to obtain the parameters a = 0.8219, b = 0.2018, the fitting results are as follows Figure 6 shown.

[0088] S7. A real-time damage prediction of a new sample of a ceramic matrix composite material is performed using a mapping model based on initial damage and process damage data. The prediction results are shown in Figure 7In. The analysis found that the process damage factor increased significantly at a specific location and was closely related to the stress-strain jitter point, showing the model's ability to accurately identify different damage modes during service. In particular, during the loading process, the model successfully identified the abnormal strain surge caused by premature fiber breakage. The deviation between the fracture strain predicted by the model and the actual measured strain was controlled within 2%, and the stress deviation was controlled within 10%, further verifying the accuracy of the model in predicting the final fracture moment of the material. The model solves the prediction problems caused by the brittleness and dispersion characteristics of ceramic matrix composites through a scientific and effective method, significantly improving the accuracy and reliability of material health monitoring and safety assessment.

[0089] The above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. The above embodiment of the present invention can also be modified in various ways. That is, all simple, equivalent changes and modifications made according to the claims and the description of the present invention application fall within the protection scope of the claims of the present invention.

Claims

1. A real-time damage prediction method for ceramic matrix composites based on deep learning, characterized in that: The steps include: S1: X-ray computed tomography is used to reconstruct the structure of ceramic matrix composites. The initial defects of ceramic matrix composites are segmented based on image segmentation methods to identify the spatial characteristics of defects. The extracted defective images are used to establish the initial damage data set. S2: Construct an initial damage model based on a U-shaped neural network, drive it in combination with physical constraints, input the initial damage data set into the initial damage model, output the initial damage factor, and predict the initial damage degree of the material; S3: Based on the acoustic emission method, the damage signal of the ceramic matrix composite material in the service process is obtained, the time domain characteristic parameters and frequency domain characteristic parameters of the damage signal are extracted, the evolution of the defect characteristics over time is characterized, the defect time characteristics are obtained, and the process damage data set is constructed; S4: Establish a noise filtering model to de-noise and normalize the process damage data set; S5: Using the improved K-means unsupervised learning algorithm, the damage pattern is extracted from the process damage data set processed by S4, and then the artificial neural network is used to establish a nonlinear mapping between the damage pattern and the degree of material process damage, and a process damage model is constructed. The output of the process damage model is the process damage factor, and the process damage degree of the ceramic matrix composite material is obtained; S6: Based on the damage accumulation theory, the initial damage model and the process damage model are coupled to construct a mapping model between the defect spatial characteristics, defect temporal characteristics and the real-time damage degree of ceramic matrix composites, and the least squares method and nonlinear optimization are used to fit the model parameters to obtain the optimal mapping model; S7: Use the trained optimal mapping model to predict the real-time damage degree of ceramic matrix composites.

2. The method for real-time damage degree prediction of ceramic matrix composite materials based on deep learning according to claim 1 is characterized in that: Step S1 specifically includes the following steps: S1.1: Obtain high-resolution three-dimensional images of the ceramic matrix composite material by X-ray computed tomography to form a reconstructed data set of the internal structure of the ceramic matrix composite material; S1.2: Based on the gray value difference of the three-dimensional image, the interactive thresholding method is used to perform preliminary segmentation of the initial defects in the ceramic matrix composite material to determine the initial damage area; S1.3: For areas that are difficult to accurately segment due to grayscale value overlap, the Interactive Top-Hat method is used to enhance the contrast of tiny structural features through morphological operations to further extract initial defect features; S1.4: Combining the results of the Interactive Thresholding and Interactive Top-Hat methods, the segmentation results of the 3D image are optimized by algorithm fusion to ensure that all initial defect spatial features in the ceramic matrix composite are accurately identified; S1.5: The 3D image segmentation results optimized by S1.4 are scaled using the nearest neighbor interpolation method and transformed to a uniform standard size to match the scanning sizes of different ceramic matrix composite samples. After the size adjustment is completed, the initial damage data set is established based on the defective section of the sample gauge section to ensure that the data between different samples are comparable.

3. The method for real-time damage degree prediction of ceramic matrix composite materials based on deep learning according to claim 2 is characterized in that: In step S1.4, the initial defect space characteristics include shape, size, and position parameters of microcracks, pores, and matrix voids.

