Multi-scale data fusion hydrogen permeation-stress prediction method and device

Abstract

The invention discloses a hydrogen permeation-stress prediction method and device based on multi-scale data fusion, and relates to the technical field of material science. The method comprises the following steps: acquiring hydrogen diffusion coefficients of hydrogen atoms in a metal pipeline material, stress field data distribution and experimental data, preprocessing data of different scales, and mapping the data to a unified input space to construct a five-dimensional tensor characteristic matrix; and constructing a forward neural network model, taking the five-dimensional tensor characteristic matrix as input, embedding a hydrogen diffusion equation and a stress balance equation into a loss function for training, and outputting a predicted key index. And constructing a reverse neural network model taking the target performance parameters as input, adding physical constraints during training, searching a variable space based on the predicted key indexes, and obtaining an optimal metal pipeline material parameter combination. The problems that in the prior art, multi-scale data fusion is difficult, complex working conditions are difficult to adapt, prediction efficiency is low, and an accurate and reliable hydrogen embrittlement prediction and prevention and control basis cannot be provided for hydrogen engineering are solved.

Description

Technical Field

[0001] This application relates to the field of materials science and technology, and in particular to a method and apparatus for predicting hydrogen permeation-stress using multi-scale data fusion. Background Technology

[0002] Hydrogen embrittlement is a failure phenomenon in which hydrogen atoms penetrate into metallic materials in a hydrogen-rich environment, accumulate under stress, and cause microcracks to initiate and propagate, ultimately reducing the material's strength and lifespan. This phenomenon poses a serious threat to the safety and reliability of pipeline steel, hydrogen storage materials, and other metallic materials during hydrogen-rich service. In industrial applications, such as hydrogen pipeline networks and hydrogen storage facilities, accurate prediction and control of hydrogen embrittlement are crucial.

[0003] Currently, research on hydrogen permeation and stress coupling mainly follows two technical approaches. The first is based on microscopic models using molecular dynamics or quantum mechanics. These models can accurately describe the diffusion mechanism of hydrogen atoms in metal lattices at the atomic scale, but the computational load is extremely large, limiting their direct application to real-time prediction of macroscopic structures due to limitations in computing resources and time. The second approach is based on finite element method (FEM) multiphysics coupling simulations. These simulations describe the interaction between hydrogen diffusion and stress fields at the macroscopic scale, and can assess key engineering parameters such as hydrogen concentration at crack tips and stress distribution. However, these models are complex and time-consuming to solve, and struggle to quickly provide predictions of hydrogen permeation and stress fields when faced with complex and variable operating conditions.

[0004] Therefore, existing technologies suffer from difficulties in multi-scale data fusion, are ill-suited to complex operating conditions, and have low prediction efficiency, thus failing to provide accurate and reliable basis for hydrogen embrittlement prediction and prevention for hydrogen-related engineering projects. Summary of the Invention

[0005] In this embodiment of the application, a hydrogen permeation-stress prediction method based on multi-scale data fusion is provided, which solves the problems of existing technologies, such as difficulty in multi-scale data fusion, inability to adapt to complex working conditions, low prediction efficiency, and inability to provide accurate and reliable basis for hydrogen embrittlement prediction and prevention for hydrogen-related projects.

[0006] In a first aspect, embodiments of this application provide a multi-scale data fusion method for hydrogen permeation-stress prediction. This method includes: acquiring the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in a metal pipe material; preprocessing the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material, mapping the preprocessed data at different scales to a unified input space, and constructing a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction; constructing a feedforward neural network model for hydrogen permeation-stress prediction, using the five-dimensional tensor feature matrix as input, embedding the hydrogen diffusion equation and stress balance equation in the loss function, and training the feedforward neural network model to learn the feature elements in the five-dimensional tensor feature matrix. The mapping relationship between elements and key indicators is established, and the predicted key indicators are output. A reverse neural network model is constructed with the target performance parameters as input. When training the reverse neural network model, physical constraints of the hydrogen diffusion equation and stress balance equation are added. Based on the predicted key indicators output by the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipe material parameter combination that satisfies the physical constraints and optimizes the target performance parameters. The feedforward neural network obtains the hydrogen permeation and stress field distribution of the material based on the feature elements in the five-dimensional tensor feature matrix. The reverse neural network model then obtains the metal pipe material parameter combination that satisfies the physical constraints and optimizes the target performance parameters.

[0007] One possible implementation also includes: comparing the predicted key indicators output by the feedforward neural network model with experimental data and evaluating the error; if the error exceeds a preset range, then adjusting the feedforward neural network model and the reverse neural network model accordingly.

[0008] In one possible implementation, obtaining the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material includes: simulating hydrogen diffusion in the metal material under different temperature conditions using molecular dynamics simulation to obtain the diffusion behavior of hydrogen atoms at different temperatures, thereby obtaining the hydrogen diffusion coefficient; performing macroscopic-scale hydrogen permeation-stress coupling simulation using finite element simulation to simulate the evolution of hydrogen concentration over time and space under different temperatures, stress states, and boundary conditions, thereby obtaining the stress field data distribution; and obtaining the mechanical property degradation index and hydrogen permeation test data of the metal pipe material under different corrosion levels, temperatures, and stress conditions through experiments, thereby obtaining experimental data.

