A method and system for predicting the corrosion fatigue life of concrete

By constructing a mapping relationship between nondestructive testing and the degree of corrosion damage, pre-training a neural network model and combining it with mechanical constraints, the problems of low accuracy and high cost in concrete fatigue life prediction in existing technologies are solved, and rapid and accurate corrosion fatigue life prediction is achieved.

CN123120160APending Publication Date: 2026-11-17EAST CHINA JIAOTONG UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611628878.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for predicting the fatigue life of concrete cannot effectively quantify the coupling effect between acid rain corrosion and cyclic loading, resulting in low prediction accuracy and high cost, making it difficult to meet the needs of rapid prediction.

Method used

By constructing a mapping relationship between nondestructive testing results and the degree of corrosion damage, the SN model is pre-trained using a pre-training dataset to establish a neural network model. The degree of corrosion damage is quickly obtained using nondestructive testing methods, and the lifetime distribution is predicted based on key factor parameters. The neural network is trained by combining mechanical monotonicity constraints and boundary condition constraints.

Benefits of technology

It enables rapid and accurate prediction of corrosion fatigue life without damaging concrete, improving prediction efficiency and accuracy, reducing costs, and the model output conforms to physical laws, with good extrapolation reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120160A_ABST
    Figure CN123120160A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for predicting the corrosion fatigue life of concrete, belonging to the field of concrete life prediction technology. The prediction method provided by this invention includes: obtaining the non-destructive testing results, corrosion damage degree, and actual fatigue life of a concrete sample; constructing a mapping relationship between the non-destructive testing results and the corrosion damage degree; pre-training an S-N model using a pre-trained dataset to construct and train a life distribution prediction model with two outputs, scale parameters and shape parameters, respectively; obtaining the non-destructive testing results of the concrete to be tested and obtaining the corrosion damage degree based on the mapping relationship; obtaining the values ​​of other key factor parameters of the concrete to be tested; inputting these values ​​into the prediction model to obtain the scale parameters and shape parameters of the concrete to be tested; and calculating the predicted fatigue life value under a given failure probability based on the inverse function of the cumulative distribution function. This invention can effectively improve the accuracy and convenience of predicting the fatigue life of concrete.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of concrete life prediction technology, and in particular to a method and system for predicting the corrosion fatigue life of concrete. Background Technology

[0002] Concrete, as one of the most widely used building materials globally, boasts significant advantages such as high strength, low cost, and ease of shaping. Its long-term service performance and durability are directly related to the safety and economy of engineering structures. When concrete is exposed to acid rain, corrosive ions in the acid rain penetrate the matrix and react with cement hydration products, leading to a gradual deterioration of concrete performance. Furthermore, during service, concrete is often subjected to repeated cyclic loads such as traffic loads and wind loads. The coupling effect of acid rain corrosion and cyclic loads significantly accelerates the initiation and propagation of microcracks within the concrete, drastically reducing fatigue life without obvious warning signs before failure, exhibiting sudden and catastrophic consequences.

[0003] Currently, there are three main categories of commonly used methods for predicting the fatigue life of concrete. The first is empirical regression methods based on the SN curve, which establish the relationship between stress level and fatigue life through indoor fatigue tests, such as the Manson-Cobb criterion and the Aas-Jakobsen formula. The second is methods based on fracture mechanics and damage mechanics, which introduce damage variables to describe microcracks within the concrete and establish damage evolution equations, such as models based on continuous damage mechanics or crack propagation models based on Paris's law. The third is methods based on finite element numerical simulation, which simulate the damage accumulation process under cyclic loading by establishing microscopic or macroscopic mechanical models of concrete. Furthermore, with the development of artificial intelligence technology, machine learning-based prediction methods have improved the accuracy of fatigue life prediction to some extent.

[0004] However, empirical formulas based on the SN curve only consider the single factor of stress level and often treat fatigue life as a deterministic variable, failing to quantify the impact of the coupling effect of environmental factors and cyclic loading on fatigue life. Furthermore, the test conditions differ significantly from the actual service environment, resulting in poor extrapolation and prediction capabilities of the formulas. While methods based on fracture mechanics and damage mechanics can introduce corrosion factors, they are difficult to accurately reflect the performance changes of concrete under the interaction of physicochemical degradation and mechanical damage during acid rain erosion. In addition, the computational cost of finite element numerical simulation is relatively high, and the coupled modeling of corrosion field and stress field is complex, making it difficult to meet the needs of rapid prediction. Therefore, there is an urgent need to provide a solution to improve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the corrosion fatigue life of concrete, which can effectively improve the accuracy and convenience of predicting the fatigue life of concrete.

[0006] In a first aspect, the present invention provides a method for predicting the corrosion fatigue life of concrete, comprising: The study aims to obtain non-destructive testing results, corrosion damage degree, and actual fatigue life of concrete samples under different loads at different corrosion ages, establish a mapping relationship between non-destructive testing results and corrosion damage degree, and identify key factors that affect actual fatigue life, including at least corrosion damage degree and load stress. The SN model was pre-trained using a pre-training dataset to construct a neural network with two outputs, namely scale parameters and shape parameters. The weights of the pre-trained model were used as at least part of the initial weights of the scale parameter output branch. The sample size of the pre-training data was larger than the data size of the actual fatigue life of concrete specimens. The parameter values ​​corresponding to key factors of concrete samples are used as input to a neural network to train the neural network and obtain a life distribution prediction model. During training, the goal is to maximize the joint likelihood of the Weibull distribution of the actual fatigue life of concrete samples under the scale parameter and shape parameter output by the neural network under the same corrosion age and the same load. The non-destructive testing results of the concrete to be tested are obtained and the degree of corrosion damage is obtained based on the mapping relationship. The other key factor parameter values ​​of the concrete to be tested are obtained. The scale parameters and shape parameters of the concrete to be tested are output based on the prediction model. The fatigue life prediction value under a given failure probability is calculated based on the inverse function of the cumulative distribution function.

