Visual feature extraction method for damage of aerospace composite material
Through integrated high-resolution non-destructive testing equipment and advanced algorithms, multi-dimensional data acquisition, preprocessing, feature extraction and visualization of aerospace composite damage detection problems are solved, and comprehensive, accurate analysis and intuitive presentation of damage information are achieved, supporting efficient maintenance and improvement of composite materials.
Patent Information
- Application Number
- CN202510443451.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-10
- Publication Date
- 2025-07-25
AI Technical Summary
The existing aerospace composite damage detection methods have problems such as limited data acquisition dimensions, poor data preprocessing effect, single feature screening model, insufficient feature fusion, single feature optimization targets and unclear visualization, resulting in insufficient accuracy and comprehensiveness of damage detection.
High-resolution non-destructive detection equipment is used to collect multiple types of data, combine wavelet transformation filtering and normalization processing, and use CNN and LSTM hybrid networks to perform initial feature screening, feature fusion is performed based on the kernel method, and the features are optimized through genetic algorithms. Finally, ray projection algorithm is used for three-dimensional visual presentation.
It realizes multi-dimensional, comprehensive and accurate damage information capture, improves the accuracy and visualization of damage characteristics, provides an intuitive damage analysis basis, and supports efficient maintenance and improvement of composite materials.
Abstract
Description
Technical Field
[0001] The present invention relates to a method for extracting visual damage features of aerospace composite materials. Background Art
[0002] In the field of aerospace, the wide application of composite materials has put forward extremely high requirements for their damage detection and evaluation technologies. However, there are many limitations in the existing methods for extracting visual damage features of aerospace composite materials.
[0003] In terms of data acquisition, traditional non-destructive testing equipment often has a single function and can only obtain a single type of data. For example, it can only collect ultrasonic data or only obtain X-ray image data. This limits the detection dimension of composite material damage and makes it difficult to comprehensively capture damage information. For example, relying solely on ultrasonic data may not be able to accurately detect abnormal temperature distributions caused by damage, while using only infrared thermography data alone is difficult to detect fine internal structural defects of the material. In addition, the acquisition parameters of existing equipment are generally low. The frequency range of ultrasonic data is narrow, and it cannot effectively detect various sizes and types of damage. The resolution of infrared thermography data and X-ray image data is low, resulting in missing image details, and tiny damages are extremely easy to be ignored, bringing great difficulties to subsequent feature extraction and analysis.
[0004] In the data preprocessing stage, the effects of common denoising algorithms are not good. Most traditional filtering algorithms cannot accurately identify and remove noise signals. While removing noise, they will also seriously lose the effective information in the original data, affecting the accuracy of subsequent analysis. Moreover, there are huge differences in the numerical range and dimension of different types of original data, but there is a lack of effective normalization processing means, resulting in difficulties for subsequent algorithms such as deep learning network models to uniformly process the data. The stability of model training is poor, the convergence speed is slow, and features cannot be efficiently extracted.
[0005] In the initial feature screening stage, the existing deep learning network model structures are single. For example, simply using a convolutional neural network can extract spatial features, but it has insufficient processing ability for time series information and is difficult to capture damage features that change over time, such as the development process of damage under dynamic loading. Moreover, there is a lack of effective anti-overfitting strategies during model training, and overfitting is likely to occur, resulting in poor generalization ability of the model and inability to accurately apply to the actual task of extracting composite material damage features.
[0006] When fusing features, the existing fusion algorithms do not fully consider the internal connections between different modal data. Simple feature splicing or weighted average methods cannot reasonably assign weights to damage features according to the similarity and correlation of different modal data, making it difficult for the fused features to comprehensively and accurately reflect the damage situation of composite materials.
[0007] In terms of feature optimization, the objective function of traditional optimization algorithms is single. Usually, it only focuses on the distinguishability of features for damage, ignoring the independence and redundancy among features. This results in a large amount of information overlap and redundant features in the selected feature set, reducing the accuracy and effectiveness of damage features and unable to provide a reliable basis for subsequent damage analysis and evaluation.
