A fatigue crack prediction method for low-carbon steel produced by arc additive manufacturing based on causality and graph attention.

Abstract

This invention provides a fatigue crack prediction method for low-carbon steel produced by arc additive manufacturing based on causality and graph attention. The method includes: extracting low-dimensional hidden features from the AE signal data of low-carbon steel materials produced by arc additive manufacturing; smoothing the low-dimensional hidden features to reduce the influence of random uncertainty; constructing a causal relationship network representing the spatial relationships of hidden features to achieve causal relationship mining of features in non-Euclidean space; achieving multi-feature information fusion in non-Euclidean space based on a graph attention network model; and exploring transfer strategies with high generalization ability based on different transfer learning strategies to achieve fatigue crack size estimation for a limited number of low-carbon steel materials produced by arc additive manufacturing. This invention automatically extracts features, integrates the temporal and spatial characteristics of the data, and can better mine key information in the data. It has good transferability and effectively solves the problem of online crack evaluation under the anisotropic conditions of material fatigue characteristics caused by the layer-by-layer melting deposition molding of arc additive manufacturing, under limited data conditions.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and in particular, it is a method for predicting fatigue cracks in low-carbon steel produced by electric arc additive manufacturing based on causality and graph attention. Background Technology

[0002] Critical components of key equipment often fail due to wear, corrosion, fatigue, and impact when operating in harsh environments. Restoring and repairing these components can restore them from their end-of-life state to a new condition capable of performing subsequent tasks, thereby reducing raw material usage and life-cycle costs. Additive remanufacturing is a technological process that employs a series of related advanced manufacturing technologies, based on failure analysis and life assessment, to remanufacture and repair damaged or scrapped components, achieving a remanufactured product with the quality of a new product. Electrical arc additive manufacturing (WAAM) is similar to conventional welding but offers higher efficiency and precision. WAAM uses a layer-by-layer construction method to generate parts, so different regions of the part experience different thermal cycles and heat accumulations, further leading to different microstructures between each deposited layer. Compared to conventionally processed materials, the unique processing conditions of WAAM result in anisotropic mechanical properties, thus fatigue assessment models in different directions lack universality. As one of the most common failure modes in structures, fatigue damage severely affects the long-term durability of WAAM structures. Therefore, achieving accurate online fatigue crack assessment that considers anisotropy in different directions is crucial for the safety, reliability, and durability of structural systems.

[0003] Acoustic emission (AE) is a promising nondestructive testing (NDT) technique widely used for damage identification in various structures. When a crack occurs or propagates, energy is released as elastic waves, which AE sensors can capture. Compared to other NDT techniques, AE is more sensitive to minor damage, and the signal itself contains damage information related to the damage mechanism and process. Despite extensive research on damage detection using AE, accurate damage estimation remains a challenge. AE signal data is inherently statistical and influenced by many factors such as material, geometry, instrumentation, and operator. Traditional methods extract features from AE signals based on expert experience to describe the damage process, such as amplitude, count, duration, energy, and rise time. However, due to the anisotropy-induced changes in AE signals, which often exhibit high uncertainty and low signal-to-noise ratio, traditional AE signal extraction and damage estimation methods struggle to achieve accurate crack size assessment. In recent years, with the acquisition of massive amounts of data and significant improvements in computer software and hardware performance, deep learning methods have rapidly developed. Unlike traditional machine learning methods, deep learning methods can automatically discover and extract information needed for specific tasks from raw data. Meanwhile, deep learning typically requires a large amount of training data. However, research and application of machine learning and deep learning methods are still limited in the problem of fatigue crack size assessment. AE signals are characterized by high sensitivity, high frequency, and high dimensionality, making it difficult for manually extracted features to meet the needs of all tasks. Therefore, traditional methods struggle to achieve accurate crack size assessment. Based on deep learning and AE signals, it is urgent and necessary to seek a fatigue crack prediction method for low-carbon steel using causality and graph attention, to automatically extract information from AE signals and achieve corresponding crack size assessment. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a method for predicting fatigue cracks in low-carbon steel produced by arc additive manufacturing based on causality and graph attention. The method includes: extracting low-dimensional hidden features from the AE signal data of low-carbon steel materials produced by arc additive manufacturing; smoothing the low-dimensional hidden features to obtain smoothed low-dimensional features; constructing a causal relationship network; obtaining the degradation crack size based on a graph attention network model; training to obtain the network weights for evaluating the degradation crack size; and using transfer learning, employing transfer strategies to fine-tune and obtain the degradation crack size of new specimens with limited data for arc additive manufacturing of low-carbon steel materials. The transfer strategies include a first transfer strategy, a second transfer strategy, a third transfer strategy, and a fourth transfer strategy. The proposed method eliminates the need for manual feature extraction, integrates the temporal and spatial characteristics of the data, and can better extract key information from the original data. It exhibits good transferability in fatigue evaluation of specimens in different orientations and can effectively solve the problem of online degradation crack size evaluation for specimens with limited monitoring data.