[0009] In one possible implementation, the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material are preprocessed. The preprocessed data at different scales are then mapped to a unified input space to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction. This includes: filling the first three channels of the five-dimensional tensor feature matrix with the VonMises equivalent stress field, hydrostatic stress field, and current hydrogen concentration field from the stress field data distribution; expanding the hydrogen diffusion coefficient to a full-image feature map with the same spatial resolution as the five-dimensional tensor feature matrix and filling it with the fourth channel; and filling the fifth channel of the five-dimensional tensor feature matrix with the experimental data to construct the five-dimensional tensor feature matrix for hydrogen permeation-stress prediction. The expression for the five-dimensional tensor feature matrix is: ;in, It is a five-dimensional array. For the set of real numbers, For batch size, The length of the time series. This represents the spatial resolution of the metal pipe material in the vertical direction. This represents the spatial resolution of the metal pipe material in the horizontal direction. This represents the number of channels.

[0010] In one possible implementation, the feedforward neural network model for hydrogen permeation-stress prediction is constructed by taking a five-dimensional tensor feature matrix as input, embedding the hydrogen diffusion equation and stress balance equation in the loss function, and training the feedforward neural network model to learn the mapping relationship between each feature element in the five-dimensional tensor feature matrix and key indicators, and outputting the predicted key indicators, including: key indicators such as hydrogen permeation flux, lag time, hydrogen concentration distribution, and stress field distribution; the feedforward neural network model includes convolutional layers, recurrent layers, and fully connected layers; in the convolutional layers, two-dimensional convolutional kernels are used to extract the spatial features of the hydrogen concentration field and stress field, and multi-channel convolution operations are used to capture stress concentration phenomena at crack tips or pits; in the recurrent layers, the feature sequence output by the convolutional layers is unfolded step-by-step to extract the spatial features of the finite element simulation hydrogen concentration field and stress field; in the fully connected layers, the spatial features extracted by the recurrent layers are mapped to the output space to generate the predicted key indicators; the expression of the loss function is: ;in, The value of the loss function. This is the residual term of the hydrogen diffusion equation. These are the weighting coefficients for the residual terms of the hydrogen diffusion equation. This is the residual term of the stress equilibrium equation. These are the weighting coefficients for the residual terms of the stress balance equation. For data fitting terms, These are the weighting coefficients of the data fitting terms; ;in, The hydrogen concentration is the output of the feedforward neural network model. For time, The rate of change of hydrogen concentration over time. The hydrogen diffusion coefficient is... Let be the partial molar volume of hydrogen. The gradient of hydrogen-induced stress, The gas constant is... Absolute temperature For gradient; ;in, For stress tensor, External force vector; ;in, These are the predicted key metric values ​​output by the feedforward neural network model. This represents the true value of the key indicator.

[0011] In one possible implementation, the construction of a reverse neural network model with target performance parameters as input involves incorporating physical constraints from the hydrogen diffusion equation and stress balance equation during training. Based on the predicted key indicators from the output of the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipe material parameter combination that satisfies the physical constraints and optimizes the target performance parameters. This includes: the target performance parameters being the maximum allowable hydrogen concentration and stress limit of the material; and the physical constraints of the hydrogen diffusion equation satisfying: The physical constraints of the stress balance equations satisfy: .

[0012] In one possible implementation, the weight coefficients of each loss term in the loss function are determined based on a preset rule. These weight coefficients include the weight coefficients of the residual terms of the hydrogen diffusion equation, the residual terms of the stress balance equation, and the data fitting term. The preset rule includes: obtaining the gradient norm of each loss term in the loss function relative to the shared weights of the fully connected layers of the feedforward neural network model; where each loss term in the loss function includes the residual terms of the hydrogen diffusion equation, the stress balance equation, and the data fitting term; obtaining the relative training rate of each loss term in the loss function by analyzing the decrease in each loss term during training; obtaining the target gradient norm of each loss term based on the relative training rate of each loss term; and updating the weight coefficients of each loss term in the loss function based on the difference between the gradient norm of each loss term and the target gradient norm.

[0013] Secondly, embodiments of this application provide a multi-scale data fusion hydrogen permeation-stress prediction device. The device includes: a data acquisition module for acquiring the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in a metal pipe material; a matrix construction module for preprocessing the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material, mapping the preprocessed data at different scales to a unified input space, and constructing a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction; and a feedforward neural network model construction module for constructing a feedforward neural network model for hydrogen permeation-stress prediction, using the five-dimensional tensor feature matrix as input, embedding the hydrogen diffusion equation and stress balance equation in the loss function, and training the feedforward neural network model to learn the various features in the five-dimensional tensor feature matrix. The system establishes a mapping relationship between feature elements and key indicators, and outputs predicted key indicators. A reverse neural network model module is constructed to build a reverse neural network model with target performance parameters as input. During training, physical constraints such as the hydrogen diffusion equation and stress balance equation are incorporated. Based on the predicted key indicators output by the feedforward neural network model, the system searches the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix to obtain the metal pipe material parameter combination that satisfies the physical constraints and optimizes the target performance parameters. A parameter combination acquisition module is used to obtain the hydrogen permeation and stress field distribution of the material from the feature elements in the five-dimensional tensor feature matrix by the feedforward neural network, and then obtain the metal pipe material parameter combination that satisfies the physical constraints and optimizes the target performance parameters through the reverse neural network model.