[0008] In terms of visual presentation, existing 3D reconstruction algorithms cannot realistically present the 3D morphology of internal damage in composite materials. The distinction of different damage features in terms of color, transparency, and shape is not clear enough, making it difficult to help users quickly identify and understand the damage situation. At the same time, the lack of interactive functions means that users cannot observe the damage visualization model from different perspectives. For complex 3D damage structures, it is difficult to comprehensively grasp the damage information, seriously restricting the formulation of subsequent decisions such as maintenance and improvement. Summary of the Invention
[0009] The purpose of the present invention is to provide a method for extracting visual damage features of aerospace composite materials to solve the problems raised in the above background technology.
[0010] To solve the above technical problems, the technical solution provided by the present invention is: a method for extracting visual damage features of aerospace composite materials, including the following steps:
[0011] Data acquisition step: Use a high-resolution non-destructive testing device to scan aerospace composite material components to obtain the original data of the components. The original data includes but is not limited to ultrasonic data, infrared thermographic data, and X-ray image data;
[0012] Data preprocessing step: Denoise the original data, remove the noise interference in the original data through a filtering algorithm based on wavelet transform, and at the same time normalize different types of original data, mapping them to a unified data range;
[0013] Initial feature screening step: Use a deep learning network model to perform preliminary feature extraction on the preprocessed data. The deep learning network model is trained with multiple sets of aerospace composite material damage samples. Its input is the preprocessed data, and the output is various possible damage features, including but not limited to preliminary feature information of cracks, delaminations, debonds, and fiber fractures;
[0014] Feature fusion step: Fuse the preliminary damage features extracted from different types of original data, adopt a feature fusion algorithm based on multi-modal data fusion, and fuse multiple preliminary damage features into a set of comprehensive damage features according to the weights and correlations of different modal data;
[0015] Feature optimization step: Use a feature optimization algorithm based on genetic algorithm to optimize the comprehensive damage features. By setting an optimization objective function, key features with high recognition of aerospace composite material damage are screened out to improve the accuracy and effectiveness of damage features.
[0016] Visualization presentation step: Map the optimized key features into three-dimensional space through a three-dimensional reconstruction algorithm to construct a visualization model of aerospace composite material damage and achieve visual display of damage. The three-dimensional reconstruction algorithm marks different damage areas with different colors and shapes according to the spatial information and damage types of the key features.
[0017] Preferably, in the data acquisition step, the high-resolution non-destructive testing equipment is an integrated device that can simultaneously collect ultrasonic data, infrared thermal image data, and X-ray image data. The frequency range of the collected ultrasonic data is [20 kHz, 10 MHz], the resolution of the infrared thermal image data is not less than 1280×1024 pixels, and the resolution of the X-ray image data is not less than 2048×1536 pixels.
[0018] Preferably, in the data preprocessing step, the wavelet basis function of the filtering algorithm based on wavelet transform is Daubechies wavelet, and the decomposition level is 4 layers. And in the normalization process, the data is normalized to the range of [0, 1].
[0019] Preferably, in the feature preliminary screening step, the deep learning network model is a hybrid network structure of a convolutional neural network (CNN) and a long short-term memory network (LSTM). It includes multiple convolutional layers, pooling layers, LSTM layers, and fully connected layers. The convolutional layers are used to extract spatial features, the LSTM layers are used to extract time series features, the pooling layers are used for dimensionality reduction and feature compression, and the fully connected layers are used to output damage features. And the early stopping method and Dropout technology are adopted in the training process of this deep learning network model to prevent overfitting.
[0020] Preferably, in the feature fusion step, the feature fusion algorithm based on multi-modal data fusion is a fusion algorithm based on kernel method. It calculates the similarity of different modal data using kernel function, and assigns different weights to the damage features of different modal data according to the similarity. The kernel function is a radial basis kernel function.
