Multi-modal fusion fatigue life prediction method based on 3D convolutional neural network

Abstract

The invention relates to a multi-modal fusion fatigue life prediction method based on a 3D convolutional neural network. The method comprises the following steps: acquiring a fatigue data set of equipment and setting a region of interest; constructing a multi-modal fusion network based on a 3D ResNet-18 network and a deep neural network, and training the multi-modal fusion network by adopting the fatigue data set to obtain a metal fatigue life prediction model; wherein the 3D ResNet-18 network comprises one layer of initial convolution, four times of downsampling and four groups of residual stacking, the fatigue data set is input into the 3D ResNet-18 network, the 3D ResNet-18 network outputs multi-scale defect features, and the multi-scale defect features and the fatigue data set are used as input of the deep neural network; and obtaining an actual fatigue life prediction result based on the metal fatigue life prediction model. Compared with the prior art, the method has the advantages that the problems that a traditional one-dimensional modal data modeling method is high in dependence on artificial feature extraction and parameter selection and insufficient in generalization are solved, and the like.

Description

Technical Field

[0001] This invention relates to the technical field of fatigue life prediction, and in particular to a multimodal fusion fatigue life prediction method based on 3D convolutional neural networks. Background Technology

[0002] As high-end equipment continues to develop towards greater precision and complexity, the demands for high reliability and safety during its service life are constantly increasing. The fatigue performance of structural materials under complex loads directly affects the reliability of equipment components. Additive manufacturing, due to its significant advantages in high design freedom for complex structures, high material utilization, and low manufacturing costs, has become a cutting-edge technology driving the leapfrog development of aerospace equipment. However, the additive manufacturing process inevitably generates inherent defects such as voids, lack of fusion, and microcracks. These defects often become sources of localized stress concentration, making fatigue cracks more likely to initiate and accelerate, thereby compromising the structural integrity and service reliability of additively manufactured components. Scientific evaluation and accurate prediction of the fatigue strength and lifespan of additively manufactured components are crucial for ensuring the service safety and lifespan reliability of equipment components.

[0003] While life prediction methods based on semi-empirical models and fracture mechanics have achieved success in specific situations, their poor integration and low transferability make it challenging to establish a universal mathematical model that can uniformly describe the complex structural and performance relationships of materials. With the improvement of computing power and the development of artificial intelligence technology, data-driven fatigue life prediction methods have been widely used. However, existing data-driven fatigue life prediction methods rely on single-modal data, extracting key parameters describing defect size, location, morphology, and distribution, and combining them with one-dimensional modal data such as load conditions and mechanical properties to construct a machine learning fatigue life prediction model for the material. But these methods rely only on explicit feature descriptions of a limited number of defects, often failing to fully capture the comprehensive effects of defects under spatial distribution, morphological complexity, and multi-scale interactions. Summary of the Invention

[0004] The purpose of this invention is to achieve collaborative modeling of cross-modal features and end-to-end prediction of fatigue life. It overcomes the problems of traditional one-dimensional modal data modeling methods, which are highly dependent on manual feature extraction and parameter selection and have insufficient generalization. This invention provides a multimodal fusion fatigue life prediction method based on 3D convolutional neural networks.

[0005] The objective of this invention can be achieved through the following technical solutions: A multimodal fusion fatigue life prediction method based on 3D convolutional neural networks, the method includes the following steps: Acquire the fatigue dataset of the equipment and set the region of interest; A multimodal fusion network was constructed based on a 3D ResNet-18 network and a deep neural network. The multimodal fusion network was trained using a fatigue dataset to obtain a metal fatigue life prediction model. The 3D ResNet-18 network consists of one initial convolutional layer, four downsampling operations, and four sets of residual stacking. The fatigue dataset is input into the 3D ResNet-18 network, and the 3D ResNet-18 network outputs multi-scale defect features. The multi-scale defect features and the fatigue dataset are used together as input to the deep neural network. The actual fatigue life prediction results are obtained based on the fatigue life prediction model for metals.

[0006] Furthermore, the 3D ResNet-18 network includes one initial convolutional layer, four downsampling operations, and four sets of residual stacking.

