X70 steel gas pipeline multi-working-condition service life prediction method based on transfer learning

Abstract

The invention belongs to the technical field of material life prediction, and particularly relates to an X70 steel gas pipeline multi-working-condition life prediction method based on transfer learning. Aiming at the problem of low prediction precision caused by dependence on single working condition data and large distribution difference between laboratory and actual service environment data in a traditional life prediction method, the invention provides a cross-domain transfer learning framework. The method comprises the following steps: acquiring mechanical property data through slow strain tensile tests of different hydrogen doping ratios in a laboratory environment, and constructing a domain self-adaptive life prediction model in combination with multi-modal feature decoupling, dynamic distribution alignment and physical constraint fusion strategies. The method significantly reduces actual scene data requirements, realizes high-precision prediction of the residual life of the pipeline under multiple working conditions, is suitable for oil and gas pipeline health monitoring and hydrogen embrittlement risk assessment, and has the advantages of high efficiency, low cost and high reliability.

Description

Technical Field

[0001] The present application relates to the technical field of material life prediction, and in particular to a multi-operating condition life prediction method for X70 steel gas pipelines based on transfer learning. Background Art

[0002] X70 pipeline steel, a high-strength, low-alloy steel, is widely used in long-distance oil and gas pipeline projects. Its service environments are often accompanied by hydrogen permeation (e.g., hydrogen energy transmission or oil and gas mixed transmission), leading to hydrogen embrittlement, significantly reducing the material's mechanical properties and accelerating pipeline failure. As a core material for oil and gas pipelines, X70 pipeline steel is susceptible to hydrogen embrittlement, leading to mechanical property degradation in hydrogen-containing environments, significantly shortening pipeline life. Traditional life prediction methods rely primarily on laboratory simulation data (e.g., yield strength and tensile strength at a fixed hydrogen doping ratio) or statistical models based on field monitoring. However, laboratory conditions struggle to replicate the dynamic complexity of actual service environments (e.g., pressure fluctuations, temperature variations, and multi-factor coupling), resulting in limited generalization of empirical models. Field data collection is costly and time-consuming, making it difficult to meet the full lifecycle monitoring requirements, especially in high-risk areas containing hydrogen. Existing machine learning methods have attempted to utilize laboratory data for modeling, but when directly applied to real-world scenarios, prediction errors exceeding 30% occur due to domain shift, failing to meet engineering accuracy requirements.

[0003] Existing research in the field of hydrogen pipeline life prediction faces two core challenges. First, the feature space misalignment between laboratory data and actual operating conditions manifests itself in the difficulty of effectively mapping the dynamic characteristics of the real-world hydrogen concentration field and multi-physics coupling effects from static hydrogen doping test data. Second, traditional transfer learning methods (such as feature matching) lack dynamic adaptability and are unable to respond in real time to changes in the service environment and the evolution of the life cycle, leading to performance degradation of monitoring models over long periods of time. Although transfer learning techniques offer new avenues for cross-domain knowledge transfer, their practical applications are still limited by two key bottlenecks. First, the data distributions of laboratory and engineering scenarios differ significantly; even pipelines under the same operating conditions can exhibit heterogeneous degradation characteristics. Second, existing deep network algorithms generally rely on the fundamental assumptions of "same distribution of training and test data" and "large data volumes." This fundamentally conflicts with the domain shift caused by data scarcity and dynamic operating conditions in real-world projects, ultimately leading to reduced accuracy in remaining useful life (RUL) prediction. Overcoming these bottlenecks urgently requires the development of a new prediction framework that integrates laboratory prior knowledge with dynamic feature adaptation to achieve both cross-domain knowledge transfer and continuous model evolution. Summary of the Invention

[0004] In view of the shortcomings of the existing method, the present application solves the problem that the conventional life model in the prior art relies on single working condition data and cannot be effectively migrated to new working conditions such as hydrogen mixing.

[0005] Therefore, the present application provides a kind of X70 steel gas pipeline multi-condition life prediction method based on transfer learning, realize the change of life under different conditions accurately, can help engineer to design X70 steel gas pipeline scheme reasonably, avoid the safety problem caused by hydrogen embrittlement and aging of pipeline.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is:

[0007] A kind of X70 steel gas pipeline multi-condition life prediction method based on transfer learning, comprising the following steps:

[0008] Step 1: install high-pressure gas cylinder, gas booster system and slow strain tensile testing machine on the slow strain rate tensile testing system, carry out slow strain tensile test under different conditions, and collect the data of yield strength, tensile strength, elongation at break and reduction of area under different conditions during the whole experiment;