[0021] Preferably, in the feature optimization step, the optimization objective function of the feature optimization algorithm based on the genetic algorithm includes the recognition rate of features for damage, the independence between features, and the redundancy of features. Among them, the recognition rate of features for damage is evaluated by the damage classification accuracy rate, the independence between features is measured by calculating the mutual information of features, and the redundancy of features is judged by calculating the correlation coefficient of features.
[0022] Preferably, in the visualization presentation step, the three-dimensional reconstruction algorithm adopts the ray casting algorithm. According to the spatial coordinates, damage types, and feature intensities of key features, different damage features are represented by different colors, transparencies, and shapes, and an interactive function is provided for users to allow them to view the damage visualization model from different perspectives.
[0023] The advantages of the present invention are as follows: First, comprehensive and accurate data collection
[0024] Multi-source data fusion collection: An integrated high-resolution non-destructive testing device that can simultaneously collect ultrasonic data, infrared thermal image data, and X-ray image data is used to obtain composite material component information from multiple physical dimensions. Different types of data reflect different characteristics of material damage. For example, ultrasonic data is sensitive to internal structure defects, infrared thermal image data can reveal abnormal temperature distributions caused by damage, and X-ray image data can clearly present the internal structure morphology. The combination of the three greatly improves the comprehensiveness of capturing damage information.
[0025] High-standard collection parameters: The frequency range of the ultrasonic data collected by the device is [20 kHz, 10 MHz], which covers a wide range from low frequency to high frequency and can detect different sizes and types of damage. The resolution of the infrared thermal image data is not less than 1280×1024 pixels, and the resolution of the X-ray image data is not less than 2048×1536 pixels. The high resolution ensures that the collected image data is rich in details and it is difficult for small damages to escape, providing a solid data foundation for subsequent feature extraction.
[0026] Second, efficient data preprocessing
[0027] Advanced wavelet denoising: A filtering algorithm based on Daubechies wavelet with 4 decomposition layers is used to denoise the original data. Daubechies wavelet has good time-frequency localization characteristics, can effectively identify and remove noise signals, and at the same time retains the effective information in the original data to the greatest extent. The 4 decomposition layers are carefully selected to analyze the signal at different scales, accurately remove noise interference, and improve the data quality.
[0028] Unified data normalization: Normalize different types of raw data to the range [0, 1], eliminating the differences in numerical range and dimension among different data types. This enables different data to be on the same standard during subsequent processing, facilitating the unified processing of data by algorithms such as deep learning network models and improving the stability and convergence speed of model training.
[0029] III. Powerful initial feature screening ability
[0030] Innovative hybrid network structure: Use a hybrid network structure of convolutional neural network (CNN) and long short-term memory network (LSTM) for initial feature extraction. The convolutional layer of CNN can effectively extract spatial features in data, capturing the characteristic patterns of composite material damage in spatial distribution, such as the direction of cracks, the shape of delamination, etc.; the LSTM layer is good at processing time series information and has good extraction ability for some damage features that change over time (such as the development process of damage under dynamic loading). The pooling layer reduces the dimension and compresses features, reducing the computational amount while retaining key features, and the fully connected layer integrates the extracted features and outputs the initial damage feature information.
[0031] Effective anti-overfitting strategy: Adopt early stopping method and Dropout technology during training to effectively prevent the overfitting phenomenon of the deep learning network model. The early stopping method stops training when the performance of the validation set no longer improves by monitoring the performance of the validation set, avoiding overfitting of the model on the training set. The Dropout technology randomly "drops out" some neurons during training, reducing the co-adaptation between neurons and enhancing the generalization ability of the model, enabling the model to better apply to the actual composite material damage feature extraction task.
[0032] IV. Scientific feature fusion strategy
[0033] Kernel-based fusion: Adopt a kernel-based feature fusion algorithm and use the radial basis kernel function to calculate the similarity of different modal data. The kernel method can map low-dimensional data to high-dimensional space, thus calculating the similarity between data more accurately. By calculating the similarity of damage features of different modal data and assigning different weights to the damage features of different modal data according to the similarity, multiple initial damage features are fused into a set of comprehensive damage features. This fusion method fully considers the internal relationship between different modal data, making the fused features more comprehensive and representative and being able to more accurately reflect the damage situation of composite materials.