4. The method for real-time damage degree prediction of ceramic matrix composite materials based on deep learning according to claim 1, characterized in that: In step S2, the physical constraints include porosity, large defects adjacent to the surface, and uniformity of defect distribution. The initial damage model adjusts the learning emphasis of these three physical constraints through a loss function, optimizes the training of the U-type neural network, and obtains the optimal initial damage model.

5. The method for real-time damage degree prediction of ceramic matrix composite materials based on deep learning according to claim 1, characterized in that: In step S3, the time domain characteristic parameters of the damage signal include amplitude, duration, rise time, ring count, rise count, energy, effective value voltage and average signal level; The frequency domain feature parameters are extracted by Fourier transform, including centroid frequency and peak frequency.

6. The method for real-time damage degree prediction of ceramic matrix composite materials based on deep learning according to claim 1, characterized in that: In step S4, the Mahalanobis distance method is combined with the characteristic parameters of the noise sample, and a noise filtering model is established by an analytical method of setting a threshold, which specifically includes the following steps: The mathematical definition of Mahalanobis distance is as follows: （1） in, D M ( n ) is the Mahalanobis distance, n is the feature vector of the data sample, μ is the mean vector of the data distribution, S is the covariance matrix of the data; Define the distribution characteristics of the noise samples: （2） （3） in, N is the number of noise samples, It is i The feature vector of the samples, and is the mean vector and co-defense difference matrix of the noise samples; The threshold is defined using the Mahalanobis distance formula: （4） in α is the confidence coefficient, usually 1; The Mahalanobis distance between the process damage data set recorded during the loading phase and the center of the noise filter model is calculated to determine whether it is noise: （5） If the above conditions are met, the sample is regarded as noise, otherwise it is regarded as valid data; For the non-noise signal filtered by the noise filter model, the Z-score standardization method is used for normalization. The specific formula is: （6） in, X and Z represents the filtered signal samples before and after normalization, μ X and σ are the sample mean and standard deviation respectively.

7. The method for real-time damage degree prediction of ceramic matrix composite materials based on deep learning according to claim 1, characterized in that: Step S5 specifically includes the following steps: S5.1: Determine the number of damage patterns and the optimal number of clusters for the process damage dataset using the elbow method and the silhouette coefficient method H , the elbow method calculates the number of each possible cluster k The corresponding total internal variance, select the point where the variance decreases significantly k As the optimal number of clusters, the specific calculation method is as follows: （7） in, is the total internal variance, k is the number of clusters considered, It is i The sample set in the cluster, m is a single sample in the cluster, It is i The centroid of the clusters; The silhouette coefficient method measures the compactness of a data point within its cluster and its separation from other clusters. The clustering effect is evaluated by the size of the silhouette coefficient. The specific calculation formula is as follows: （8） in, It is a sample i The average distance to other samples in its cluster, It is a sample i The average distance to all samples in the nearest cluster, G is the number of clusters; S5.2: Optimize the K-means unsupervised learning algorithm by selecting new centroids based on the probability of the distance between the sample and the selected centroid, optimize the initial conditions of clustering and the global optimality of the results, and use the automatic variable weighting method to optimize the clustering feature weights. The weight adjustment uses the following formula: （9） in, It is i The weight of the feature, It is i The feature in j The variance in the clusters; S5.3: Using the results in steps S5.1 and S5.2, an artificial neural network is used to learn the nonlinear relationship between different damage modes and the degree of material process damage. The input layer receives the clustering results obtained from the improved K-means unsupervised learning algorithm, and the output layer is used to predict the process damage factor of the material. A process damage model is established, and the process damage factor is used to represent the degree of process damage of different damage modes to the ceramic matrix composite material. S5.4: Physical constraint optimization model, combined with Pearson correlation analysis, establishes a loss function to ensure the positive correlation between the process damage factor and the damage signal amplitude. Through iterative training, the network weights and biases are systematically adjusted to obtain the optimal process damage model.

8. The method for real-time damage degree prediction of ceramic matrix composite materials based on deep learning according to claim 1, characterized in that: In step S6, based on the damage accumulation theory, linear superposition is performed to couple the initial damage model and the process damage model, specifically: （10） in, x is the initial damage factor, y is the process damage factor, a and b are the fitting parameters; The least square method and nonlinear optimization are used to fit the mapping model parameters and the optimal mapping model is obtained.

Images