[0007] Optionally, the non-destructive testing method used to obtain the non-destructive testing results of the concrete sample is the same as that used to obtain the non-destructive testing results of the concrete to be tested, and the non-destructive testing method includes at least one of ultrasonic testing, impact rebound testing, resistance / conductivity testing, and radiation imaging testing.

[0008] Optionally, the degree of corrosion damage is obtained based on the changes in parameters of concrete samples before and after loading different loads at different corrosion ages. The parameters include at least one of compressive strength, mass, corrosion depth, elastic modulus, crack density, and saturated surface water absorption rate.

[0009] Optionally, the non-destructive testing results are used as input and the degree of corrosion damage is used as output to perform regression fitting to obtain a mapping relationship. The regression fitting includes at least one of linear regression, multinomial regression, neural network regression, and support vector regression.

[0010] Optionally, when determining the key factors that affect the actual fatigue life, the process includes: obtaining the characteristic factors that affect the actual fatigue life, evaluating the importance of the characteristic factors to obtain evaluation values, and taking the characteristic factors with evaluation values ​​not less than a threshold as key factors; wherein, the threshold is the minimum value of the evaluation value corresponding to the degree of corrosion damage and the load stress, and the characteristic factors include the degree of corrosion damage, the load stress, the pH value of the corrosive medium, and the corrosion age.

[0011] Optionally, when using the random forest algorithm to evaluate the importance of feature factors, the process includes: constructing multiple decision trees to form a random forest; obtaining a first score based on the change in impurity of the feature factors when splitting at decision tree nodes; permuting the values ​​of the feature factors in the samples outside the decision tree bag and obtaining a second score based on the increase in error before and after the permutation; and normalizing and weighting the first and second scores to obtain the evaluation value of the feature factors.

[0012] Optionally, when the evaluation value of corrosion age is not less than a threshold, the Spearman correlation coefficient between corrosion age and corrosion damage degree is calculated. If the correlation coefficient is not less than a preset threshold, the corrosion age is discarded; otherwise, the corrosion age is used as a key factor. The calculation of the Spearman correlation coefficient includes:

[0013] in: The Spearman correlation coefficient is the relationship between corrosion age and corrosion damage degree. For the first Corrosion age of each concrete sample The order of things; For the first The degree of corrosion damage of each concrete sample The order of things; This represents the total number of concrete samples.

[0014] Optionally, the neural network can be trained based on mechanical monotonicity constraints and boundary condition constraints; The mechanical monotonicity constraint includes at least one of the following: when the input load stress increases and other characteristics remain unchanged, the output scale parameter decreases monotonically; when the input corrosion damage degree increases and other characteristics remain unchanged, the output scale parameter decreases monotonically. Boundary condition constraints include at least one of the following: positive constraints on the output scale parameters and / or shape parameters; when the input load stress converges to the static strength of the concrete specimen, the output scale parameters converge to 1; when the input corrosion damage degree converges to 0, the output scale parameters converge to the reference life.

[0015] Optionally, training the neural network includes: Obtain representative values ​​of key factors in the concrete sample group. The sample group is a combination of concrete samples with the same corrosion age and the same load. Input the representative values ​​of the parameters into the neural network, and output the scale parameters and shape parameters of the sample group; A loss function is constructed with the objective of maximizing the joint likelihood of the actual fatigue life of concrete specimens in the sample group under dimensional and shape parameters. The lifetime distribution prediction model is obtained by backpropagating the neural network based on minimizing the loss function and updating the weights until convergence.

[0016] Optionally, when constructing the loss function with the objective of maximizing the joint likelihood, it includes:

[0017] in: For the first Shape parameters of the concrete sample group; For the first Dimensional parameters of each concrete sample group; For the first The first concrete sample group The actual fatigue life of a concrete sample; For the first The total number of concrete samples in a concrete sample group; The total number of concrete sample groups; This represents the total number of concrete samples.

[0018] Optionally, when calculating the fatigue life prediction value for a given failure probability based on the inverse function of the cumulative distribution function, the following steps are included:

[0019] in: Given the failure probability; These are the dimensional parameters of the concrete to be tested; The shape parameters of the concrete to be tested; Given a failure probability The predicted fatigue life value at that time.

[0020] Secondly, the present invention also provides a concrete corrosion fatigue life prediction system, comprising: The data processing module is used to obtain the non-destructive testing results, corrosion damage degree and actual fatigue life of concrete samples under different loads at different corrosion ages, construct the mapping relationship between non-destructive testing results and corrosion damage degree, and determine the key factors that affect the actual fatigue life; the key factors include at least the corrosion damage degree and the load stress. The pre-training module is used to pre-train the SN model using a pre-training dataset, construct a neural network with two outputs, namely scale parameters and shape parameters, and use the weights of the pre-trained model as at least part of the initial weights of the scale parameter output branch. The model training module is used to train the neural network by taking the parameter values ​​of the concrete sample corresponding to the key factors as input to the neural network to obtain the life distribution prediction model. During training, the goal is to maximize the joint likelihood of the Weibull distribution of the actual fatigue life of the concrete sample under the scale parameter and shape parameter of the neural network output under the same corrosion age and the same load. The life prediction module is used to obtain the non-destructive testing results of the concrete to be tested and obtain the degree of corrosion damage based on the mapping relationship. It also obtains the values ​​of other key factor parameters of the concrete to be tested, outputs the scale parameters and shape parameters of the concrete to be tested based on the prediction model, and calculates the fatigue life prediction value under a given failure probability based on the inverse function of the cumulative distribution function.

