Additive manufacturing metal fatigue life prediction method based on multi-dimensional defect features and ensemble learning
By employing multidimensional defect features and ensemble learning methods, the accuracy and robustness issues of fatigue life prediction for additive manufacturing components were addressed, achieving high-precision life prediction and process optimization guidance.
Patent Information
- Application Number
- CN202610038078.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-13
- Publication Date
- 2026-04-17
AI Technical Summary
Existing technologies struggle to accurately predict the fatigue life of additively manufactured components, particularly due to the single dimension of defect feature extraction, the simple model structure, and the lack of high-dimensional heterogeneous feature screening, resulting in insufficient prediction accuracy and robustness.
A multidimensional defect feature and ensemble learning approach is adopted to obtain the three-dimensional morphology of internal defects in components through X-ray computed tomography, calculate multidimensional geometric features, screen key features, and use a two-layer stacking ensemble learning model to predict fatigue life.
It significantly improves the accuracy and robustness of fatigue life prediction, is suitable for small sample scenarios, provides efficient and objective component life assessment, and can reveal the degree of influence of defect attributes on fatigue life.
Smart Images
Figure CN121884018A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of metal manufacturing evaluation and analysis technology, and in particular to a method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and integrated learning. Background Technology
[0002] Metal additive manufacturing technology, especially laser powder bed fusion, is widely used in critical fields such as aerospace and biomedicine due to its ability to efficiently form complex geometries. However, this process inevitably introduces internal manufacturing defects, such as porosity and lack of fusion. These defects become stress concentration points under cyclic loading, easily inducing fatigue cracks, leading to early component failure, and causing significant dispersion in fatigue life data, severely restricting the reliability and engineering application of additively manufactured components.
[0003] To ensure the service safety of additively manufactured components, accurate prediction of their fatigue life is crucial. Traditional life prediction methods mainly rely on classical fracture mechanics theory or empirical models based on a large amount of experimental data. The former usually requires simplified assumptions about the defect shape and cannot accurately characterize the irregular three-dimensional morphology of real defects; the latter heavily depends on experimental databases of specific materials and processes, resulting in poor generalization ability.
[0004] With the development of high-resolution X-ray computed tomography (CT) technology, it is now possible to non-destructively obtain accurate three-dimensional morphology and spatial distribution of internal defects in components. Recently, researchers have begun to explore the complex mapping relationship between defect features and fatigue life by introducing machine learning algorithms. However, existing solutions still have significant limitations: First, feature extraction is often limited to a single dimension, focusing primarily on defect size and failing to systematically integrate key geometric attributes such as the spatial location and complex morphology of defects; second, the model structure is simple, typically employing a single machine learning model, which struggles to fully learn complex nonlinear relationships in small sample data, resulting in insufficient prediction accuracy and robustness; third, there is a lack of effective screening and optimization mechanisms for high-dimensional heterogeneous features, which may lead to model overfitting. Summary of the Invention
[0005] Based on existing technical problems, this invention proposes a method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning. This method can automatically extract multidimensional three-dimensional defect features, intelligently screen key influencing factors, and use a robust ensemble learning framework to perform high-precision life prediction.
[0006] The present invention proposes a method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning, including S1, three-dimensional defect data acquisition and extraction;
[0007] S2, Calculation of multi-dimensional defect geometric features;
[0008] S3. Feature dataset construction and standardization processing;
[0009] S4. Key Feature Subset Filtering;
[0010] S5. Lifetime prediction based on a two-layer stacking ensemble learning model;
[0011] S6. Prediction result output and evaluation;
[0012] S7. Validity verification.
[0013] Preferably, in step S1, the additive manufacturing metal sample is first scanned using an X-ray computed tomography system to reconstruct the internal three-dimensional structure of the sample, and then an image processing algorithm is used to identify and separate individual manufacturing defects from the reconstructed volume to obtain a three-dimensional digital model of each defect.
[0014] Preferably, in step S2, for each 3D model of a defect in step S1, three types of geometric features are calculated in parallel:
[0015] S21. Dimensional characteristics; calculate the volume, surface area, and square root of the projected area of the defect on the plane perpendicular to the principal stress direction;
[0016] S22, Location characteristics; Calculate the shortest spatial distance from the defect centroid to the nearest free surface of the specimen;
[0017] S23. Morphological characteristics; calculate the sphericity and aspect ratio of the defect.
