Creep residual life prediction method based on physical information and recurrent neural network
By dividing the data of high-temperature metallic materials into static and dynamic features, constructing a physical information neural network and combining it with a recurrent neural network, the problems of insufficient prediction accuracy and physical understanding in existing technologies are solved, and high-precision creep remaining life prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- EAST CHINA UNIV OF SCI & TECH
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-10
Smart Images
Figure CN122369720A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of predicting the remaining creep life of high-temperature metallic materials, and in particular to a method for predicting the remaining creep life based on physical information and recurrent neural networks. Background Technology
[0002] High-temperature metallic materials are widely used in aerospace, energy, and chemical equipment. Under prolonged high temperatures and loads, they are prone to creep. The remaining creep life directly affects the operational safety and structural integrity of equipment. Therefore, predicting the remaining creep life of high-temperature metallic materials is an important research topic in the field of structural integrity. Accurate prediction of remaining creep life provides a scientific basis for equipment maintenance, repair, and replacement, effectively preventing safety accidents caused by material creep failure and reducing operation and maintenance costs.
[0003] Currently, traditional methods for predicting creep remaining life mainly rely on empirical formulas and creep constitutive models. These methods are mostly derived from experimental data under specific materials and specific working conditions, and have problems such as poor prediction accuracy and limited versatility. They are difficult to adapt to the creep remaining life prediction needs of different materials and different high-temperature load conditions, and cannot meet the high standards of prediction accuracy and applicability required in engineering practice.
[0004] With the rapid development of machine learning technology, its powerful data mining and fitting capabilities have provided a new technical approach for predicting the remaining creep life of high-temperature metallic materials, gradually becoming a research hotspot in this field. However, most existing machine learning-based remaining creep life prediction models are purely data-driven. These models rely excessively on a large amount of labeled data and lack the integration of physical laws during the creep process of high-temperature metallic materials. This results in prediction results lacking physical understanding and failing to effectively handle different types of features, ultimately leading to relatively limited accuracy in remaining creep life prediction and making it difficult to meet the needs of practical engineering applications.
[0005] In view of the shortcomings of the existing technology, there is an urgent need for a creep remaining life prediction method that can take into account different types of feature processing and improve prediction accuracy. Summary of the Invention
[0006] The purpose of this invention is to provide a creep remaining life prediction method based on physical information and recurrent neural networks, so as to take into account different types of characteristics and achieve accurate prediction of creep remaining life of high-temperature metallic materials.
[0007] To achieve the above objectives, the present invention provides a creep remaining lifetime prediction method based on physical information and recurrent neural networks, comprising: S1, acquire relevant data of the target high-temperature metallic material and classify it into static and dynamic characteristics; S2, calculate the correlation between static characteristics and creep fracture life; S3, the positive or negative correlation coefficient between static features and creep fracture life is used as a physical loss term and incorporated into the loss function of the machine learning model to construct a physical information neural network; S4, normalize both static and dynamic features; S5. The normalized static features are input into the constructed physical information neural network and trained and adjusted. The trained physical information neural network is used to obtain the predicted creep fracture life of high-temperature metal materials. S6. The normalized dynamic features are input into the recurrent neural network and trained and the parameters are adjusted. The trained recurrent neural network is used to obtain the predicted proportion of creep remaining lifetime to creep fracture lifetime. S7, multiply the predicted creep fracture life by the predicted ratio to obtain the predicted creep remaining life of high-temperature metallic materials.
[0008] Furthermore, in step S1: Relevant data include initial material composition, heat treatment conditions (normalizing temperature, normalizing time, tempering temperature, tempering time, etc.), test conditions (test temperature, stress, etc.), creep strain, creep remaining life, and creep fracture life. Static characteristics include material composition, heat treatment conditions, and test conditions, with the corresponding target characteristic being creep rupture life. The dynamic characteristics include creep strain, and the corresponding target characteristic is the ratio of the remaining creep lifetime to the creep fracture lifetime.
[0009] Furthermore, the formula for calculating the correlation between static characteristics and creep fracture life in step S2 is as follows: In the formula, X i It is a static feature; Y is the creep fracture life output by static features; Cov() is the covariance; Var() is the variance.
[0010] Furthermore, the expression for the physical loss term in step S3 is: In the formula, λ This refers to the physical loss weighting coefficient; To predict creep fracture life; X i It is a static feature; s k Static features k The sign of the correlation coefficient between static characteristics and creep fracture life. kWhen it is positively correlated with creep fracture life, s k =+1, when static features k When it is negatively correlated with creep fracture life, s k =-1.
