A data-driven creep life prediction method for chromium-molybdenum-vanadium steel

Abstract

The present invention discloses a data-driven creep life prediction method for chromium-molybdenum-vanadium steel, comprising: Step S1, data collection, converting the collected data into a structured digital format and manually annotating the data; Step S2, using a generative adversarial network algorithm based on hyperbolic space embedding for sample generation, and then expanding the data; Step S3, using a neural network algorithm based on dynamic population evolution optimization as a feature extraction model to train the expanded data; Step S4, using an autoencoder neural network based on feature refinement as a feature reduction model to train the data after feature extraction; Step S5, using a fractional-order neural network based on quantum state transfer as a classifier model to train the data after dimensionality reduction; Step S6, using the fully trained model to process and predict new samples. The present invention improves the prediction ability of the creep life of materials.

Description

Technical Field

[0001] The present invention relates to a method for predicting the creep life of chromium-molybdenum-vanadium steel based on data driving. Background Art

[0002] Due to its excellent high-temperature performance and creep resistance, chromium-molybdenum-vanadium steel is widely used in the manufacture of equipment under high-temperature and high-pressure environments, such as power station boilers, petrochemical equipment, etc. However, as the equipment operates under high-temperature stress conditions for a long time, the material will undergo creep phenomena, resulting in performance degradation and even failure. Therefore, accurately predicting the creep life of chromium-molybdenum-vanadium steel is of great significance for ensuring the safe operation of equipment.

[0003] Traditional creep life prediction methods mainly rely on experimental data or empirical formulas. However, creep experiments are time-consuming and costly, and it is difficult to carry out comprehensively in practical applications. In addition, the material is affected by various factors during the creep process, such as temperature, stress, microstructural changes, etc., resulting in low accuracy of prediction results. In recent years, with the development of data-driven methods, technologies such as machine learning have gradually been introduced into the field of material life prediction. Through the mining and analysis of large-scale data, the accuracy and generalization ability of prediction can be effectively improved. However, the data characteristics in the field of materials science are complex, and the problem of data scarcity still exists. How to improve the performance of the prediction model under limited data has become a technical problem to be solved urgently.

[0004] The existing technologies mainly have the following problems:

[0005] 1. When traditional generative adversarial networks generate material data, it is difficult to capture the hierarchical structure and complex geometric characteristics of the material data. The generated data samples cannot maintain consistent quality at different scales, resulting in poor data augmentation effects and unable to meet the requirements of materials science for high-quality data.

[0006] 2. Most of the existing feature extraction models are based on fixed evolutionary rules and lack the adaptive ability to the real-time changes of data characteristics, resulting in an inflexible feature extraction process and prone to problems such as gradient disappearance or gradient explosion, affecting the efficiency and stability of model training. At the same time, it is difficult for the model to effectively extract key features from complex material data, and the generalization ability is limited.

[0007] 3. When existing feature dimension reduction technologies process high-dimensional data, there are often problems of information loss, especially it is difficult to effectively retain the core features of material data during the dimension reduction process. Redundant features and noise are difficult to be fully weakened, resulting in insufficient conciseness of the features after dimension reduction, affecting the input quality of the subsequent classifier model and the overall prediction accuracy.

[0008] 4. Most existing classifier models rely on traditional activation functions and neural network structures, and are unable to finely control the flow of information. The activation function lacks the ability to dynamically adjust, making it difficult to capture the non-linear changes in complex material data, resulting in insufficient accuracy in predicting material life and limited processing ability of the model.

[0009] Therefore, a data-driven creep life prediction method for chromium molybdenum vanadium steel is provided. Summary of the Invention

[0010] The purpose of the present invention is to provide a data-driven creep life prediction method for chromium molybdenum vanadium steel to overcome the existing defects, and improve the prediction ability of material creep life.

[0011] The technical solution to achieve the above purpose is:

[0012] A data-driven creep life prediction method for chromium molybdenum vanadium steel, comprising:

[0013] Step S1, data acquisition, converting the acquired data into a structured digital format and manually annotating the data;

[0014] Step S2, using a generative adversarial network algorithm based on hyperbolic space embedding for sample generation, and then expanding the data;

[0015] Step S3, using a neural network algorithm based on dynamic population evolution optimization as a feature extraction model to train the expanded data;

[0016] Step S4, using an autoencoder neural network based on feature refinement as a feature dimensionality reduction model to train the data after feature extraction;

[0017] Step S5, using a fractional-order neural network based on quantum state transfer as a classifier model to train the data after dimensionality reduction;

[0018] Step S6, using the fully trained feature extraction model, feature dimensionality reduction model and classifier model to process and predict new samples, and then obtaining the classification result of predicting the creep life of chromium molybdenum vanadium steel;

[0019] In the said step S2, the training process of the generative adversarial network algorithm based on hyperbolic space embedding is as follows:

[0020] Step S21, initializing the network parameters of the generator and discriminator; the initialization method is expressed as:

[0021]

[0022] In the formula, W cDenote the weight matrix of the generative adversarial network as \(b\). c Denote the bias vector of the generative adversarial network; \(\sim\) means following a specific distribution. Denote the normal distribution. Denote the normal distribution with a mean of 0 and a variance of 1.

[0023] In step S22, map the original data into the hyperbolic space, and retain the inherent hierarchical structure and complex characteristics of the data through the mapping function; let the hyperbolic space be Perform a non-linear transformation using the hyperbolic tangent function, expressed as:

[0024]

[0025] In the formula, \(X\) c is the original data vector. Denote the \(i\)-th component of the original data vector, \(\alpha\) c is the expansion parameter used to adjust the sensitivity of the embedding, \(d\) new is the embedding dimension. Denote the data vector in the mapped hyperbolic space. is the first parameter for adjusting the mapping intensity, used to control the distribution density of the data in the hyperbolic space. is the second parameter for adjusting the mapping intensity, used to control the structure of the data in the hyperbolic space.

[0026] And adopt a dynamic adjustment mechanism so that the embedding dimension can be adjusted according to the real-time characteristics of the data. The calculation method is expressed as:

[0027] d new =\(\delta\) c (\(X\) c , \(\theta\) d ) = \(d\) base + round(\(\theta\) d ·(var(\(X\) c ) + skew(\(X\) c )));

[0028] In the formula, \(\delta\) c () is the dynamic dimension adjustment function, \(\theta\) d is the adjustment coefficient used to control the sensitivity of the dimension adjustment, \(d\) base is the basic embedding dimension, var(\(X\) c ) and skew(\(X\) c ) respectively represent the variance and skewness of the data, used to reflect the distribution characteristics of the data.