[0018] Preferably, in step S3, for each sample, a critical defect is determined according to a preset rule, the multidimensional feature parameters of the critical defect are combined into a feature vector, and the vector corresponds to the actual fatigue life value obtained by the sample through experiments, together forming a sample data, and the features of all samples in the dataset are standardized.
[0019] Preferably, in step S4, a tree-based feature importance assessment method is used to calculate the contribution of each feature to lifetime prediction, and features with high importance are selected to form an optimized feature subset.
[0020] Preferably, the two-level Stacking ensemble regression model of S5 includes;
[0021] The first layer contains multiple heterogeneous regression models as base learners;
[0022] The second layer uses a meta-learner, and takes the prediction results of all base learners in the first layer as input. It learns the optimal weighted fusion coefficients through training and outputs the final fatigue life prediction value.
[0023] Preferably, the specific steps for training and inference of the integrated learning model include:
[0024] S51. Standardize the constructed complete feature dataset;
[0025] S52. Based on the standardized data, use the random forest model to analyze the contribution of each feature to lifetime prediction, and select and record several features and their indices.
[0026] S53. Using the selected subset of key features as input and the corresponding fatigue life as output, train a Stacking ensemble learning model. During training, cross-validation is used to determine the model parameters.
[0027] S54. After the model training is completed, save the complete ensemble model, all parameters required for feature standardization, and the list of key feature indices.
[0028] S55. When making predictions, the original characteristics of new samples are preprocessed using the saved standardized parameters.
[0029] S56. Based on the saved key feature index, extract the corresponding data from the preprocessed features to form the final model input vector;
[0030] S57. Input the input vector into the loaded ensemble model, perform prediction, and directly output the lifetime value.
[0031] Preferably, in step S6, after the defect features of the new sample are processed in the same way, they are input into the trained ensemble learning model to obtain its fatigue life prediction value. At the same time, the confidence interval or uncertainty measure of the prediction can be output.
[0032] Preferably, in step S6, for a new sample, steps S1 to S2 are repeated to extract its features, and the saved trained model and key parameters required for data processing are used. The key parameters include standardized parameters and feature indexes. After processing, the data is input into the model to obtain the lifetime prediction value.
[0033] Preferably, in step S7, the predicted results are compared with the experimental values and the predicted results using a single model, and then the coefficient of determination R² and root mean square error RMSE are used as evaluation indicators.
[0034] The beneficial effects of this invention are as follows:
[0035] 1. High prediction accuracy: By comprehensively considering multi-dimensional three-dimensional features such as defect size, location, and shape, and by employing ensemble learning to integrate the advantages of multiple algorithms, the prediction accuracy is significantly improved, making it particularly suitable for small sample scenarios.
[0036] 2. Strong robustness: The feature selection and ensemble learning framework effectively reduces the model's sensitivity to data noise and irrelevant features, enhancing the stability of prediction results and its adaptability to different batches of data.
[0037] 3. Clear physical meaning: The selected features have clear physical or geometric meanings. The model can not only provide predicted values, but also reveal the degree of influence of different defect attributes on fatigue life through feature importance analysis, which helps to provide guidance for process optimization.
[0038] 4. High degree of automation and good practicality: The process from feature extraction to life prediction is programmable and automated, and it is easy to integrate into the quality control system of additive manufacturing, so as to achieve efficient and objective rapid assessment of component life. Attached Figure Description
[0039] Figure 1 This is a schematic diagram of an additive manufacturing metal fatigue life prediction method based on multidimensional defect features and ensemble learning proposed in this invention.
[0040] Figure 2 This is a flowchart of the three-dimensional defect data extraction process for an additive manufacturing metal fatigue life prediction method based on multi-dimensional defect features and ensemble learning proposed in this invention.
[0041] Figure 3 This is a schematic diagram of a two-layer Stacking ensemble learning model for an additive manufacturing metal fatigue life prediction method based on multidimensional defect features and ensemble learning proposed in this invention.
[0042] Figure 4 This is a comparison chart showing the prediction performance of an additive manufacturing metal fatigue life prediction method based on multidimensional defect features and ensemble learning proposed in this invention. Detailed Implementation
[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0044] Reference Figures 1-4 A method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning includes S1, defect three-dimensional data acquisition and extraction, whereby the defect three-dimensional data includes:
[0045] 1. CT scan data: ;
[0046] in, This is 3D grayscale data from a CT scan. For three-dimensional spatial coordinate indexes.
