Steel structure and welding seam corrosion fatigue life prediction method based on neural network

Abstract

This invention provides a neural network-based method for predicting the corrosion fatigue life of steel structures and their welds, belonging to the technical field of corrosion fatigue life prediction for steel structures and their welds. To overcome the shortcomings of existing steel structure corrosion fatigue prediction schemes in terms of generalization ability and prediction accuracy, the method preprocesses the collected multi-source corrosion fatigue data; constructs a main fatigue life prediction network; constructs a corrosion-induced stress amplification coefficient subnetwork ℓ₁; and constructs data loss functions respectively. L data Monotonic physical constraint loss function L MR Continuous damage mechanical loss function L CDM The model is trained; the life prediction of steel structures and their welded structures is performed based on transfer learning; the present invention is applied to the prediction of corrosion fatigue life of steel structures and their welds.

Description

Technical Field

[0001] This invention provides a method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks, belonging to the technical field of predicting the corrosion fatigue life of steel structures and their welds. Background Technology

[0002] Currently, the prediction of corrosion fatigue life of steel mainly employs empirical formulas, statistical models based on SN curves, continuous damage mechanics (CDM) models, and purely data-driven machine learning models. Empirical formulas and CDM models typically rely on a large amount of fatigue test data for parameter calibration, with model parameters being fixed constants. Their applicability is limited to specific materials and working conditions, making it difficult to adapt to different materials, corrosion states, and loading conditions. These models heavily depend on single or small-scale test data, resulting in insufficient predictive stability and reliability. While data-driven methods such as artificial neural networks can fit complex nonlinear relationships, they lack physical constraints. Purely data-driven models lack physical consistency and are prone to producing results that violate monotonicity and damage evolution laws.

[0003] In summary, in the prediction of corrosion fatigue life of welded structures, due to the scarcity of experimental data and the limited sample size, it is difficult to directly establish a high-precision model. Existing prediction methods have significant shortcomings in terms of generalization ability and prediction accuracy. Summary of the Invention

[0004] To address the technical problems existing in the background art, the present invention adopts the following technical solution: a method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks, comprising the following prediction steps:

[0005] Step 1: Preprocess the collected multi-source corrosion fatigue data;

[0006] Step 2: Construct the master fatigue life prediction network;

[0007] Step 3: Constructing a subnetwork of corrosion-induced stress amplification factor ;

[0008] Step 4: Construct the data loss function L separately. data Monotonic physical constraint loss function L MR Continuous damage mechanical loss function L CDM And train the model;

[0009] Step 5: Predict the lifespan of steel structures and their welded structures based on transfer learning.

[0010] The specific method for step one is as follows:

[0011] Step 1.1: Collect fatigue test data of different steels and their welded joints under corrosive environments and cyclic loading, including: maximum stress σ. max Minimum stress σ min Loading frequency f, corrosion mass loss rate η, material yield strength f y Elastic modulus E, alloy element content w Ae Residual element content w Re And the corresponding fatigue life N f ;

[0012] Step 1.2: Use Spearman correlation coefficient to perform correlation analysis on the input features, filter out or merge highly correlated variables, and obtain a feature subset with low redundancy and clear physical meaning. The expression for the correlation coefficient ρ is:

[0013] (1);

[0014] Among them, R xi and R yi Representing the original dataset x respectively i and y i rank sequence;

[0015] Step 1.3: The selected features are dimensionless using Z-score standardization. Z-score standardization transforms the original features into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows:

[0016] (2);

[0017] Where: x represents the original feature observation value; σ is the standardized feature value; μ is the arithmetic mean of the original dataset for this feature; σ is the standard deviation of the original dataset for this feature.

[0018] The specific method for step two is as follows:

[0019] Step 2.1: Construct a multi-layer fully connected neural network as the main prediction network. The input is the standardized feature vector, and the output is the predicted fatigue life. ;

[0020] The output h of the neurons in the hidden layer is calculated using the following formula:

[0021] (3);

[0022] Where, x i Represents the input features, w ji It is the connection weight between the i-th neuron in the previous layer and the j-th neuron in the current layer, bj f(x) is the bias term, and f(x) is the activation function.