[0005] This invention provides a method for predicting fatigue cracks in low-carbon steel produced by arc additive manufacturing based on causality and graph attention, comprising the following steps:

[0006] S1. Extracting low-dimensional hidden features from AE signal data of low-carbon steel material produced by arc additive manufacturing: The original AE signal data of low-carbon steel material produced by arc additive manufacturing is collected as input data. A one-dimensional convolutional autoencoder is used to reconstruct the AE signal data of low-carbon steel material produced by arc additive manufacturing in the sample dimension, and low-dimensional hidden features that can characterize the state of high-dimensional AE signals are learned. The one-dimensional convolutional autoencoder includes convolutional layers and deconvolutional layers. The convolutional layers include the first N-1 convolutional layers and the Nth convolutional layer. The deconvolutional layers include the first N-1 deconvolutional layers and the Nth deconvolutional layer. The first N-1 convolutional layers all use the ReLU activation function, the first N-1 deconvolutional layers all use the ReLU activation function, and the Nth deconvolutional layer uses the Sigmoid activation function.

[0007] S2. Smoothing low-dimensional hidden features to obtain smooth low-dimensional features: The low-dimensional hidden features for low-carbon steel material obtained by the one-dimensional convolutional autoencoder are smoothed in the feature dimension to obtain smooth low-dimensional features.

[0008] S3. Constructing a causal relationship network: Using a causal discovery algorithm, we can mine the causal relationships between the smooth, low-dimensional features of low-carbon steel materials produced by arc additive manufacturing, and construct a causal relationship network.

[0009] S4. Obtaining the degradation crack size of low-carbon steel material produced by arc additive manufacturing based on a graph attention network model: Using smoothed low-dimensional feature data and a causal relationship network as input, the information between nodes and edges in the causal relationship network graph is aggregated through a graph attention layer to construct a crack representation vector for low-carbon steel material produced by arc additive manufacturing. This vector is then mapped to the degradation crack size sample space through a fully connected layer to obtain the degradation crack size. All graph attention layers use the ReLU activation function, and the last fully connected layer uses the Sigmoid activation function.

[0010] S5. Training to obtain the weights of the degradation crack size assessment network for low-carbon steel material produced by arc additive manufacturing: Based on the AE signal data of the low-carbon steel material specimen produced by arc additive manufacturing, training is performed to obtain the weights of the degradation crack size assessment network.

[0011] S6. Obtaining the degradation crack size of new specimens with limited data on arc-added low-carbon steel based on transfer learning: For the AE signal data of new specimens with limited data on arc-added low-carbon steel, based on different transfer learning strategies, the weights and results of the degradation crack size evaluation network obtained in step S5 are transferred to the crack size prediction of the new specimens, and the degradation crack size on the new specimens is obtained through fine-tuning. In this process, the most effective and generalizable model transfer strategy is explored to achieve the evaluation of degradation crack size of arc-added steel under small sample conditions;

[0012] In S61, the weights of the degradation crack size assessment network are directly transferred to the new specimen to reduce the dimensionality of the original AE signal data of the arc additive low carbon steel material based on the one-dimensional convolutional autoencoder trained in step S1. The extracted low-dimensional hidden features are smoothed based on the data in step S2 to obtain smooth low-dimensional features.

[0013] S62. In the weights of the degradation crack size assessment network, the causal relationship network trained in step S3 is directly transferred to the new specimen as the input of the graph attention network model in step S4.

[0014] In S63, in the weighting of the degradation crack size assessment network, after the graph attention network model and fully connected layer trained in step S4 are transferred to the new specimen, the AE signal data of the new specimen are used for fine-tuning training using a transfer strategy: the transfer strategy includes the first transfer strategy, the second transfer strategy, the third transfer strategy and the fourth transfer strategy.

[0015] Furthermore, in step S63, the first transfer strategy only fine-tunes the fully connected network and is named 2L; the second transfer strategy fine-tunes the last layer of the fully connected network and the graph attention network and is named 2L+1G; the third transfer strategy fine-tunes the last two layers of the fully connected network and the graph attention network and is named 2L+2G; and the fourth transfer strategy fine-tunes the fully connected network and the graph attention network and is named 2L+3G.

[0016] Preferably, in step S1, the one-dimensional convolutional autoencoder has 6 convolutional layers and 6 deconvolutional layers; the kernel size of the first convolutional layer and the sixth deconvolutional layer is set to 80; the kernel size of the second, third, and fourth convolutional layers and the third, fourth, and fifth deconvolutional layers is set to 60; the kernel size of the fifth convolutional layer and the second deconvolutional layer is set to 40; and the kernel size of the sixth convolutional layer and the first deconvolutional layer is set to 20; the stride size of the second, third, fourth, fifth, and sixth convolutional layers and the first, second, third, fourth, and fifth deconvolutional layers is set to 2; and the stride size of the first convolutional layer and the sixth deconvolutional layer is set to 2. The size is set to 4; the number of output channels for the first and fifth convolutional layers and the first and fifth deconvolutional layers is set to 8, the number of output channels for the second and fourth convolutional layers and the second and fourth deconvolutional layers is set to 16, the number of output channels for the third convolutional layer and the third deconvolutional layer is set to 32, and the number of output channels for the sixth convolutional layer and the sixth deconvolutional layer is set to 1; in step S4, the graph attention network model has 3 graph attention layers and 2 fully connected layers, the number of output channels for the first graph attention layer is set to 2, and the number of output channels for the second and third graph attention layers is set to 3; the number of attention mechanism heads for the first, second, and third graph attention layers is set to 10.