[0034] V. Precise feature optimization
[0035] Multi-dimensional optimization objective: For the feature optimization algorithm based on genetic algorithm, its optimization objective function covers the recognition rate of features for damage, the independence between features, and the redundancy of features. The recognition rate of features for damage is evaluated by the damage classification accuracy to ensure that the selected features can effectively distinguish different types of damage; the independence between features is calculated using mutual information to avoid information overlap between features; the redundancy of features is judged by calculating the correlation coefficient to remove redundant features. This multi-dimensional optimization strategy can screen out key features with high recognition rate for aerospace composite material damage, greatly improving the accuracy and effectiveness of damage features, and providing a more reliable basis for subsequent damage analysis and evaluation.
[0036] VI. Intuitive Visualization Presentation
[0037] Advanced 3D reconstruction algorithm: The ray casting algorithm is used for 3D reconstruction. According to the spatial coordinates, damage types, and feature intensities of key features, different damage features are represented by different colors, transparencies, and shapes. The ray casting algorithm can vividly present the three-dimensional morphology of internal damage in the composite material, making the damage situation clear at a glance. Different visualization attributes (color, transparency, shape) correspond to different damage features, which helps users quickly identify and understand the type, location, and severity of damage.
[0038] Convenient interaction function: An interaction function is provided for users, allowing them to view the damage visualization model from different perspectives. This function enables users to observe the damage situation of the composite material comprehensively and in-depth. Especially for complex three-dimensional damage structures, users can analyze from different angles through operations such as rotation and scaling, so as to master the damage information more comprehensively and provide strong support for subsequent decisions such as repair and improvement. Specific Embodiment
[0039] For the purposes of the following detailed description, it should be understood that the present invention may take various alternative variations and step sequences, unless expressly specified to the contrary. In addition, except in any operating examples, or otherwise indicated, all numbers expressing, for example, quantities of ingredients used in the specification and claims are to be understood as being modified in all instances by the term "about". Accordingly, unless indicated to the contrary, the numerical parameters set forth in the following specification and attached claims are approximations that may vary depending upon the desired properties sought to be obtained by the present invention. At the very least, and not as an attempt to limit the application of the doctrine of equivalents to the scope of the claims, each numerical parameter should at least be construed in light of the number of reported significant digits and by applying ordinary rounding techniques.
[0040] Although the numerical ranges and parameters that describe the broad scope of the present invention are approximations, the numerical values set forth in the specific examples are reported as precisely as possible. However, any numerical value inherently contains certain errors necessarily resulting from the standard deviation found in its respective test measurements.
[0041] In addition, it should be understood that any numerical range described herein is intended to include all sub-ranges subsumed therein. For example, the range of "1 to 10" is intended to include all sub-ranges between (and including) the minimum value of 1 and the maximum value of 10 described, that is, having a minimum value equal to or greater than 1 and a maximum value equal to or less than 10.
[0042] Example 1: Detection of Wing Composite Material Damage
[0043] Data acquisition: An integrated high-resolution non-destructive testing device that can simultaneously acquire ultrasonic data, infrared thermal image data, and X-ray image data is used to scan the aerospace composite material components of a certain type of aircraft wing. The frequency range of the acquired ultrasonic data is in the range of [20 kHz, 10 MHz], the resolution of the infrared thermal image data is 1280×1024 pixels, and the resolution of the X-ray image data is 2048×1536 pixels. Through multi-source data acquisition, the original information of the wing composite material components is comprehensively obtained.
[0044] Data preprocessing: A filtering algorithm based on Daubechies wavelet with 4 decomposition levels is used to denoise the original data, remove noise interference, and at the same time normalize different types of original data to the range of [0, 1] to provide high-quality data for subsequent processing.