[0047] 2. Image segmentation: ;
[0048] in, It is a binary image (1 represents a defect, 0 represents the matrix). This is the image segmentation threshold.
[0049] threshold The Otsu method is used to automatically determine the value; the calculation formula is as follows:
[0050] ;
[0051] in, For grayscale values less than or equal to The pixel ratio (defect area). For grayscale values greater than The pixel ratio (base area). The average gray level of the defect area. The average gray level of the matrix region.
[0052] 3. Defect model extraction: ;
[0053] in, Let k be the set of voxels for the k-th defect. For the defect number, The coordinates of the defect voxel are given.
[0054] First, an X-ray computed tomography (CT) system is used to scan the additively manufactured metal sample to reconstruct its internal three-dimensional structure. Then, image processing algorithms are used to identify and separate individual manufacturing defects from the reconstructed volume, obtaining a three-dimensional digital model of each defect.
[0055] like Figure 2 As shown, S1 is the foundation for all subsequent analyses, and its core lies in obtaining accurate three-dimensional morphology of defects non-destructively. First, a set of additively manufactured metal samples are prepared, and their experimental fatigue life data is obtained through fatigue testing. Then, CT scanning is used to obtain internal grayscale data of the samples, identify and separate all internal defects, and subsequently, image segmentation techniques, such as thresholding, are employed. The process involves segmentation, separating defects from the matrix material, and then using three-dimensional connected domain analysis to identify individual defects. Based on the maximum projected area criterion, a critical defect is determined for each sample. Finally, a three-dimensional digital model, such as a mesh model, is output for each defect that can be used for geometric calculations.
[0056] S2, Multi-dimensional defect geometric feature calculation; For each defect's 3D model in S1, three types of geometric features are calculated in parallel:
[0057] S21. Dimensional characteristics: The volume, surface area, and square root of the projected area of the defect on the plane perpendicular to the principal stress direction can be obtained by directly measuring the model entity.
[0058] S22, Location characteristics; Calculate the shortest spatial distance from the defect centroid to the nearest free surface of the specimen, which can be achieved by calculating the spatial distance from the model centroid to the specimen surface;
[0059] S23. Morphological characteristics: Calculate the sphericity and aspect ratio of the defect. The shape complexity and anisotropy of the defect can be described by parameters such as sphericity and aspect ratio.
[0060] This yields six parameters: volume, surface area, square root of projected area, distance to the nearest free surface, sphericity, and aspect ratio.
[0061] 1. Volume: ;
[0062] in, For the defect volume, The total number of defective voxels. Physical volume of a single voxel .
[0063] 2. Surface area: ;
[0064] in, This represents the surface area of the defect. This represents the number of triangular mesh faces. For the first A triangular facet.
[0065] 3. Square root of the projected area: ;
[0066] in, The square root of the projected area. Let be the projected area of the defect on the plane perpendicular to the principal stress direction.
[0067] 4. Distance to the nearest free surface: ;
[0068] in, Let be the distance from the defect centroid to the nearest free surface. Defect centroid coordinates , For the first sample Any point on the surface, For Euclidean distance.
[0069] 5. Sphericity: ;
[0070] in, Sphericity ( ), For the defect volume, The area represents the defect surface area.
[0071] 6. Aspect Ratio: ;
[0072] in, Aspect ratio, Let the length of the defect be along the longest principal axis. The length of the defect along the shortest principal axis.
[0073] The features S21-S23 above together constitute the key indicator system for assessing the severity of defects. It has the advantage of high prediction accuracy: by comprehensively considering multi-dimensional three-dimensional features such as defect size, location, and morphology, and employing ensemble learning to integrate the advantages of multiple algorithms, it significantly improves prediction accuracy, making it particularly suitable for small sample scenarios.
[0074] S3. Feature Dataset Construction and Standardization: For each sample, a "critical defect" is determined according to a preset rule (such as maximum projected area). The critical defect selection rule is: assuming the sample contains... There are 1 defect, and the critical defect index is:
[0075] ;
[0076] in, This is an index for critical defects. For the first The projected area of a defect.
[0077] The six multidimensional feature parameters of the critical defect are combined into a feature vector, which is then correlated with the actual fatigue life value obtained experimentally for the specimen, forming a sample dataset. The features of all samples in the dataset are standardized to eliminate the influence of differences in feature dimensions and numerical ranges on model training.