[0007] Furthermore, after the fatigue dataset is input into the 3D ResNet-18 network for initial convolution and downsampling, the regions of interest in the images in the fatigue dataset are first processed by initial convolution, the image resolution is reduced to 112×112 pixels, the number of feature channels is increased to 64, and the depth is kept at 270. Then, the images after initial convolution are downsampled 4 times to obtain shallow generalized features.

[0008] Furthermore, shallow generalization features are used as inputs to the residuals, and the convolution kernel depths of the four residuals are different.

[0009] Furthermore, the 3D ResNet-18 network also includes a network output section. The output of the residual is used as the input to the network output section. After passing through the average pooling layer and fully connected layer of the network output section, multi-scale defect features are obtained.

[0010] Furthermore, the multi-scale defect features are one-dimensional feature vectors of length 128.

[0011] Furthermore, a deep neural network consists of one input layer, seven hidden layers, and one output layer.

[0012] Furthermore, each hidden layer of a deep neural network contains 64 neurons.

[0013] Furthermore, the deep neural network outputs fatigue life prediction values. y pred The loss function of a deep neural network minimizes the fatigue life prediction value. y pred Compared with the true value y test The deviation.

[0014] Furthermore, the fatigue dataset includes process parameters, mechanical properties, and load conditions.

[0015] Compared with the prior art, the present invention has the following beneficial effects: A fatigue life prediction model based on 3D convolutional neural networks is constructed, which integrates multimodal data. This model achieves collaborative modeling of cross-modal features and end-to-end prediction of fatigue life by fusing 3D defect voxels of the region of interest of the specimen with one-dimensional modal data such as process parameters, mechanical properties, and load conditions. This overcomes the problems of traditional one-dimensional modal data modeling methods, which are highly dependent on manual feature extraction and parameter selection and have insufficient generalization. Attached Figure Description

[0016] Figure 1 3D voxels representing the region of interest for a Ti-6Al-4V fatigue specimen; Figure 2 This describes the defect feature extraction process based on 3D ResNet-18. Figure 3 A fatigue life model based on multimodal data fusion; Figure 4 The fatigue life of Ti-6Al-4V is predicted by a multimodal fusion network based on a 3D convolutional neural network. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0018] This invention proposes a multimodal fusion fatigue life prediction method based on 3D convolutional neural networks, applied to the field of mechanical structure strength. Addressing the limitations of existing technologies that rely on single-modal data to build machine learning fatigue life prediction models, this invention utilizes 3D defect voxels of regions of interest (ROIs) in additive manufacturing metal materials, along with one-dimensional modal data such as process parameters, mechanical properties, and load conditions. A multimodal fusion model is established based on 3D convolutional neural networks and deep neural networks. The 3D convolutional neural network adaptively extracts multi-scale defect features from the ROIs of the material. High-dimensional feature vectors representing defect information, along with process parameters, mechanical properties, and load conditions, are used as inputs to the deep neural network fusion layer, achieving joint modeling of cross-modal features and ultimately obtaining the predicted fatigue life value. This multimodal data fusion-based fatigue life prediction method achieves collaborative modeling and information complementarity from multi-source heterogeneous data, overcoming the limitations of single-modal data information in accurately constructing machine learning fatigue life models.

[0019] This invention constructs a fatigue life prediction model based on 3D convolutional neural networks and multimodal data fusion. This model integrates 3D defect voxels in the region of interest of the specimen with one-dimensional modal data such as process parameters, mechanical properties, and load conditions. This achieves collaborative modeling of cross-modal features and end-to-end prediction of fatigue life, overcoming the problems of traditional one-dimensional modal data modeling methods, such as strong dependence on manual feature extraction and parameter selection, and insufficient generalization.

[0020] Ti-6Al-4V samples were prepared using laser powder bed fusion additive manufacturing with different process parameters (laser power, scanning speed, layer thickness, and scanning spacing). Prior to fatigue testing, high-resolution computed tomography (CT) was employed to perform non-destructive testing on the Ti-6Al-4V samples to obtain 3D defect voxels in the region of interest (ROI). The RIO was located in the middle of the working section of the fatigue sample (±4 mm), forming an hourglass-shaped scanning area to ensure the capture of defect features in the main stress-bearing areas. Figure 1 The results show the three-dimensional defect voxel reconstruction of the region of interest (ROI) of the Ti-6Al-4V specimen. Quasi-static tensile tests were performed on the Ti-6Al-4V specimen to obtain its ultimate tensile strength and elongation. Multi-stress fatigue tests with a stress ratio of 0.1 were conducted on the CT-tested Ti-6Al-4V specimen at an ambient temperature of 250℃ to obtain its fatigue life. This led to the establishment of a fatigue dataset containing one-dimensional modal data (process parameters, mechanical properties, and load conditions) and 3D defect voxel data for additively manufactured Ti-6Al-4V.