[0023] Step 2.2: Use the exponential linear unit, which has self-normalization properties, as the activation function. The expression is:

[0024] (4);

[0025] Step 2.3: Train the model using the Adam optimizer. The learning rate scheduling is implemented using a cosine annealing strategy, expressed as:

[0026] (5);

[0027] Among them, ι t Indicates the learning rate in the current round, ι max It is the initial maximum learning rate, ι min It is the minimum learning rate at the lower bound of annealing, where t represents the current iteration round, and t max Indicates the maximum number of iterations;

[0028] Step 2.4: Using the experimental lifetime N f With predicted lifespan The mean square error between them is used as the data loss function L data The expression is:

[0029] (6).

[0030] The specific method for step three is as follows:

[0031] Step 3.1: Constructing the corrosion stress amplification factor subnetwork Its inputs are the corrosion damage parameter η and the alloy element content w. Ae and residual element content w Re The expression is:

[0032] ξ cor = (η, w Ae , w Re (7);

[0033] Step 3.2: Use the subnetwork to output the corrosion stress amplification factor ξ cor This characterizes the localized stress concentration effect caused by corrosion pitting;

[0034] Step 3.3: For ξ cor Apply physical constraints to satisfy ξ cor ≥ 1, and monotonically increases with η, the expression is:

[0035] (8);

[0036] Step 3.4: Utilizing ξ cor The nominal stress is converted into equivalent stress for subsequent fatigue damage calculation. The expression for equivalent stress is:

[0037] σ eq = ξ cor σ n (9).

[0038] In step four, a monotonic physical constraint loss function L is constructed. MR The specific method is as follows:

[0039] The physical laws are transformed into mathematical constraints using a penalty function, which incurs a penalty only when the law is violated. The expression is as follows:

[0040] (10);

[0041] Construct a monotonic physical constraint loss function L MR The expression is:

[0042] (11).

[0043] In step four, the continuous damage mechanical loss function L is constructed. CDM The specific method is as follows:

[0044] Step 4.1: The fatigue damage index D is defined based on the reduction of the effective load-bearing area, and its expression is:

[0045] (12);

[0046] Where S is the effective bearing area and S0 is the undamaged cross-sectional area;

[0047] A fatigue damage evolution equation is established. Under high-cycle fatigue conditions, the expression for the damage increment ΔD / ΔN per cycle under different stress ratios R is as follows:

[0048] (13);

[0049] (14);

[0050] Therefore, given the initial condition N f = 0, D = 0 and termination condition N f = N R D = 1, the number of cycles N required to achieve fatigue failure is calculated. R The calculation formula is:

[0051] (15);

[0052] To facilitate the calibration of CDM model parameters, each term in equation (15) is defined as B = q + 1;

[0053] Step 4.2: Construct the parametric subnetwork With loading frequency f, yield strength f y And the elastic modulus E is used as input, and the CDM model parameters A and B are output for calibration. The expression is:

[0054] (16);

[0055] Step 4.3: Utilize the subnetting method from Step 3 Calculated maximum equivalent stress σ eqM and minimum equivalent stress σ eqm Calculate the theoretical fatigue life N CDM The calculation formula is:

[0056] (17);

[0057] Step 4.4: Using the theoretical lifespan N of the CDM CDM With predicted lifespan The mean square error between them is used as the theoretical loss function L in CDM. CDM This is used to constrain the consistency between the neural network prediction results and the theoretical damage evolution law. The calculation formula is:

[0058] (18);

[0059] Step 4.5: Reduce data loss L data Monotonicity constraint loss L MR and CDM loss L CDM The weighted combination forms the total loss function L. total The expression is:

[0060] (19);

[0061] Step 4.6: The backpropagation algorithm is used to jointly train the parameters of the main network and each sub-network, and an early stopping mechanism is introduced. When the verification loss does not improve within one-tenth of the maximum number of training epochs, the training will automatically terminate. Then, the weights K of each loss are adaptively determined by the Bayesian optimization method.