[0014] One or more technical solutions provided in the embodiments of this application have at least the following technical effects: This application provides a multi-scale data fusion method for hydrogen permeation-stress prediction. It acquires hydrogen diffusion coefficients, stress field distribution data, and experimental data from hydrogen atoms in metal pipe materials. These data at different scales are preprocessed and mapped to a unified input space to construct a five-dimensional tensor feature matrix. A feedforward neural network model is constructed, using the five-dimensional tensor feature matrix as input. The hydrogen diffusion equation and stress balance equation are embedded in the loss function for training, outputting key predicted indicators. A reverse neural network model is constructed, using target performance parameters as input. Physical constraints are incorporated during training, and the variable space is searched based on the predicted key indicators to obtain the optimal combination of metal pipe material parameters. This method solves the problems of existing technologies, such as difficulty in multi-scale data fusion, inability to adapt to complex working conditions, and low prediction efficiency, failing to provide accurate and reliable prediction and prevention of hydrogen embrittlement in hydrogen-related engineering projects. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a multi-scale data fusion method for predicting hydrogen permeation-stress, provided as an embodiment of this application; Figure 2 A schematic diagram of a five-dimensional tensor structure provided in an embodiment of this application; Figure 3 A schematic diagram of a hydrogen permeation-stress prediction device with multi-scale data fusion provided in this application embodiment; Figure 4 This is a schematic diagram of a physical device for hydrogen permeation-stress prediction based on multi-scale data fusion, provided as an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] The following description of some technologies involved in the embodiments of this application is provided to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, some descriptions of well-known functions and structures are omitted in the following description.

[0019] This application provides a multi-scale data fusion method for predicting hydrogen permeation-stress, such as... Figure 1 As shown, the method includes steps S101 to S105. Wherein, Figure 1 This is merely one execution order shown in the embodiments of this application and does not represent the only execution order for a multi-scale data fusion hydrogen permeation-stress prediction method. Where the final result can be achieved, Figure 1 The steps shown can be performed in parallel or in reverse order.

[0020] S101: Obtain the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material.

[0021] Obtain the hydrogen diffusion coefficient, stress field distribution, and experimental data of hydrogen atoms in the metal pipe material, including the following:

[0022] Hydrogen diffusion in metallic materials was simulated using molecular dynamics simulations at different temperature conditions to obtain the diffusion behavior of hydrogen atoms at different temperatures, thereby obtaining the hydrogen diffusion coefficient.

[0023] Specifically, the molecular dynamics simulation software can be LAMMPS.

[0024] The evolution of hydrogen concentration over time and space under different temperatures, stress states, and boundary conditions was simulated using macroscopic hydrogen permeation-stress coupling simulation through finite element method (FEM) simulation, in order to obtain stress field data distribution.

[0025] Specifically, a permeation-stress coupling model of X80 pipeline steel under hydrogen exposure in a hydrogen-contaminated environment can be constructed in the finite element simulation software (COMSOL Multiphysics). Preset conditions such as microcracks, boundary hydrogen concentration, and infiltration flux can be set to obtain the hydrogen concentration distribution and stress field evolution of X80 pipeline under different temperatures and stress states.

[0026] Experiments were conducted to obtain mechanical property degradation indices and hydrogen permeation test data for metal pipe materials under different corrosion levels, temperatures, and stress conditions, in order to obtain experimental data.

[0027] Specifically, a constant temperature water bath corrosion simulation device and a hydrogen-filled tensile test can be used to measure the yield strength, elongation and other mechanical properties of samples with different corrosion levels and hydrogen filling times, and the hydrogen hysteresis time and permeation flux can be obtained through a hydrogen permeation test.

[0028] S102: The hydrogen diffusion coefficient, stress field data distribution and experimental data of hydrogen atoms in metal pipe materials are preprocessed, and the preprocessed data of different scales are mapped to a unified input space to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction.

[0029] Specifically, preprocessing includes normalization and noise reduction.

[0030] The hydrogen diffusion coefficient, stress field distribution, and experimental data of hydrogen atoms in metal pipe materials are preprocessed. The preprocessed data at different scales are mapped to a unified input space to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction, including the following contents.

[0031] Figure 2 This is a schematic diagram of a five-dimensional tensor structure provided in an embodiment of this application. Figure 2Each layer in the diagram represents a channel within a five-dimensional tensor feature matrix. The first layer, the first channel, features a topological curve pattern resembling map contour lines, representing the Von Mises equivalent stress field in the stress field data distribution, comprehensively reflecting the stress state within the material. The second layer, the second channel, features a dense grid texture, representing the hydrostatic stress field and reflecting the hydrostatic pressure experienced by the material. The third layer, the third channel, features gradually changing profile fill lines, representing the current hydrogen concentration field, directly related to the distribution of hydrogen atoms. The fourth layer, the fourth channel, is presented as a blank and uniform solid plane, representing the hydrogen diffusion coefficient. The spatially expanded full-map feature map is filled here, ensuring continuity and consistency across the entire material space. The fifth layer, the fifth channel, has many discrete small black dots scattered across its surface. These dots represent experimental data points, containing the material's actual performance under different real-world conditions, such as mechanical property degradation indices and hydrogen permeation test data.