[0009] Step 2: according to the collected experimental data, time domain and frequency domain multi-sensor feature extraction is carried out, multi-dimensional multi-sensor feature vector is formed, and normalization processing is carried out through gradual domain adaptation strategy; dynamically adjust the target domain data statistics to approximate the actual distribution;

[0010] Step 3: according to the slow strain tensile test (SSRT) data and micro fracture mechanism analysis, the corresponding residual life label is obtained, the slow strain tensile mechanical response data under different gas environments (pure methane, pure nitrogen, nitrogen mixed with hydrogen) are respectively taken as source domain data set and target domain data set, a multi-modal feature decoupling transfer learning network life prediction model based on domain self-adaptation is built, the domain classification loss, dynamic distribution alignment (DDA) loss and time degradation feature reconstruction loss are introduced to reduce the difference between the two domains, the residual life prediction model is iteratively trained by using stochastic gradient descent (SGD) combined with momentum optimization algorithm, and finally the residual life prediction result of the target domain gas pipeline is output.

[0011] Preferably, in step 1, the test gas environment includes pure methane, pure nitrogen and nitrogen mixed with hydrogen; the slow strain rate tensile testing system includes high-pressure gas cylinder, gas booster system and slow strain tensile testing machine, and the test conditions cover different hydrogen mixing ratio, pressure and standing time, and the multi-condition mechanical response data is obtained.

[0012] Preferably, the multi-condition slow strain tensile mechanical property parameters are normalized, and the formula is as follows:

[0013]

[0014] Where x is the original characteristic parameter (including yield strength, tensile strength, elongation at break, hydrogen-induced crack growth rate, etc.), μ source and σ source Respectively represent the mean and standard deviation of the corresponding features of the source domain (pure methane / nitrogen environment) data. For the target domain (nitrogen-doped hydrogen environment), a progressive domain adaptation strategy is adopted to dynamically adjust the normalization parameters of the target domain data during the training process. Specifically, the progressive domain adaptation strategy in step 2 uses the sliding average method to dynamically update the target domain normalization parameters. The formula is:

[0015]

[0016] In the formula, μ batch and σ batch is the mean and standard deviation of the target domain data of the current batch, μ target and σ target is the overall mean and standard deviation of the target domain, α is the momentum coefficient (taken as 0.9), and the real distribution of the target domain is gradually approached by sliding average.

[0017] Preferably, the temporal feature extraction module adopts a bidirectional LSTM network combined with a strain rate self-attention mechanism to extract the temporal degradation features of the slow strain curve; the domain adversarial adaptation module realizes the feature distribution alignment of the source domain and the target domain through a gradient reversal layer and a multi-core MMD loss.

[0018] Preferably, the physical constraint decoupling module embeds residual constraints of the hydrogen embrittlement kinetic equation, including the Paris formula for crack propagation and the Fick law for hydrogen diffusion, which is:

[0019]

[0020] Among them, L physics Physical constraint loss, is the crack growth rate, which represents the growth of the crack length a in each load cycle, that is, da / dN, C and m are the crack growth coefficients, ΔK is the stress intensity factor amplitude, C H is the hydrogen concentration, t is the time, D H is the hydrogen diffusion coefficient, is the Laplace operator of hydrogen concentration, that is, the divergence of the concentration gradient.

[0021] Preferably, the lifespan prediction model based on domain adaptive multimodal feature decoupling transfer learning includes:

[0022] Temporal feature extraction module: A bidirectional LSTM network is used to extract temporal degradation features (such as strength decay rate and accumulated plastic deformation) from slow strain tensile curves, and a strain rate self-attention mechanism is embedded to enhance the characteristic representation of the hydrogen-induced crack sensitive stage;

[0023] Domain adversarial adaptation module: Constructs a conditional domain classifier (Conditional Domain Adversarial Network) and uses a gradient reversal layer to align the feature distributions of the source domain (pure methane / nitrogen) and the target domain (nitrogen mixed with hydrogen) at the same strain stage;

[0024] Multi-scale distribution alignment module:

[0025] Macroscale: The inter-domain distribution differences of mechanical performance parameters (yield strength, tensile strength) are measured based on Wasserstein distance;

[0026] Microscale: Using Graph Convolutional Networks (GCN) to model the topological similarity of hydrogen-induced crack propagation paths;

[0027] Life prediction and hydrogen embrittlement assessment dual-task module:

[0028] Main task branch: Predict the remaining lifespan through fully connected layer regression, and use Huber loss as the loss function to enhance outlier robustness;

[0029] Auxiliary task branch: Based on the gating mechanism, the fracture SEM image features are integrated to output the hydrogen embrittlement sensitivity coefficient (HICIndex);

[0030] Parameter dynamic migration module: The parameters of the source domain pre-trained model are initialized to the target domain network through the residual parameterization transfer (ResPT) strategy, retaining the general mechanical laws while adaptively adjusting the parameters related to the hydrogen environment.