[0029] In step S23, in the hyperbolic space, the generator creates new data samples by optimizing the loss function and generates fake data that is difficult for the discriminator to distinguish; the loss function of the generator is expressed as:

[0030]

[0031] Wherein, G c is the generator, D c is the discriminator, G c () is the generator function, represents the generated fake data, D c () is the discriminator function, z c is the input noise distribution, represents the loss function of the generator, φ c () is the composite function, θ c are the parameters of the generator, represents the regularization term of conditional dependence, C c is the conditional variable, that is, the data variable;

[0032] The calculation method of the regularization term of conditional dependence is expressed as:

[0033]

[0034] Wherein, λ r is the regularization coefficient, used to control the influence intensity of the regularization term, ∥∥ is the L2 norm, represents the expectation;

[0035] The composite function combines the generated data and parameters, and is used to increase the complexity and adjustability of the data generation process. The calculation method is expressed as:

[0036]

[0037] Wherein, Sig() is the Sigmoid activation function, μ c is the adjustment term that increases the sensitivity of the model to noise;

[0038] Step S24, improve the accuracy of the generator to identify true and false samples by training the discriminator; the loss function of the discriminator is expressed as:

[0039]

[0040] Wherein, represents the loss function of the discriminator, D c (X c ) represents the evaluation of the discriminator on the real data, X c represents the real sample, D c (G c (z c )) represents the evaluation of the discriminator on the generated data, ψ c () is the discriminant result processing function, γ cTo control the slope of the function, which is used to enhance the model's judgment ability for classifying true and false data;

[0041] The discrimination result processing function uses the slope parameter of the control function to adjust the discrimination sensitivity, and the calculation method is expressed as:

[0042]

[0043] In the formula, Dcs represents the discrimination result of the discriminator, that is, D c (G c (z c ));

[0044] Step S25, while generating data at each step, optimize the generated data at multiple scales through the feature pyramid structure to capture data features at different scales; that is, the optimized data is expressed as:

[0045]

[0046] In the formula, P c () is the feature pyramid function, λ k is the hierarchical weight, conv k () is the basic convolution operation of the k-th layer, conv' k () is the auxiliary convolution operation of this layer, which is used to capture additional detailed features, is the optimized generated data, Kcs is the number of convolutional layers, ξ k is the parameter that controls the contribution of the auxiliary convolution;

[0047] Step S26, during the iteration process, optimize the generator and the discriminator until convergence; the optimization method of the parameters is expressed as:

[0048]

[0049] In the formula, η c is the learning rate of the generative adversarial network, and are the gradients of the weight and the bias respectively, is the total loss function, including the sum of the losses of the generator and the discriminator, is the weight of the generative adversarial network before update, is the bias of the generative adversarial network before update, is the weight of the generative adversarial network after update, is the bias of the generative adversarial network after update;

[0050] Step S27, repeat and iterate the steps S21 - S26 until the preset iteration stop condition is met, which indicates that the training of the generative adversarial network algorithm based on hyperbolic space embedding is completed; wherein, the preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times;

[0051] In the step S5, the training process of the fractional-order neural network based on quantum state transfer is as follows:

[0052] Step S51, initialize the fractional-order neural network, define the neurons of the fractional-order neural network as quantum neurons, and set parameters for each quantum neuron in the fractional-order neural network; wherein, the parameters include the angles and phases of quantum gates, as well as the neural network weights and biases, and the initialization process is expressed as:

[0053]

[0054] In the formula, represents a uniform distribution, θ u and φ u are the angles and phases of the quantum gate respectively, uniformly distributed in [0, 2π], W u and b u are the weights and biases of the fractional-order neural network, following a normal distribution with a mean of 0 and a standard deviation of 1;

[0055] Step S52, in the forward propagation stage, the input feature dimension-reduced data is first encoded into qubits, and each qubit is operated through a quantum gate to achieve non-classical information processing; the calculation method of the conversion function of the quantum gate is expressed as:

[0056]

[0057] In the formula, G u () represents the quantum gate operation, i * is the imaginary unit, X u is the data input to the fractional-order neural network, that is, the feature dimension-reduced data, θ u and φ u regulate the angles and phases of the gate, so that X u experiences a conversion in the complex number domain to simulate the superposition effect of quantum states;

[0058] Step S53, adopt a fractional-order activation function, the order of which is dynamically adjusted according to the error in the training process, and is used to optimize the learning process and enhance the non-linear expression ability of the model;

[0059] Specifically, let the fractional-order derivative be where α uis a dynamically adjusted fractional-order parameter that enables the activation function to respond more flexibly to changes in the input. The resulting fractional-order derivative can not only reflect local changes but also capture the overall trend, making it suitable for processing data with highly variable non-linearity. The calculation method is expressed as:

[0060]

[0061] In the formula, Sig() is the Sigmoid activation function, represents the fractional-order derivative of f u , α u is the fractional-order parameter, m is the next integer after the integer part of α u , Γ is the gamma function used for normalization, and Sig (m) () is the m-th derivative of the Sigmoid activation function;

[0062] Step S54, use the gradient information of the quantum state to calculate the derivative of the loss function, and update the weights and biases of the network using the fractional-order derivative; the update increment of the fractional-order neural network parameters is expressed as:

[0063]

[0064] In the formula, ΔW u is the weight update increment of the fractional-order neural network, Δb u is the bias update increment of the fractional-order neural network, η u is the learning rate of the fractional-order neural network, is the loss function of the fractional-order neural network, and the loss function adopts the cross-entropy loss function;

[0065] During the calculation of the gradient of the loss function in backpropagation, the quantum gradient is enhanced through the quantum state interference effect. Then, the calculation method of the gradient of the loss function with respect to the weight is expressed as:

[0066]

[0067] In the formula, represents the quantum gradient enhancement term, which enhances or weakens the effect of the gradient through the phase interference of the quantum state. ⊙ represents the Hadamard product;

[0068] Step S55, repeat the forward propagation and backpropagation. In each iteration, update the weight and bias parameters of the fractional-order neural network; it is expressed as:

[0069]

[0070] In the formula, represents the weight of the fractional-order neural network at the t-th iteration, denotes the bias of the fractional-order neural network in the \(t\)-th iteration, denotes the weights of the fractional-order neural network in the \((t + 1)\)-th iteration, denotes the bias of the fractional-order neural network in the \((t + 1)\)-th iteration, and \(R()\) is the recoding function;

[0071] The recoding function is a composite function of adaptive learning rate adjustment and parameter range limitation, and its calculation method is expressed as:

[0072] \(R(p,\Delta p,\eta)=p+\eta\) 0 \(\cdot f(\Delta p,p,\eta)\);

[0073] In the formula, \(f()\) is the modulation function, which is used to adjust the learning rate \(\eta\) of the recoding function according to the current parameter \(p\), the update amount \(\Delta p\) and the feedback of the learning environment. \(\eta\) 0 is the initial learning rate of the recoding function;

[0074] Based on the gradient magnitude of the parameter and the previous update history, the calculation method of the modulation function is expressed as:

[0075]

[0076] In the formula, \(g()\) is the adjustment function for the learning rate, which increases or decreases the learning rate depending on the past update efficiency. \(\epsilon\) is a small constant to prevent division by zero;

[0077] After each iteration, according to the training accuracy \(P\) of the model u , dynamically adjust \(g(\eta)\), which is expressed as:

[0078] \(g(\eta)=\eta\cdot\exp(-\gamma\cdot(1 - P\) u ));

[0079] In the formula, \(\gamma\) is the sensitivity hyperparameter that controls the learning rate adjustment;

[0080] Step S56, repeat the iteration of steps S51 - S55 until the preset iteration stop condition is met, which means that the training of the fractional-order neural network based on quantum state transfer is completed; among them, the preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0081] Preferably, in step S1, the data acquisition channels include laboratory creep experiments, publicly available material databases, finite element simulation results, and microstructural characterization, and then data is obtained. Among them,

[0082] The laboratory creep experiment records the creep life and material behavior by applying different temperatures and stresses to the material samples under controlled conditions;

[0083] The publicly available material database is used to provide detailed data of historical creep experiments;

[0084] The finite element simulation results are used to generate data points that have not been actually tested under specific conditions;

[0085] Microstructural characterization is used to describe in detail the internal structure changes of materials through techniques such as scanning electron microscopy, electron backscatter diffraction, or transmission electron microscopy;

[0086] The attributes of the data, i.e., the data variables, include: stress-strain rate R a , temperature D a , test duration T a , pre-strain P a , material composition ratio M a , grain size G a , dislocation density B a , number of creep cavities H a , remaining life F a and microstructure evolution index L a .