[0017] Preferably, in step S1, the number of hidden layer output dimensions of the one-dimensional convolutional autoencoder is much smaller than the number of hidden layer input dimensions, forcing the one-dimensional convolutional autoencoder to learn the low-dimensional hidden features that best represent the original AE signal data.

[0018] Preferably, the data smoothing in step S2 uses the Savitzky-Golay filter; and the causal discovery algorithm in step S3 uses the GOLEM algorithm.

[0019] Preferably, the polynomial order of the Savitzky-Golay filter is set to 3, and the smoothing window length is set to a positive odd integer.

[0020] Preferably, in step S4, the optimizer in the graph attention network model is Adam, and the loss function is the mean absolute error (MAE).

[0021] Compared with the prior art, the technical effects of the present invention are as follows:

[0022] 1. This invention proposes a fatigue crack prediction method for low-carbon steel produced by arc additive manufacturing based on causality and graph attention. The proposed method eliminates the need for manual feature extraction, automatically extracting features from the AE (autocorrelation) signal of the low-carbon steel material produced by arc additive manufacturing. It then uses a causal discovery algorithm and a graph neural network to assess the size of the degradation crack. For example, in step S61, the degradation crack size assessment network weights, the one-dimensional convolutional autoencoder trained in step S1 is directly transferred to a new specimen for dimensionality reduction of the original AE signal data of the low-carbon steel material produced by arc additive manufacturing. No fine-tuning of the one-dimensional convolutional autoencoder is required. The extracted low-dimensional hidden features are smoothed based on the data from step S2 to obtain smoothed low-dimensional features.

[0023] 2. The present invention proposes a fatigue crack prediction method for low-carbon steel produced by electric arc additive manufacturing based on causality and graph attention. It mines the causal network spatial structure between low-dimensional features through a causal discovery algorithm and aggregates information of multiple features in the causal network based on a graph neural network model. Unlike existing technical methods (such as convolutional neural networks) that only consider the temporal correlation between data, the proposed method integrates the temporal and spatial characteristics of the data, realizes the fusion of features in non-Euclidean space, and can better mine key information in the original data.

[0024] 3. The present invention proposes a fatigue crack prediction method for low-carbon steel produced by arc additive manufacturing based on causality and graph attention. The proposed method solves the problem of fatigue anisotropy in different directions and has good transferability between models with different directions. When the framework trained on a certain specimen data is transferred to a new specimen, only the graph neural network model needs to be fine-tuned to realize the assessment of the degradation crack size on the new specimen. It can effectively solve the problem of online fatigue crack size assessment of low-carbon steel specimens produced by arc additive manufacturing with limited monitoring data.

[0025] 4. This invention proposes a fatigue crack assessment method for low-carbon steel samples in different directions using arc additive manufacturing. The proposed method has good transferability. The fatigue crack assessment method obtained on a sample in one direction can be used to assess fatigue cracks on a sample in another direction by only fine-tuning the graph neural network model. This method can effectively solve the problem of fatigue crack assessment with anisotropic characteristics caused by the special process of layer-by-layer deposition in arc additive manufacturing. Attached Figure Description

[0026] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0027] Figure 1 This is a flowchart of the fatigue crack prediction method for low carbon steel based on causality and graph attention in arc additive manufacturing according to the present invention.

[0028] Figure 2 This is a schematic diagram of the fatigue crack prediction method in a specific embodiment of the present invention;

[0029] Figure 3a and Figure 3b These are schematic diagrams of the original AE signal and the signal reconstructed by a one-dimensional convolutional autoencoder for a certain set of samples according to the present invention.

[0030] Figure 4a and Figure 4b These are schematic diagrams of a certain dimension feature before and after smoothing in a specific embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram illustrating the construction of a corresponding causal relationship network in a specific embodiment of the present invention;

[0032] Figure 6 This is a schematic diagram of the crack prediction result obtained after training in a specific embodiment of the present invention;

[0033] Figure 7a This is a comparison diagram of experimental results and comparison methods for transferring the weights of the degraded crack size evaluation network trained on specimen 1 to a new specimen 2 in a specific embodiment of the present invention.

[0034] Figure 7b This is a comparison chart of experimental results and comparison methods for transferring the weights of the degraded crack size evaluation network trained on specimen 1 to a new specimen 3 in a specific embodiment of the present invention.

[0035] Figure 7c This is a comparison chart of experimental results and comparison methods for transferring the weights of the degradation crack size evaluation network trained on specimen 1 to a new specimen 4 in a specific embodiment of the present invention.

[0036] Figure 7d This is a comparison diagram of experimental results and comparison methods for transferring the weights of the degraded crack size evaluation network trained on specimen 1 to a new specimen 5 in a specific embodiment of the present invention.