[0078] Step S3 serves as a bridge connecting the raw data with the machine learning model. The key is to transform the geometric information of the 3D defects into structured data. For each specimen, the "critical defect" (usually the largest or closest to the surface) most likely to cause failure is selected from all defects, and its characteristics represent the defect state of that specimen. Subsequently, all features are standardized to make features at different scales comparable, laying the foundation for subsequent model training.
[0079] S4. Key Feature Subset Selection: A tree-based feature importance assessment method is used to calculate the contribution of each feature to lifetime prediction. Based on this, the features with the highest importance ranking are selected to form an optimized feature subset with strong discriminative power and low redundancy.
[0080] The feature selection method of this invention aims to improve model efficiency and generalization ability. Specifically, it employs a random forest regression algorithm, evaluating the importance of a feature during training based on the reduction in impurity caused by splitting at decision tree nodes. Features with importance scores exceeding a set threshold or ranking in the top K are selected to form an optimized feature subset. This process effectively eliminates noisy and redundant features, highlighting core influencing factors.
[0081] S5. Lifetime prediction based on a two-layer stacking ensemble learning model: Construct a two-layer stacking ensemble regression model, and the two-layer stacking ultimately predicts:
[0082] ;
[0083] in, This is the final predicted fatigue life value. For the input feature vector, For the first The predicted output of each base learner For bias terms, The weighted fusion coefficient, The number of base learners.
[0084] The first layer contains multiple heterogeneous regression models as base learners, such as random forests, gradient boosting regression trees, and fully connected neural networks, which make preliminary predictions based on the optimized feature subset. The second layer employs a meta-learner, such as a linear regression model, which takes the predictions from all the base learners in the first layer as input, learns the optimal weighted fusion coefficients through training, and outputs the final fatigue life prediction value.
[0085] The specific steps of training and inference for ensemble learning models are as follows:
[0086] S51. Standardize the constructed complete feature dataset to eliminate the influence of units.
[0087] S52. Based on the standardized data, use the random forest model to analyze the contribution of each feature to lifetime prediction, and select and record the most important features and their indices.
[0088] S53. Using the selected subset of key features as input and the corresponding fatigue life as output, train a Stacking ensemble learning model. During training, cross-validation is used to determine the model parameters.
[0089] S54. After model training is complete, save the complete ensemble model, all parameters required for feature standardization, and a list of key feature indices. Feature standardization (Z-score standardization) is as follows:
[0090] ;
[0091] in, These are the standardized eigenvalues. These are the original eigenvalues. The mean of this feature. This represents the standard deviation of the feature.
[0092] S55. When making predictions, the original characteristics of new samples are preprocessed using the saved standardized parameters.
[0093] S56. Based on the saved key feature index, extract the corresponding data from the preprocessed features to form the final model input vector.
[0094] S57. Input the input vector into the loaded ensemble model, perform prediction, and directly output the lifetime value.
[0095] Step S5 has the advantage of strong robustness: the feature selection and ensemble learning framework effectively reduces the model's sensitivity to data noise and irrelevant features, and enhances the stability of the prediction results and its adaptability to different batches of data.
[0096] S6. Prediction Result Output and Evaluation: After processing the defect features of the new sample in the same way, input them into the trained ensemble learning model to obtain its fatigue life prediction value. Simultaneously, the prediction confidence interval or uncertainty measure can be output. For a new sample, repeat steps S1 to S2 to extract its features, and use the saved trained model and key parameters required for data processing, including standardized parameters and feature indices, to process the data and input it into the model to obtain the life prediction value.
[0097] The two-layer Stacking ensemble learning model structure proposed in step S6 above is as follows: Figure 3 As shown, its core lies in integrating the advantages of different base learners. Multiple heterogeneous base learners in the first layer model the problem from different hypothesis spaces, providing diverse initial predictions; the meta-learner in the second layer learns how to optimally combine these initial predictions to form the final decision. This structure prevents data leakage through cross-validation training, ensuring the model's robustness and significantly improving prediction accuracy and stability with small sample sizes compared to a single model.
[0098] It has the characteristic of clear physical meaning: the selected features have clear physical or geometric meanings. The model can not only provide predicted values, but also reveal the degree of influence of different defect attributes on fatigue life through feature importance analysis, which helps to provide guidance for process optimization.