[0021] The method for predicting the corrosion fatigue life of concrete provided by this invention has at least one of the following beneficial technical effects compared with the prior art: 1. By establishing a mapping relationship between non-destructive testing results and the degree of corrosion damage in advance, only the same non-destructive testing method needs to be used to test the concrete under test in practical applications. The mapping relationship can be used to quickly obtain the degree of corrosion damage of the concrete under test without damaging it. This helps to conduct batch screening and real-time evaluation of the concrete under test during on-site implementation, which can effectively improve the prediction efficiency and reduce the prediction cost when predicting the corrosion life of concrete. 2. By using a pre-training dataset with a sample size larger than that of actual concrete specimens to pre-train the SN model, a pre-trained model with prior physical knowledge of stress and fatigue life can be obtained. The weights of the pre-trained model are used as the weights of the scale parameter output branch in the neural network. Thus, the actual fatigue life and key factor parameters of the concrete specimens are used to train the neural network, enabling the model to converge quickly with small sample data and avoid overfitting, which helps to improve the stability and generalization ability of the prediction model. 3. By training the neural network through mechanical monotonicity constraints and boundary condition constraints, the neural network can be trained and optimized under the constraints of physical laws. In this way, when the number of actual fatigue life samples of concrete specimens is small, the model overfitting can be effectively suppressed and the accuracy of the prediction results can be improved. This makes the output of the prediction model more consistent with the actual fatigue life distribution law reflected by the sample data. At the same time, the use of physical constraints can make the model output within a physically reasonable parameter space, avoiding violations of mechanical common sense, which helps to improve the physical consistency and extrapolation reliability of the prediction model. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a method for predicting the corrosion fatigue life of concrete according to the present invention. Figure 2 This is a schematic diagram of the structure of a concrete corrosion fatigue life prediction system according to the present invention.

[0023] Explanation of reference numerals in the attached figures: 101. Data processing module; 102. Pre-training module; 103. Model training module; 104. Lifespan prediction module. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art to which this invention pertains.

[0025] See Figure 1 This invention provides a method for predicting the corrosion fatigue life of concrete, comprising the following steps: S1. Obtain the non-destructive testing results, corrosion damage degree and actual fatigue life of concrete samples under different loads at different corrosion ages, construct the mapping relationship between non-destructive testing results and corrosion damage degree, and determine the key factors that affect the actual fatigue life. S2. The SN model is pre-trained using a pre-training dataset to construct a neural network with two outputs, namely scale parameters and shape parameters. The weights of the pre-trained model are used as at least part of the initial weights of the scale parameter output branch. S3. Use the parameter values ​​of the concrete sample corresponding to the key factors as the input of the neural network, and train the neural network to obtain the life distribution prediction model. S4. Obtain the non-destructive testing results of the concrete to be tested and obtain the degree of corrosion damage based on the mapping relationship. Obtain the other key factor parameter values ​​of the concrete to be tested. Output the scale parameters and shape parameters of the concrete to be tested based on the prediction model. Calculate the fatigue life prediction value under a given failure probability based on the inverse function of the cumulative distribution function.

[0026] In fact, in step S1, after loading concrete samples with different loads at different corrosion ages, non-destructive testing is performed to obtain non-destructive testing results, and corrosion damage detection is performed to obtain the degree of corrosion damage and actual fatigue life. Then, the non-destructive testing results and the degree of corrosion damage are mapped and correlated. In practical applications, after testing the concrete to be tested using the same non-destructive testing method, the degree of corrosion damage of the concrete to be tested can be quickly obtained by using the mapping relationship, thus avoiding the use of destructive means to cause structural damage to the concrete to be tested.

[0027] In some embodiments, the non-destructive testing method used in step S1 includes at least one of ultrasonic testing, impact rebound testing, resistance / conductivity testing, and radiation imaging testing. In fact, by using non-destructive testing methods to test concrete samples / concrete to be tested, non-destructive testing data can be obtained without damaging the concrete. This data can indirectly reflect the internal changes of concrete under corrosive and loaded environments, such as internal cracks, surface hardness, corrosive media, and damage conditions.

[0028] Furthermore, employing multiple non-destructive testing (NDT) methods for comprehensive testing of concrete can improve the mapping accuracy and robustness of NDT results to the degree of corrosion damage, thereby enhancing the predictive accuracy of the concrete under test during application. Simultaneously, it can reduce the uncertainty arising from using a single testing method, providing a more comprehensive and accurate reflection of the internal damage state of concrete under the coupled effects of corrosion and load. Specifically, when using multiple NDT methods for comprehensive testing, regression fitting can be performed on the various NDT results and the degree of corrosion damage to obtain a mapping relationship.

[0029] In fact, the corrosion age in step S1 refers to the length of time that the concrete sample is continuously exposed to the corrosive environment. The corrosion age can be used to characterize the cumulative effect of corrosion on the concrete sample. During the corrosion process, the corrosive medium will gradually penetrate into the interior of the concrete sample and produce physicochemical reactions with the concrete, causing structural changes such as increased porosity, microcrack initiation and propagation, surface spalling and decreased strength inside the concrete structure. Moreover, the structural changes caused by corrosion will continue to intensify with the increase of the corrosion age.

[0030] Furthermore, the corrosive environment can be at least one of the following: chloride ion solution environment, sulfate solution environment, acid solution environment, wet-dry cycle environment, salt-freeze cycle environment, electrolytically accelerated corrosion environment, simulated seawater environment, simulated acid rain environment, and marine tidal environment. In fact, during the simulated corrosion of concrete samples, multiple environments can be used for combined corrosion to more realistically simulate the combined effects of various corrosion mechanisms such as corrosive medium penetration and chemical corrosion in actual engineering scenarios.

[0031] In fact, the load in step S1 refers to the external force applied to the concrete sample, which can effectively simulate the repeated stress state that concrete experiences during engineering service. Indeed, when obtaining the degree of corrosion damage of concrete samples at different corrosion ages and under different loads, the load and corrosion effects act simultaneously on the concrete sample. Through stress-corrosion coupling, the degree of corrosion damage can be directly affected. For example, under high loads, the initiation and growth of microcracks caused by corrosion will be further promoted, causing the corrosive medium to penetrate further into the interior of the concrete sample.