[0062] The specific method for step five is as follows:

[0063] Step 5.1: Use the structural steel corrosion fatigue model as the source domain model and the welded structure data as the target domain;

[0064] Step 5.2: Introduce a sub-network of weld geometric stress amplification factor into the target domain model. The inputs include weld width, bevel angle, and bevel type; the stress amplification factor ξ caused by weld geometry is calculated. geom and its equivalent stress σ eq The calculation formula is:

[0065] (20);

[0066] (twenty one);

[0067] (twenty two);

[0068] Consistent with step four, the data loss function L is then constructed respectively. data Monotonic physical constraint loss function L MR Continuous damage mechanical loss function L CDM and its equivalent stress σ eq ;

[0069] Step 5.3: Freeze the learned network parameters in the source domain, fine-tune only the layers related to welding features, and then select the best-performing combination sequence.

[0070] Step 5.4: Predict the corrosion fatigue life of welded structures under small sample conditions.

[0071] The beneficial effects of this invention compared to the prior art are as follows: The method for predicting the corrosion fatigue life of steel structures and their welds provided by this invention mainly adopts multi-physical constraint embedding, which significantly improves the physical consistency and reliability of the prediction results. It achieves adaptive learning of corrosion and damage parameters through sub-networks, avoiding the limitations of manual calibration in traditional models, and has good generalization ability under multiple steel types and multiple corrosion conditions. This invention adopts the transfer learning method, which can effectively solve the problem of scarce corrosion fatigue data for welded structures. It is suitable for the full life assessment and remaining life prediction of engineering structures and has good engineering application prospects. Attached Figure Description

[0072] The present invention will be further described below with reference to the accompanying drawings:

[0073] Figure 1 This is a schematic diagram of the multi-subnet physical information neural network used in this invention;

[0074] Figure 2 This is a schematic diagram of the principle of the physical information neural network using multi-subnet transfer learning in this invention;

[0075] Figure 3 This is a flowchart of the steps of the present invention;

[0076] Figure 4The diagram shows the results of lifetime prediction using two types of neural networks in this invention. Detailed Implementation

[0077] like Figures 1 to 4 As shown, this invention provides a method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks. It involves a steel corrosion fatigue life prediction scheme that integrates physical mechanisms, data-driven approaches, and transfer learning. By embedding multi-physics sub-networks and physical constraint loss functions into the neural network, it achieves high-precision prediction of the life of structural steel and welded structures under corrosion-fatigue coupling.

[0078] The meanings of the parameters appearing in this invention are as follows:

[0079] N f Fatigue life (number of failure cycles);

[0080] σ a For stress amplitude, σ max For the maximum stress, σ min The minimum stress is Δσ, and the stress range is Δσ.

[0081] R is the stress ratio; f is the loading frequency; η is the mass loss rate after corrosion; f u The ultimate strength; f y E is the yield strength; R² is the elastic modulus; RMSE is the root mean square error; ρ is the correlation coefficient.

[0082] w P For phosphorus content, w S For sulfur content, w Cr For chromium content, w Ni For nickel content, w Cu This refers to the copper content;

[0083] α is the butt weld bevel angle 1, β is the butt weld bevel angle 2, and d α d represents the weld width corresponding to the bevel angle α of the butt weld. β This refers to the weld width corresponding to the bevel angle β of the butt weld.

[0084] v type For bevel type; w Re For residual elements; w Ae It is an alloying element;

[0085] Q, q, v, R p v These are the parameters of the CDM theoretical equations;

[0086] l t The learning rate for the current step size, l max l is the initial maximum learning rate.min To minimize the learning rate, t max L is the maximum number of iterations. data Let L be the data loss function. MR Let L be the monotonic regularization loss function. CDM The loss function is based on continuous loss mechanics;

[0087] This is a subnetwork for the stress amplification factor induced by corrosion. This is a subnetwork for stress amplification factors induced by weld geometry. Subnetwork for calibrating the mechanical parameters of continuous damage; ξ cor ξ is the corrosion-induced stress amplification factor; geom σ is the stress amplification factor induced by weld geometry. eqM The maximum equivalent stress; σ eqm The minimum equivalent stress;

[0088] A represents the parameters of the continuous damage mechanics theoretical equation; B represents the parameters of the continuous damage mechanics theoretical equation; K1 represents the weight of the data loss function; K2 represents the weight of the monotonic regularized loss function; K3 represents the weight of the loss function based on continuous loss mechanics; ψ represents the transfer learning configuration.