[0032] The VonMises equivalent stress field, hydrostatic stress field, and current hydrogen concentration field from the stress field data distribution are respectively filled into the first three channels of the five-dimensional tensor feature matrix.

[0033] Specifically, stress field data distribution contains rich mechanical information about materials under stress, which is crucial for accurately predicting hydrogen permeation and stress field changes. The Von Mises equivalent stress field comprehensively reflects the internal stress state of the material, the hydrostatic stress field reflects the hydrostatic pressure the material experiences, and the current hydrogen concentration field is directly related to the distribution of hydrogen atoms. When constructing the five-dimensional tensor feature matrix, these three key stress field data are precisely filled into the first three channels of the matrix. Specifically, according to the spatial coordinate correspondence of the matrix, the Von Mises equivalent stress value, hydrostatic stress value, and current hydrogen concentration value are sequentially filled into the corresponding positions of the corresponding channels, ensuring the complete preservation and accurate representation of stress field information.

[0034] The hydrogen diffusion coefficient is expanded into a full-image feature map with the same spatial resolution as the five-dimensional tensor feature matrix and filled into the fourth channel of the five-dimensional tensor feature matrix.

[0035] Specifically, the hydrogen diffusion coefficient is a crucial parameter describing the ability of hydrogen atoms to diffuse within a material, reflecting the ease or difficulty of hydrogen diffusion. Since the hydrogen diffusion coefficient is typically a scalar value measured under specific conditions, and the five-dimensional tensor characteristic matrix possesses spatial resolution, it needs to be expanded into a full-map characteristic map with the same spatial resolution as the five-dimensional tensor characteristic matrix in order to enable the analysis of the hydrogen diffusion coefficient within a unified spatial framework, alongside other information such as stress field data. This involves appropriately expanding the hydrogen diffusion coefficient spatially so that it has a corresponding value at every spatial location, and then filling the fourth channel of the five-dimensional tensor characteristic matrix with the expanded full-map characteristic map. This step ensures the continuity and consistency of the hydrogen diffusion coefficient across the entire material space.

[0036] Experimental data were populated into the fifth channel of the five-dimensional tensor feature matrix to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction.

[0037] Specifically, experimental data is crucial for validating and calibrating the model, encompassing the actual performance of materials under various real-world conditions. Experimental data, including mechanical property degradation indices and hydrogen permeation test data of metal pipe materials under different corrosion levels, temperatures, and stress conditions, obtained through experiments, are organized according to predetermined data formats and specifications and then populated into the fifth channel of the five-dimensional tensor feature matrix. During the population process, it is essential to ensure that the experimental data corresponds to the data in other channels of the matrix in both time and space, enabling the model to comprehensively analyze the influence of various factors on hydrogen permeation and stress fields. By populating the matrix with experimental data, the five-dimensional tensor feature matrix not only includes theoretical simulation data but also incorporates actual experimental data, improving the model's reliability and generalization ability.

[0038] The expression for the eigenvalue matrix of the five-dimensional tensor is: .in, It is a five-dimensional array. For the set of real numbers, For batch size, The length of the time series. This represents the spatial resolution of the metal pipe material in the vertical direction. This represents the spatial resolution of the metal pipe material in the horizontal direction. This represents the number of channels.

[0039] Specifically, the batch size specifies the number of data samples input simultaneously in a single processing step, improving data processing efficiency. The time series length reflects the changes in data over time, enabling the model to capture the dynamic evolution of hydrogen permeation and stress fields over time. This application sets the number of channels to 5.

[0040] S103: Construct a feedforward neural network model for hydrogen permeation-stress prediction. Take the five-dimensional tensor feature matrix as input, embed the hydrogen diffusion equation and stress balance equation in the loss function, train the feedforward neural network model to learn the mapping relationship between each feature element in the five-dimensional tensor feature matrix and key indicators, and output the predicted key indicators.

[0041] A feedforward neural network model for hydrogen permeation-stress prediction is constructed. A five-dimensional tensor feature matrix is ​​used as input, and the hydrogen diffusion equation and stress balance equation are embedded in the loss function. The feedforward neural network model is trained to learn the mapping relationship between each feature element in the five-dimensional tensor feature matrix and key indicators, and outputs the predicted key indicators, including the following:

[0042] Key indicators include hydrogen permeation flux, lag time, hydrogen concentration distribution, and stress field distribution.

[0043] Feedforward neural network models include convolutional layers, recurrent layers, and fully connected layers.

[0044] Two-dimensional convolution kernels are used in the convolutional layer to extract the spatial features of the hydrogen concentration field and stress field, and multi-channel convolution operations are used to capture stress concentration phenomena at crack tips or pits.