[0011] Furthermore, in step S3, the physical information neural network includes an input layer, a hidden layer, and an output layer. The number of hidden layers, the number of neurons in the hidden layers, the learning rate, and the physical loss weight coefficients are determined by grid search or Bayesian optimization.
[0012] Furthermore, the normalization formula in step S4 is: In the formula, b i These are the normalized eigenvalues; B i For eigenvalues; B max and B min These represent the maximum and minimum values of the static / dynamic features.
[0013] Furthermore, in step S6, the parameters of the recurrent neural network include the input layer, hidden layer, and output layer. The number of hidden layers, the number of neurons in the hidden layers, and the learning rate are determined by grid search or Bayesian optimization.
[0014] The present invention has the following beneficial effects: (1) This invention clearly divides the relevant data into static features and dynamic features, realizes the accurate classification of features, avoids the interference of model training caused by the mixing of different types of features, lays the foundation for subsequent modular processing of features and improves the model's targeting, and clarifies the target variables corresponding to each feature, ensuring the accurate correspondence between the data and the prediction target and reducing the impact of invalid data.
[0015] (2) This invention incorporates the positive and negative correlation coefficient between static features and creep fracture life into the physical information neural network in the form of physical loss terms, which effectively solves the problem of lack of physical interpretability in traditional pure data-driven models, making the model prediction results conform to the basic physical laws of material creep and avoiding prediction results that violate physical common sense. At the same time, the model parameters are determined by the grid search method to ensure optimal parameter configuration, taking into account the model's fitting accuracy and generalization ability, and solving the problem of poor prediction stability caused by traditional model parameters being set based on experience.
[0016] (3) This invention predicts creep fracture life by incorporating the positive and negative correlation coefficient between static features and creep fracture life into a physical information neural network in the form of a physical loss term, and predicts the proportion of creep remaining life to creep fracture life by inputting dynamic features into a recurrent neural network. This not only leverages the advantages of physical information neural networks in static feature processing and physical law integration, but also utilizes the strengths of recurrent neural networks in dynamic time-series feature processing, effectively solving the problem that existing models cannot take into account different types of features. At the same time, step-by-step prediction reduces the prediction difficulty of a single model, improves the overall prediction accuracy, and the prediction process is traceable and interpretable, fully meeting the high standards of prediction accuracy and applicability required by engineering practice. Attached Figure Description
[0017] Figure 1 This is a flowchart of the creep remaining life prediction process of the present invention.
[0018] Figure 2 This is a Pearson correlation coefficient thermogram of static characteristics and creep fracture life in Example 1.
[0019] Figure 3 This is a comparison chart of the predicted remaining creep lifetime and the actual remaining creep lifetime in Example 1. Detailed Implementation
[0020] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. However, these embodiments are not intended to limit the present invention. Any similar structures and similar variations of the present invention should be included in the protection scope of the present invention. The commas in the present invention all indicate the relationship between and. The English letters in the present invention are case-sensitive.
[0021] like Figure 1 As shown, this invention provides a method for predicting creep remaining lifetime based on physical information and recurrent neural networks, comprising: S1, acquire relevant data of the target high-temperature metallic material and classify it into static and dynamic characteristics; Relevant data include initial material composition, heat treatment conditions (normalizing temperature, normalizing time, tempering temperature, tempering time, etc.), test conditions (test temperature, stress, etc.), creep strain, creep remaining life, and creep fracture life. Static characteristics include material composition, heat treatment conditions, and test conditions, with the corresponding target characteristic being creep rupture life. The dynamic characteristics include creep strain, and the corresponding target characteristic is the ratio of the remaining creep lifetime to the creep fracture lifetime.
[0022] S2, calculate the correlation between static characteristics and creep fracture life, using the following formula: In the formula, X i It is a static feature; Y The creep fracture lifetime is the output of static features; Cov() is the covariance; Var() is the variance. By using a quantitative formula to accurately calculate the correlation between static features and creep fracture lifetime, this method overcomes the limitations of traditional methods that rely solely on experience to judge feature correlation. It achieves a quantitative expression of the correlation, providing a scientific and traceable basis for the subsequent construction of physical loss terms. Simultaneously, it can screen out static features with high correlation to the prediction target, indirectly improving model training efficiency and prediction accuracy.
[0023] S3, the sign of the correlation coefficient between static features and creep fracture life is included as a physical loss term in the loss function of the machine learning model. Specifically, the sign of the correlation coefficient between highly correlated static features and creep fracture life is selected as a physical loss term and included in the loss function of the machine learning model to construct a physical information neural network. The expression for the physical loss term is: In the formula, λ This refers to the physical loss weighting coefficient; To predict creep fracture life; X i It is a static feature; s k Static features k The sign of the correlation coefficient between static characteristics and creep fracture life. k When it is positively correlated with creep fracture life, s k =+1, when static features k When it is negatively correlated with creep fracture life, s k =-1.