[0045] Initial feature screening: The preprocessed data is input into a hybrid network structure composed of a convolutional neural network (CNN) and a long short-term memory network (LSTM). The convolutional layer of the CNN extracts spatial features such as crack orientation and delamination shape, and the LSTM layer captures features such as the change of damage over time due to the force during flight. Through the pooling layer for dimensionality reduction and feature compression, and then the fully connected layer outputs the initial damage features, such as the initial feature information indicating that cracks and delaminations may exist in some areas of the wing. During the training process, the early stopping method and Dropout technology are used to prevent overfitting and ensure the generalization ability of the model.
[0046] Feature fusion: Using a kernel method fusion algorithm based on a radial basis kernel function, calculate the similarity of damage features of different modality data, assign weights to the damage features of different modality data according to the similarity, and fuse the initial damage features into comprehensive damage features to more comprehensively reflect the damage situation of the wing composite material.
[0047] Feature Optimization: Based on the genetic algorithm, the recognition ability of features for damage is evaluated by the damage classification accuracy. The independence between features is calculated through mutual information, and the feature redundancy is judged by the correlation coefficient. The comprehensive damage features are optimized to screen out the key features with high recognition ability for wing damage.
[0048] Visualization Presentation: The ray casting algorithm is used for 3D reconstruction. According to the spatial coordinates, damage types and feature intensities of the key features, the wing damage areas are marked with different colors, transparencies and shapes. For example, cracks are shown as red slender lines, and the delaminated areas are represented by semi-transparent blue blocks. Users can view the wing damage visualization model from different perspectives through the interactive function, comprehensively master the damage information, and provide strong support for the maintenance decision-making of the wing.
[0049] Example 2: Damage Assessment of Satellite Solar Panel Composite Materials
[0050] Data Acquisition: For the aerospace composite materials of satellite solar panels, an integrated high-resolution non-destructive testing device is used for scanning to obtain ultrasonic, infrared thermography and X-ray image data. Ensure that the ultrasonic data frequency is in [20kHz, 10MHz], the infrared thermography data resolution is not less than 1280×1024 pixels, and the X-ray image data resolution is not less than 2048×1536 pixels to obtain comprehensive original data of the solar panel composite materials.
[0051] Data Preprocessing: A filtering algorithm based on the 4-layer decomposition of Daubechies wavelet is used to remove the noise in the original data, and the data is normalized to the range of [0, 1] to prepare high-quality data for subsequent analysis.
[0052] Initial Feature Screening: The preprocessed data is input into the hybrid network model of CNN and LSTM. During the operation of the satellite, the damage situation of the solar panel changes with time due to space environmental factors. The LSTM layer effectively extracts these time series features, combines with the spatial features extracted by the CNN convolutional layer, and outputs the preliminary damage features through the pooling layer and the fully connected layer. For example, the preliminary signs of fiber fracture and debonding in the solar panel are found. The early stopping method and Dropout technology are used to prevent the model from overfitting during training.
[0053] Feature Fusion: Based on the fusion algorithm of the kernel method, the similarity of damage features of different modal data is calculated according to the radial basis kernel function, and corresponding weights are assigned to fuse and obtain the comprehensive damage features, presenting the damage features of the solar panel comprehensively.
[0054] Feature Optimization: Using the feature optimization algorithm based on the genetic algorithm, the comprehensive damage features are optimized from multiple dimensions of the recognition ability, independence and redundancy of features for damage, and the key features are screened out to improve the accuracy and effectiveness of the solar panel damage features.
[0055] Visual presentation: The optimized key features are three-dimensionally reconstructed through the ray casting algorithm, and the damage of the solar panel is presented with different colors, transparencies, and shapes. For example, the fiber breakage is shown as yellow flashing points, and the debonding area is represented by green irregular shapes. Through the interactive function, users can view the visual model of the solar panel damage from various angles, providing detailed basis for evaluating the performance of the solar panel and formulating repair plans.
[0056] As mentioned above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.