[0045] Specifically, the weights of the convolution kernel are updated during training using the backpropagation algorithm to minimize the spatial error between the predicted field and the real field.

[0046] In the recurrent layer, the feature sequence output by the convolutional layer is unfolded by time step to extract the spatial features of the hydrogen concentration field and stress field in the finite element simulation.

[0047] Specifically, the recurrent layer unfolds the feature sequence output by the convolutional layer step by step to capture the sequential features of hydrogen concentration and stress changes over time. The recurrent layer adopts a Long Short-Term Memory (LSTM) network structure, utilizing its memory units and gating mechanism to learn the historical memory effect of hydrogen diffusion. During training, the error gradient flows backward along the time axis, updating the weights of the forget gate, input gate, and output gate.

[0048] In the fully connected layer, the spatial features extracted by the recurrent layer are mapped to the output space to generate key predictive metrics.

[0049] Specifically, the training of the fully connected layer is carried out using the conventional gradient descent method, which is the starting point for error backpropagation.

[0050] The expression for the loss function is: .in, The value of the loss function. This is the residual term of the hydrogen diffusion equation. These are the weighting coefficients for the residual terms of the hydrogen diffusion equation. This is the residual term of the stress equilibrium equation. These are the weighting coefficients for the residual terms of the stress balance equation. For data fitting terms, These are the weighting coefficients for the data fitting terms.

[0051] .in, The hydrogen concentration is the output of the feedforward neural network model. For time, The rate of change of hydrogen concentration over time. The hydrogen diffusion coefficient is... Let be the partial molar volume of hydrogen. The gradient of hydrogen-induced stress, The gas constant is... Absolute temperature For gradient.

[0052] .in, For stress tensor, This is the external body force vector.

[0053] .in, These are the predicted key metric values ​​output by the feedforward neural network model. This represents the true value of the key indicator.

[0054] The weighting coefficients for each loss term in the loss function are determined based on preset rules. These weighting coefficients include those for the residual terms of the hydrogen diffusion equation, the residual terms of the stress balance equation, and the data fitting term.

[0055] The default rules include the following.

[0056] Obtain the gradient norm of each term in the loss function relative to the shared weights of the fully connected layers of the feedforward neural network model. The loss terms in the loss function include the residual terms of the hydrogen diffusion equation, the residual terms of the stress balance equation, and the data fitting term.

[0057] By analyzing the decrease of each loss during training, the relative training rate of each loss in the loss function is obtained, and the target gradient norm of each loss is obtained based on the relative training rate of each loss.

[0058] Based on the difference between the gradient norm of each loss term and the target gradient norm, the weight coefficients of each loss term in the loss function are updated.

[0059] The update formula for the weight coefficients of each loss term in the loss function is: .in, For the updated time At that time, the first The weighting coefficients of the loss term, In time At that time, the first The weighting coefficients of the loss term, For learning rate, Indicates the first Weighting coefficients of the loss Perform differentiation. For the current number The gradient norm of the loss term, Let the target gradient norm be... For the first The value of the loss, For the average relative loss, This is a hyperparameter.

[0060] S104: Construct a reverse neural network model with the target performance parameters as input. When training the reverse neural network model, physical constraints such as the hydrogen diffusion equation and stress balance equation are added. Based on the key indicators predicted by the output of the feedforward neural network model, search the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix to obtain the metal pipe material parameter combination that satisfies the physical constraints and makes the target performance parameters optimal.

[0061] It should be noted that both the feedforward neural network model and the reverse neural network model in this application can be deep fully connected deep neural networks (DNN).

[0062] A reverse neural network model is constructed with the target performance parameters as input. When training the reverse neural network model, physical constraints such as the hydrogen diffusion equation and the stress balance equation are added. Based on the key indicators predicted by the output of the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipe material parameter combination that satisfies the physical constraints and makes the target performance parameters optimal, including the following:

[0063] The target performance parameters include the maximum allowable hydrogen concentration and stress limit of the material.

[0064] Specifically, the target performance parameters are pre-set based on the safety standards for pipeline materials and engineering requirements. The input to the structure of the inverse neural network model is the target performance parameters, and the output is the optimal parameters of the metal pipeline material, such as temperature, stress state, corrosion level, and hydrogen charging time.

[0065] Specifically, to train the inverse neural network model and incorporate physical constraints from the hydrogen diffusion equation and stress balance equation, a large-scale sample dataset was constructed. The dataset contains 20,000 samples, generated in the parameter space using Latin hypercube sampling. For each sample, a forward FEM model (forward finite element model) was used to calculate the corresponding performance metrics to ensure data accuracy and reliability. The dataset was split strictly according to machine learning standards. The training set comprised 70% (14,000 samples) and was used for gradient updates of the inverse neural network parameters. The validation set comprised 20% (4,000 samples) and was used to monitor for overfitting and implement early stopping strategies. The test set comprised 10% (2,000 samples) and was not used for training; it was used solely for the final evaluation of the model's generalization ability and prediction accuracy.

[0066] The physical constraints of the hydrogen diffusion equation are satisfied: .