[0089] Furthermore, the present invention provides a method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks, which specifically includes the following prediction steps:

[0090] Step 1: Preprocessing of multi-source corrosion fatigue data:

[0091] Step 1.1: Data Acquisition:

[0092] Fatigue test data of different steels and their welded joints under corrosive environments and cyclic loading were collected, including: maximum stress σ. max Minimum stress σ min Loading frequency f, corrosion mass loss rate η, material yield strength f y Elastic modulus E, alloy element content w Ae Residual element content w Re And the corresponding fatigue life N f .

[0093] Step 1.2: Feature Correlation Analysis and Dimensionality Reduction:

[0094] Spearman correlation coefficient is used to perform correlation analysis on the input features, filtering out or merging highly correlated variables to obtain a feature subset with low redundancy and clear physical meaning. The expression for the correlation coefficient ρ is:

[0095] (1);

[0096] Among them, R xi and R yi Representing the original dataset x respectively i and y i The rank sequence.

[0097] Step 1.3: Data standardization processing:

[0098] In corrosion fatigue life prediction models, input parameters (such as stress, mass loss rate, and strength) have different physical meanings and magnitudes. Directly inputting heterogeneous data can lead to model training being dominated by large numerical features, resulting in bias. Therefore, the selected features need to be dimensionlessly processed using the Z-score normalization method to eliminate the influence of differences in the dimensions and numerical scales of different physical quantities on model training.

[0099] The Z-score method is used to transform the original features into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows:

[0100] (2);

[0101] Where: x represents the original feature observation value; σ is the standardized feature value; μ is the arithmetic mean of the original dataset for this feature; σ is the standard deviation of the original dataset for this feature.

[0102] The above derivation and processing procedure is as follows:

[0103] Mean calculation: For all sample values ​​x of a certain feature i Summing and averaging the results determines the central location of the distribution.

[0104] Bias quantization: Calculate the distance (x - μ) between each sample and the center position to eliminate positional offset.

[0105] Scaling transformation: Dividing the deviation by the standard deviation σ eliminates the influence of dimensions, making features with different units (such as MPa, Hz, %) on the same order of magnitude; this processing ensures the balanced contribution of each feature weight during model training, significantly improving the convergence speed and generalization ability of the neural network.

[0106] Step 2: Construct the master fatigue life prediction network:

[0107] Step 2.1: Construct a multi-layer fully connected neural network as the main prediction network. The input is the standardized feature vector, and the output is the predicted fatigue life. ;

[0108] The output h of the neurons in the hidden layer is represented by equation (3):

[0109] (3);

[0110] Where, x i Represents the input features, w ji It is the connection weight between the i-th neuron in the previous layer and the j-th neuron in the current layer, b j f(x) is the bias term, and f(x) is the activation function.

[0111] Step 2.2: Use the self-normalized exponential linear unit (SELU) as the activation function (Equation (4)) to improve the stability of model training, with predefined constants λ=1.0507 and a=1.6733.

[0112] (4);

[0113] Step 2.3: The model is trained using the Adam optimizer, and the learning rate is scheduled using a cosine annealing strategy, as shown in equation (5):

[0114] (5);

[0115] Among them, ι t Indicates the learning rate in the current round, ι max It is the initial maximum learning rate, ι min It is the minimum learning rate at the lower bound of annealing, where t represents the current iteration round, and t max This indicates the maximum number of iterations.

[0116] Step 2.4: Using the experimental lifetime N f With predicted lifespan The mean square error between them is used as the data loss function L data (6):

[0117] (6).

[0118] Step 3: Construct the corrosion stress amplification factor sub-network:

[0119] To transform static physical formulas into dynamic models that can adapt to different operating conditions, the system embeds three key subnets that run in parallel with the main prediction network and provide the necessary parameters for physical constraints.

[0120] Constructing a subnetwork of corrosion-induced stress amplification factor hour:

[0121] Inputs: Corrosion mass loss rate η, alloying element w Ae and residual element w Re .

[0122] Function: Outputs corrosion stress amplification factor, used to amplify the nominal stress σ n Convert to equivalent stress σ considering corrosion damage eq (9).