[0021] This paper utilizes 3D convolutional neural networks to identify and extract defect information. In the 3D convolutional neural network architecture, 3D convolution is used as the core operator. Through stereo perception kernel, local perception and parameter sharing mechanisms are used to extract features from voxelized data, capturing the spatial correlation between adjacent positions in the input data and revealing the spatial layout and morphological features of defects. Pooling layers further reduce the feature dimension by downsampling the output of convolutional layers, while enhancing the model's translation invariance to the input data and improving its adaptability to defects of different scales and locations. Fully connected layers map the extracted high-dimensional features to the final target output, realizing the mapping from 3D defect images to high-dimensional matrix information.

[0022] A 3D ResNet-18 convolutional neural network was established to identify and extract multi-scale features of defects in the region of interest of Ti-6Al-4V, such as... Figure 2As shown, since 3D defect voxel data cannot be recognized by 3D-CNN, the 3D defect voxel data is converted into two-dimensional images for processing through slicing operations. The region of interest of each sample is reconstructed into 270 slices, each slice with a resolution of 1360×1360 pixels. Therefore, each fatigue sample will contain 270 defect images. The defect images are processed by 3DResNet-18 convolutional layers and downsampling to progressively extract defect features at different scales. The established 3D ResNet-18 network includes one initial convolutional layer, four downsampling operations, and four sets of residual stacking. First, the input defect images are preprocessed to unify the image resolution to 224×224 pixels. After the initial convolutional layer processing, the image resolution is reduced to 112×112 pixels, the number of feature channels is increased to 64, and the depth remains at 270. Subsequently, the first downsampling is performed through a 3×3×3 convolutional kernel with a stride of 2 to further compress the spatial dimension, reducing the feature mapping dimension to 54×54 and the depth to 135. In this stage, the network mainly focuses on extracting shallow generalization features, such as the edge contours and texture features of defects.

[0023] Next, the network input is processed through four sets of residual blocks to achieve multi-layer feature extraction from local geometry to global morphology. Each set of residual blocks undergoes spatial downsampling and semantic abstraction step by step through convolutional kernels of different depths, enabling the network to capture more discriminative defect patterns while maintaining stable gradient propagation. As the network depth increases, the spatial dimension of the feature mapping decreases sequentially to 28×28×68, 14×14×34, and 7×7×17, while the number of feature channels increases progressively to 128, 256, and 512. This hierarchical feature extraction mechanism allows the model to form a deep understanding of the morphological complexity, spatial distribution patterns, and scale-dependent features of defects while preserving global spatial information.

[0024] Finally, the network output is converted into a one-dimensional feature vector of length 128 through an average pooling layer and a fully connected layer, realizing the mapping from high-dimensional spatial features to compact semantic expression.

[0025] Furthermore, a multimodal fusion network was constructed based on 3D ResNet-18 and deep neural networks to predict the fatigue life of Ti-6Al-4V, such as... Figure 3 As shown, the model framework consists of two core modules. The 3D ResNet-18 module is responsible for adaptively extracting multi-scale defect features of the region of interest from the 3D defect voxels reconstructed from CT scans, capturing high-dimensional semantic information of the defects in terms of morphology, size, spatial location, and distribution. The deep neural network module is used to fuse structured information from different sources, integrating the high-dimensional semantic features of the defects with one-dimensional modal data such as process parameters, mechanical properties, and load conditions, to achieve joint modeling of cross-modal features and end-to-end prediction of fatigue life.

[0026] During model training, forward propagation is used to calculate the error between predicted and true values, and backpropagation is used to continuously optimize network weight parameters. The 3D ResNet-18 module's parameter updates employ the Adam optimization algorithm, with a batch size of 8 samples. The initial learning rate is set to 0.04, and it decreases by 0.92 times the previous value after every 10 training epochs, enabling the model to converge quickly in the early stages and achieve fine-grained parameter optimization in later stages. The total number of iterations is set to 2000.