[0037] Figure 7e This is a comparison chart showing the experimental results and comparison method of transferring the weights of the degradation crack size evaluation network trained on specimen 1 to a new specimen 6 in a specific embodiment of the present invention. Detailed Implementation

[0038] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] Figure 1 This invention illustrates a fatigue crack prediction method for low-carbon steel produced by arc additive manufacturing based on causality and graph attention. The method includes the following steps:

[0041] S1. Extracting low-dimensional hidden features from AE signal data of low-carbon steel material produced by arc additive manufacturing: The original AE signal data of low-carbon steel material produced by arc additive manufacturing is collected as input data. A one-dimensional convolutional autoencoder is used to reconstruct the AE signal data of low-carbon steel material produced by arc additive manufacturing in the sample dimension, and low-dimensional hidden features that can characterize the state of high-dimensional AE signal are learned.

[0042] The experiment involved sampling the arc-generated additive material from different directions to prepare standard fatigue specimens. The samples were then subjected to fatigue loading using an MTS servo-hydraulic fatigue testing machine until fracture. During the experiment, a microscope was used to record the crack length, and an acoustic emission device was used to collect ultrasonic signals.

[0043] The testing system can be divided into three parts: a fatigue testing system, a crack propagation imaging system, and an acoustic emission monitoring system. The fatigue testing system mainly refers to the fatigue testing machine, which cyclically loads the specimen under the set loading conditions and outputs information on force and the number of cycles. The crack propagation imaging system mainly includes an electron microscope, a support, and a monitor, used to observe and record the crack size and shape at different stages during the fatigue test. In addition, in-situ acoustic emission monitoring and acquisition equipment is used to acquire in-situ acoustic emission signal data throughout the fatigue process.

[0044] Two acoustic emission sensors were attached to the specimen surface. Due to the small overall size of the fatigue specimen and the close proximity of the sensors to the crack tip, traditional tape wrapping methods could affect the stress on the specimen. Therefore, appropriate clamps were used for fixation based on the size of the specimen and sensor probes. The sensor probes were fixed to the specimen surface without affecting the stress on the specimen or crack propagation. A grease coupling agent was used to provide reliable acoustic coupling between the sample and the sensors. Both sensors employed independent channels and threshold triggering methods. Appropriate thresholds were set to filter noise between background noise and crack-related ultrasonic signals. In this embodiment, according to the verification procedure of ASTM E976, three lead-breaking tests were performed before each test to ensure the sensitivity between the sensor and the material surface. The one-dimensional convolutional autoencoder had six convolutional layers and six deconvolutional layers; the first five convolutional layers and the first five deconvolutional layers all used the ReLU activation function, while the sixth deconvolutional layer used the Sigmoid activation function.

[0045] The kernel size of the first convolutional layer and the sixth deconvolutional layer is set to 80. The kernel size of the second, third, and fourth convolutional layers and the third, fourth, and fifth deconvolutional layers is set to 60. The kernel size of the fifth convolutional layer and the second deconvolutional layer is set to 40. The kernel size of the sixth convolutional layer and the first deconvolutional layer is set to 20.

[0046] The stride size of the second, third, fourth, fifth, and sixth convolutional layers and the first, second, third, fourth, and fifth deconvolutional layers is set to 2, while the stride size of the first convolutional layer and the sixth deconvolutional layer is set to 4.

[0047] The number of output channels for the first and fifth convolutional layers and the first and fifth deconvolutional layers is set to 8 each; the number of output channels for the second and fourth convolutional layers and the second and fourth deconvolutional layers is set to 16 each; the number of output channels for the third convolutional layer and the third deconvolutional layer is set to 32 each; and the number of output channels for the sixth convolutional layer and the sixth deconvolutional layer is set to 1 each.

[0048] The number of dimensions of the hidden layer output of a one-dimensional convolutional autoencoder is much smaller than the number of dimensions of the input AE signal, enabling the one-dimensional convolutional autoencoder to learn the low-dimensional hidden features that best represent the original AE signal data.

[0049] S2. Smoothing low-dimensional hidden features to obtain smooth low-dimensional features: The low-dimensional hidden features for low-carbon steel material obtained by one-dimensional convolutional autoencoder are smoothed in the feature dimension to reduce the impact of randomness on the quality of the proposed features and obtain smooth low-dimensional features.

[0050] Data smoothing employs a Savitzky-Golay filter. The Savitzky-Golay filter is a finite digital filter based on least squares estimation. The polynomial order of the Savitzky-Golay filter is set to 3, and the smoothing window length is set to a positive odd integer, which can be determined based on the number of samples. A larger window length results in a more pronounced smoothing effect.

[0051] S3. Constructing a causal relationship network: The causal discovery algorithm is used to mine the causal relationships between the smooth low-dimensional features of arc-additive low-carbon steel materials and construct a causal relationship network. The causal discovery algorithm adopts the GOLEM algorithm to mine the spatial relationships between the smooth low-dimensional features of arc-additive low-carbon steel, so as to more reasonably characterize the mutual influence relationship between different features and form a feature representation in non-Euclidean space.