[0031] Preferably, the domain-adaptive multimodal feature decoupling transfer learning lifespan prediction model includes: a multimodal encoder module, a domain adversarial feature aligner, a physical constraint decoupler, and a cross-modal fusion decision module;

[0032] The multimodal encoder module includes a bidirectional GRU layer, a multi-head self-attention layer, a ViT block embedding layer, and a hierarchical graph convolution layer;

[0033] The domain adversarial feature aligner includes a dynamic gradient reversal layer and a multi-core MMD alignment layer;

[0034] The multi-physics constraint decoupler includes a Gram-Schmidt orthogonal decomposition layer and a hydrogen embrittlement dynamics residual constraint layer;

[0035] The multi-cross-modal fusion decision module includes a multi-head cross attention layer and a gated fusion fully connected layer;

[0036] First, preprocessed slow-strain tensile time series data (yield strength, tensile strength, and strain rate) are fed into a feature extractor consisting of a bidirectional GRU layer and a multi-head self-attention layer to learn the time-dependent pattern of material strength degradation and key failure stage characteristics. Simultaneously, SEM image data are fed into a ViT block embedding layer and a hierarchical graph convolution layer to extract deep spatial topological information about crack propagation. Subsequently, the hydrogen concentration field features output by the GAT layer of the physical prior encoder, the 32-dimensional feature vector output by the compression layer of the mechanical temporal encoder, and the 256-dimensional feature vector output by the pooling layer of the image encoder are fed into a domain adversarial feature aligner. Dynamic gradient inversion layers and multi-core MMD alignment layers are used to eliminate the distribution differences between the source domain (pure methane / nitrogen) and the target domain (nitrogen doped with hydrogen). Finally, the fused features generated by the multi-head cross-attention layer in the cross-modal fusion decision module are fed into a regression module consisting of gated fusion fully connected layers to obtain the final remaining life prediction results. A hydrogen embrittlement kinetic residual constraint layer ensures that the predicted values ​​conform to the physical constraints of the Paris formula and Fick's law.

[0037] Preferably, during the model training process, the loss objective function of the life prediction model based on domain adaptive multimodal feature decoupling transfer learning includes:

[0038] Main task regression loss (L regression ): Adaptive Huber loss is used to dynamically balance the mean square error and absolute error:

[0039]

[0040] y is the actual remaining life, is the predicted value; δ decays linearly from the initial value 10 to 5 (decaying by 0.5 per epoch).

[0041] Domain adversarial loss (L domain ): Contains intra-modal adversarial loss and cross-modal distribution alignment loss:

[0042]

[0043] Gradient reversal adversarial loss:

[0044]

[0045] Multi-core MMD loss:

[0046]

[0047] φ k is the Gaussian kernel function (σ=1,5,10), β=0.3

[0048] (3) Feature orthogonality constraint loss (L orthogonal ): Forced mechanical feature Z mech and hydrogen damage characteristics Z damage Orthogonal:

[0049]

[0050] Orthogonalized weight λ ortho =0.5

[0051] (4) Physical mechanism constraint loss (L physics ):Embedded hydrogen embrittlement kinetic equation residual:

[0052]

[0053] Crack growth coefficient C = 2.5 × 10-11, m = 3.2; hydrogen diffusion coefficient DH = 1.8 × 10 -9 m 2 / s

[0054] (5) Auxiliary classification loss (L class ): Cross entropy loss for hydrogen embrittlement damage level classification:

[0055]

[0056] y c is the actual injury level (0-no injury, 1-mild, 2-severe); p c Predict probabilities for classification

[0057] (6) Regularization loss (L reg ):Weight decay prevents overfitting:

[0058]

[0059] The final optimized loss objective function is obtained. The loss weight configurations are shown in Table 1 and the expression is:

[0060] L total =L regression +λ adv L domain +λ ortho L orthogonal +λ physics L physics +λ class L class +λ reg L reg (12)

[0061] Among them, L total is the total loss, λ regression is the regression loss weight, L regression is the regression loss, λadv is the domain adversarial loss weight, L domain is the domain adversarial loss, λ physics is the physical loss weight, L physics is the physical constraint loss, λ ortho is the orthogonal loss weight, L orthogonal is the orthogonality constraint loss, λ reg is the regression loss weight, L reg is the regularization loss. The loss weight configuration is shown in Table 1.

[0062] Table 1 Loss weight configuration

[0063]

[0064] Preferably, the model initializes the target domain network through a residual parameterization migration strategy, retains the general mechanical law parameters of the source domain, and dynamically adjusts the hydrogen environment related parameters.