[0123] Step 3.1: Constructing the corrosion stress amplification factor subnetwork Its inputs are the corrosion damage parameter η and the alloy element content w. Ae and residual element content w Re The expression is:

[0124] ξ cor = (η, w Ae , w Re (7);

[0125] Step 3.2: Subnetwork outputs corrosion stress amplification factor ξ cor , used to characterize the local stress concentration effect caused by corrosion pits;

[0126] Step 3.3: For ξ cor Apply physical constraints to satisfy ξ cor ≥ 1, and monotonically increases with η, the expression is:

[0127] (8);

[0128] Step 3.4: Utilizing ξ cor The nominal stress is converted into equivalent stress for subsequent fatigue damage calculation. The expression for equivalent stress is:

[0129] σ eq = ξ cor σ n (9).

[0130] Step 4: Construct the multi-physics constraint loss function, including the data loss function L. data Monotonic physical constraint loss function L MR Continuous damage mechanical loss function L CDM And train the model, including:

[0131] Constructing the data loss function L data ;

[0132] Construct a monotonic physical constraint loss function L MR This is used to constrain the monotonic relationship of predicted fatigue life with respect to key physical variables, including:

[0133] (1) Fatigue life decreases as the maximum stress increases;

[0134] (2) Fatigue life increases with increasing secondary maximum stress;

[0135] (3) Fatigue life decreases as the degree of corrosion damage increases;

[0136] (4) Fatigue life increases with the increase of material strength.

[0137] (5) During the subnet construction process ξ cor The relationship is positively correlated with the mass loss rate η.

[0138] Based on extensive experimental evidence, the fatigue life N was determined. f Monotonic relationships with key variables include:

[0139] Stress constraint: with the maximum stress σ max Increased crack initiation and propagation accelerate, leading to a decrease in lifetime (the first derivative is negative); and this decreasing trend tends to level off in the long-life region (the second derivative is positive).

[0140] Corrosion constraint: An increase in the mass loss rate η leads to stress concentration and significantly reduces fatigue performance.

[0141] Material constraint: Material yield strength f y The higher the level, the stronger the ability to resist fatigue damage, and the longer the lifespan.

[0142] Subnet constraint: Corrosion stress amplification factor ξ cor It should increase monotonically as η increases.

[0143] The above physical laws are transformed into mathematical constraints using a penalty function, which incurs a penalty only when the laws are violated. The expression is as follows:

[0144] (10);

[0145] Finally, the monotonic physical constraint loss function L is constructed. MR The expression is:

[0146] (11);

[0147] Constructing the continuous damage mechanical loss function L CDM ,include:

[0148] Step 4.1: The fatigue damage index D is defined based on the reduction of the effective load-bearing area, and its expression is:

[0149] (12);

[0150] Where S is the effective load-bearing area and S0 is the undamaged cross-sectional area.

[0151] Within the framework of irreversible thermodynamics and CDM theory, a fatigue damage evolution equation can be established. Under high-cycle fatigue conditions, the expression for the damage increment ΔD / ΔN per cycle under different stress ratios R is:

[0152] (13);

[0153] (14);

[0154] Therefore, given the initial condition N f = 0, D = 0 and termination condition N f = N R D = 1, the number of cycles N required to achieve fatigue failure is calculated. R The calculation formula is:

[0155] (15);

[0156] Subsequently, to facilitate the calibration of the continuous damage mechanics (CDM) model parameters, the terms in equation (15) are defined as follows: B = q+1. However, the evolution of fatigue damage is not only affected by degradation caused by external corrosion, but also strongly depends on the contribution of material constitutive properties to the CDM parameters.

[0157] Therefore, in order to further improve the model framework and ensure that CDM theory maintains accuracy and universality under complex service conditions, it is necessary to construct a parametric subnetwork. A data-driven approach is used to calibrate core CDM parameters A and B.