[0052] S4. Obtaining the Degradation Crack Size of Low-Carbon Steel Material Based on a Graph Attention Network Model: Using smoothed low-dimensional feature data and a causal relationship network as input, multi-dimensional feature information within the causal relationship network is fused. Information between nodes and edges in the causal relationship network is aggregated through a graph attention layer to construct a crack representation vector for the low-carbon steel material produced by arc additive manufacturing. This vector is then mapped to the degradation crack size sample space through a fully connected layer to obtain the degradation crack size. All graph attention layers use the ReLU activation function, and the last fully connected layer uses the Sigmoid activation function. Specifically:

[0053] The graph attention network model has 3 graph attention layers and 2 fully connected layers. The graph attention layers aggregate the information between nodes and edges in the causal relationship network graph, and the fully connected layers map it to the degradation crack size sample space of the low carbon steel material produced by arc additive manufacturing to obtain the degradation crack size.

[0054] All graph attention layers use the ReLU activation function, and the last fully connected layer uses the Sigmoid activation function.

[0055] The number of output channels for the first graph attention layer is set to 2, and the number of output channels for the second and third graph attention layers is set to 3 each; the number of attention mechanism heads for the first, second, and third graph attention layers is set to 10 each.

[0056] The optimizer used in the graph attention network model is Adam, and the loss function is the mean absolute error (MAE).

[0057] Existing methods mostly rely on convolutional operations to fuse multiple feature information, but they cannot effectively determine the feature fusion strategy and require training complex models to achieve high crack assessment accuracy. Unlike existing methods, this patent uses causal relationships to characterize the true mutual influence between features. This characterization is more effective and concise, significantly improving the quality of degradation features in low-carbon steel materials produced by arc additive manufacturing. Simultaneously, causal relationships can better enhance the multi-feature fusion efficiency of graph attention networks, enabling information fusion between adjacent feature nodes on the causal relationship network, resulting in a more effective fusion strategy. Utilizing the directed acyclic causal relationships between multiple features significantly reduces the parameters of the graph attention network, minimizing the risk of model overfitting and improving model training efficiency and stability.

[0058] S5. Training to obtain the weights of the degradation crack size assessment network for arc additive low-carbon steel material: Based on the AE signal data of the arc additive low-carbon steel material specimen, training is performed to obtain the weights of the degradation crack size assessment network.

[0059] S6. Obtaining the degradation crack size of a new specimen with limited data on arc-added low-carbon steel based on transfer learning: For the AE signal data of the new specimen with limited data on arc-added low-carbon steel, based on transfer learning, the weights and results of the degradation crack size evaluation network obtained in step S5 are transferred to the crack size prediction of the new specimen, and the degradation crack size on the new specimen is obtained through fine-tuning. This invention, through a thorough study of different transfer strategies, proposes several transfer strategies that are effective and have good generalization ability in crack evaluation of arc-added low-carbon steel.

[0060] In S61, the weights of the degradation crack size assessment network are directly transferred to the new specimen to reduce the dimensionality of the original AE signal data of the low carbon steel material trained in step S1. There is no need to fine-tune the one-dimensional convolutional autoencoder. The extracted low-dimensional hidden features are smoothed based on the data in step S2 to obtain smooth low-dimensional features.

[0061] In step S62, the weights of the degradation crack size assessment network are directly transferred to the new specimen as the input of the graph attention network model in step S4, based on the causal relationship network trained in step S3.

[0062] In S63, in the weighting of the degradation crack size assessment network, after the graph attention network model and fully connected layer trained in step S4 are transferred to the new specimen, the AE signal data of the new specimen are used for fine-tuning training using a transfer strategy: the transfer strategy includes the first transfer strategy, the second transfer strategy, the third transfer strategy and the fourth transfer strategy.

[0063] The first transfer strategy only fine-tunes the fully connected network, named 2L; the second transfer strategy fine-tunes the last layer of both the fully connected network and the graph attention network, named 2L+1G; the third transfer strategy fine-tunes the last two layers of both the fully connected network and the graph attention network, named 2L+2G; and the fourth transfer strategy fine-tunes both the fully connected network and the graph attention network, named 2L+3G.

[0064] In one specific embodiment, the fatigue crack prediction method of the present invention is further verified by combining AE signal data collected from 5 arc additive manufacturing low-carbon steel material specimens and 1 hot-rolled low-carbon steel material specimen.

[0065] Table 1 shows the detailed information of the six specimens in the experiment. The fatigue crack prediction method proposed in this invention was trained on specimen 1. The trained framework was transferred to specimens 2-6, which are made of the same material and in the same direction, or are made of the same material but in a different direction, or are made of different materials, to achieve the corresponding assessment of the size of the degradation crack.