[0067] Specifically, in the inverse neural network model, by minimizing the residual of the equation, it is ensured that the hydrogen concentration distribution predicted by the model satisfies the physical laws of hydrogen diffusion.

[0068] The physical constraints of the stress balance equations are satisfied: .

[0069] Specifically, during the training of the inverse neural network model, the residual of the equation is forced to approach zero to ensure that the stress field distribution predicted by the model conforms to the principle of mechanical equilibrium.

[0070] Based on the predicted key indicators output by the feedforward neural network model, the inverse neural network model needs to search within the variable space comprised of adjustable feature parameters in the five-dimensional tensor feature matrix. These adjustable feature parameters include temperature, stress state, corrosion level, and hydrogen charging time, which together constitute the variable space. The inverse neural network model continuously adjusts these feature parameters, using optimization algorithms (such as gradient descent, reinforcement learning, or genetic algorithms) to search the variable space to find the optimal combination of metal pipe material parameters that satisfies physical constraints and optimizes the target performance parameters.

[0071] After thorough training and optimization, the inverse neural network model will output a set of metal pipe material parameters that satisfy all physical constraints and achieve optimal performance in the target parameters. These parameters can be directly applied to engineering practice to guide the safe use of materials and structural design, ensuring the reliability and durability of metal pipes in hydrogen-rich environments.

[0072] S105: The feedforward neural network obtains the hydrogen permeation and stress field distribution of the material based on the feature elements in the five-dimensional tensor feature matrix, and obtains the combination of metal pipe material parameters that satisfies physical constraints and optimizes the target performance parameters through the inverse neural network model.

[0073] This application also includes: comparing the key indicators predicted by the feedforward neural network model with experimental data and evaluating the error; if the error exceeds the preset range, then adjusting the feedforward neural network model and the reverse neural network model accordingly.

[0074] Specifically, the evaluation metrics used in this application can be root mean square error and / or coefficient of determination. For feedforward neural network models, adjustments can include network structural parameters, weight coefficients, etc. For example, structural parameters such as the number of neurons and connection methods in convolutional, recurrent, and fully connected layers can be adjusted, as well as weight coefficients can be adjusted through optimization algorithms to improve the model's ability to extract spatial and temporal features, thereby improving prediction accuracy. Similarly, for inverse neural network models, their parameters also need to be adjusted. Inverse neural network models take target performance parameters as input and output the optimal combination of metal pipe material parameters. When the error exceeds a preset range, the optimization algorithm parameters used to search for optimal parameters in the inverse neural network model need to be adjusted, such as the exploration rate in reinforcement learning, and the crossover rate and mutation rate in genetic algorithms.

[0075] This application also provides a multi-scale data fusion hydrogen permeation-stress prediction device 300, such as... Figure 3 As shown, the device includes: a data acquisition module 301, a matrix construction module 302, a feedforward neural network model construction module 303, a reverse neural network model construction module 304, and a parameter combination acquisition module 305.

[0076] The data acquisition module 301 is used to acquire the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material.

[0077] The matrix construction module 302 is used to preprocess the hydrogen diffusion coefficient, stress field data distribution and experimental data of hydrogen atoms in metal pipe materials. The preprocessed data of different scales are mapped to a unified input space to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction.

[0078] The forward neural network model module 303 is used to construct a forward neural network model for hydrogen permeation-stress prediction. It takes the five-dimensional tensor feature matrix as input, embeds the hydrogen diffusion equation and stress balance equation into the loss function, trains the forward neural network model to learn the mapping relationship between each feature element in the five-dimensional tensor feature matrix and key indicators, and outputs the predicted key indicators.

[0079] The reverse neural network model construction module 304 is used to construct a reverse neural network model with the target performance parameters as input. When training the reverse neural network model, physical constraints such as the hydrogen diffusion equation and the stress balance equation are added. Based on the predicted key indicators output by the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipe material parameter combination that satisfies the physical constraints and makes the target performance parameters optimal.

[0080] The parameter combination module 305 is used by the feedforward neural network to obtain the hydrogen permeation and stress field distribution of the material based on the feature elements in the five-dimensional tensor feature matrix, and to obtain the metal pipe material parameter combination that satisfies the physical constraints and makes the target performance parameters optimal through the inverse neural network model.

[0081] In practical applications, this application utilizes a physical device for hydrogen permeation-stress prediction based on multi-scale data fusion, such as... Figure 4 As shown, the physical device includes: a smart sensor housing 401, an embedded computing module 402, a high-performance processor chip 403, a multi-physics integrated probe 404, and a metal pipe wall 405.