[0065] On the other hand, the present application also provides a pipeline health monitoring system, characterized by comprising:

[0066] Input unit, real-time input of X70 steel gas pipeline pressure, temperature and hydrogen concentration data;

[0067] and a monitoring unit, which is deployed with a model trained using the method described above, and outputs a remaining life prediction result and a hydrogen embrittlement risk index based on the input data of the input unit.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] Multimodal physics embedding architecture: The first three-modal decoupling encoding that integrates mechanics, image processing, and physics equations, breaking through the limitations of a single data source.

[0070] Dynamic adversarial migration mechanism: achieves progressive domain alignment and hybrid distribution matching to solve the problem of cross-operation data drift;

[0071] Small sample generalization capability: Through residual parameter migration and physical constraint pre-training, the dependence on labeled data is significantly reduced;

[0072] Industrial-grade real-time inference: Optimizes model structure and loss function to meet millisecond-level response requirements in engineering scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0074] Figure 1 This is a flow chart of a multi-condition life prediction method for X70 steel gas pipelines based on transfer learning;

[0075] Figure 2 The normalized pipeline life prediction results are compared in the following graph. DETAILED DESCRIPTION

[0076] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0077] The present application will be further described below with reference to the accompanying drawings and examples, which will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be pointed out that those skilled in the art can make several changes and improvements without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0078] The present application aims at the problems that it is difficult to obtain sufficient sample size of full life cycle experimental data under different variable conditions in practical application, and there are significant domain differences in material degradation mechanism under different variable environments (such as pure methane / nitrogen and nitrogen mixed with hydrogen), etc. A kind of X70 steel gas pipeline multi-condition life prediction method based on transfer learning is proposed. By constructing a cross-domain feature decoupling-physical constraint fusion architecture, the following innovative solutions are realized.

[0079] Example one: a kind of X70 steel gas pipeline multi-condition life prediction method based on transfer learning, as shown in Figure 1 The method comprises the following steps:

[0080] Step 1, multi-source data acquisition and cross-domain feature construction:

[0081] 1. Multi-source data acquisition and preprocessing

[0082] A high-pressure gas cylinder, a gas booster system and a slow strain rate tensile testing machine are installed on a slow strain rate tensile testing system to carry out slow strain rate tensile tests under different conditions, and the data of yield strength, tensile strength, elongation and reduction of area under different conditions during the whole experiment are collected, as shown in Table 1.

[0083] Table 1: Summary of X70 steel gas pipeline multi-condition mechanical property data

[0084]

[0085]

[0086] Source domain data: mechanical response data (such as tensile strength, yield strength) under pure methane and pure nitrogen environment, for example: pure nitrogen environment: tensile strength 724.284 MPa, yield strength 662.024 MPa.

[0087] Target domain data: Mechanical response data in a nitrogen-hydrogen mixed environment, for example: in a 20% hydrogen-doped environment: tensile strength 719.047 MPa, yield strength 644.441 MP.

[0088] 2. Perform time-domain and frequency-domain multi-sensor feature extraction based on the collected experimental data, form multi-dimensional multi-sensor feature vectors, and perform normalization processing.

[0089] The obtained multi-condition slow strain tensile mechanical property parameters are normalized, and the formula is as follows:

[0090]

[0091] where x is the original feature parameter (including yield strength, tensile strength, elongation at break, hydrogen-induced crack propagation rate, etc.), and μsource and σsource represent the mean and standard deviation of the source domain (pure methane / nitrogen environment) data corresponding to the feature. For the target domain (nitrogen-hydrogen mixed environment), a gradual domain adaptation strategy is adopted to dynamically adjust the normalization parameters of the target domain data during training: Specifically, the gradual domain adaptation strategy uses a moving average method to dynamically update the target domain normalization parameters, and the formula is:

[0092]

[0093] where μbatch and σbatch are the statistics of the current batch of target domain data, and α is the momentum coefficient (0.9), which gradually approaches the real distribution of the target domain through moving average.

[0094] Specifically:

[0095] Data standardization: Standardize the features so that their mean is 0 and standard deviation is 1

[0096] scaler_X = StandardScaler()

[0097] X_scaled = scaler_X.fit_transform(X)

[0098] Data normalization: Normalize the target variable y so that its range is between [0, 1]

[0099] scaler_y = MinMaxScaler()

[0100] y_scaled = scaler_y.fit_transform(y.reshape(-1, 1)).flatten() reshape into two-dimensional data and then convert back to one-dimensional data

[0101] A progressive domain adaptation strategy is used to dynamically adjust the mean and standard deviation of the target domain data:

[0102]

[0103] Among them, α = 0.9, source domain parameters μsource = 724.284, σsource = 5.2, target domain batch parameters μbatch = 719.047, σbatch = 4.8.

