Aircraft support tensile strength evaluation method based on deep learning
By extracting the geometric and material features of aircraft supports using deep learning methods, a tensile strength prediction model is constructed. This solves the problems of high evaluation cost, low efficiency, and narrow applicability in existing technologies, and achieves high-precision and rapid support strength evaluation and self-optimization capabilities, supporting real-time monitoring and preventive maintenance of aircraft.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 商飞软件有限公司
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies for evaluating the tensile strength of aircraft supports suffer from high costs, low efficiency, limited accuracy, or narrow applicability. They also struggle to comprehensively handle the complex nonlinear interactions between multiple sources of information, failing to meet the demands of modern aircraft for lightweight and high-strength design.
A deep learning-based approach is adopted to extract the geometric and material features of the aircraft support through three-dimensional convolutional neural networks and deep neural networks. By combining transfer learning and active learning strategies, a tensile strength prediction model is constructed to achieve non-destructive assessment and high-precision prediction.
It enables high-precision and rapid assessment of the tensile strength of aircraft brackets, reduces testing costs, expands the scope of application, and has self-optimization capabilities. It supports real-time monitoring and prediction of strength degradation trends, providing a scientific basis for preventive maintenance of aircraft.
Smart Images
Figure CN121960136A_ABST
Abstract
Description
A Deep Learning-Based Method for Evaluating the Tensile Strength of Aircraft Supports Technical Field
[0001] This invention belongs to the field of deep learning technology, specifically a deep learning-based method for evaluating the tensile strength of aircraft brackets. Background Technology
[0002] As a critical load-bearing component in aircraft structure, the tensile strength of aircraft brackets directly affects the structural safety and service life of the entire aircraft. Currently, the assessment of this strength mainly relies on four traditional methods, but all of them have significant limitations. First, destructive testing (such as according to GB / T 228.1 or ASTM E8 standards) requires tensile testing of the sample until failure, which is costly and sacrifices the tested component, making it unsuitable for real-time monitoring of in-service brackets. Second, while finite element analysis does not require sample destruction, the accuracy of its results is highly dependent on the precision of the numerical model and the rationality of the boundary conditions, resulting in high computational complexity. For brackets with complex geometries and non-uniform material distribution, it is difficult to accurately simulate real performance. Third, prediction methods based on empirical formulas are usually only applicable to specific structures or materials, with weak generalization ability, making it difficult to cope with new materials and innovative designs. Fourth, while non-destructive testing techniques such as ultrasonic and X-ray testing can maintain sample integrity, they mainly detect internal defects, and the information obtained lacks a precise and universally applicable quantitative relationship model between tensile strength and the tensile strength.
[0003] In summary, existing technologies generally suffer from problems such as high evaluation costs, low efficiency, limited accuracy, or narrow applicability. Moreover, most of them analyze single factors such as geometry or materials in isolation, making it difficult to comprehensively handle the complex nonlinear interactions between multiple sources of information. They cannot meet the urgent needs of modern aircraft for high-efficiency, high-precision, and strong generalization capabilities in the context of lightweight and high-strength design. Summary of the Invention
[0004] The technical solution adopted in this invention is as follows: A method for evaluating the tensile strength of aircraft brackets based on deep learning, comprising the following steps:
[0005] Step 1: Data acquisition and preprocessing. Obtain the geometric shape data of the aircraft support through the 3D model, and collect the material properties of the support material. Clean, standardize and normalize the acquired data.
[0006] Step 2: Feature extraction network construction. Two parallel feature extraction channels are designed: the geometric feature extraction channel uses a three-dimensional convolutional neural network to process the geometric shape data of the support frame and extract the influence of the geometric features of the support frame on the strength; the material property extraction channel uses a deep neural network to process the basic material property data and extract the material composition features. The features extracted by each channel are integrated through a feature fusion layer to form a comprehensive feature representation.
[0007] Step 3: Construction and training of the tensile strength prediction model. Based on the fused feature representation, a prediction network with multiple fully connected layers is designed to finally output the predicted tensile strength value of the support. Data augmentation technology is used to expand the limited training data. Transfer learning strategy and active learning strategy are introduced for model training. Batch gradient descent and Adam optimizer are used for model training to minimize the root mean square error between the predicted strength and the actual strength, and an early stopping strategy is used to avoid overfitting.
[0008] Step 4: Model evaluation and optimization. The Monte Carlo dropout method is used to estimate the uncertainty of the prediction results. Sensitivity analysis is performed to analyze the influence of each input feature on the prediction results and identify the most critical factors for the tensile strength of the stent. Based on the evaluation results and sensitivity analysis, the model structure and hyperparameters are optimized.