[0024] The physical information neural network consists of an input layer, hidden layers, and an output layer. The number of hidden layers, the number of neurons in the hidden layers, the learning rate, and the physical loss weight coefficients are determined by grid search or Bayesian optimization.
[0025] S4 normalizes both static and dynamic features. This standardization eliminates dimensional differences between static and dynamic features (such as differences in the numerical ranges of material composition and creep strain), preventing features with large dimensions from dominating model training. This ensures that each feature has equal weight during model training, accelerates model convergence, reduces gradient vanishing or exploding problems during training, and improves model training efficiency and stability. The normalization formula is: In the formula, b i These are the normalized eigenvalues; B i For eigenvalues; B max and B min These represent the maximum and minimum values of the static / dynamic features.
[0026] S5. The normalized static features are input into the constructed physical information neural network and trained and adjusted. The trained physical information neural network is used to obtain the predicted creep fracture life of high-temperature metal materials. S6. The normalized dynamic features are input into the recurrent neural network and trained and the parameters are adjusted. The trained recurrent neural network is used to obtain the predicted proportion of creep remaining lifetime to creep fracture lifetime. The parameters of a recurrent neural network include an input layer, hidden layers, and an output layer. The number of hidden layers, the number of neurons in the hidden layers, and the learning rate are determined by grid search or Bayesian optimization.
[0027] S7, multiply the predicted creep fracture life by the predicted ratio to obtain the predicted creep remaining life of high-temperature metallic materials.
[0028] Example 1 This embodiment uses P91 steel as an example to predict its remaining creep life. The specific method includes: S1. Based on publicly available literature, the initial material composition (C, Si, Mn, N, Al, Cr, Mo, Nb, Ni, V), heat treatment conditions (normalizing temperature (NTem), normalizing time (NTim), tempering temperature (TTem), tempering time (TTim)), test conditions (test temperature (Tem), stress (Stress)), creep strain, creep remaining life, and creep rupture life (CRT) of P91 steel were obtained. The initial material composition (C, Si, Mn, N, Al, Cr, Mo, Nb, Ni, V), heat treatment conditions (normalizing temperature (NTem), normalizing time (NTim), tempering temperature (TTem), tempering time (TTim)), and test conditions (test temperature (Tem), stress (Stress)) were classified as static characteristics, with the corresponding target characteristic being creep rupture life (CRT). Creep strain was classified as a dynamic characteristic, with the corresponding target characteristic being the ratio of creep remaining life to creep rupture life.
[0029] S2, calculate the correlation between static characteristics and creep fracture life, using the following formula: In the formula, Xi For C, Si, Mn, N, Al, Cr, Mo, Nb, Ni, V, normalizing temperature (NTem), normalizing time (NTim), tempering temperature (TTem), tempering time (TTim), test temperature (Tem), and stress (Stress). Y The creep fracture life (CRT) is the static feature output; Cov() is the covariance; Var() is the variance.
[0030] like Figure 2 As shown, according to the correlation analysis, the contents of C, Si, Mo and Ni, normalizing temperature (NTem), tempering temperature (TTem), and stress have a relatively large impact on creep fracture life. Moreover, the contents of C and Mo, tempering temperature (TTem), and stress are negatively correlated with creep fracture life, while the contents of Si and Ni, normalizing temperature (NTem) and creep fracture life are positively correlated. S3, the positive or negative correlation coefficient between static features and creep fracture life is used as a physical loss term and incorporated into the loss function of the machine learning model to construct a physical information neural network; The expression for the physical loss term is: Specifically, when the static characteristics are C and Mo content, tempering temperature (TTem), and stress, s k =-1, when the static characteristics are Si and Ni content, and normalizing temperature (NTem), s k =+1.
[0031] S4. Normalize both static and dynamic features using the following formula: In the formula, b i These are the normalized eigenvalues; B i For eigenvalues; B max and B min These represent the maximum and minimum values of the static / dynamic features.
[0032] S5. The normalized static features are input into the constructed physical information neural network and trained and adjusted. The trained physical information neural network is used to obtain the predicted creep fracture life of high-temperature metal materials. The optimal structure of the physical information neural network is one input layer, four hidden layers, and one output layer. The output layer is used to output the predicted creep fracture life. Since there are seven main static features, the number of neurons in the input layer is seven. The number of neurons in each hidden layer was determined to be 12, 6, 6, and 2 using a grid search method. The optimal learning rate is 0.01, and the physical loss weight coefficient is... λ The optimal value is 0.1.