[0104] Step 2: Multimodal transfer learning model training:

[0105] 1. Template architecture design

[0106] The corresponding remaining life labels are obtained based on slow strain tensile test (SSRT) data and micro-fracture mechanism analysis. The slow strain tensile mechanical response data under different gas environments (pure methane, pure nitrogen, and nitrogen mixed with hydrogen) are used as the source domain dataset and target domain dataset, respectively. A multimodal feature decoupling transfer learning network life prediction model based on domain adaptation is constructed. The difference between the two domains is narrowed by introducing domain classification loss, dynamic distribution alignment (DDA) loss, and temporal degradation feature reconstruction loss. The stochastic gradient descent (SGD) combined with the momentum optimization algorithm is used to iteratively train the remaining life prediction model, and finally the remaining life prediction results of the gas pipeline in the target domain are output.

[0107] Temporal feature extraction module: A bidirectional LSTM network is used to extract temporal degradation features (such as strength decay rate and accumulated plastic deformation) from the slow strain tensile curve, and a strain rate self-attention mechanism is embedded to enhance the characteristic representation of the hydrogen-induced crack sensitive stage. The temporal feature extraction module uses a bidirectional LSTM network combined with a strain rate self-attention mechanism to extract the temporal degradation features of the slow strain curve.

[0108] Image feature extraction module: ViT block embedding layer processes the fracture SEM image and extracts the topological structure of hydrogen-induced cracks (such as fractal dimension 1.78).

[0109] Domain Adversarial Adaptation Module: Builds a Conditional Domain Adversarial Network (CDAN) and uses a gradient reversal layer to align the feature distributions of the source domain (pure methane / nitrogen) and the target domain (nitrogen doped with hydrogen) at the same strain stage. This layer aligns the feature distributions of the source and target domains, dynamically adjusting weights to focus on key features (such as yield strength). The CDAN uses a gradient reversal layer and multi-core MMD loss to align the feature distributions of the source and target domains.

[0110] Multi-scale distribution alignment module:

[0111] Macro-scale: measure the distribution difference of mechanical property parameters (yield strength, tensile strength) between domains based on Wasserstein distance;

[0112] Micro-scale: model the topological structure similarity of hydrogen-induced crack propagation path using graph convolution network (GCN);

[0113] Life prediction and hydrogen embrittlement evaluation dual-task module:

[0114] Main task branch: predict the remaining life through a fully connected layer, and use Huber loss to enhance the robustness of abnormal values;

[0115] Auxiliary task branch: fuse the fracture SEM image features based on the gating mechanism, and output the hydrogen embrittlement sensitive coefficient (HICIndex);

[0116] Parameter dynamic transfer module: initialize the target domain network through residual parameter transfer (ResPT) strategy based on the source domain pre-training model parameters, while retaining the general mechanical law and adaptively adjusting the hydrogen environment related parameters.

[0117] Preferably, the multi-modal feature decoupling transfer learning life prediction model based on domain adaptation includes: a multi-modal encoder module, a domain adversarial feature aligner, a physical constraint decoupler, and a cross-modal fusion decision module.

[0118] The multi-modal encoder module includes a bidirectional GRU layer, a multi-head self-attention layer, a ViT block embedding layer, and a hierarchical graph convolution layer.

[0119] The domain adversarial feature aligner includes a dynamic gradient reversal layer and a multi-kernel MMD alignment layer.

[0120] The multi-physical constraint decoupler includes a Gram-Schmidt orthogonal decomposition layer and a hydrogen embrittlement dynamics residual constraint layer. The physical constraint decoupling module embeds the residual constraint of the hydrogen embrittlement dynamics equation, including the crack propagation Paris formula and the hydrogen diffusion Fick law, and the formula is:

[0121]

[0122] where L physics Physical constraint loss, is the crack propagation rate, which represents the increase of crack length a in each load cycle cycle, i.e. da / dN, C and m are crack propagation coefficients, ΔK is the stress intensity factor amplitude, C H is the hydrogen concentration, t is the time, D H is the hydrogen diffusion coefficient, is the Laplacian of hydrogen concentration, i.e. the divergence of concentration gradient.