[0087] Preferably, in the step S1, the structured digital formats include CSV and HDF5 formats.

[0088] Preferably, in the step S1, the categories of manual annotation include: remaining life ≤ 25%, remaining life 25% - 75%, and remaining life ≥ 75%, a total of three categories.

[0089] Preferably, in the step S3, the training process of the neural network algorithm based on dynamic population evolution optimization is as follows:

[0090] Step S31, according to the initialization method of the bionic algorithm, in the initialization stage, an initial population is generated; each individual represents a configuration of network weights, and the population size is set to N p , the weights and biases of the i-th individual are initialized, expressed as:

[0091]

[0092] In the formula, W pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual, represents the weight matrix of the i-th individual in the initial state, represents the bias of the i-th individual in the initial state, σ 2 represents the initialized variance, represents a normal distribution with a mean of 0 and a variance of σ 2 ; is a normal distribution;

[0093] Step S32: For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function. Specifically, for the i-th individual, calculate the loss on the training data set using its weights and biases, which is expressed as:

[0094]

[0095] where L pi is the loss of the neural network corresponding to the i-th individual, mps represents the number of samples input in the current batch, l() represents the composite loss function, fsig() represents the neural network model function, represents the feature of the j-th sample, represents the label of the j-th sample;

[0096] The composite loss function includes a regularization term, and its calculation method is expressed as:

[0097]

[0098] where MSE() is the mean squared error function, Reg(W pi ) is the regularization term, and λ ps is the regularization parameter;

[0099] The calculation method of the regularization term is expressed as:

[0100]

[0101] where W pi,k is the k-th weight of the neural network corresponding to the i-th individual;

[0102] Step S33: According to the fitness of the individuals, select the best-performing individuals from the current population for retention as candidate solutions for the next generation. Specifically, perform selection based on the fitness of the individuals, and excellent individuals have a higher probability of being selected. The calculation method of the probability of an individual being selected is expressed as:

[0103]

[0104] where P select (i) represents the probability of the i-th individual being selected, γ pse is the parameter controlling the selection pressure, and L pk is the loss of the neural network corresponding to the k-th individual;

[0105] Step S34: Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange some genes to generate new offspring. Specifically, the crossover operation randomly selects two individuals for gene exchange, which is expressed as:

[0106] W′ pi = α pcs W p1 +(1 - α pcs )W p2 ;

[0107] b' pi = α pcs b p1 +(1 - α pcs )b p2 ;

[0108] Wherein, α pcs is the crossover rate, W p1 is the weight of the neural network corresponding to the first selected individual, b p1 is the bias of the neural network corresponding to the first selected individual, W p2 is the weight of the neural network corresponding to the second selected individual, b p2 is the bias of the neural network corresponding to the second selected individual, W' pi is the weight of the neural network corresponding to the individual after the crossover operation, b' pi is the bias of the neural network corresponding to the individual after the crossover operation;

[0109] The mutation operation performs a small random perturbation on the weights of the newly generated individual, which is expressed as:

[0110]

[0111] Wherein, τ 2 represents the variance of the mutation, W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, b″ pi is the bias of the neural network corresponding to the individual after the mutation operation;

[0112] Step S35, repeat and iterate the said steps S31 - S34 until the preset iteration stop condition is satisfied, which means that the neural network algorithm training based on dynamic population evolution optimization is completed; wherein, the preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0113] Preferably, in the said step S4, the training process of the auto - encoder neural network based on feature refinement is as follows:

[0114] Step S41, the encoder adopts a multi - layer non - linear mapping structure to map the high - dimensional data to the initial low - dimensional feature space; assume the data input to the auto - encoder neural network is X r , then the mapping of the high - dimensional data to the initial low - dimensional feature space is expressed as:

[0115] Zr = Sig enc (W r X r + b r );

[0116] In the formula, Z r is the initial low-dimensional feature, W r is the weight matrix of the encoder, b r is the bias vector of the encoder, and Sig enc () is the multi-layer Sigmoid activation function of the encoder;

[0117] Step S42, after the low-dimensional feature is generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space; the calculation method of the initial weight matrix is expressed as:

[0118] A r = diag(α r );

[0119] In the formula, A r is the initial weight matrix, α r is the feature weight vector, and each element α r in α r,i is initialized to the same value, indicating that all features have the same importance in the initial stage; diag() is a function to extract the diagonal elements of the matrix;

[0120] The adjusted feature can be expressed as:

[0121] Z' r = A r Z r ;

[0122] In the formula, Z' r is the feature representation after the feature weight adjustment;

[0123] Step S43, the feature adjustment module recursively optimizes the initially generated low-dimensional features, that is, in each round of iteration, the feature adjustment module adjusts the weights of each feature according to the performance of the previous round of features, gradually enhancing the features with important influences and gradually weakening the redundant or noisy features; in the t-th round of iteration, the weight update rule is as follows:

[0124]

[0125] In the formula, represents the feature weight of the (t + 1)-th round of iteration, represents the feature weight of the t-th round of iteration, η r is the learning rate of the autoencoder neural network, and L r () is the loss function of the autoencoder neural network, Yr is the label data, indicating the gradient of the loss function with respect to the feature weights;

[0126] In step S44, the decoder remaps the low-dimensional features back to the high-dimensional space; the reconstruction process of the decoder is expressed as:

[0127] X' r = Sig dec (W' r Z' r + b' r );

[0128] where X' r is the reconstructed high-dimensional data, W' r is the weight matrix of the decoder, b' r is the bias vector of the decoder, and Sig dec () is the multi-layer Sigmoid activation function of the decoder;

[0129] In step S45, repeat the steps S41 - S44 until the preset stop iteration condition is satisfied, which means the training of the autoencoder neural network based on feature refinement is completed; among them, the preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0130] Preferably, in step S6, the collected original data is input into the trained feature extraction and feature dimensionality reduction model for feature processing, and the processed features are input into the classifier model for the training of the classifier, and then the classification result is obtained; among them, the classification categories include: remaining life ≤ 25%, remaining life 25% - 75%, and remaining life ≥ 75% three categories.

[0131] The beneficial effects of the present invention are:

[0132] 1) Data augmentation is achieved through the generative adversarial network based on hyperbolic space embedding, solving the problem of insufficient model generalization ability due to the scarcity of materials science data; the use of hyperbolic space embedding enables the generative adversarial network to capture the complex hierarchical structure and irregular geometric characteristics in the materials data more naturally; during the generation process, the generative adversarial network combines the multi-scale optimization of the feature pyramid, making the generated samples have high quality at different scales, further enhancing the richness and diversity of the data;

[0133] 2) In the feature extraction stage, a neural network algorithm based on dynamic population evolution optimization is adopted, and the evolution rules are automatically adjusted according to the data characteristics through a self-correction mechanism. This model can adjust the crossover and mutation operations according to the dynamic changes of the current data, ensuring the flexibility of feature extraction and solving the problem that traditional neural networks are vulnerable to gradient disappearance and gradient explosion. Through this adaptive mechanism, the model can more efficiently extract effective features from complex material data, enhancing the stability and efficiency of feature extraction.