[0009] Furthermore, in step one: data acquisition and preprocessing, the geometric data includes information on size, shape, thickness variation, and hole distribution. The basic information of the material includes material type and chemical composition. The material type is aluminum alloy or titanium alloy, and the chemical composition is the percentage content of the main elements.
[0010] Furthermore, in step one: data acquisition and preprocessing, the preprocessing includes eliminating outliers and noise, and unifying data from different sources into a standardized format.
[0011] Furthermore, in step two: the feature extraction network construction, the three-dimensional convolutional neural network structure of the geometric feature extraction channel includes convolutional layers, pooling layers, and fully connected layers, and the deep neural network of the material property extraction channel includes multiple hidden layers.
[0012] Furthermore, in step three, the construction and training of the tensile strength prediction model, the data augmentation technique expands the limited training data through parameter perturbation, geometric transformation, and random noise addition.
[0013] Furthermore, in step three: the construction and training of the tensile strength prediction model, the transfer learning strategy is to use the model parameters pre-trained on a large amount of general structural strength data to transfer them to the specific task of aircraft brackets.
[0014] Furthermore, in step three, the construction and training of the tensile strength prediction model, the active learning strategy is to design an uncertainty sampling strategy, which selectively selects the most informative samples from a large amount of unlabeled data for labeling and model updating.
[0015] Furthermore, in step three: the construction and training of the tensile strength prediction model, the model training uses root mean square error as the loss function and L2 regularization.
[0016] Furthermore, in step four: model evaluation and optimization, the model evaluation adopts the cross-validation method and calculates the root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination.
[0017] Furthermore, in step four: model evaluation and optimization, the network structure is adjusted based on the results of the sensitivity analysis to enhance the ability to extract important features, and grid search is used to optimize the network depth, width, regularization strength, and learning rate hyperparameters.
[0018] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0019] (1) This invention can comprehensively analyze multi-dimensional information such as the geometry and material composition of aircraft brackets through deep learning models, establish complex nonlinear mapping relationships, and thus achieve high-precision prediction of the tensile strength of the brackets. Its accuracy is better than that of traditional empirical formulas or finite element analysis methods.
[0020] (2) The present invention realizes non-destructive evaluation. There is no need to perform destructive tensile testing on the actual support. Strength evaluation can be completed based on its three-dimensional model and material data, which significantly reduces the testing cost and makes it possible to monitor and evaluate the strength of key supports of in-service aircraft.
[0021] (3) The present invention can simultaneously consider and integrate the comprehensive influence of geometric features and material properties on tensile strength, overcoming the limitations of traditional methods that often isolate and analyze a single factor, making the evaluation results closer to engineering practice and more comprehensive.
[0022] (4) By introducing a transfer learning strategy and a variety of data augmentation techniques, this invention effectively solves the problem of scarcity of high-quality labeled data in the aviation field, improves the model’s generalization ability for aircraft supports of different models and materials, and has a wider range of applications.
[0023] (5) The well-trained model can achieve rapid reasoning and output tensile strength prediction and uncertainty assessment in near real time after obtaining input data, providing efficient tool support for engineering decision-making and greatly improving assessment efficiency.
[0024] (6) The active learning strategy and incremental model update mechanism adopted in this invention enable the system to continuously learn from newly generated data, thereby continuously optimizing and improving prediction performance during application, and possessing good adaptability and continuous evolution capability.
[0025] (7) This invention can not only assess the current static strength of the support, but the feature-strength relationship revealed by its model can also help analyze key design parameters and provide a basis for predicting the degradation trend of support strength under specific load spectrum or environment, thereby providing a scientific reference for the preventive maintenance and life management of aircraft. Attached Figure Description
[0026] Figure 1 is a schematic diagram of the process of the present invention;
[0027] Figure 2 is a schematic diagram of the feature extraction network of the present invention. Detailed Implementation
[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0029] Example 1
[0030] Referring to Figures 1-2: A deep learning-based method for evaluating the tensile strength of aircraft brackets includes the following steps:
[0031] Step 1: Data Acquisition and Preprocessing
[0032] Data acquisition is fundamental to the method of this invention, primarily encompassing the acquisition of geometric data and material property data. Geometric data acquisition involves directly exporting the 3D model data of the scaffold from CATIA and converting it to a standard format. The 3D model is then converted into a 3D array through voxelization or into a 2D image sequence through cross-sectional processing, serving as input to a convolutional neural network. Material property data acquisition records the basic parameters of the scaffold material, including material type and chemical composition.