[0158] Step 4.2: Sub-network With loading frequency f, yield strength f y The elastic modulus E is taken as input, and the CDM model parameters A and B are output, expressed as follows:

[0159] (16);

[0160] Step 4.3: Reuse the subnetting method from Step 3 Calculated maximum equivalent stress σ eqM and minimum equivalent stress σ eqm Calculate the theoretical fatigue life N CDM The calculation formula is:

[0161] (17);

[0162] Step 4.4: Using the theoretical lifespan N of the CDM CDM With predicted lifespan The mean square error between them is used as the theoretical loss function L in CDM.CDM This is used to constrain the consistency between the neural network prediction results and the theoretical damage evolution law. The calculation formula is:

[0163] (18);

[0164] Step 4.5: Reduce data loss L data Monotonicity constraint loss L MR and CDM loss L CDM The weighted combination forms the total loss function L. total The expression is:

[0165] (19);

[0166] Step 4.6: The backpropagation algorithm is used to jointly train the parameters of the main network and each sub-network, and an early stopping mechanism is introduced. When the validation loss does not improve within one-tenth of the maximum training epochs, the training will automatically terminate. Then, the weights K of each loss are adaptively determined using the Bayesian optimization method.

[0167] Step 5: Predict the lifespan of steel structures and their welded structures based on transfer learning, including:

[0168] Step 5.1: Use the structural steel corrosion fatigue model as the source domain model and the welded structure data as the target domain;

[0169] Step 5.2: Introduce a sub-network of weld geometric stress amplification factor into the target domain model. The inputs include weld width, bevel angle, and bevel type; the stress amplification factor ξ caused by weld geometry is calculated. geom and its equivalent stress σ eq The calculation formula is:

[0170] (20);

[0171] (twenty one);

[0172] (twenty two);

[0173] Consistent with step four, the data loss function L is then constructed respectively. data Monotonic physical constraint loss function L MR Continuous damage mechanical loss function L CDM and its equivalent stress σ eq The calculation is updated to Equation (22).

[0174] Step 5.3: Freeze the learned network parameters in the source domain and fine-tune only the layers related to welding features; essentially, this is a random combination, and then select the combination sequence with the best training performance. The training combination sequence information is shown in the table below.

[0175]

[0176] Step 5.4: Predict the corrosion fatigue life of welded structures under small sample conditions.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks, characterized in that: The prediction steps include the following: Step 1: Preprocess the collected multi-source corrosion fatigue data; Step 2: Construct the master fatigue life prediction network; Step 3: Constructing a subnetwork of corrosion-induced stress amplification factor ; Step 4: Construct the data loss function L separately. data Monotonic physical constraint loss function L MR Continuous damage mechanical loss function L CDM And train the model; Step 5: Predict the lifespan of steel structures and their welded structures based on transfer learning.

2. The method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks according to claim 1, characterized in that: The specific method for step one is as follows: Step 1.1: Collect fatigue test data of different steels and their welded joints under corrosive environments and cyclic loading, including: maximum stress σ. max Minimum stress σ min Loading frequency f, corrosion mass loss rate η, material yield strength f y Elastic modulus E, alloy element content w Ae Residual element content w Re And the corresponding fatigue life N f ; Step 1.2: Use Spearman correlation coefficient to perform correlation analysis on the input features, filter out or merge highly correlated variables, and obtain a feature subset with low redundancy and clear physical meaning. The expression for the correlation coefficient ρ is: （1）； Among them, R xi and R yi Representing the original dataset x respectively i and y i rank sequence; Step 1.3: The selected features are dimensionless using Z-score standardization. Z-score standardization transforms the original features into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: （2）； Where: x represents the original feature observation value; σ is the standardized feature value; μ is the arithmetic mean of the original dataset for this feature; σ is the standard deviation of the original dataset for this feature.