[0134] 3) A self-encoding neural network based on feature refinement is used to perform dimensionality reduction on the extracted high-dimensional features. The autoencoder maps high-dimensional data to a low-dimensional feature space through multi-layer non-linear mapping. At the same time, a feature adjustment module is added, and through recursive feature adaptive optimization, the dimensionality-reduced feature space is dynamically adjusted, enabling important features to be strengthened and secondary or redundant features to be gradually weakened, ensuring that the features after dimensionality reduction can retain data information and have good simplicity.

[0135] 4) In the classification stage, a fractional-order neural network based on quantum state transfer is used. The quantum state transfer mechanism can make the information flow and processing process in the neural network more accurate. The fractional-order activation function dynamically adjusts the order of the activation function according to the error in the training process, enabling the model to more flexibly capture the non-linear changes of complex material data, thereby improving the accuracy of life prediction. Through the gradient enhancement of the quantum state, the model further optimizes the learning efficiency during the training process and improves the prediction ability of the creep life of materials. Description of the Drawings

[0136] Figure 1 is a flowchart of a data-driven creep life prediction method for chromium molybdenum vanadium steel according to the present invention;

[0137] Figure 2 is a training flowchart of a generative adversarial network algorithm based on hyperbolic space embedding in the present invention;

[0138] Figure 3 is a training flowchart of a neural network algorithm based on dynamic population evolution optimization in the present invention;

[0139] Figure 4 is a training flowchart of a self-encoding neural network based on feature refinement in the present invention;

[0140] Figure 5 is a training flowchart of a fractional-order neural network based on quantum state transfer in the present invention. Detailed Embodiments

[0141] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0142] The present invention will be further described below in conjunction with the accompanying drawings.

[0143] As Figure 1 shown, a data-driven creep life prediction method for chromium molybdenum vanadium steel includes:

[0144] Step S1, data acquisition, converting the collected data into a structured digital format and manually annotating the data.

[0145] In the embodiment, the channels for data acquisition include laboratory creep experiments, publicly available material databases, finite element simulation results, and microstructural characterization, so as to obtain data. Among them,

[0146] The laboratory creep experiment records the creep life and material behavior by applying different temperatures and stresses to the material samples under controlled conditions;

[0147] The publicly available material database is used to provide detailed data of historical creep experiments;

[0148] The finite element simulation results are used to generate data points that have not been actually tested under specific conditions;

[0149] The microstructural characterization is used to describe in detail the internal structure changes of the material through techniques such as scanning electron microscopy, electron backscatter diffraction, or transmission electron microscopy;

[0150] The attributes of the data, that is, the data variables include: stress strain rate R a , temperature D a , test duration T a , pre-strain P a , material composition ratio M a , grain size G a , dislocation density B a , number of creep cavities H a , remaining life F a and tissue evolution index L a; It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In practical applications, the attributes of data are usually more than 10 attributes, and the number of data attributes may reach dozens.

[0151] In the embodiment, the structured digital formats include CSV and HDF5 formats for easy processing and analysis.

[0152] In the embodiment, the categories of manual annotation include three categories: remaining life ≤ 25%, remaining life 25% - 75%, and remaining life ≥ 75%.

[0153] Step S2, use the generative adversarial network algorithm based on hyperbolic space embedding to generate samples, and then expand the data.

[0154] In the present invention, the acquisition, annotation, and preprocessing of training data are time-consuming and laborious, and insufficient training samples are likely to lead to poor generalization ability of the model and affect the accuracy of the model at the same time; the present invention uses the generative adversarial network algorithm based on hyperbolic space embedding to generate samples, and then realizes data expansion; on the basis of the traditional generative adversarial network, the present invention embeds the operations of the generator and the discriminator into the hyperbolic space. Since the hyperbolic space can more naturally represent the hierarchical structure and irregular geometric characteristics of data, it is particularly suitable for materials science data with complex physical properties and organizational structure characteristics.

[0155] As Figure 2 shown, the training process of the generative adversarial network algorithm based on hyperbolic space embedding is as follows:

[0156] Step S21, initialize the network parameters of the generator and the discriminator; the initialization method is expressed as:

[0157]

[0158] In the formula, W c represents the weight matrix of the generative adversarial network, and b c represents the bias vector of the generative adversarial network; ~ represents being subject to a specific distribution, represents the normal distribution, represents the normal distribution with a mean of 0 and a variance of 1;

[0159] Step S22, map the original data into the hyperbolic space, and retain the internal hierarchical structure and complex characteristics of the data through the mapping function; assume the hyperbolic space is Use the hyperbolic tangent function for nonlinear conversion, which is expressed as:

[0160]

[0161] In the formula, X c is the original data vector, Xci represents the i-th component of the original data vector, α c is an expansion parameter used to adjust the sensitivity of the embedding, d new is the embedding dimension, represents the data vector in the mapped hyperbolic space, α ci is the first parameter for adjusting the mapping strength, used to control the distribution density of the data in the hyperbolic space, β ci is the second parameter for adjusting the mapping strength, used to control the structure of the data in the hyperbolic space;

[0162] In one embodiment, in the field of materials science, the complexity of data can vary greatly due to different experimental conditions and sample differences. Therefore, a dynamic adjustment mechanism is adopted to enable the embedding dimension to be adjusted according to the real-time characteristics of the data, improving the flexibility and accuracy of data representation. The calculation method is expressed as:

[0163] d new =δ c (X c ,θ d )=d base +round(θ d ·(var(X c )+skew(X c )));

[0164] In the formula, δ c () is the dynamic dimension adjustment function, θ d is the adjustment coefficient used to control the sensitivity of dimension adjustment, d base is the basic embedding dimension, var(X c ) and skew(X c ) respectively represent the variance and skewness of the data, used to reflect the distribution characteristics of the data. Among them, d base is set to 3, θ d is set to 2;

[0165] Step S23. In the hyperbolic space, the generator creates new data samples by optimizing the loss function and generates fake data that is difficult for the discriminator to distinguish; the loss function of the generator is expressed as:

[0166]

[0167] In the formula, G c is the generator, D c is the discriminator, G c () is the generator function, represents the generated fake data, D c () is the discriminator function, z c is the input noise distribution, Denotes the loss function of the generator, φ c () is a composite function, θ c are the parameters of the generator, Denotes the regularization term of conditional dependence, C c is the conditional variable, i.e., the data variable;

[0168] The calculation method of the regularization term of conditional dependence is expressed as:

[0169]

[0170] In the formula, λ r is the regularization coefficient, set to 0.2, used to control the influence intensity of the regularization term, ∥∥ is the L2 norm, Denotes the expectation;

[0171] The composite function combines the generated data and parameters to increase the complexity and adjustability of the data generation process. The calculation method is expressed as:

[0172]

[0173] In the formula, Sig() is the Sigmoid activation function, μ c is the adjustment term to increase the sensitivity of the model to noise, set to 0.1;

[0174] Step S24, improve the accuracy of the generator in identifying true and false samples by training the discriminator; the loss function of the discriminator is expressed as:

[0175]

[0176] In the formula, Denotes the loss function of the discriminator, D c (X c ) represents the evaluation of the discriminator on the real data, X c represents the real sample, D c (G c (z c )) represents the evaluation of the discriminator on the generated data, ψ c () is the discriminant result processing function, γ c is the slope of the control function, used to enhance the judgment ability of the model in classifying true and false data;

[0177] The discriminant result processing function uses the slope parameter of the control function to adjust the discrimination sensitivity. The calculation method is expressed as:

[0178]

[0179] In the formula, Dcs represents the discriminant result of the discriminator, i.e., D c(G c (z c )),γ c is set to 1;

[0180] Step S25, while generating data at each step, optimize the generated data at multiple scales through the feature pyramid structure to ensure that the generated data has high quality at different scales and capture data features at different scales; that is, the optimized data is represented as:

[0181]

[0182] In the formula, P c () is the feature pyramid function, λ k is the hierarchical weight, conv k () is the basic convolution operation of the k-th layer, conv' k () is the auxiliary convolution operation of this layer, used to capture additional detailed features, is the optimized generated data, Kcs is the number of convolutional layers, ξ k is the parameter that controls the contribution of the auxiliary convolution;

[0183] Step S26, during the iteration process, optimize the generator and discriminator until convergence; the optimization method of the parameters is represented as:

[0184]

[0185] In the formula, η c is the learning rate of the generative adversarial network, set to 0.01, and are the gradients of the weight and bias respectively, is the total loss function, including the sum of the losses of the generator and discriminator, is the weight of the generative adversarial network before update, is the bias of the generative adversarial network before update, is the weight of the generative adversarial network after update, is the bias of the generative adversarial network after update;

[0186] Step S27, repeat the iterative steps S21 - S26 until the preset stop iteration condition is satisfied, which means that the training of the generative adversarial network algorithm based on hyperbolic space embedding is completed; among them, the preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0187] Step S3, use the neural network algorithm based on dynamic population evolution optimization as the feature extraction model to train the feature extraction model for the augmented data.

[0188] The present invention uses a 6-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as vanishing gradients, exploding gradients, or getting stuck in local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention uses a neural network algorithm based on dynamic population evolution optimization as the feature extraction model. In the traditional population evolution algorithm, the evolution of all individuals is based on fixed rules. The present invention uses a self-correction mechanism to automatically adjust the evolution rules according to the characteristics of the current training data, and adjusts the probabilities of crossover and mutation according to the change trend of the loss function, so that the algorithm can more flexibly adapt to different data distributions and improve the generalization ability and training efficiency of the model.

[0189] As Figure 3 shown, the training process of the neural network algorithm based on dynamic population evolution optimization is as follows:

[0190] Step S31, according to the initialization method of the bionic algorithm, in the initialization stage, an initial population is generated; each individual represents a configuration of network weights. Let the population size be N p , initialize the weights and biases for the i-th individual, expressed as:

[0191]

[0192] In the formula, W pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual, represents the weight matrix of the i-th individual in the initial state, represents the bias of the i-th individual in the initial state, σ 2 represents the initialized variance, set to 0.01, represents a normal distribution with a mean of 0 and a variance of σ 2 ; is a normal distribution;

[0193] Step S32, for each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function; specifically, for the i-th individual, calculate the loss on the training data set using its weights and biases, expressed as:

[0194]

[0195] In the formula, L pi is the loss of the neural network corresponding to the i-th individual, mps represents the number of samples input in the current batch, l() represents the composite loss function, and fsig() represents the neural network model function, Denote the features of the j-th sample, Denote the label of the j-th sample;

[0196] The composite loss function includes a regularization term, which can enhance the generalization ability of the model. The calculation method is expressed as:

[0197]

[0198] In the formula, MSE() is the mean squared error function, Reg(W pi ) is the regularization term, and λ ps is the regularization parameter;

[0199] The calculation method of the regularization term is expressed as:

[0200]

[0201] In the formula, W pi,k is the k-th weight of the neural network corresponding to the i-th individual;

[0202] Step S33: According to the fitness of the individuals, select the best-performing individual from the current population for retention as the candidate solution for the next generation; specifically, select based on the fitness of the individuals. Excellent individuals have a higher probability of being selected. The calculation method of the probability of an individual being selected is expressed as:

[0203]

[0204] In the formula, P select (i) represents the probability that the i-th individual is selected, γ pse is the parameter controlling the selection pressure, set to 2, and L pk is the loss of the neural network corresponding to the k-th individual;

[0205] Step S34: Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange part of their genes to generate new offspring, and the mutation operation randomly changes part of the genes in an individual to increase the diversity of the population; specifically, the crossover operation randomly selects two individuals for gene exchange, expressed as:

[0206] W′ pi =α pcs W p1 +(1 - α pcs )W p2 ;

[0207] b' pi =α pcs b p1 +(1 - α pcs )b p2 ;

[0208] In the formula, α pcs is the crossover rate, set to 0.3, W p1 is the weight of the neural network corresponding to the first selected individual, b p1 is the bias of the neural network corresponding to the first selected individual, W p2 is the weight of the neural network corresponding to the second selected individual, b p2 is the bias of the neural network corresponding to the second selected individual, W' pi is the weight of the neural network corresponding to the individual after the crossover operation, b' pi is the bias of the neural network corresponding to the individual after the crossover operation;

[0209] The mutation operation performs a small random perturbation on the weights of the newly generated individuals, expressed as:

[0210]

[0211] In the formula, τ 2 represents the variance of the mutation, set to 0.04, W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, b″ pi is the bias of the neural network corresponding to the individual after the mutation operation;

[0212] Step S35, repeat the iterative steps S31 - S34 until the preset iteration stop condition is met, which means the training of the neural network algorithm based on dynamic population evolution optimization is completed; among them, the preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0213] Step S4, use the self - encoding neural network based on feature refinement as the feature dimensionality reduction model to train the data after feature extraction with the feature dimensionality reduction model.

[0214] The present invention uses the self - encoding neural network based on feature refinement as the dimensionality reduction model. The self - encoding neural network based on feature refinement consists of three parts: an encoder, a decoder, and a feature adjustment module. The encoder is responsible for mapping the high - dimensional input data to the low - dimensional feature space, and the decoder is used to reconstruct the reduced - dimension features back to the original space to ensure the reversibility of the dimensionality reduction process; the feature adjustment module dynamically adjusts the reduced - dimension feature space through recursive feature self - adaptive optimization, so that important features can be strengthened and secondary features are gradually weakened, so that the reduced - dimension feature representation has good simplicity while retaining data information.

[0215] As Figure 4 shown, the training process of the self - encoding neural network based on feature refinement is as follows:

[0216] Step S41: The encoder adopts a multi-layer non-linear mapping structure to map high-dimensional data to an initial low-dimensional feature space. Let the data input to the auto-encoder neural network be X r , then the mapping of high-dimensional data to the initial low-dimensional feature space is expressed as:

[0217] Z r = Sig enc (W r X r + b r );

[0218] In the formula, Z r is the initial low-dimensional feature, W r is the weight matrix of the encoder, b r is the bias vector of the encoder, and Sig enc () is the multi-layer Sigmoid activation function of the encoder;

[0219] Step S42: After the low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. When the module is initialized, the same initial weight is assigned to all features for subsequent step-by-step adjustment according to the feature contribution. The calculation method of the initial weight matrix is expressed as:

[0220] A r = diag(α r );

[0221] In the formula, A r is the initial weight matrix, α r is the feature weight vector, and each element α r in α r,i is initialized to the same value, indicating that all features have the same importance in the initial stage; diag() is a function to extract the diagonal elements of the matrix;

[0222] The adjusted features can be expressed as:

[0223] Z′ r = A r Z r ;

[0224] In the formula, Z' r is the feature representation after feature weight adjustment;

[0225] Step S43: The feature adjustment module performs recursive optimization on the preliminarily generated low-dimensional features. That is, in each round of iteration, the feature adjustment module adjusts the weights of each feature according to the performance of the previous round of features, gradually enhancing the features with important influence and gradually weakening the redundant or noisy features. Let the weight update rule in the t-th round of iteration be as follows:

[0226]

[0227] In the formula, represents the feature weight of the (t + 1)-th round of iteration,

[0228] represents the feature weight of the t-th round of iteration, and η r is the learning rate of the autoencoder neural network, set to 0.01, and L r () is the loss function of the autoencoder neural network, and Y r is the label data, represents the gradient of the loss function with respect to the feature weight;

[0229] Step S44. To ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that no important information is lost during the dimensionality reduction process; the reconstruction process of the decoder is expressed as:

[0230] X' r = Sig dec (W' r Z' r + b' r );

[0231] In the formula, X' r is the reconstructed high-dimensional data, W' r is the weight matrix of the decoder, b' r is the bias vector of the decoder, and Sig dec () is the multi-layer Sigmoid activation function of the decoder;

[0232] Step S45. Repeat the iterative steps S41 - S44 until the preset stop iteration condition is satisfied, which means that the training of the autoencoder neural network based on feature refinement is completed; among them, the preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0233] Step S5. Use the fractional-order neural network based on quantum state transfer as the classifier model to train the classifier model for the dimensionality-reduced data.