[0033] Preprocessing the collected data is a crucial step in ensuring effective model training. Preprocessing includes outlier detection and handling, using statistical methods to identify and process abnormal data; data standardization, applying Min-Max standardization to each feature to ensure it falls within the range [0,1]; feature encoding, encoding categorical features (material type); and data augmentation, expanding geometric data by adding minor perturbations, rotations, scaling, etc. These preprocessing steps effectively improve data quality, providing a solid foundation for subsequent model training.
[0034] Step 2: Feature Extraction Network Construction
[0035] The feature extraction network is the core of this invention. Two parallel feature extraction channels are designed to process geometric data and material property data respectively. The geometric feature extraction channel uses a 3D-CNN to process voxelized geometric data. The network structure includes a three-layer convolutional-pooling combination, specifically: the first convolutional layer has 32 3×3×3 convolutional kernels with a stride of 1, using the ReLU activation function; the first pooling layer is a 2×2×2 max pooling layer with a stride of 2; the second convolutional layer has 64 3×3×3 convolutional kernels with a stride of 1, using the ReLU activation function; the second pooling layer is a 2×2×2 max pooling layer with a stride of 2; the third convolutional layer has 128 3×3×3 convolutional kernels with a stride of 1, using the ReLU activation function; the third pooling layer is a 2×2×2 max pooling layer with a stride of 2; a global average pooling layer compresses the feature map into a fixed-length feature vector; and a fully connected layer has 128 neurons, uses the ReLU activation function, and outputs a geometric feature vector.
[0036] The material property extraction channel uses a fully connected neural network to process material property data, including: an input layer that receives material property parameters; a first hidden layer with 64 neurons using the ReLU activation function; a second hidden layer with 32 neurons using the ReLU activation function; and an output layer with 16 neurons using a linear activation function, outputting a material feature vector.
[0037] The feature fusion layer employs a feature concatenation method, directly concatenating the geometric feature vector (128-dimensional) and the material feature vector (16-dimensional) into a longer vector (144-dimensional). This combined feature vector (128-dimensional) is then output through a dimension unification layer (fully connected layer, 128 neurons, ReLU activation function). In this way, the model can simultaneously consider the influence of the scaffold's geometry and material properties on its tensile strength.
[0038] Step 3: Construction and Training of Tensile Strength Prediction Model
[0039] Based on the fused feature vector, this invention constructs a tensile strength prediction network, comprising: an input layer receiving a 128-dimensional comprehensive feature vector; a first hidden layer with 64 neurons, using the ReLU activation function with a dropout rate of 0.3; a second hidden layer with 32 neurons, using the ReLU activation function with a dropout rate of 0.3; and an output layer with 1 neuron, using a linear activation function, outputting the predicted tensile strength value (MPa).
[0040] To improve the generalization ability and prediction accuracy of the model, this invention implements a transfer learning strategy. The model is pre-trained on a dataset containing more than 50 different structural components. The weights of the convolutional layers of the geometric feature extraction network are retained, and only the weights of the fully connected layers and the prediction network are fine-tuned. A small learning rate (0.0001) is used for fine-tuning training to prevent overfitting.
[0041] Furthermore, this invention implements an active learning strategy, including: training an initial model using 20 labeled samples (with known tensile strength values); predicting unlabeled samples and calculating prediction uncertainty using Monte Carlo Dropout; selecting the 5 samples with the highest uncertainty for actual testing to obtain their tensile strength values; adding new labeled samples to the training set and retraining the model; repeating the above steps twice, resulting in a final training set containing 30 samples. This strategy maximizes the use of limited labeled samples and improves the model's training performance.
[0042] The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 16, and the root mean square error (RMSE) as the loss function. L2 regularization (weight decay coefficient of 0.0001) and Dropout (rate of 0.3) were used to avoid overfitting. The maximum number of training epochs was 500, and an early stopping strategy was adopted (training was stopped if the validation set loss did not improve for 30 consecutive epochs).
[0043] Step 4: Model Evaluation and Optimization
[0044] This invention employs a comprehensive evaluation method to verify the model's performance. Five-fold cross-validation is used to evaluate model performance, and the following metrics are calculated: Root Mean Square Error (RMSE): √(Σ(predicted value - actual value)² / n), assessing the absolute difference between predicted and actual values; Mean Absolute Error (MAE): Σ|predicted value - actual value| / n, providing an intuitive understanding of the error; Mean Absolute Percentage Error (MAPE): Σ(|predicted value - actual value| / actual value) / n×100%, assessing the relative error level; Coefficient of Determination (R²)... 2 ): 1 - residual sum of squares / total sum of squares, measures the model's ability to explain data variability.