[0032] In some embodiments, when applying loads to concrete samples at different corrosion ages in step S1, cyclic loading can be used. Specifically, the load can be characterized by maximum stress, minimum stress, the stress ratio of maximum stress to minimum stress, and stress level, where the stress level is the ratio of maximum stress to the static strength of the concrete sample. This allows concrete samples at the same corrosion age to be subjected to cyclic loading, and fatigue coupling experiments are conducted under the effects of load and corrosion.

[0033] In fact, the degree of corrosion damage in step S1 can characterize the extent of structural changes and performance degradation of the concrete specimen under corrosive environment and load, and is used to characterize the damage magnitude of the concrete specimen under the coupled effects of corrosion and load. Furthermore, in step S1, the degree of corrosion damage can be obtained based on the parameter changes of the concrete specimen before and after loading different loads at different corrosion ages. The parameters used include at least one of compressive strength, mass, corrosion depth, elastic modulus, crack density, and saturated surface water absorption rate.

[0034] In some embodiments, when the change of any parameter before and after the coupling effect of corrosion and load is used to characterize the degree of corrosion damage, it can be expressed as:

[0035] in, This represents the degree of corrosion damage to the concrete sample after the coupled action of corrosion age T and load S. These are the parameter values ​​of the concrete sample after the combined action of corrosion age T and load S. This represents the average parameter value of the concrete sample under standard curing conditions. In fact, when calculating the parameter changes of the concrete sample before and after coupled corrosion in step S1, the concrete samples used are all standard-sized samples made of concrete materials of the same grade and proportion using the same process.

[0036] In some embodiments, when the changes of multiple parameters before and after the coupling effect of corrosion and load are used to characterize the overall degree of corrosion damage, it can be expressed as:

[0037]

[0038] Where D represents the comprehensive corrosion damage degree of the concrete sample after the coupled action of corrosion age T and load S. The concrete specimen under the coupled action of corrosion age T and load S on the [number]th [period]. The degree of corrosion damage corresponds to each parameter. For the first Each parameter corresponds to a weight representing the degree of corrosion damage. This represents the total number of parameters used to characterize the overall degree of corrosion damage.

[0039] Furthermore, when using the changes of multiple parameters before and after the coupling effect of corrosion and load to characterize the comprehensive corrosion damage degree, the weight of each parameter corresponding to the corrosion damage degree is independent of each other. The weight can be determined using at least one of the commonly used methods in this field: equal weighting method, expert scoring method, entropy weighting method, principal component analysis method, random forest importance evaluation method, and analytic hierarchy process. In fact, the weights influence the evaluation tendency represented by the comprehensive corrosion damage degree. For example, when the weight value of tensile strength corresponding to the corrosion damage degree is high, the comprehensive corrosion damage degree better reflects the loss of structural strength in the concrete sample.

[0040] In some embodiments, when obtaining the actual fatigue life of the concrete sample under corrosion coupling in step S1, commonly used methods for determining concrete fatigue life in the art can be used. For example, the method described in Chapter 13 of GB / T 50082-2024 "Standard for Test Methods of Long-Term Performance and Durability of Concrete" for compressive fatigue deformation test can be used to determine the actual fatigue life of the concrete sample. In practice, when determining the actual fatigue life of the concrete sample under corrosion coupling, the concrete sample after being subjected to stress coupling in a corrosive environment can be transferred to a standard environment for actual fatigue life determination.

[0041] In fact, after the concrete sample is subjected to corrosion coupling under a load in the corrosive environment in step S1, the concrete sample has already suffered corrosion damage. The concrete sample is then transferred to the standard environment and cyclic load is applied until the concrete sample fails. The number of cycles of the concrete sample in the standard environment is recorded as the remaining fatigue life of the concrete sample after corrosion damage, and the remaining fatigue life is taken as the actual fatigue life of the concrete sample after corrosion damage.

[0042] In some embodiments, when constructing the mapping relationship between non-destructive testing results and corrosion damage degree in step S1, the non-destructive testing results can be used as input, and the corrosion damage degree as output. Regression fitting is then performed to obtain the mapping relationship. In practice, during the construction of the mapping relationship, the non-destructive testing results of the concrete sample under the coupled effects of corrosion and load can be used as input features, and the corresponding corrosion damage degree can be used as the output label. The input features and output labels are combined to construct a sample dataset for regression fitting. Specifically, the regression fitting method used includes at least one of linear regression, multinomial regression, neural network regression, and support vector regression.

[0043] Furthermore, in step S1, multiple non-destructive testing (NDT) results of the concrete sample can be collected, and these results can be weighted and combined to obtain a comprehensive NDT result, thus establishing a mapping relationship between the comprehensive NDT result and the degree of corrosion damage. In practice, the multiple NDT results need to be normalized beforehand, and when weighting the results, the weight of each result can be determined using at least one of the following methods: equal weighting, entropy weighting, principal component analysis, random forest importance evaluation, or analytic hierarchy process (AHP), with the sum of all weights being 1.

[0044] Furthermore, when multiple nondestructive testing results are weighted and combined to obtain a comprehensive nondestructive testing result, it can be expressed as:

[0045]

[0046] in, The concrete specimen under the coupled action of corrosion age T and load S on the [number]th [period]. One non-destructive testing result, For the first The weight of each nondestructive testing result, This represents the comprehensive non-destructive testing results of concrete samples after the coupled action of corrosion age T and load S. This represents the total number of nondestructive testing results used to characterize the overall nondestructive testing results.