3. The method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks according to claim 2, characterized in that: The specific method for step two is as follows: Step 2.1: Construct a multi-layer fully connected neural network as the main prediction network. The input is the standardized feature vector, and the output is the predicted fatigue life. ; The output h of the neurons in the hidden layer is calculated using the following formula: （3）； Where, x i Represents the input features, w ji It is the connection weight between the i-th neuron in the previous layer and the j-th neuron in the current layer, b j f(x) is the bias term, and f(x) is the activation function. Step 2.2: Use the exponential linear unit, which has self-normalization properties, as the activation function. The expression is: （4）； Step 2.3: Train the model using the Adam optimizer. The learning rate scheduling is implemented using a cosine annealing strategy, expressed as: （5）； Among them, ι t Indicates the learning rate for the current round, ι max It is the initial maximum learning rate, ι min It is the minimum learning rate at the lower bound of annealing, where t represents the current iteration round, and t max Indicates the maximum number of iterations; Step 2.4: Using the experimental lifetime N f With predicted lifespan The mean square error between them is used as the data loss function L data The expression is: （6）。 4. The method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks according to claim 3, characterized in that: The specific method for step three is as follows: Step 3.1: Constructing the corrosion stress amplification factor subnetwork Its inputs are the corrosion damage parameter η and the alloy element content w. Ae and residual element content w Re The expression is: x cor = (h, w Ae , w Re )(7); Step 3.2: Use the subnetwork to output the corrosion stress amplification factor ξ cor This characterizes the localized stress concentration effect caused by corrosion pitting; Step 3.3: For ξ cor Apply physical constraints to satisfy ξ cor ≥ 1, and monotonically increases with η, the expression is: （8）； Step 3.4: Utilizing ξ cor The nominal stress is converted into equivalent stress for subsequent fatigue damage calculation. The expression for equivalent stress is: s eq = ξ cor s n (9).

5. The method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks according to claim 4, characterized in that: In step four, a monotonic physical constraint loss function L is constructed. MR The specific method is as follows: The physical laws are transformed into mathematical constraints using a penalty function, which incurs a penalty only when the law is violated. The expression is as follows: （10）； Construct a monotonic physical constraint loss function L MR The expression is: （11）。 6. The method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks according to claim 5, characterized in that: In step four, the continuous damage mechanical loss function L is constructed. CDM The specific method is as follows: Step 4.1: The fatigue damage index D is defined based on the reduction of the effective load-bearing area, and its expression is: （12）； Where S is the effective bearing area and S0 is the undamaged cross-sectional area; A fatigue damage evolution equation is established. Under high-cycle fatigue conditions, the expression for the damage increment ΔD / ΔN per cycle under different stress ratios R is as follows: （13）； （14）； Therefore, given the initial condition N f = 0, D = 0 and termination condition N f = N R D = 1, the number of cycles N required to achieve fatigue failure is calculated. R The calculation formula is: （15）； To facilitate the calibration of CDM model parameters, the terms in equation (15) are defined as follows: B = q + 1; Step 4.2: Construct the parametric subnetwork With loading frequency f, yield strength f y And the elastic modulus E is used as input, and the CDM model parameters A and B are output for calibration. The expression is: （16）； Step 4.3: Utilize the subnetting method from Step 3 Calculated maximum equivalent stress σ eqM and minimum equivalent stress σ eqm Calculate the theoretical fatigue life N CDM The calculation formula is: （17）； Step 4.4: Using the theoretical lifespan N of the CDM CDM With predicted lifespan The mean square error between them is used as the theoretical loss function L in CDM. CDM This is used to constrain the consistency between the neural network prediction results and the theoretical damage evolution law. The calculation formula is: （18）； Step 4.5: Reduce data loss L data Monotonicity constraint loss L MR and CDM loss L CDM The weighted combination forms the total loss function L. total The expression is: （19）； Step 4.6: The backpropagation algorithm is used to jointly train the parameters of the main network and each sub-network, and an early stopping mechanism is introduced. When the verification loss does not improve within one-tenth of the maximum number of training epochs, the training will automatically terminate. Then, the weights K of each loss are adaptively determined by the Bayesian optimization method.

7. The method for predicting the corrosion fatigue life of steel structures and their welds based on neural networks according to claim 6, characterized in that: The specific method for step five is as follows: Step 5.1: Use the structural steel corrosion fatigue model as the source domain model and the welded structure data as the target domain; Step 5.2: Introduce a sub-network of weld geometric stress amplification factor into the target domain model. The inputs include weld width, bevel angle, and bevel type; the stress amplification factor ξ caused by weld geometry is calculated. geom and its equivalent stress σ eq The calculation formula is: （20）； （21）； （22）； Consistent with step four, the data loss function L is then constructed respectively. data Monotonic physical constraint loss function L MR Continuous damage mechanical loss function L CDM and its equivalent stress σ eq ; Step 5.3: Freeze the learned network parameters in the source domain, fine-tune only the layers related to welding features, and then select the best-performing combination sequence. Step 5.4: Predict the corrosion fatigue life of welded structures under small sample conditions.

Images