[0082] The intelligent sensor housing 401 is fixed to the metal pipe wall 405 by clamping or magnetic attraction to achieve real-time hydrogen embrittlement risk prediction. A multiphysics integrated probe 404 is used to acquire experimental data of the metal pipe material in real time. This experimental data includes key information such as hydrogen permeation flux signals and local strain signals. The acquired experimental data will serve as input to the fifth channel of the five-dimensional tensor feature matrix, providing a basis for calibration of the real physical field for subsequent neural network-based model predictions, ensuring that the predicted key indicators closely match the actual situation. An embedded computing module 402 is located inside the intelligent sensor housing 401, integrating a memory and a high-performance processor chip 403. The memory plays a crucial role, storing a pre-built molecular dynamics (MD) hydrogen diffusion coefficient database and a finite element method (FEM) stress field simulation database, as well as the parameters of the trained neural network model. The high-performance processor chip 403 receives real-time experimental data from the multiphysics integrated probe 404, and combines it with the internally stored hydrogen diffusion coefficient and stress field data distribution. By preprocessing these data from different sources and at different scales, including normalization and denoising, they are mapped to a unified input space, thereby constructing a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction. The high-performance processor chip 403 runs a feedforward neural network model, which outputs the predicted key indicators in real time based on constraints (hydrogen diffusion equation and stress balance equation).

[0083] Furthermore, when abnormal indicators are detected, the high-performance processor chip 403 can activate the inverse neural network model. During the training of the inverse neural network model, physical constraints from the hydrogen diffusion equation and stress balance equation are incorporated. Based on the set target performance parameters (such as the maximum allowable hydrogen concentration) and the predicted key indicators output by the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipeline material parameter combination that satisfies the physical constraints and optimizes the target performance parameters. This results in the output of optimal operating condition adjustment suggestions (such as reducing the transport pressure or adjusting the temperature), thereby achieving intelligent prevention and control of pipeline hydrogen embrittlement risk.

[0084] Some modules in the apparatus described in this application can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0085] The apparatus or module described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. For ease of description, the above apparatus is described by dividing it into various modules according to their functions. When implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0086] The methods, apparatus, or modules described in this application can be implemented in a computer-readable program code manner. The controller can be implemented in any suitable manner, for example, as a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of a memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code manner, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included within it for implementing various functions can also be considered as structures within the hardware component. Alternatively, the device used to implement various functions can be viewed as either a software module that implements the method or a structure within a hardware component.

[0087] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, or it can be embodied in the process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in the embodiments of this application.

[0088] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations.

[0089] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A multi-scale data fusion method for predicting hydrogen permeation-stress, characterized in that, include: To obtain the hydrogen diffusion coefficient, stress field distribution, and experimental data of hydrogen atoms in metal pipe materials; The hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in metal pipe materials are preprocessed, and the preprocessed data at different scales are mapped to a unified input space to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction. A feedforward neural network model for hydrogen permeation-stress prediction is constructed. The five-dimensional tensor feature matrix is ​​used as input, and the hydrogen diffusion equation and stress balance equation are embedded in the loss function. The feedforward neural network model is trained to learn the mapping relationship between each feature element in the five-dimensional tensor feature matrix and key indicators, and output the predicted key indicators. A reverse neural network model with target performance parameters as input is constructed. When training the reverse neural network model, physical constraints of hydrogen diffusion equation and stress balance equation are added. Based on the key indicators predicted by the output of the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipe material parameter combination that satisfies the physical constraints and makes the target performance parameters optimal. The feedforward neural network obtains the hydrogen permeation and stress field distribution of the material based on the feature elements in the five-dimensional tensor feature matrix, and obtains the combination of metal pipe material parameters that satisfies physical constraints and optimizes the target performance parameters through the inverse neural network model.

2. The hydrogen permeation-stress prediction method based on multi-scale data fusion according to claim 1, characterized in that, Also includes: The key metrics predicted by the feedforward neural network model are compared with the experimental data and the error is evaluated. If the error exceeds the preset range, the feedforward neural network model and the reverse neural network model are adjusted accordingly.

3. The hydrogen permeation-stress prediction method based on multi-scale data fusion according to claim 1, characterized in that, The acquisition of hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material includes: Hydrogen diffusion in metallic materials was simulated using molecular dynamics simulations at different temperature conditions to obtain the diffusion behavior of hydrogen atoms at different temperatures, thereby obtaining the hydrogen diffusion coefficient. The evolution of hydrogen concentration over time and space under different temperatures, stress states and boundary conditions was simulated by finite element simulation on a macroscopic scale to obtain stress field data distribution. Experiments were conducted to obtain mechanical property degradation indices and hydrogen permeation test data for metal pipe materials under different corrosion levels, temperatures, and stress conditions, in order to obtain experimental data.

4. The hydrogen permeation-stress prediction method based on multi-scale data fusion according to claim 1, characterized in that, The process involves preprocessing hydrogen diffusion coefficients, stress field distribution data, and experimental data of hydrogen atoms in the metal pipe material. The preprocessed data at different scales are then mapped to a unified input space to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction, including: The VonMises equivalent stress field, hydrostatic stress field and current hydrogen concentration field in the stress field data distribution are respectively filled into the first three channels of the five-dimensional tensor feature matrix. The hydrogen diffusion coefficient is extended to a full-image feature map with the same spatial resolution as the five-dimensional tensor feature matrix and filled into the fourth channel of the five-dimensional tensor feature matrix. Experimental data were populated into the fifth channel of the five-dimensional tensor feature matrix to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction. The expression for the eigenvalue matrix of the five-dimensional tensor is: ;in, It is a five-dimensional array. For the set of real numbers, For batch size, The length of the time series. This represents the spatial resolution of the metal pipe material in the vertical direction. This represents the spatial resolution of the metal pipe material in the horizontal direction. This represents the number of channels.