[0047] Specifically, in step S1, key factors that influence actual fatigue life are identified. This helps to eliminate redundant and irrelevant input features during neural network training, reducing the input dimensionality and multicollinearity between features. It also reduces the risk of overfitting when the actual fatigue life sample size is small, improving the training efficiency and convergence speed of the neural network, and enhancing the accuracy and stability of the prediction model for corrosion fatigue life. Furthermore, the key factors include at least the degree of corrosion damage and the load stress.

[0048] In some embodiments, determining key factors in step S1 includes: acquiring characteristic factors that influence actual fatigue life, evaluating the importance of the characteristic factors to obtain evaluation values, and selecting characteristic factors with evaluation values ​​not less than a threshold as key factors. In fact, by evaluating the importance of characteristic factors, the degree of influence of characteristic factors on actual fatigue life can be objectively quantified, thereby selecting characteristic factors that have a significant impact on actual fatigue life as key factors, thus avoiding the limitations of experience-based or subjective judgments.

[0049] Specifically, when acquiring characteristic factors that influence actual fatigue life, these factors include at least the degree of corrosion damage, load stress, pH value of the corrosive medium, and corrosion age. Furthermore, when setting the threshold for the evaluation values, the minimum value corresponding to the degree of corrosion damage and load stress is used as the threshold, thus ensuring that the degree of corrosion damage and load stress are included as key factors. In addition, using thresholds for screening can eliminate characteristic factors with poor correlation to actual fatigue life, thereby reducing the input dimensions of the neural network.

[0050] Furthermore, when using the random forest algorithm to evaluate the importance of feature factors, the process includes: constructing multiple decision trees to form a random forest; obtaining a first score based on the change in impurity of the feature factors when splitting at decision tree nodes; permuting the values ​​of feature factors in the samples outside the decision tree bag and obtaining a second score based on the increase in error before and after the permutation; and normalizing and weighting the first and second scores to obtain the evaluation value of the feature factors.

[0051] In fact, the first score is based on the change in impurity generated when a node splits, which reflects the contribution of the feature factor in reducing the uncertainty of the target value of the node sample. In regression random forest, this can be specifically manifested as the reduction in mean squared error. The second score is based on the random permutation of feature values ​​in the out-of-bag samples, which can reflect the degree of error increase after the feature factor is perturbed, thus measuring the true impact of the feature factor on the generalization ability.

[0052] Furthermore, normalizing and weighting the first and second scores can combine their advantages, avoiding biases in the importance of feature factors caused by a single indicator. This helps improve the robustness and reliability of feature factor evaluation values, which can be expressed as follows:

[0053]

[0054] in, For the first Evaluation values ​​of each characteristic factor For the first The first score after normalization of each feature factor For the first The second score after normalization of each feature factor As the weight of the first score, The weighting for the second score.

[0055] In some embodiments, after evaluating the importance of the feature factors, when the evaluation value of the corrosion age is not less than the threshold, the Spearman correlation coefficient between the corrosion age and the degree of corrosion damage is calculated. When the absolute value of the correlation coefficient is not less than the preset threshold, it can be considered that the degree of corrosion damage can fully characterize the impact of the corrosion age on the actual fatigue life. Therefore, the corrosion age can be removed from the key factors. Otherwise, the corrosion age is retained in the key factors.

[0056] In fact, the preset threshold is a preset correlation threshold between corrosion age and corrosion damage degree. When the absolute value of the correlation coefficient is not less than the preset threshold, the corrosion age and corrosion damage degree are highly correlated. This indicates that the corrosion damage degree can characterize the impact of corrosion age on actual fatigue life to a certain extent. At this time, removing the corrosion age helps to reduce the input dimension of the neural network and avoid feature collinearity affecting the generalization ability of the prediction model. However, when the absolute value of the correlation coefficient is less than the preset threshold, it indicates that there are still independent effects in the corrosion age that cannot be characterized by the corrosion damage degree. Therefore, retaining the corrosion age as a key factor helps to improve the prediction accuracy of the prediction model.

[0057] Furthermore, when calculating the Spearman correlation coefficient between corrosion age and corrosion damage degree, it can be expressed as:

[0058] in, The Spearman correlation coefficient is the relationship between corrosion age and the degree of corrosion damage. For the first Corrosion age of each concrete sample Rank among all concrete samples For the first The degree of corrosion damage of each concrete sample Rank among all concrete samples This represents the total number of concrete samples.

[0059] Specifically, when calculating the Spearman correlation coefficient, concrete samples with different corrosion ages and different loads can be calculated uniformly. For example, a total of [number] samples can be collected. Corrosion age, The number of concrete samples for each group of corrosion coupling tests is [number] under each load. ,but ,Will The corrosion ages of the concrete samples are sorted in descending order to obtain the rank of the concrete samples. Furthermore, when concrete samples have the same corrosion age, the corresponding ranks can be averaged, and the order of corrosion age and corrosion damage degree is the same when obtaining the rank.

[0060] In fact, the pre-training dataset used in step S2 is an existing large-sample stress-life dataset, which contains multiple load stress levels and their corresponding fatigue life data under standard curing conditions. Furthermore, the sample size in the pre-training dataset is larger than the actual fatigue life data of the concrete specimens used to train the neural network. Further, the SN model used in step S2 is an empirical model characterizing the mapping relationship between load stress and fatigue life; its input is load stress, and its output is fatigue life.

[0061] In some embodiments, step S2 may employ methods such as least squares, maximum likelihood estimation, and gradient descent to pre-train the SN model using a pre-training dataset, enabling the SN model to learn the mapping relationship between load stress and fatigue life under standard maintenance conditions. Furthermore, the SN model used may be one of the following: a power function SN model, a log-linear SN model, an exponential SN model, or a polynomial SN model.

[0062] For example, when the SN model adopts the log-linear SN model, it can be expressed as:

[0063] in, For fatigue life, , All are fitted parameters. For load stress, and Negative values ​​are used to indicate load stress. The larger the value, the longer the fatigue life. The smaller the value, the better. Furthermore, after determining the fitting parameters of the SN model through pre-training, the parameters of the SN model can be used as at least part of the initial weights of the scale parameter output branch in the dual-output neural network, so that the neural network can converge quickly under small sample conditions and maintain physical consistency.