[0099] S7. Validation: To verify the comprehensive performance advantages of this invention, its prediction results are compared with experimental values and prediction results using a single model (such as Support Vector Regression, SVR). The coefficient of determination R² and root mean square error RMSE are used as the main evaluation indicators. Figure 4 As shown, the predicted values of the method of the present invention are closer to the diagonal than the experimental values, the R² is higher, and the prediction error of the test samples is significantly smaller than that of the traditional single model, which proves its superiority in prediction accuracy and generalization ability.
[0100] This invention is highly automated and practical: the process from feature extraction to life prediction is programmable and automated, and it is easy to integrate into the quality control system of additive manufacturing, so as to achieve efficient and objective rapid assessment of component life.
[0101] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for additive manufacturing metal fatigue life prediction based on multi-dimensional defect features and ensemble learning, characterized in that: Includes S1, defect 3D data acquisition and extraction; S2, Calculation of multi-dimensional defect geometric features; S3. Feature dataset construction and standardization processing; S4. Key Feature Subset Filtering; S5. Lifetime prediction based on a two-layer stacking ensemble learning model; S6. Prediction result output and evaluation; S7. Validity verification.
2. The method of claim 1, wherein: In step S1, an X-ray computed tomography system is first used to scan the additively manufactured metal sample to reconstruct the internal three-dimensional structure of the sample. Then, an image processing algorithm is used to identify and separate individual manufacturing defects from the reconstructed volume to obtain a three-dimensional digital model of each defect.
3. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 1, characterized in that: In S2, for each 3D model of a defect in S1, three types of geometric features are calculated in parallel: S21. Dimensional characteristics; Calculate the volume, surface area, and square root of the projected area of the defect onto a plane perpendicular to the principal stress direction; S22, Location characteristics; Calculate the shortest spatial distance from the centroid of the defect to the nearest free surface of the specimen; S23. Morphological characteristics; Calculate the sphericity and aspect ratio of the defect.
4. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 1, characterized in that: In step S3, for each sample, a critical defect is determined according to a preset rule. The multidimensional feature parameters of the critical defect are combined into a feature vector, which corresponds to the actual fatigue life value obtained by the sample through experiments, and together they constitute a sample data. The features of all samples in the dataset are standardized.
5. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 1, characterized in that: In step S4, a tree-based feature importance assessment method is used to calculate the contribution of each feature to lifetime prediction. Based on this, features with the highest importance ranking are selected to form an optimized feature subset.
6. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 1, characterized in that: The two-level Stacking ensemble regression model of S5 includes: The first layer contains multiple heterogeneous regression models as base learners; The second layer uses a meta-learner, and takes the prediction results of all base learners in the first layer as input. It learns the optimal weighted fusion coefficients through training and outputs the final fatigue life prediction value.
7. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 6, characterized in that: The specific steps for training and inference of the integrated learning model include: S51. Standardize the constructed complete feature dataset; S52. Based on the standardized data, use the random forest model to analyze the contribution of each feature to lifetime prediction, and select and record several features and their indices. S53. Using the selected subset of key features as input and the corresponding fatigue life as output, train a Stacking ensemble learning model. During training, cross-validation is used to determine the model parameters. S54. After the model training is completed, save the complete ensemble model, all parameters required for feature standardization, and the list of key feature indices. S55. When making predictions, the original characteristics of new samples are preprocessed using the saved standardized parameters. S56. Based on the saved key feature index, extract the corresponding data from the preprocessed features to form the final model input vector; S57. Input the input vector into the loaded ensemble model, perform prediction, and directly output the lifetime value.
8. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 1, characterized in that: In step S6, after the defect features of the new sample are processed in the same way, they are input into the trained ensemble learning model to obtain its fatigue life prediction value. At the same time, the confidence interval or uncertainty measure of the prediction can be output.
9. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 8, characterized in that: In step S6, for a new sample, steps S1 to S2 are repeated to extract its features. The system utilizes the saved trained model and key parameters required for data processing, including standardized parameters and feature indices. After processing, the data is input into the model to obtain lifetime prediction values.
10. The method for predicting the fatigue life of additive manufacturing metals based on multidimensional defect features and ensemble learning according to claim 1, characterized in that: In step S7, the predicted results are compared with the experimental values and the predicted results using a single model, and then the coefficient of determination R² and root mean square error RMSE are used as evaluation indicators.