[0234] Input the dimensionality-reduced data into the classifier for training the classifier model. The present invention uses the fractional-order neural network based on quantum state transfer as the classification algorithm. On the basis of the traditional fractional-order neural network, a quantum state transfer mechanism is adopted, enabling the network to more precisely control the flow and processing of information with the help of quantum bit states, thereby improving the accuracy of life prediction.

[0235] As Figure 5 shown, the training process of the fractional-order neural network based on quantum state transfer is as follows:

[0236] Step S51, initialize the fractional-order neural network, define the neurons of the fractional-order neural network as quantum neurons, and set parameters for each quantum neuron in the fractional-order neural network; among them, the parameters include the angle and phase of the quantum gate, as well as the neural network weights and biases. The initialization process is expressed as:

[0237]

[0238] In the formula, represents a uniform distribution, θ u and φ u are respectively the angle and phase of the quantum gate, uniformly distributed in [0, 2π], W u and b u are the weights and biases of the fractional-order neural network, following a normal distribution with a mean of 0 and a standard deviation of 1;

[0239] Step S52, in the forward propagation stage, the input data after feature dimensionality reduction is first encoded into qubits, and each qubit is operated through a quantum gate to achieve non-classical information processing; the calculation method of the transfer function of the quantum gate is expressed as:

[0240]

[0241] In the formula, G u () represents the quantum gate operation, i * is the imaginary unit, X u is the data input to the fractional-order neural network, that is, the data after feature dimensionality reduction, θ u and φ u regulate the angle and phase of the gate, so that X u experiences a transformation in the complex domain to simulate the superposition effect of quantum states;

[0242] Step S53, adopt a fractional-order activation function, the order of which is dynamically adjusted according to the error in the training process, to optimize the learning process and enhance the non-linear expression ability of the model;

[0243] Specifically, let the fractional derivative be D u αu , where α u is the dynamically adjusted fractional-order parameter, so that the activation function can respond more flexibly to the changes in the input. The obtained fractional derivative can not only reflect local changes, but also capture the overall trend, and is suitable for processing data with highly variable non-linearity. The calculation method is expressed as:

[0244]

[0245] In the formula, Sig() is the Sigmoid activation function, Denotes the fractional derivative of f u , where α u is the fractional-order parameter, m is the next integer of the integer part of α u , Γ is the gamma function for normalization, and Sig (m) () is the m-th derivative of the Sigmoid activation function;

[0246] In step S54, use the gradient information of the quantum state to calculate the derivative of the loss function, and update the weights and biases of the network using fractional derivatives, ensuring that the fractional neural network can effectively learn the patterns in the data while maintaining quantum characteristics; the update increment of the fractional neural network parameters is expressed as:

[0247]

[0248] In the formula, ΔW u is the weight update increment of the fractional neural network, Δb u is the bias update increment of the fractional neural network, η u is the learning rate of the fractional neural network, set to 0.1, is the loss function of the fractional neural network, and the loss function adopts the cross-entropy loss function;

[0249] In the process of calculating the gradient of the loss function in backpropagation, the quantum gradient is enhanced through the quantum state interference effect, and the calculation method of the gradient of the loss function with respect to the weight is expressed as:

[0250]

[0251] In the formula, represents the quantum gradient enhancement term, which enhances or weakens the effect of the gradient through the phase interference of the quantum state, and ⊙ represents the Hadamard product;

[0252] In step S55, repeat the forward propagation and backpropagation. In each iteration, update the weight and bias parameters of the fractional neural network; it is expressed as:

[0253]

[0254] In the formula, represents the weight of the fractional neural network in the t-th iteration, represents the bias of the fractional neural network in the t-th iteration, represents the weight of the fractional neural network in the (t + 1)-th iteration, represents the bias of the fractional neural network in the (t + 1)-th iteration, and R() is the recoding function;

[0255] The recoding function can achieve real-time adjustment of the parameter update strategy during training through the dynamic parameter recoding mechanism, improve the model's adaptability to material behavior changes under different experimental conditions, and thus achieve higher prediction accuracy and stability in materials science applications. The recoding function is a composite function of adaptive learning rate adjustment and parameter range limitation, and its calculation method is expressed as:

[0256] R(p,Δp,η)=p+η 0 ·f(Δp,p,η);

[0257] In the formula, f() is a modulation function used to adjust the learning rate η of the recoding function according to the current parameter p, the update amount Δp, and the feedback of the learning environment. η 0 is the initial learning rate of the recoding function, and η 0 is set to 0.01;

[0258] Based on the gradient magnitude of the parameter and the previous update history, the calculation method of the modulation function is expressed as:

[0259]

[0260] In the formula, g() is an adjustment function for the learning rate, increasing or decreasing the learning rate depending on the past update efficiency. ∈ is a small constant to prevent division by zero, and ∈ is set to 0.01;

[0261] After each iteration, according to the training accuracy P of the model u , dynamically adjust g(η), which is expressed as:

[0262] g(η)=η·exp(-γ·(1-P u ));

[0263] In the formula, γ is a sensitivity hyperparameter that controls the adjustment of the learning rate, and is set to 0.5;

[0264] Step S56, repeat the iterative steps S51 - S55 until the preset stop iteration condition is met, which means the training of the fractional-order neural network based on quantum state transfer is completed; among them, the preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0265] Step S6, use the fully trained feature extraction model, feature dimensionality reduction model, and classifier model to process and predict new samples, and then obtain the classification result of predicting the creep life of chromium molybdenum vanadium steel.

[0266] In the embodiment, the collected original data is input into the trained feature extraction and feature dimensionality reduction model for feature processing, and the processed features are input into the classifier model for classifier training, thereby obtaining the classification result; wherein, the classification categories include: remaining life ≤ 25%, remaining life 25% - 75%, and remaining life ≥ 75%.