[0064] In fact, by constructing a neural network with two outputs, namely scale parameters and shape parameters, and the neural network includes a shared feature extraction sub-network, a scale parameter output branch, and a shape parameter output branch, the shared feature extraction sub-network can perform nonlinear feature mapping on the input parameters and obtain a shared feature vector, while the scale parameter output branch and the shape parameter output branch can use the shared feature vector as input to output the corresponding scale parameters and shape parameters, respectively.

[0065] Furthermore, both the scale parameter output branch and the shape parameter output branch contain at least one fully connected layer. The output layer can use softplus activation functions, exponential activation functions, etc., and ensures that the output scale and shape parameters are both positive. In addition, the weights of the pre-trained SN model are used as part of the initial weights of the scale parameter output branch, so that the scale parameter output branch is located near the parameter space that conforms to the mapping law between load stress and fatigue life in the early stage of training.

[0066] In fact, the scale parameter and shape parameter in neural networks are two statistical parameters in the Weibull distribution, respectively. These parameters can be used to characterize the probability distribution characteristics of concrete fatigue life. Specifically, the scale parameter characterizes the overall level of fatigue life; its magnitude reflects the central location and order of magnitude of the life distribution. For example, a larger scale parameter indicates a longer fatigue life for concrete with the same failure probability. Furthermore, the shape parameter characterizes the degree of dispersion of fatigue life. A larger shape parameter indicates a more concentrated and less dispersed fatigue life, while a smaller shape parameter indicates a greater dispersion.

[0067] In fact, when training the neural network in step S3, concrete samples with the same corrosion age and the same load can be divided into the same group. The representative values ​​of the parameters of the concrete samples in the same group on the key factors are used as the input of the neural network. The neural network outputs the corresponding scale parameters and shape parameters. The goal is to maximize the joint likelihood of the Weibull distribution of the actual fatigue life of each concrete sample in the group under the scale parameters and shape parameters. The weights of the neural network are updated until convergence, thereby obtaining the life distribution prediction model.

[0068] Furthermore, by using the joint likelihood maximization of the Weibull distribution as the objective during neural network training, the central location and dispersion of the life distribution can be constrained simultaneously. This allows the neural network to learn the central tendency of fatigue life and characterize the randomness of the life distribution, which is more consistent with the distribution patterns of real concrete samples. In addition, by directly outputting the scale and shape parameters of the Weibull distribution through the neural network, error propagation and fitting biases generated during staged prediction and distribution fitting can be avoided, which helps to improve the accuracy and precision of the prediction model.

[0069] In some embodiments, training the neural network in step S3 includes: Obtain representative values ​​of key factors in the concrete sample group. The sample group is a combination of concrete samples with the same corrosion age and the same load. Input the representative values ​​of the parameters into the neural network, and output the scale parameters and shape parameters of the sample group; A loss function is constructed with the objective of maximizing the joint likelihood of the actual fatigue life of concrete specimens in the sample group under dimensional and shape parameters. The lifetime distribution prediction model is obtained by backpropagating the neural network based on minimizing the loss function and updating the weights until convergence.

[0070] In practice, by obtaining representative values ​​of key factors in a concrete sample group, it is possible to ensure that multiple actual fatigue life samples of the same concrete sample group correspond to the same set of scale and shape parameters. This also reduces individual differences among concrete samples within the group, thereby improving model stability when training neural networks under small sample conditions. Specifically, the representative values ​​can be the average values ​​of the key factors for each concrete sample within the group.

[0071] In some embodiments, when training a neural network and constructing a loss function with the objective of maximizing joint likelihood, it can be expressed as:

[0072] in, For the first Shape parameters of the concrete sample group For the first Dimensional parameters of each concrete sample group For the first The first concrete sample group The actual fatigue life of the concrete sample. For the first The total number of concrete samples in each concrete sample group. The total number of concrete sample groups. This represents the total number of concrete samples.

[0073] In fact, the loss function constructed when training the neural network is the negative logarithm of the joint likelihood of the Weibull distribution. By minimizing the loss function, the joint likelihood can be maximized, so that the scale parameter and shape parameter output by the neural network fit the true distribution characteristics of the actual fatigue life in the sample group to the greatest extent. This makes full use of the sample data of the actual fatigue life in the sample group, which is beneficial for parameter estimation of the life distribution under small sample conditions and improves the stability and accuracy of the estimation.

[0074] In some embodiments, the neural network can be trained in step S3 based on mechanical monotonicity constraints and boundary condition constraints. This allows fatigue mechanics laws and physical boundaries to be introduced into the loss function through a penalty term, enabling the neural network to be trained and its weights updated within a parameter space that conforms to natural physical laws, thus avoiding data overfitting or outputting results that do not conform to natural laws. In practice, since the number of actual fatigue life samples of concrete specimens is relatively small, the use of mechanical monotonicity constraints and boundary condition constraints allows the neural network to quickly learn the distribution law of the actual fatigue life.

[0075] Furthermore, the mechanical monotonicity constraint includes at least one of the following: when the input load stress increases while other characteristics remain unchanged, the output scale parameter monotonically decreases; when the input corrosion damage degree increases while other input characteristics remain unchanged, the output scale parameter monotonically decreases. In reality, the greater the load stress applied to the concrete specimen, the faster the microcrack propagation rate inside the concrete specimen, leading to a shorter fatigue life. Furthermore, the greater the degree of corrosion damage to the concrete specimen, the more severe the structural damage, and therefore the shorter the fatigue life.