[0045] To quantify the reliability of the prediction, this invention performs 20 forward propagations using Monte Carlo Dropout to calculate the mean and standard deviation of the prediction results, quantify the prediction uncertainty, and provide a 95% confidence interval, thus providing a more reliable basis for engineering decisions.
[0046] Furthermore, this invention also optimizes the model by identifying key features through feature sensitivity analysis and analyzing the impact of geometric and material features on the prediction results. Based on the analysis results, the network structure is adjusted to enhance the extraction capability of important features, and grid search is used to optimize key hyperparameters, including network depth, width, regularization strength, and learning rate.
[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A deep learning-based method for evaluating the tensile strength of aircraft brackets, characterized in that, Includes the following steps: Step 1: Data Acquisition and Preprocessing. Geometric shape data of the aircraft support structure is obtained through a 3D model. Simultaneously, material properties of the support structure are collected. The acquired data is cleaned, standardized, and normalized. Step 2: Feature Extraction Network Construction. Two parallel feature extraction channels are designed: the geometric feature extraction channel uses a 3D convolutional neural network to process the geometric shape data of the support structure and extract the influence of geometric features on strength; the material property extraction channel uses a deep neural network to process basic material property data and extract material composition features. A feature fusion layer integrates the features extracted from each channel to form a comprehensive feature representation. Step 3: Construction and training of the tensile strength prediction model. Based on the fused feature representation, a prediction network containing multiple fully connected layers is designed to output the predicted tensile strength value of the stent. Data augmentation techniques are used to expand the limited training data. Transfer learning and active learning strategies are introduced for model training. Batch gradient descent and Adam optimizer are used to minimize the root mean square error between the predicted strength and the actual strength, and an early stopping strategy is used to avoid overfitting. Step 4: Model evaluation and optimization. The Monte Carlo dropout method is used to estimate the uncertainty of the prediction results. Sensitivity analysis is performed to analyze the influence of each input feature on the prediction results and identify the most critical factors for the tensile strength of the stent. Based on the evaluation results and sensitivity analysis, the model structure and hyperparameters are optimized.
2. The method for evaluating the tensile strength of aircraft brackets based on deep learning according to claim 1, characterized in that, In step one: data acquisition and preprocessing, the geometric data includes information on size, shape, thickness variation, and hole distribution. The basic information of the material includes the material type and chemical composition. The material type is aluminum alloy or titanium alloy, and the chemical composition is the percentage content of the main elements.
3. The method for evaluating the tensile strength of aircraft brackets based on deep learning according to claim 1, characterized in that, In step one: data acquisition and preprocessing, the preprocessing includes eliminating outliers and noise, and unifying data from different sources into a standardized format.
4. The method for evaluating the tensile strength of aircraft brackets based on deep learning according to claim 1, characterized in that, In step two In the construction of the feature extraction network, the three-dimensional convolutional neural network structure of the geometric feature extraction channel includes convolutional layers, pooling layers, and fully connected layers, and the deep neural network of the material property extraction channel includes multiple hidden layers.
5. The method for evaluating the tensile strength of aircraft brackets based on deep learning according to claim 1, characterized in that, In step three, during the construction and training of the tensile strength prediction model, the data augmentation technique expands the limited training data through parameter perturbation, geometric transformation, and random noise addition.
6. The method for evaluating the tensile strength of aircraft brackets based on deep learning according to claim 1, characterized in that, In step three, during the construction and training of the tensile strength prediction model, the transfer learning strategy involves using model parameters pre-trained on a large amount of general structural strength data to transfer them to the specific task of aircraft brackets.
7. The method for evaluating the tensile strength of aircraft brackets based on deep learning according to claim 1, characterized in that, In step three, during the construction and training of the tensile strength prediction model, the active learning strategy involves designing an uncertainty sampling strategy to selectively pick the most informative samples from a large amount of unlabeled data for labeling and model updating.
8. The method for evaluating the tensile strength of aircraft brackets based on deep learning according to claim 1, characterized in that, In step three, the tensile strength prediction model is constructed and trained. The model training uses root mean square error as the loss function and L2 regularization.
9. The method for evaluating the tensile strength of an aircraft support based on deep learning according to claim 1, characterized in that, In step four: model evaluation and optimization, the model evaluation adopts the cross-validation method and calculates the root mean square error, mean absolute error, mean absolute percentage error and coefficient of determination.
10. The method for evaluating the tensile strength of an aircraft support based on deep learning according to claim 1, characterized in that, In step four: model evaluation and optimization, the network structure is adjusted based on the results of the sensitivity analysis to enhance the ability to extract important features, and grid search is used to optimize the network depth, width, regularization strength and learning rate hyperparameters.