[0267] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven creep life prediction method for chromium-molybdenum-vanadium steel, characterized in that: include: Step S1, data collection, converting the collected data into a structured digital format and manually annotating the data; Step S2, using a generative adversarial network algorithm based on hyperbolic space embedding to generate samples and then expand the data; Step S3, using a neural network algorithm based on dynamic population evolution optimization as a feature extraction model, and training the feature extraction model on the expanded data; Step S4, using an autoencoder neural network based on feature refinement as a feature dimension reduction model, and training the feature dimension reduction model on the data after feature extraction; Step S5, using a fractional-order neural network based on quantum state transfer as a classifier model, and training the classifier model on the reduced-dimensional data; Step S6, using the fully trained feature extraction model, feature dimension reduction model and classifier model to process and predict new samples, and then obtain a classification result for predicting the creep life of chromium-molybdenum-vanadium steel; In step S2, the training process of the generative adversarial network algorithm based on hyperbolic space embedding is as follows: Step S21, initialize the network parameters of the generator and the discriminator; the initialization method is expressed as: Where W c represents the weight matrix of the generative adversarial network, b c represents the bias vector of the generative adversarial network; ~ represents compliance with a specific distribution, represents a normal distribution, represents a normal distribution with a mean of 0 and a variance of 1; Step S22, mapping the original data into a hyperbolic space, retaining the intrinsic hierarchical structure and complex characteristics of the data through a mapping function; using a hyperbolic tangent function for nonlinear transformation, expressed as: Where, X c is the original data vector, represents the i-th component of the original data vector, α c is an extended parameter used to adjust the sensitivity of embedding, d new is the embedding dimension, represents the data vector in the mapped hyperbolic space, The first parameter for adjusting the mapping intensity is used to control the distribution density of data in the hyperbolic space. A second parameter for adjusting the mapping strength is used to control the structure of the data in the hyperbolic space; A dynamic adjustment mechanism is used so that the embedding dimension can be adjusted according to the real-time characteristics of the data. The calculation method is expressed as: d new =d c (X c ,the d )=d base +round(θ d ·(var(X c )+skew(X c ))); In the formula, δ c () is the dynamic dimension adjustment function, θ d is the adjustment coefficient, which is used to control the sensitivity of dimension adjustment, d base is the base embedding dimension, var(X c ) and skew(X c ) represent the variance and skewness of the data, respectively, and are used to reflect the distribution characteristics of the data; Step S23, in the hyperbolic space, the generator creates new data samples by optimizing the loss function, and generates false data that is difficult for the discriminator to distinguish through the generator; the loss function of the generator is expressed as: In the formula, G c is the generator, D c is the discriminator, G c () is the generator function, represents the generated fake data, D c () is the discriminator function, z c is the input noise distribution, represents the loss function of the generator, φ c () is a composite function, θ c are the parameters of the generator, represents the regularization term of conditional dependence, C c is the conditional variable, i.e., the data variable; The calculation method of the conditionally dependent regularization term is expressed as: In the formula, λ r is the regularization coefficient, which is used to control the influence of the regularization term, |||| is the L2 norm, express expectations; The composite function combines generated data and parameters to increase the complexity and adjustability of the data generation process. The calculation method is expressed as: Where Sig() is the sigmoid activation function, μ c It is a regulation term that increases the model’s sensitivity to noise; Step S24, improve the accuracy of the generator in identifying true and false samples by training the discriminator; the loss function of the discriminator is expressed as: In the formula, Denotes the loss function of the discriminator, D c (X c ) represents the evaluation of the discriminator on the original data vector, D c (G c (z c )) represents the evaluation of the discriminator on the generated data, ψ c () is the discriminant result processing function, γ c To control the slope of the function, it is used to enhance the model's ability to classify true and false data; The discrimination result processing function uses the slope parameter of the control function to adjust the discrimination sensitivity. The calculation method is expressed as: In the formula, Dcs represents the discriminant result, that is, D c (G c (z c )); Step S25, while generating data in each step, the generated data is optimized at multiple scales through the feature pyramid structure to capture data features at different scales; that is, the optimized data is represented as: Where P c () is the feature pyramid function, λ k is the layer weight, conv k () is the basic convolution operation of the kth layer, conv′ k () is the auxiliary convolution operation of this layer, which is used to capture additional detail features. is the optimized generated data, Kcs is the number of convolutional layers, ξ k Parameters to control the contribution of auxiliary convolution; Step S26, in the iterative process, optimize the generator and the discriminator until convergence; the parameter optimization method is expressed as: Where η c is the learning rate of the generative adversarial network, and are the gradients of weight and bias respectively, is the total loss function, including the sum of the losses of the generator and the discriminator, is the weight of the generative adversarial network before updating, is the bias of the generative adversarial network before updating, is the updated weight of the generative adversarial network, is the bias of the updated generative adversarial network; Step S27, repeatedly iterating the steps S21 to S26 until a preset stop iteration condition is met, indicating that the training of the generative adversarial network algorithm based on hyperbolic space embedding is completed; wherein the preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times; In step S5, the fractional-order neural network training process based on quantum state transfer is as follows: Step S51, initialize the fractional-order neural network, define the neurons of the fractional-order neural network as quantum neurons, and set parameters for each quantum neuron in the fractional-order neural network; wherein the parameters include the angle and phase of the quantum gate, and the neural network weight and bias, and the initialization process is expressed as: In the formula, represents uniform distribution, θ u and φ u are the angle and phase of the quantum gate, evenly distributed in [0, 2π], W u and b u are the weights and biases of the fractional-order neural network, following a normal distribution with a mean of 0 and a standard deviation of 1; Step S52, in the forward propagation stage, the input feature-reduced data is first encoded into quantum bits, and each quantum bit is operated through a quantum gate to achieve non-classical information processing; the calculation method of the quantum gate conversion function is expressed as: In the formula, G u () represents quantum gate operation, i * is the imaginary unit, X u is the data input to the fractional-order neural network, that is, the data after feature dimension reduction, θ u and φ u Adjust the angle and phase of the door so that X u undergoes transformations in the complex number domain to simulate superposition effects of quantum states; Step S53, using a fractional-order activation function, whose order is dynamically adjusted according to the error in the training process, to optimize the learning process and enhance the nonlinear expression ability of the model; Specifically, let the fractional derivative be where α u It is a dynamically adjusted fractional-order parameter that enables the activation function to respond more flexibly to changes in the input. The obtained fractional-order derivative can not only reflect local changes, but also capture the overall trend. It is suitable for processing highly nonlinear data. The calculation method is expressed as: Where Sig() is the Sigmoid activation function, Indicates f u The fractional derivative of u is the fractional order parameter, m is α u The next integer in the integer part of is the gamma function, used for normalization, Sig (m) () is the m-th order derivative of the Sigmoid activation function; Step S54, using the gradient information of the quantum state to calculate the derivative of the loss function, and using the fractional-order derivative to update the weights and biases of the network; the update increment of the fractional-order neural network parameters is expressed as: In the formula, ΔW u is the weight update increment of the fractional-order neural network, Δb u is the bias update increment of the fractional-order neural network, η u is the learning rate of the fractional-order neural network, is the loss function of the fractional-order neural network. Use cross entropy loss function; In the gradient calculation process of the loss function in back propagation, the quantum gradient is enhanced by the quantum state interference effect, and the gradient calculation method of the loss function with respect to the weight is expressed as: In the formula, represents the quantum gradient enhancement term, which enhances or weakens the gradient effect through the phase interference of quantum states, and ⊙ represents the Hadamard product; Step S55, repeating forward propagation and backward propagation, in each iteration, updating the weight and bias parameters of the fractional-order neural network; expressed as: In the formula, represents the weight of the fractional-order neural network at the tth iteration, represents the bias of the fractional-order neural network at the tth iteration, represents the weight of the fractional-order neural network at the t+1th iteration, represents the bias of the fractional-order neural network at the t+1th iteration, and R() is the recoding function; The recoding function is a composite function of adaptive learning rate adjustment and parameter range restriction, and the calculation method is expressed as: R (p, Δp, η) = p + η0·f (Δp, p, η); Where f() is a modulation function, which is used to adjust the learning rate η of the recoding function according to the current parameter p, the update amount Δp and the feedback of the learning environment, and η0 is the initial learning rate of the recoding function; Based on the gradient amplitude of the parameters and the previous update history, the modulation function is calculated as: In the formula, g() is the adjustment function of the learning rate, increasing or decreasing the learning rate, depending on the past update efficiency, ∈ is a small constant to prevent division by zero; After each iteration, according to the model's training accuracy P u , dynamically adjust g(η), expressed as: g(η)=η·exp(-γ·(1-P u )); Where γ is the sensitivity hyperparameter that controls the learning rate adjustment; Step S56, repeatedly iterate the steps S51 to S55 until a preset stop iteration condition is met, which means that the fractional-order neural network training based on quantum state transfer is completed; wherein the preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