[0076] Furthermore, the boundary condition constraints include at least one of the following: positive constraints on the output scale parameters and / or shape parameters; the output scale parameters converge to 1 when the input load stress converges to the static strength of the concrete specimen; the output scale parameters converge to the reference life when the input corrosion damage degree converges to 0. Specifically, the reference life is the predicted life of the pre-trained model under the same load stress, or the actual fatigue life of uncorroded concrete under the same load stress.

[0077] In practice, when the load stress on the concrete specimen converges to the static strength of the concrete specimen, it can be considered that the concrete specimen fails during the first cycle. At this point, the output dimensional parameter converges to 1. Conversely, when the degree of corrosion damage converges to 0, the concrete specimen can be considered to have not suffered corrosion damage, and the output dimensional parameter should be equivalent to the baseline life under non-corrosion conditions. In fact, imposing boundary conditions and mechanical monotonicity constraints during training ensures that the output of the prediction model conforms to fundamental physical laws and helps improve the extrapolation prediction capability of the model.

[0078] Specifically, when obtaining the non-destructive testing results of the concrete to be tested in step S4, the same non-destructive testing method as in step S1 needs to be used to ensure that the non-destructive testing results of the concrete to be tested can be mapped to the degree of corrosion damage of the concrete to be tested using the mapping relationship obtained in step S1. This allows for the rapid acquisition of the degree of corrosion damage for predicting corrosion fatigue life without damaging the concrete to be tested. Specifically, the strength grade of the concrete to be tested is the same as or the difference between the strength grade of the concrete sample and the strength grade of the concrete sample is within a preset range.

[0079] In some embodiments, when obtaining the non-destructive testing results of the concrete to be tested in step S4, multiple non-destructive sampling can be performed on the concrete to be tested, and the results from multiple points can be combined to obtain the non-destructive testing results. For example, multiple sampling points can be arranged on the surface or target testing area of ​​the concrete to be tested in a preset grid, randomly or in a partitioned manner. After collecting data point by point, obviously abnormal test values ​​are removed, and the remaining test values ​​are processed by arithmetic average, weighted average, etc., to obtain the non-destructive testing results.

[0080] In fact, obtaining the other key factor parameter values ​​of the concrete to be tested in step S4 means obtaining the parameter values ​​corresponding to all key factors in the key factor set except for the degree of corrosion damage. For example, if the key factors determined in step S1 include: degree of corrosion damage, load stress, characteristic factor A, characteristic factor B, and characteristic factor C, then the other key factor parameter values ​​refer to the parameter values ​​corresponding to load stress, characteristic factor A, characteristic factor B, and characteristic factor C.

[0081] Furthermore, in step S4, the parameter values ​​of all key factors are input into the prediction model. The prediction model outputs the scale parameters and shape parameters of the concrete to be tested, thereby obtaining the Weibull distribution that the fatigue life of the concrete to be tested follows. This can reflect the concentration trend and dispersion of the fatigue life of the concrete to be tested in the environment. The predicted fatigue life of the concrete to be tested is calculated using the inverse function of the cumulative distribution function under a given failure probability.

[0082] In some embodiments, when calculating the predicted fatigue life value under a given failure probability based on the inverse function of the cumulative distribution function in step S4, it can be expressed as:

[0083] in: Given a failure probability, These are the dimensional parameters of the concrete to be tested. The shape parameters of the concrete to be tested are: Given a failure probability The predicted fatigue life of the concrete under test can be obtained through inverse function calculations. The probability of fatigue failure occurring before the next cycle is: .

[0084] In fact, see Figure 2 The present invention also provides a prediction system for implementing the prediction method in any of the above embodiments, comprising: The data processing module 101 is used to acquire the non-destructive testing results, corrosion damage degree and actual fatigue life of concrete samples under different loads at different corrosion ages, construct the mapping relationship between non-destructive testing results and corrosion damage degree, and determine the key factors that affect the actual fatigue life; the key factors include at least the corrosion damage degree and the load stress. The pre-training module 102 is used to pre-train the SN model using a pre-training dataset, construct a neural network with two outputs, namely scale parameters and shape parameters, and use the weights of the pre-trained model as at least part of the initial weights of the scale parameter output branch. The model training module 103 is used to train the neural network by taking the parameter values ​​of the concrete sample corresponding to the key factors as input to the neural network to obtain the life distribution prediction model. During training, the goal is to maximize the joint likelihood of the Weibull distribution of the actual fatigue life of the concrete sample under the scale parameter and shape parameter output by the neural network under the same corrosion age and the same load. The life prediction module 104 is used to obtain the non-destructive testing results of the concrete to be tested and obtain the degree of corrosion damage based on the mapping relationship, obtain the other key factor parameter values ​​of the concrete to be tested, output the scale parameters and shape parameters of the concrete to be tested based on the prediction model, and calculate the fatigue life prediction value under a given failure probability based on the inverse function of the cumulative distribution function.

[0085] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A method of predicting the corrosion fatigue life of concrete, characterized by, include: The non-destructive testing results, corrosion damage degree, and actual fatigue life of concrete samples under different loads at different corrosion ages were obtained, and the mapping relationship between non-destructive testing results and corrosion damage degree was constructed. Identify the key factors that affect actual fatigue life. These key factors should include at least the degree of corrosion damage and the load stress. The SN model is pre-trained using a pre-training dataset to construct a neural network with two outputs, namely scale parameters and shape parameters. The weights of the pre-trained model are used as at least part of the initial weights of the scale parameter output branch. The sample size of the pre-training data is larger than the data size of the actual fatigue life of concrete specimens. The parameter values ​​corresponding to key factors of concrete samples are used as input to a neural network to train the neural network and obtain a life distribution prediction model. During training, the goal is to maximize the joint likelihood of the Weibull distribution of the actual fatigue life of concrete samples under the scale parameter and shape parameter output by the neural network under the same corrosion age and the same load. The non-destructive testing results of the concrete to be tested are obtained and the degree of corrosion damage is obtained based on the mapping relationship. The other key factor parameter values ​​of the concrete to be tested are obtained. The scale parameters and shape parameters of the concrete to be tested are output based on the prediction model. The fatigue life prediction value under a given failure probability is calculated based on the inverse function of the cumulative distribution function.