[0027] The deep neural network structure consists of one input layer, seven hidden layers, and one output layer. The high-dimensional defect vector, process parameters, mechanical properties, and load conditions extracted by the 3D ResNet-18 module are used as the input layer. The mechanical properties and load conditions include ultimate tensile strength, elongation, stress amplitude, and temperature. The target variable for the output layer is fatigue life. Each hidden layer contains 64 neurons. During training, the Adam optimization algorithm is used with a learning rate of 0.001 and 1000 iterations. The ReLU activation function is used, and a loss function is defined. L ( y test , y pred To minimize the predicted value y pred Compared with the true value y test To mitigate the bias, the model updates parameters and optimizes performance by minimizing the loss function.

[0028] (1) Ti-6Al-4V fatigue samples were randomly divided into training and testing sets, with 80% of the samples used for model training and the remaining 20% ​​used for model testing. The fatigue life prediction results of this method on the testing set are compared with experimental results. Figure 4 As shown in the figure, the comparison between the experimental results and the predicted results shows that the multimodal fusion fatigue life prediction method based on 3D convolutional neural network has high prediction accuracy. 62% of the prediction results are within the 1.5x error band, and 83% of the prediction results are within the 2x error band. It also shows good prediction performance in different life ranges.

[0029] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A multimodal fusion fatigue life prediction method based on 3D convolutional neural networks, characterized in that, The method includes the following steps: Acquire the fatigue dataset of the equipment and set the region of interest; A multimodal fusion network was constructed based on a 3D ResNet-18 network and a deep neural network. The multimodal fusion network was trained using a fatigue dataset to obtain a metal fatigue life prediction model. The 3D ResNet-18 network consists of one initial convolutional layer, four downsampling operations, and four sets of residual stacking. The fatigue dataset is input into the 3D ResNet-18 network, and the 3D ResNet-18 network outputs multi-scale defect features. The multi-scale defect features and the fatigue dataset are used together as input to the deep neural network. The actual fatigue life prediction results are obtained based on the metal fatigue life prediction model.

2. The multimodal fusion fatigue life prediction method based on 3D convolutional neural networks according to claim 1, characterized in that, The 3D ResNet-18 network includes one initial convolutional layer, four downsampling operations, and four sets of residual stacks.

3. The multimodal fusion fatigue life prediction method based on 3D convolutional neural networks according to claim 1, characterized in that, After the fatigue dataset is input into the 3D ResNet-18 network for initial convolution and downsampling, the regions of interest in the images in the fatigue dataset are first processed by initial convolution, the image resolution is reduced to 112×112 pixels, the number of feature channels is increased to 64, and the depth is kept at 270. Then, the images after initial convolution are downsampled 4 times to obtain shallow generalized features.

4. The multimodal fusion fatigue life prediction method based on 3D convolutional neural networks according to claim 3, characterized in that, Shallow generalization features are used as input to the residuals, and the convolution kernels of the four residuals have different depths.

5. The multimodal fusion fatigue life prediction method based on 3D convolutional neural networks according to claim 4, characterized in that, The 3D ResNet-18 network also includes a network output section. The output of the residual is used as the input of the network output section. After passing through the average pooling layer and fully connected layer of the network output section, multi-scale defect features are obtained.

6. The multimodal fusion fatigue life prediction method based on 3D convolutional neural networks according to claim 5, characterized in that, The multi-scale defect features are one-dimensional feature vectors of length 128.

7. The multimodal fusion fatigue life prediction method based on 3D convolutional neural networks according to claim 1, characterized in that, A deep neural network consists of one input layer, seven hidden layers, and one output layer.

8. The multimodal fusion fatigue life prediction method based on 3D convolutional neural network according to claim 7, characterized in that, Each hidden layer of a deep neural network contains 64 neurons.

9. The multimodal fusion fatigue life prediction method based on 3D convolutional neural network according to claim 8, characterized in that, Deep neural networks output fatigue life prediction values y pred The loss function of a deep neural network minimizes the fatigue life prediction value. y pred Compared with the true value y test The deviation.

10. The multimodal fusion fatigue life prediction method based on 3D convolutional neural networks according to claim 1, characterized in that, The fatigue dataset includes process parameters, mechanical properties, and load conditions.

Images