[0066]

[0067] Table 1

[0068] The fatigue crack prediction method for low-carbon steel based on causality and graph attention proposed in this invention is as follows: Figure 2 As shown, the specific implementation steps are as follows:

[0069] S1. Extracting low-dimensional hidden features from AE signal data of low-carbon steel material produced by arc additive manufacturing: The original AE signal data of low-carbon steel material produced by arc additive manufacturing is collected as input data. A one-dimensional convolutional autoencoder is used to reconstruct the AE signal data of low-carbon steel material produced by arc additive manufacturing in the sample dimension, and low-dimensional hidden features that can characterize the state of high-dimensional AE signal are learned.

[0070] The AE signal of specimen 1 was used as input, and a one-dimensional convolutional autoencoder (architecture shown in Table 2) was used to reduce its dimensionality. The output dimension of the hidden layer was set to 48, so the original 9996-dimensional AE signal was reduced to 48-dimensional by the one-dimensional convolutional autoencoder. Figure 3a and Figure 3b (The horizontal axis in the figure represents the sampling time, and the vertical axis represents the signal amplitude.) The figures show the original AE signal and the signal reconstructed by a one-dimensional convolutional autoencoder for a given set of samples. It can be observed that the one-dimensional convolutional autoencoder can reconstruct the input data well and has a certain noise reduction effect. Therefore, the low-dimensional features output by its hidden layer contain a lot of key information that can characterize the original data.

[0071] S2. Smoothing low-dimensional hidden features to obtain smooth low-dimensional features: The low-dimensional hidden features for low-carbon steel material obtained by one-dimensional convolutional autoencoder are smoothed in the feature dimension to obtain smooth low-dimensional features.

[0072] The low-dimensional hidden features output by the one-dimensional convolutional autoencoder are smoothed using a Savitzky-Golay filter with a polynomial order of 3 and a window length of 751. Figure 4a and Figure 4b (The horizontal axis in the figure represents the sampling time, and the vertical axis represents the signal amplitude.) The figure shows a one-dimensional feature before and after smoothing. It can be seen that the feature trend is more obvious after smoothing by the Savitzky-Golay filter.

[0073]

[0074] Table 2

[0075] S3. Constructing a causal relationship network: The causal discovery algorithm is used to mine the causal relationships between the smooth low-dimensional features of low-carbon steel materials produced by arc additive manufacturing, and to construct a causal relationship network; the causal discovery algorithm adopts the GOLEM algorithm.

[0076] For the smoothed 48-dimensional features, the GOLEM causal discovery algorithm is used to mine the causal relationships between these features and construct the corresponding causal relationship network, such as... Figure 5 As shown.

[0077] S4. Obtain the degradation crack size of low-carbon steel material by arc additive manufacturing based on graph attention network model: Using smooth low-dimensional feature data and causal relationship network as input, the information between nodes and edges in the causal relationship network graph is aggregated through graph attention layer, and mapped to the degradation crack size sample space of low-carbon steel material by fully connected layer to obtain the degradation crack size.

[0078] The architecture of the graph attention network model used is shown in Table 3.

[0079] S5. Training to obtain the network weights for evaluating the magnitude of degradation cracks in low-carbon steel materials produced by arc additive manufacturing. During the training process, the Adam optimizer was used, with a training set ratio of 0.7, a batch size of 500, an initial learning rate of 0.0001, 3000 iterations, and the mean absolute error (MAE) loss function. The crack prediction results for low-carbon steel materials produced by arc additive manufacturing after training are as follows: Figure 6 As shown.

[0080] To verify the effectiveness of the framework of this invention, the following three comparison methods are used:

[0081] One approach is using a Convolutional Neural Network (CNN). The original AE signal data is used as input, and the CNN directly predicts the crack size in low-carbon steel produced by arc additive manufacturing. The framework of the CNN in this embodiment is shown in Table 4.

[0082]

[0083] Table 3

[0084]

[0085]

[0086] Table 4

[0087] The second approach involves a one-dimensional convolutional autoencoder combined with an artificial neural network. The smoothed 48-dimensional features output from the one-dimensional convolutional autoencoder are used as input, and an artificial neural network (ANN) is employed to predict the crack length of low-carbon steel materials produced by arc additive manufacturing.

[0088] The third approach is a correlation network plus a graph attention network. A smoothed 48-dimensional correlation network is constructed based on the Pearson correlation coefficient. Then, a graph attention network is used to predict the crack length of low-carbon steel materials produced by arc additive manufacturing. The difference from the framework proposed in this invention is that the causal relationship network is replaced with a correlation network.

[0089] The final comparison results between the proposed method and the comparative method are shown in Table 5. It can be found that the fatigue crack prediction method for low carbon steel based on causality and graph attention proposed in this invention performs best in the crack size prediction experiment of specimen 1, which also verifies the effectiveness of the proposed method.

[0090]

[0091] Table 5

[0092] S6. Based on transfer learning, obtain the degradation crack size of a new specimen of low-carbon steel material with limited data.

[0093] The weights of the degradation crack size assessment network trained on specimen 1 were transferred to specimens 2-6, and degradation crack size assessment on the new specimens was achieved through appropriate fine-tuning training.