Claims
1. A method for extracting visual characteristics of aerospace composite material damage, characterized in that, The following steps are involved: Data acquisition step: using high-resolution nondestructive testing equipment to scan aerospace composite components to obtain raw data of the components, including but not limited to ultrasonic data, infrared thermal imaging data, and X-ray image data; Data preprocessing step: denoising the raw data, removing noise interference in the raw data through a wavelet transform-based filtering algorithm, and normalizing different types of raw data to map them to a unified data range; Initial feature screening step: preliminary feature extraction of the preprocessed data using a deep learning network model, the deep learning network model is trained on multiple sets of aerospace composite material damage samples, the input of which is the preprocessed data, and the output is a variety of possible damage features, including but not limited to preliminary feature information of cracks, delamination, debonding and fiber breakage; Feature fusion step: The preliminary damage features extracted from different types of raw data are fused, and a feature fusion algorithm based on multimodal data fusion is used to fuse multiple preliminary damage features into a set of comprehensive damage features according to the weights and correlations of different modal data; Feature optimization step: optimizing the comprehensive damage features by using a feature optimization algorithm based on a genetic algorithm, and screening out key features with high recognition of damage to aerospace composite materials by setting an optimization objective function, so as to improve the accuracy and effectiveness of the damage features; Visualization presentation step: Map the optimized key features to three-dimensional space through a three-dimensional reconstruction algorithm, construct a visualization model of aerospace composite material damage, and realize visualization of damage. The three-dimensional reconstruction algorithm marks different damaged areas with different colors and shapes according to the spatial information of the key features and the damage type.
2. The method for extracting visual characteristics of aerospace composite material damage according to claim 1, characterized in that In the data acquisition step, the high-resolution nondestructive testing equipment is an integrated device that can simultaneously acquire ultrasonic data, infrared thermal imaging data and X-ray image data. The frequency range of the acquired ultrasonic data is [20kHz, 10MHz], the resolution of the infrared thermal imaging data is not less than 1280×1024 pixels, and the resolution of the X-ray image data is not less than 2048×1536 pixels.
3. The aerospace composite material damage visualization feature extraction method according to claim 2, wherein, In the data preprocessing step, the wavelet basis function of the wavelet transform-based filtering algorithm is Daubechies wavelet, the number of decomposition layers is 4, and in the normalization process, the data is normalized to the range of [0, 1].
4. The aerospace composite material damage visualization feature extraction method according to claim 3, wherein In the feature initial screening step, the deep learning network model is a hybrid network structure of a convolutional neural network (CNN) and a long short-term memory network (LSTM), which includes multiple convolutional layers, pooling layers, LSTM layers and fully connected layers. The convolutional layer is used to extract spatial features, the LSTM layer is used to extract time series features, the pooling layer is used for dimensionality reduction and feature compression, and the fully connected layer is used to output damage features. In addition, the deep learning network model uses early stopping and Dropout technology to prevent overfitting during training.
5. The method for extracting visual characteristics of aerospace composite material damage according to claim 4, wherein In the feature fusion step, the feature fusion algorithm based on multi-modal data fusion is a fusion algorithm based on the kernel method. It calculates the similarity of different modal data using the kernel function, and assigns different weights to the damage features of different modal data according to the similarity. The kernel function is the radial basis kernel function.
6. The method for extracting the visual damage characteristics of aerospace composite materials according to claim 5, characterized in that In the feature optimization step, the optimization objective function of the feature optimization algorithm based on the genetic algorithm includes the discrimination of features for damage, the independence between features, and the redundancy of features. Among them, the discrimination of features for damage is evaluated by the damage classification accuracy rate, the independence between features is measured by calculating the mutual information of features, and the redundancy of features is judged by calculating the correlation coefficient of features.
7. The method for extracting visual characteristics of aerospace composite material damage according to claim 6, characterized in that, In the visualization presentation step, the three-dimensional reconstruction algorithm uses the ray casting algorithm. According to the spatial coordinates, damage types, and feature intensities of the key features, different damage features are represented by different colors, transparencies, and shapes, and an interactive function is provided for users, allowing users to view the damage visualization model from different perspectives.