[0123] The multi-cross modality fusion decision module includes a multi-head cross attention layer and a gated fusion fully connected layer;

[0124] First, preprocessed slow-strain tensile time series data (yield strength, tensile strength, and strain rate) are fed into a feature extractor consisting of a bidirectional GRU layer and a multi-head self-attention layer to learn the time-dependent pattern of material strength degradation and key failure stage characteristics. Simultaneously, SEM image data are fed into a ViT block embedding layer and a hierarchical graph convolution layer to extract deep spatial topological information about crack propagation. Subsequently, the hydrogen concentration field features output by the GAT layer of the physical prior encoder, the 32-dimensional feature vector output by the compression layer of the mechanical temporal encoder, and the 256-dimensional feature vector output by the pooling layer of the image encoder are fed into a domain adversarial feature aligner. Dynamic gradient inversion layers and multi-core MMD alignment layers are used to eliminate the distribution differences between the source domain (pure methane / nitrogen) and the target domain (nitrogen doped with hydrogen). Finally, the fused features generated by the multi-head cross-attention layer in the cross-modal fusion decision module are fed into a regression module consisting of gated fusion fully connected layers to obtain the final remaining life prediction results. A hydrogen embrittlement kinetic residual constraint layer ensures that the predicted values ​​conform to the physical constraints of the Paris formula and Fick's law.

[0125] 2. Hybrid Loss Function and Training Strategy

[0126] During the model training process, the loss objective function of the lifespan prediction model based on domain adaptive multimodal feature decoupling transfer learning includes:

[0127] Main task regression loss (Lregression): Adaptive Huber loss is used to dynamically balance the mean square error and absolute error:

[0128]

[0129] y is the actual remaining life, is the predicted value; δ decays linearly from the initial value 10 to 5 (decaying by 0.5 per epoch).

[0130] Domain adversarial loss (L domain ): Contains intra-modal adversarial loss and cross-modal distribution alignment loss:

[0131]

[0132] Gradient reversal adversarial loss:

[0133]

[0134] Multi-core MMD loss:

[0135]

[0136] φk is the Gaussian kernel function (σ=1,5,10), β=0.3

[0137] (3) Feature orthogonality constraint loss (L orthogonal ):Forced mechanical characteristics Z mech and hydrogen damage characteristics Z damage

[0138] Orthogonal:

[0139]

[0140] Orthogonalized weight λ ortho =0.5

[0141] (4) Physical mechanism constraint loss (L physics ):Embedded hydrogen embrittlement kinetic equation residual:

[0142]

[0143] Crack growth coefficient C = 2.5 × 10-11, m = 3.2; hydrogen diffusion coefficient DH = 1.8 × 10 -9 m 2 / s

[0144] (5) Auxiliary classification loss (L class ): Cross entropy loss for hydrogen embrittlement damage level classification:

[0145]

[0146] y c is the actual injury level (0-no injury, 1-mild, 2-severe); p c Predict probabilities for classification

[0147] (6) Regularization loss (L reg ):L2 weight decay prevents overfitting:

[0148]

[0149] The final optimized loss objective function is obtained, and the loss weight configurations are shown in Table 1. The expression is:

[0150] L total =L regression +λ adv L domain +λ ortho L orthogonal +λ physics L physics +λ class L class +λ reg L reg

[0151] Specifically, in this embodiment, the total loss function is:

[0152] L total =0.5L regression +0.3L domain +0.1L orthogonal +0.1L physics

[0153] Regression loss (Lregression): Adaptive Huber loss, δ decays from 10 to 5.

[0154] Physical constraints (Lphysics): Embedded Paris formula (C = 2.5 × 10-11, m = 3.2) and Fick's law (DH = 1.8 × 10-9m 2 / s).

[0155] First, preprocessed slow-strain tensile time series data (yield strength, tensile strength, and strain rate) are fed into a feature extractor consisting of a bidirectional GRU layer and a multi-head self-attention layer to learn the time-dependent pattern of material strength degradation and key failure stage characteristics. Simultaneously, SEM image data are fed into a ViT block embedding layer and a hierarchical graph convolution layer to extract deep spatial topological information about crack propagation. Subsequently, the hydrogen concentration field features output by the GAT layer of the physical prior encoder, the 32-dimensional feature vector output by the compression layer of the mechanical temporal encoder, and the 256-dimensional feature vector output by the pooling layer of the image encoder are fed into a domain adversarial feature aligner. Dynamic gradient inversion layers and multi-core MMD alignment layers are used to eliminate the distribution differences between the source domain (pure methane / nitrogen) and the target domain (nitrogen doped with hydrogen). Finally, the fused features generated by the multi-head cross-attention layer in the cross-modal fusion decision module are fed into a regression module consisting of gated fusion fully connected layers to obtain the final remaining life prediction results. A hydrogen embrittlement kinetic residual constraint layer ensures that the predicted values ​​conform to the physical constraints of the Paris formula and Fick's law.