[0033] S6. The normalized dynamic features are input into the recurrent neural network and trained and the parameters are adjusted. The trained recurrent neural network is used to obtain the predicted proportion of creep remaining lifetime to creep fracture lifetime. The optimal structure of the recurrent neural network was determined by grid search to be one input layer, three hidden layers, and one output layer. The output layer is used to output the predicted proportion of the remaining creep lifetime to the creep fracture lifetime. The number of neurons in the three hidden layers are 16, 8, and 4, respectively, and the optimal learning rate is 0.005.
[0034] S7. Multiply the predicted creep fracture life by the predicted ratio to obtain the predicted remaining creep life of the high-temperature metallic material. Compare the predicted remaining creep life with the actual remaining creep life of P91 steel. Figure 3 As shown in the figure, the data points almost perfectly fit the 1:1 reference line, indicating that the deviation between the predicted creep remaining life and the actual creep remaining life is extremely small; all points fall within the ±1.5 error band, proving that the model's prediction accuracy meets or even exceeds the conventional requirements in engineering.
[0035] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
Claims
1. A creep remaining lifetime prediction method based on physical information and recurrent neural networks, characterized in that, include: S1, acquire relevant data of the target high-temperature metallic material and classify it into static and dynamic characteristics; S2, calculate the correlation between static characteristics and creep fracture life; S3, the positive or negative correlation coefficient between static features and creep fracture life is used as a physical loss term and incorporated into the loss function of the machine learning model to construct a physical information neural network; S4, normalize both static and dynamic features; S5. The normalized static features are input into the constructed physical information neural network and trained and adjusted. The trained physical information neural network is used to obtain the predicted creep fracture life of high-temperature metal materials. S6. The normalized dynamic features are input into the recurrent neural network and trained and the parameters are adjusted. The trained recurrent neural network is used to obtain the predicted proportion of creep remaining lifetime to creep fracture lifetime. S7, multiply the predicted creep fracture life by the predicted ratio to obtain the predicted creep remaining life of high-temperature metallic materials.
2. The creep remaining lifetime prediction method based on physical information and recurrent neural networks according to claim 1, characterized in that, In step S1: Relevant data include initial material composition, heat treatment conditions, test conditions, creep strain, creep remaining life, and creep fracture life; among which, heat treatment conditions include normalizing temperature, normalizing time, tempering temperature, and tempering time; test conditions include test temperature and stress. Static characteristics include material composition, heat treatment conditions, and test conditions, with the corresponding target characteristic being creep rupture life. The dynamic characteristics include creep strain, and the corresponding target characteristic is the ratio of the remaining creep lifetime to the creep fracture lifetime.
3. The creep remaining lifetime prediction method based on physical information and recurrent neural networks according to claim 1, characterized in that, The formula for calculating the correlation between static characteristics and creep fracture life in step S2 is as follows: In the formula, X i It is a static feature; Y is the creep fracture life output by static features; Cov() is the covariance; Var() is the variance.
4. The creep remaining lifetime prediction method based on physical information and recurrent neural networks according to claim 1, characterized in that, The expression for the physical loss term in step S3 is: In the formula, λ This refers to the physical loss weighting coefficient; To predict creep fracture life; X i It is a static feature; s k Static features k The sign of the correlation coefficient between static characteristics and creep fracture life. k When it is positively correlated with creep fracture life, s k =+1, when static features k When it is negatively correlated with creep fracture life, s k =-1.
5. The creep remaining lifetime prediction method based on physical information and recurrent neural networks according to claim 1, characterized in that, In step S3, the physical information neural network includes an input layer, a hidden layer, and an output layer. The number of hidden layers, the number of neurons in the hidden layers, the learning rate, and the physical loss weight coefficients are determined by grid search or Bayesian optimization.
6. The creep remaining lifetime prediction method based on physical information and recurrent neural networks according to claim 1, characterized in that, The normalization formula in step S4 is: In the formula, b i These are the normalized eigenvalues; B i For eigenvalues; B max and B min These represent the maximum and minimum values of the static / dynamic features.
7. The creep remaining lifetime prediction method based on physical information and recurrent neural networks according to claim 1, characterized in that, In step S6, the recurrent neural network includes an input layer, a hidden layer, and an output layer. The number of hidden layers, the number of neurons in the hidden layers, and the learning rate are determined by grid search or Bayesian optimization.