2. The data-driven creep life prediction method for chromium-molybdenum-vanadium steel according to claim 1 is characterized in that: In step S1, the data collection channels include laboratory creep experiments, published material databases, finite element simulation results, and microstructure characterization, thereby obtaining data, wherein: Laboratory creep experiments record creep life and material behavior by subjecting material specimens to varying temperatures and stresses under controlled conditions; Publicly available material databases are used to provide detailed data from historical creep experiments; Finite element simulation results are used to generate data points that are not actually tested under specific conditions; Microstructural characterization is used to describe in detail the internal structural changes of materials by using scanning electron microscopy, electron backscatter diffraction or transmission electron microscopy techniques; The attributes of the data, that is, the data variables include: stress strain rate R a , Temperature D a , test duration T a , pre-strain variable P a , material composition ratio M a , grain size G a , dislocation density B a , creep void number H a , Residual life F a and organizational evolution index L a .

3. The data-driven creep life prediction method for chromium-molybdenum-vanadium steel according to claim 1 is characterized in that: In the step S1, the structured digital format includes CSV and HDF5 formats.

4. The data-driven creep life prediction method for chromium-molybdenum-vanadium steel according to claim 1 is characterized in that: In the step S1, the manually labeled categories include three categories: remaining life ≤ 25%, remaining life 25%-75% and remaining life ≥ 75%.

5. The method for predicting creep life of chromium-molybdenum-vanadium steel based on data drive according to claim 1, characterized in that: In step S3, the training process of the neural network algorithm based on dynamic population evolution optimization is as follows: Step S31, according to the bionic algorithm initialization method, in the initialization stage, an initial population is generated; each individual represents a configuration of network weights, and the population size is set to N p , initialize the weights and biases for the i-th individual, expressed as: In the formula, Wpi is the weight of the neural network corresponding to the i-th individual, bpi is the bias of the neural network corresponding to the i-th individual, represents the weight matrix of the i-th individual in the initial state, represents the bias of the i-th individual in the initial state, σ 2 represents the initialization variance, Indicates that the mean is 0 and the variance is σ 2 Normal distribution of is a normal distribution; Step S32, for each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function; specifically, for the i-th individual, use its weight and bias to calculate the loss on the training data set, expressed as: Where, L pi is the loss of the neural network corresponding to the i-th individual, mps represents the number of samples input in the current batch, l() represents the composite loss function, fsig() represents the neural network model function, represents the characteristics of the jth sample, represents the label of the jth sample; The composite loss function contains a regularization term, which is calculated as follows: Where MSE() is the mean square error function, Reg(W pi ) is the regularization term, λ ps is the regularization parameter; The calculation method of the regularization term is expressed as: Where W pi，k is the kth weight of the neural network corresponding to the i-th individual; Step S33, according to the fitness of the individual, select the best performing individual from the current population to be retained as the candidate solution for the next generation; specifically, based on the fitness of the individual, the excellent individual has a higher probability of being selected, and the calculation method of the probability of the individual being selected is expressed as: Where P select (i) represents the probability of the i-th individual being selected, γ pse is the parameter controlling the selection pressure, L pk is the loss of the neural network corresponding to the kth individual; Step S34, generating new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange some genes to generate new offspring. Specifically, the crossover operation randomly selects two individuals for gene exchange, which is expressed as: W′ pi =a pcs W p1 +(1-a pcs )W p2 ; b′ pi =a pcs b p1 +(1-a pcs )b p2 ; In the formula, α pcs is the crossover rate, W p1 is the weight of the neural network corresponding to the first individual selected, b p1 is the bias of the neural network corresponding to the first individual selected, W p2 is the weight of the neural network corresponding to the second individual selected, b p2 is the bias of the neural network corresponding to the second individual selected, W′ pi is the weight of the neural network corresponding to the individual after the crossover operation, b′ pi is the bias of the neural network corresponding to the individual after the crossover operation; The mutation operation performs a small random perturbation on the newly generated individual weights, expressed as: In the formula, τ 2 represents the variance of variation, W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, b″ pi is the bias of the neural network corresponding to the individual after the mutation operation; Step S35, repeatedly iterate the steps S31-S34 until the preset stop iteration condition is met, which means that the training of the neural network algorithm based on dynamic population evolution optimization is completed; wherein, the preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

6. The method for predicting creep life of chromium-molybdenum-vanadium steel based on data drive according to claim 1, characterized in that: In step S4, the training process of the autoencoder neural network based on feature refinement is as follows: Step S41: The encoder uses a multi-layer nonlinear mapping structure to map high-dimensional data to an initial low-dimensional feature space; assuming that the data input to the autoencoder neural network is X r , then mapping the high-dimensional data to the initial low-dimensional feature space is expressed as: Z r =Sig enc (W r X r +b r ); In the formula, Z r is the initial low-dimensional feature, W r is the weight matrix of the encoder, b r is the encoder bias vector, Sig enc () is the multi-layer Sigmoid activation function of the encoder; Step S42, after the low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space; the calculation method of the initial weight matrix is ​​expressed as: A r =diag(a r ); In the formula, A r is the initial weight matrix, α r is the feature weight vector, α r Each element α in r，i Initialized to the same value, indicating that all features have the same importance in the initial stage; diag() is a function for extracting the diagonal elements of the matrix; The adjusted features can be expressed as: Z′ r =A r From r ; In the formula, Z′ r is the feature representation after feature weight adjustment; In step S43, the feature adjustment module recursively optimizes the initially generated low-dimensional features, that is, in each round of iteration, the feature adjustment module adjusts the weights of each feature according to the performance of the features in the previous round, gradually enhances the features with important influence, and gradually weakens the redundant or noise features; suppose that in the tth round of iteration, the weight update rule is as follows: η r is the learning rate L of the autoencoder neural network r () is the loss function of the autoencoder neural network, Y r is the label data, Represents the gradient of the loss function with respect to the feature weight; Step S44, the decoder remaps the low-dimensional features back to the high-dimensional space; the reconstruction process of the decoder is expressed as: X′ r =Sig dec (W′ r Z′ r +b′ r ); In the formula, X′ r is the reconstructed high-dimensional data, W′ r is the weight matrix of the decoder, b′ r is the bias vector of the decoder, Sigde c () is the multi-layer Sigmoid activation function of the decoder; Step S45, repeatedly iterate the steps S41 to S44 until a preset stop iteration condition is met, indicating that the training of the autoencoder neural network based on feature refinement is completed; wherein the preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

7. The data-driven creep life prediction method for chromium-molybdenum-vanadium steel according to claim 1 is characterized in that: In step S6, the collected raw data is input into the trained feature extraction and feature dimension reduction model for feature processing, and the processed features are input into the classifier model for classifier training to obtain classification results; wherein the classification categories include: remaining life ≤ 25%, remaining life 25%-75% and remaining life ≥ 75%.

Images