5. The hydrogen permeation-stress prediction method based on multi-scale data fusion according to claim 1, characterized in that, The constructed feedforward neural network model for hydrogen permeation-stress prediction takes a five-dimensional tensor feature matrix as input, embeds the hydrogen diffusion equation and stress balance equation into the loss function, and trains the feedforward neural network model to learn the mapping relationship between each feature element in the five-dimensional tensor feature matrix and key indicators, and outputs the predicted key indicators, including: Key indicators include hydrogen permeation flux, lag time, hydrogen concentration distribution, and stress field distribution. The feedforward neural network model includes convolutional layers, recurrent layers, and fully connected layers; Two-dimensional convolution kernels are used in the convolutional layer to extract the spatial features of hydrogen concentration field and stress field, and multi-channel convolution operation is used to capture stress concentration at crack tip or pit. In the recurrent layer, the feature sequence output by the convolutional layer is unfolded by time step to extract the spatial features of the hydrogen concentration field and stress field in the finite element simulation. In the fully connected layer, the spatial features extracted by the recurrent layer are mapped to the output space to generate key predictive metrics. The expression for the loss function is: ;in, The value of the loss function. This is the residual term of the hydrogen diffusion equation. These are the weighting coefficients for the residual terms of the hydrogen diffusion equation. This is the residual term of the stress equilibrium equation. These are the weighting coefficients for the residual terms of the stress balance equation. For data fitting terms, These are the weighting coefficients of the data fitting terms; ;in, The hydrogen concentration is the output of the feedforward neural network model. For time, The rate of change of hydrogen concentration over time. The hydrogen diffusion coefficient is... Let be the partial molar volume of hydrogen. The gradient of hydrogen-induced stress, The gas constant is... Absolute temperature For gradient; ;in, For stress tensor, External force vector; ;in, These are the predicted key metric values ​​output by the feedforward neural network model. This represents the true value of the key indicator.

6. The hydrogen permeation-stress prediction method based on multi-scale data fusion according to claim 5, characterized in that, The construction of the inverse neural network model, with the target performance parameters as input, incorporates physical constraints such as the hydrogen diffusion equation and stress balance equation during training. Based on the predicted key indicators output by the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipe material parameter combination that satisfies the physical constraints and optimizes the target performance parameters. This includes: Target performance parameters include the maximum allowable hydrogen concentration and stress limit of the material; The physical constraints of the hydrogen diffusion equation are satisfied: ; The physical constraints of the stress balance equations are satisfied: .

7. The hydrogen permeation-stress prediction method based on multi-scale data fusion according to claim 5, characterized in that, The weighting coefficients of each loss term in the loss function are determined based on preset rules; among them, the weighting coefficients of each loss term in the loss function include the weighting coefficients of the residual terms of the hydrogen diffusion equation, the weighting coefficients of the residual terms of the stress balance equation, and the weighting coefficients of the data fitting terms. The preset rules include: Obtain the gradient norm of each term in the loss function relative to the shared weights of the fully connected layers of the feedforward neural network model; where each term in the loss function includes the residual terms of the hydrogen diffusion equation, the residual terms of the stress balance equation, and the data fitting term; By analyzing the decrease of each loss during the training process, the relative training rate of each loss in the loss function is obtained, and the target gradient norm of each loss is obtained based on the relative training rate of each loss. Based on the difference between the gradient norm of each loss term and the target gradient norm, the weight coefficients of each loss term in the loss function are updated.

8. A multi-scale data fusion hydrogen permeation-stress prediction device, characterized in that, The device performs the method as described in any one of claims 1 to 7, including: The data acquisition module is used to acquire the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in the metal pipe material. A matrix module is constructed to preprocess the hydrogen diffusion coefficient, stress field data distribution, and experimental data of hydrogen atoms in metal pipe materials. The preprocessed data at different scales are mapped to a unified input space to construct a five-dimensional tensor feature matrix for hydrogen permeation-stress prediction. A feedforward neural network model module is constructed to build a feedforward neural network model for hydrogen permeation-stress prediction. The five-dimensional tensor feature matrix is ​​used as input, and the hydrogen diffusion equation and stress balance equation are embedded in the loss function to train the feedforward neural network model to learn the mapping relationship between each feature element in the five-dimensional tensor feature matrix and key indicators, and output the predicted key indicators. A reverse neural network model module is constructed to build a reverse neural network model with target performance parameters as input. When training the reverse neural network model, physical constraints such as hydrogen diffusion equation and stress balance equation are added. Based on the predicted key indicators output by the feedforward neural network model, the variable space composed of adjustable feature parameters in the five-dimensional tensor feature matrix is ​​searched to obtain the metal pipe material parameter combination that satisfies the physical constraints and makes the target performance parameters optimal. The parameter combination module is used to obtain the hydrogen permeation and stress field distribution of the material by the feedforward neural network based on the feature elements in the five-dimensional tensor feature matrix, and to obtain the metal pipe material parameter combination that satisfies the physical constraints and optimizes the target performance parameters through the inverse neural network model.

Images