2. The prediction method according to claim 1, characterized in that: The non-destructive testing (NDT) methods used to obtain the NDT results of the concrete samples are the same as those used to obtain the NDT results of the concrete to be tested. The NDT methods include at least one of ultrasonic testing, impact rebound testing, resistance / conductivity testing, and radiation imaging testing. And / or, the degree of corrosion damage is obtained based on the changes in parameters of concrete samples before and after loading different loads at different corrosion ages, wherein the parameters include at least one of compressive strength, mass, corrosion depth, elastic modulus, crack density, and saturated surface water absorption rate; And / or, using the non-destructive testing results as input and the degree of corrosion damage as output, a regression fitting is performed to obtain a mapping relationship, wherein the regression fitting includes at least one of linear regression, multinomial regression, neural network regression, and support vector regression.

3. The prediction method of claim 1, wherein, When determining the key factors that affect the actual fatigue life, the process includes: obtaining the characteristic factors that affect the actual fatigue life, evaluating the importance of the characteristic factors to obtain evaluation values, and taking the characteristic factors with evaluation values ​​not less than a threshold as key factors; wherein, the threshold is the minimum value of the evaluation value corresponding to the degree of corrosion damage and the load stress, and the characteristic factors include the degree of corrosion damage, the load stress, the pH value of the corrosive medium, and the corrosion age.

4. The prediction method of claim 3, wherein, When using the random forest algorithm to evaluate the importance of feature factors, the process includes: constructing multiple decision trees to form a random forest; obtaining a first score based on the change in impurity of the feature factor when splitting at a decision tree node; permuting the values ​​of the feature factors in the samples outside the decision tree bag and obtaining a second score based on the increase in error before and after the permutation; and normalizing and weighting the first and second scores to obtain the evaluation value of the feature factor.

5. The prediction method of claim 3, wherein: When the evaluation value of corrosion age is not less than the threshold, the Spearman correlation coefficient between corrosion age and corrosion damage degree is calculated. If the correlation coefficient is not less than the preset threshold, the corrosion age is discarded; otherwise, the corrosion age is used as a key factor. The calculation of the Spearman correlation coefficient includes: , wherein: Spearman correlation coefficient of corrosion age and corrosion damage degree; is the rank of corrosion age of the th concrete sample ; is the rank of corrosion damage degree of the th concrete sample ; is the total number of concrete samples.

6. The prediction method of claim 1, wherein: The neural network is trained based on mechanical monotonicity constraints and boundary condition constraints; The mechanical monotonicity constraint includes at least one of the following: when the input load stress increases and other characteristics remain unchanged, the output scale parameter decreases monotonically; when the input corrosion damage degree increases and other characteristics remain unchanged, the output scale parameter decreases monotonically. Boundary condition constraints include at least one of the following: positive constraints on the output scale parameters and / or shape parameters; when the input load stress converges to the static strength of the concrete specimen, the output scale parameters converge to 1; when the input corrosion damage degree converges to 0, the output scale parameters converge to the reference life.

7. The prediction method of claim 1, wherein, Training a neural network includes: Obtain representative values ​​of key factors in the concrete sample group. The sample group is a combination of concrete samples with the same corrosion age and the same load. Input the representative values ​​of the parameters into the neural network, and output the scale parameters and shape parameters of the sample group; A loss function is constructed with the objective of maximizing the joint likelihood of the actual fatigue life of concrete specimens in the sample group under dimensional and shape parameters. The lifetime distribution prediction model is obtained by backpropagating the neural network based on minimizing the loss function and updating the weights until convergence.

8. The prediction method of claim 7, wherein, When constructing a loss function with the objective of maximizing joint likelihood, the following should be included: , wherein: is the shape parameter for the th concrete specimen group; is the size parameter for the th concrete specimen group; is the actual fatigue life of the th concrete specimen in the th concrete specimen group; is the total number of concrete specimens in the th concrete specimen group; is the total number of concrete specimen groups; is the total number of concrete specimens.

9. The prediction method of claim 1, wherein, When calculating the fatigue life prediction value for a given failure probability based on the inverse function of the cumulative distribution function, the following steps are included: , in: Given the failure probability; These are the dimensional parameters of the concrete to be tested; The shape parameters of the concrete to be tested; Given a failure probability The predicted fatigue life value at that time.

10. A prediction system for implementing the prediction method as described in any one of claims 1 to 9, characterized in that, include: The data processing module is used to obtain the non-destructive testing results, corrosion damage degree and actual fatigue life of concrete samples under different loads at different corrosion ages, construct the mapping relationship between non-destructive testing results and corrosion damage degree, and determine the key factors that affect the actual fatigue life. Key factors include at least the degree of corrosion damage and the load stress; The pre-training module is used to pre-train the SN model using a pre-training dataset, construct a neural network with two outputs, namely scale parameters and shape parameters, and use the weights of the pre-trained model as at least part of the initial weights of the scale parameter output branch. The model training module is used to train the neural network by taking the parameter values ​​of the concrete sample corresponding to the key factors as input to the neural network to obtain the life distribution prediction model. During training, the goal is to maximize the joint likelihood of the Weibull distribution of the actual fatigue life of the concrete sample under the scale parameter and shape parameter of the neural network output under the same corrosion age and the same load. The life prediction module is used to obtain the non-destructive testing results of the concrete to be tested and obtain the degree of corrosion damage based on the mapping relationship. It also obtains the values ​​of other key factor parameters of the concrete to be tested, outputs the scale parameters and shape parameters of the concrete to be tested based on the prediction model, and calculates the fatigue life prediction value under a given failure probability based on the inverse function of the cumulative distribution function.