[0156] During the model training process, the loss objective function of the lifespan prediction model based on domain adaptive multimodal feature decoupling transfer learning includes:

[0157] Main task regression loss (L regression ): Adaptive Huber loss is used to dynamically balance the mean square error and absolute error:

[0158]

[0159] y is the actual remaining life, is the predicted value; δ decays linearly from the initial value 10 to 5 (decaying by 0.5 per epoch).

[0160] Domain adversarial loss (L domain ): Contains intra-modal adversarial loss and cross-modal distribution alignment loss:

[0161]

[0162] Gradient reversal adversarial loss:

[0163]

[0164] Multi-core MMD loss:

[0165]

[0166] φ k is the Gaussian kernel function (σ=1,5,10), β=0.3

[0167] (3) Feature orthogonality constraint loss (L orthogonal ): Forced mechanical feature Z mech and hydrogen damage characteristics Z damage Orthogonal:

[0168]

[0169] Orthogonalized weight λ ortho =0.5

[0170] (4) Physical mechanism constraint loss (L physics ):Embedded hydrogen embrittlement kinetic equation residual:

[0171]

[0172] Crack growth coefficient C = 2.5 × 10-11, m = 3.2; hydrogen diffusion coefficient DH = 1.8 × 10 -9 m 2 / s

[0173] (5) Auxiliary classification loss (L class ): Cross entropy loss for hydrogen embrittlement damage level classification:

[0174]

[0175] y c is the actual injury level (0-no injury, 1-mild, 2-severe); p c Predict probabilities for classification

[0176] (6) Regularization loss (L reg ):Weight decay prevents overfitting:

[0177]

[0178] The final optimized loss objective function is obtained. The loss weight configurations are shown in Table 1 and the expression is:

[0179] L total =L regression +λ adv L domain +λ ortho L orthogonal +λ physics L physics +λ class L class +λ reg L reg (12)

[0180] Among them, L total is the total loss, λ regression is the regression loss weight, L regression is the regression loss, λ adv is the domain adversarial loss weight, L domain is the domain adversarial loss, λ physics is the physical loss weight, L physics is the physical constraint loss, λ ortho is the orthogonal loss weight, L orthogonal is the orthogonality constraint loss, λ reg is the regression loss weight, L reg is the regularization loss. The loss weight configuration is shown in Table 2.

[0181] Table 2 Loss weight configuration

[0182]

[0183] The model initializes the target domain network through a residual parameterization migration strategy, retains the general mechanical law parameters of the source domain, and dynamically adjusts the hydrogen environment related parameters.

[0184] 3. Training parameters

[0185] Optimizer: Stochastic Gradient Descent (SGD) with momentum (β = 0.9), initial learning rate 0.01, decayed by 50% every 10 epochs.

[0186] Batch size: 32 groups in the source domain, 16 groups in the target domain, and 100 rounds of training.

[0187] Step 3: Cross-domain life prediction and engineering verification:

[0188] 1. Model Validation and Performance Evaluation

[0189] Use the target domain test set (e.g., "weld 6MPa static 48h" data, tensile strength 679.577MPa, yield strength 539.648MPa) to verify the model:

[0190] Table 2 Model validation results

[0191]

[0192] 2. Project deployment and real-time monitoring

[0193] The model is integrated into the pipeline health monitoring system, and field data (such as pressure fluctuation ±6MPa and temperature change ±15℃) is input in real time to output:

[0194] Remaining life prediction: The current remaining life of the target pipeline is 18.3 years (95% confidence level).

[0195] Hydrogen embrittlement risk assessment: HIC Index = 0.42 (threshold > 0.5 indicates high risk), indicating that the risk is controllable.

[0196] 3. Key advantages

[0197] Data efficiency: The target domain only requires 50 sets of labeled data, which is 70% less than traditional methods.

[0198] Dynamic response: When the hydrogen concentration increases from 10% to 30%, the prediction error is stable within ±5%.

[0199] Real-time performance: Single inference takes less than 50ms, supporting millisecond-level working condition response.

[0200] Embodiment 2: A pipeline health monitoring system, characterized by comprising:

[0201] Input unit, real-time input of X70 steel gas pipeline pressure, temperature and hydrogen concentration data;

[0202] and a monitoring unit, which is deployed with a model trained using the method described in Example 1 and outputs a remaining life prediction result and a hydrogen embrittlement risk index based on input data of the input unit.

[0203] This method solves the distribution difference problem between laboratory data and actual working conditions through multimodal feature fusion and dynamic domain adaptation. It achieves an average error of ≤3.5% in multi-scenario life prediction of X70 steel gas pipelines, which is significantly better than the 30% error of traditional models and provides reliable protection for engineering safety.