[0094] To verify the effectiveness of the proposed causal and graph attention-based fatigue crack prediction method for low-carbon steel in arc additive manufacturing during the migration process, the following three comparison methods are used in this embodiment:

[0095] One approach is retraining. The weight parameters of the 3-layer graph attention network and 2-layer fully connected layer trained on test piece 1 are not transferred. Instead, the 3-layer graph attention network and 2-layer fully connected layer are retrained on the new test piece starting from the initial weights.

[0096] The second is the Convolutional Neural Network (CNN). The CNN trained on test piece 1 will be transferred to the new test piece, and all network layers will be fine-tuned.

[0097] The third method is direct transfer without training (NoTrain). The proposed framework trained on test piece 1 is directly transferred to the new test piece without any fine-tuning training.

[0098] During the training process on the new test piece, the batch size was set to 100, the initial learning rate was set to 0.00005, and the remaining parameters were the same as those used in the training process on test piece 1. In addition, to more comprehensively compare the effectiveness of the proposed method, experiments were conducted on the new test piece using different training set ratios (0.1-0.9).

[0099] Finally, the experimental results and comparison methods for transferring the weights of the degradation crack size evaluation network trained on specimen 1 to new specimens 2-6 are obtained. Figures 7a-7e As shown, the fatigue crack prediction method based on causal networks and graph attention networks trained on specimen 1 still performs well when transferred to new specimens made of the same or different materials as the training specimen. Satisfactory results can be obtained by using only 20% to 30% of the new specimen data for fine-tuning the graph attention network model. Furthermore, among the four transfer strategies proposed in this paper, 2L+1G, 2L+2G, and 2L+3G exhibit good and stable performance overall. On the other hand, from the perspective of improving model efficiency by fine-tuning with as few layers as possible, the 2L+1G transfer strategy is more recommended.

[0100] In summary, the results from the embodiments above demonstrate that the method proposed in this invention can not only accurately predict crack size in low-carbon steel materials produced by arc additive manufacturing, but also, through framework transfer, assess fatigue crack size in new specimens. Furthermore, the proposed framework can be applied not only to new specimens with the same material and AE signal direction as the training specimen, but also to new specimens with different AE signal directions and different materials. This excellent transferability is significant for achieving online crack size prediction for specimens with limited data.

[0101] This invention proposes a fatigue crack prediction method for low-carbon steel produced by arc additive manufacturing based on causality and graph attention. The proposed method eliminates the need for manual feature extraction, automatically extracting features from the AE (autocorrelation) signals of the low-carbon steel material. It utilizes a causal discovery algorithm and graph neural networks to characterize and fuse the relationships between crack features in fatigue specimens, enabling the assessment of degradation crack size in specimens with different orientations. The causal discovery algorithm mines the causal network spatial structure between low-dimensional features, and the graph neural network model aggregates the information contained within the causal network. Unlike existing methods such as convolutional neural networks that only consider the temporal correlation between data, the proposed method integrates the temporal and spatial characteristics of the data, enabling better extraction of key information from the original data. The proposed method exhibits good transferability; when a framework trained on one specimen is transferred to a new specimen, only minor adjustments to the graph neural network model are needed to assess the degradation crack size on the new specimen. This effectively solves the problem of online degradation crack size assessment for low-carbon steel specimens produced by arc additive manufacturing with limited monitoring data.

[0102] Finally, it should be noted that the above embodiments are for illustration only and not for limiting the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention without departing from the spirit and scope of the present invention. Any modifications or partial substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting fatigue cracks in low-carbon steel produced by arc additive manufacturing based on causality and graph attention, characterized in that, It includes the following steps: S1. Extracting low-dimensional hidden features from AE signal data of low-carbon steel material produced by arc additive manufacturing: The original AE signal data of low-carbon steel material produced by arc additive manufacturing is collected as input data. A one-dimensional convolutional autoencoder is used to reconstruct the AE signal data of low-carbon steel material produced by arc additive manufacturing in the sample dimension, and low-dimensional hidden features that can characterize the state of high-dimensional AE signals are learned. The one-dimensional convolutional autoencoder includes convolutional layers and deconvolutional layers. The convolutional layers include the first N-1 convolutional layers and the Nth convolutional layer. The deconvolutional layers include the first N-1 deconvolutional layers and the Nth deconvolutional layer. The first N-1 convolutional layers all use the ReLU activation function, the first N-1 deconvolutional layers all use the ReLU activation function, and the Nth deconvolutional layer uses the Sigmoid activation function. S2. Smoothing low-dimensional hidden features to obtain smooth low-dimensional features: The low-dimensional hidden features for low-carbon steel material obtained by one-dimensional convolutional autoencoder are smoothed in the feature dimension to reduce the impact of randomness on the quality of the proposed features and obtain smooth low-dimensional features. S3. Constructing a causal relationship network: Using a causal discovery algorithm, we can mine the causal relationships between the smooth, low-dimensional features of low-carbon steel materials produced by arc additive manufacturing, and construct a causal relationship network. S4. Obtaining the degradation crack size of low-carbon steel material produced by arc additive manufacturing based on a graph attention network model: Using smoothed low-dimensional feature data and a causal relationship network as input, the information between nodes and edges in the causal relationship network graph is aggregated through a graph attention layer to construct a crack representation vector for low-carbon steel material produced by arc additive manufacturing. This vector is then mapped to the degradation crack size sample space through a fully connected layer to obtain the degradation crack size. All graph attention layers use the ReLU activation function, and the last fully connected layer uses the Sigmoid activation function. S5. Training to obtain the weights of the degradation crack size assessment network for low-carbon steel material produced by arc additive manufacturing: Based on the AE signal data of the low-carbon steel material specimen produced by arc additive manufacturing, training is performed to obtain the weights of the degradation crack size assessment network. S6. Based on transfer learning, obtain the degradation crack size of the new specimen with limited data of arc additive low carbon steel material: For the AE signal data of the new specimen with limited data of arc additive low carbon steel material, based on transfer learning, transfer the weights and results of the degradation crack size evaluation network of arc additive low carbon steel material obtained in step S5 to the crack size prediction of the new specimen, and obtain the degradation crack size on the new specimen through fine-tuning; In S61, the weights of the degradation crack size assessment network are directly transferred to the new specimen to reduce the dimensionality of the original AE signal data of the arc additive low carbon steel material based on the one-dimensional convolutional autoencoder trained in step S1. The extracted low-dimensional hidden features are smoothed based on the data in step S2 to obtain smooth low-dimensional features. S62. In the weights of the degradation crack size assessment network, the causal relationship network trained in step S3 is directly transferred to the new specimen as the input of the graph attention network model in step S4. In S63, in the weighting of the degradation crack size assessment network, after the graph attention network model and fully connected layer trained in step S4 are transferred to the new specimen, the AE signal data of the new specimen are used for fine-tuning training using a transfer strategy: the transfer strategy includes the first transfer strategy, the second transfer strategy, the third transfer strategy and the fourth transfer strategy.

2. The method for predicting fatigue cracks in low-carbon steel using arc additive manufacturing based on causality and graph attention according to claim 1, characterized in that, In step S63, the first transfer strategy only fine-tunes the fully connected network and is named 2L; the second transfer strategy fine-tunes the last layer of the fully connected network and the graph attention network and is named 2L+1G; the third transfer strategy fine-tunes the last two layers of the fully connected network and the graph attention network and is named 2L+2G; and the fourth transfer strategy fine-tunes the fully connected network and the graph attention network and is named 2L+3G.

3. The method for predicting fatigue cracks in low-carbon steel using arc additive manufacturing based on causality and graph attention, as described in claim 1, is characterized in that... In step S1, the one-dimensional convolutional autoencoder has six convolutional layers and six deconvolutional layers. The kernel size of the first convolutional layer and the sixth deconvolutional layer is set to 80. The kernel size of the second, third, and fourth convolutional layers and the third, fourth, and fifth deconvolutional layers is set to 60. The kernel size of the fifth convolutional layer and the second deconvolutional layer is set to 40. The kernel size of the sixth convolutional layer and the first deconvolutional layer is set to 20. The stride size of the second, third, fourth, fifth, and sixth convolutional layers and the first, second, third, fourth, and fifth deconvolutional layers is set to 2. The stride size of the first convolutional layer and the sixth deconvolutional layer is set to 2. All are set to 4; the number of output channels for the first and fifth convolutional layers and the first and fifth deconvolutional layers are both set to 8, the number of output channels for the second and fourth convolutional layers and the second and fourth deconvolutional layers are both set to 16, the number of output channels for the third convolutional layer and the third deconvolutional layer are both set to 32, and the number of output channels for the sixth convolutional layer and the sixth deconvolutional layer are both set to 1; in step S4, the graph attention network model has 3 graph attention layers and 2 fully connected layers, the number of output channels for the first graph attention layer is set to 2, and the number of output channels for the second and third graph attention layers is set to 3; the number of attention mechanism heads for the first, second, and third graph attention layers is set to 10.

4. The method for predicting fatigue cracks in low-carbon steel using arc additive manufacturing based on causality and graph attention, as described in claim 1, is characterized in that... In step S1, the number of hidden layer output dimensions of the one-dimensional convolutional autoencoder is much smaller than the number of hidden layer input dimensions, forcing the one-dimensional convolutional autoencoder to learn low-dimensional hidden features that best represent the original AE signal data.

5. The method for predicting fatigue cracks in low-carbon steel using arc additive manufacturing based on causality and graph attention according to claim 1, characterized in that, In step S2, the data smoothing uses the Savitzky-Golay filter; in step S3, the causal discovery algorithm uses the GOLEM algorithm.

6. The method for predicting fatigue cracks in low-carbon steel using arc additive manufacturing based on causality and graph attention according to claim 5, characterized in that, The polynomial order of the Savitzky-Golay filter is set to 3, and the smoothing window length is set to a positive odd integer.

7. The method for predicting fatigue cracks in low-carbon steel using arc additive manufacturing based on causality and graph attention according to claim 1, characterized in that, In step S4, the optimizer in the graph attention network model is Adam, and the loss function is the mean absolute error (MAE).

Images