Claims

1. A multi-condition life prediction method for X70 steel gas pipeline based on transfer learning, characterized in that: The following steps are involved: Step 1: In a slow strain rate tensile test system, collect mechanical property data of X70 steel gas pipeline under different gas environments, including tensile strength, yield strength, elongation at break, and reduction of area; Step 2: Extract multi-sensor features in the time and frequency domains from the mechanical property data to generate multi-dimensional and multi-modal feature vectors, normalize them using a progressive domain adaptation strategy, and dynamically adjust the target domain data statistics to approximate the actual distribution; Step 3: Based on the source domain data and the target domain data, a domain-adaptive multimodal transfer learning network model is constructed. The model includes a temporal feature extraction module, a domain adversarial adaptation module, a physical constraint decoupling module, and a cross-modal fusion decision module. It is iteratively trained by fusing domain classification loss, dynamic distribution alignment loss, and physical mechanism constraint loss to output the remaining life prediction results of the gas pipeline in the target domain.

2. The multi-condition life prediction method for X70 steel gas pipeline based on transfer learning according to claim 1 is characterized in that: In step 1, the test gas environment includes pure methane, pure nitrogen, and nitrogen mixed with hydrogen; the slow strain rate tensile testing system includes a high-pressure gas storage cylinder, a gas pressurization system, and a slow strain tensile testing machine. The test conditions cover different hydrogen mixing ratios, pressures, and static times to obtain mechanical response data under multiple working conditions.

3. The multi-condition life prediction method for X70 steel gas pipeline based on transfer learning according to claim 1 is characterized in that: The progressive domain adaptation strategy in step 2 uses the sliding average method to dynamically update the target domain normalization parameters. The formula is: Among them, μ batch and σ batch is the mean and standard deviation of the target domain data of the current batch, μ target and σ target is the overall mean and standard deviation of the target domain, and α is the momentum coefficient.

4. The multi-condition life prediction method for X70 steel gas pipeline based on transfer learning according to claim 1 is characterized in that: The temporal feature extraction module adopts a bidirectional LSTM network combined with a strain rate self-attention mechanism to extract the temporal degradation features of the slow strain curve; the domain adversarial adaptation module realizes the feature distribution alignment between the source domain and the target domain through a gradient reversal layer and a multi-core MMD loss.

5. The multi-condition life prediction method for X70 steel gas pipeline based on transfer learning according to claim 1 is characterized in that: The physical constraint decoupling module embeds the residual constraints of the hydrogen embrittlement dynamics equation, including the Paris formula for crack propagation and the Fick law for hydrogen diffusion, which are as follows: Among them, L physics Physical constraint loss, is the crack growth rate, which represents the growth of the crack length a in each load cycle, that is, da / dN, C and m are the crack growth coefficients, ΔK is the stress intensity factor amplitude, C H is the hydrogen concentration, t is the time, D H is the hydrogen diffusion coefficient, is the Laplace operator of hydrogen concentration, that is, the divergence of the concentration gradient.

6. The multi-condition life prediction method for X70 steel gas pipeline based on transfer learning according to claim 1 is characterized in that: The cross-modal fusion decision module fuses mechanical timing features, SEM image topological features and hydrogen concentration field features through a multi-head cross-attention layer, and inputs the gated fusion fully connected layer to generate the remaining life prediction value.

7. The multi-condition life prediction method for X70 steel gas pipeline based on transfer learning according to claim 1 is characterized in that: The loss function of the model includes adaptive Huber regression loss, domain adversarial loss, feature orthogonal constraint loss and physical mechanism constraint loss. The total loss function is: L total =λ regression L regression +λ adv L domain +λ physics L physics +λ ortho L orthogonal +λ reg L reg Among them, L total is the total loss, λ regression is the regression loss weight, L regression is the regression loss, λ adv is the domain adversarial loss weight, L domain is the domain adversarial loss, λ physics is the physical loss weight, L physics is the physical constraint loss, λ ortho is the orthogonal loss weight, L orthogonal is the orthogonality constraint loss, λ reg is the regression loss weight, L reg is the regularization loss.

8. The multi-condition life prediction method for X70 steel gas pipeline based on transfer learning according to claim 1 is characterized in that: The model initializes the target domain network through a residual parameterization migration strategy, retains the general mechanical law parameters of the source domain, and dynamically adjusts the hydrogen environment related parameters.

9. A pipeline health monitoring system, characterized in that: include: Input unit, real-time input of X70 steel gas pipeline pressure, temperature and hydrogen concentration data; and a monitoring unit, which is equipped with a model trained by the method according to any one of claims 1 to 8, and outputs a remaining life prediction result and a hydrogen embrittlement risk index according to the input data of the input unit.

Images