A method for predicting creep properties of metal structural materials based on hard-constrained neural network model

Abstract

The present invention discloses a method for predicting the creep properties of metal structural materials based on a hard-constrained neural network model. The steps include: S1, establishing the constraints of the first derivative and the second derivative of the creep strength creep life curve; S2, establishing a hard-constrained neural network model in combination with the constraints, including the establishment of a network structure and the derivation of the network structure, a constraint loss function, etc.; S3, setting the structure, input and output parameters, training methods, etc. of the hard-constrained neural network model, fitting experimental data, obtaining fitting results and prediction results, and comparing them with experimental data; S4, finally analyzing the accuracy of the obtained results. The method of the present invention can be used to predict the long-term creep properties of most commercial austenitic stainless steels, nickel-based alloys, high-chromium steels, and new materials such as high-temperature alloys that are currently in the research and development stage, and the results are stable and reliable.

Description

Technical Field

[0001] The invention relates to the technical field of creep performance analysis of high-temperature pressure-bearing metal structural materials for power plants, and in particular to a method for predicting the creep performance of metal structural materials based on a hard-constrained neural network model. Background Art

[0002] High chromium steel, such as 9-12% Cr steel, austenitic stainless steel and nickel-based alloy, is an important metal structural material currently used in the construction of thermal power plants and nuclear power plants. These materials are often used in important components of power plants, such as nuclear reactors, heat exchangers, superheaters and reheaters. These components usually need to withstand high temperature and high pressure for a long time under service conditions. Creep performance is an important indicator of the long-term service performance of high-temperature metal structural materials. Accurate characterization of creep performance usually requires long-term experiments, which is not only costly, but also not conducive to the research and development and application of materials. Therefore, extrapolating and predicting the long-term creep performance of materials is a very important issue. At present, the creep models that are more widely used in the world include Norton formula, time-temperature parameter method, etc. These simple and easy-to-use models usually involve a large number of fitting parameters or adjustable parameters, but usually do not have actual physical meaning. Therefore, their reliability will be uncertain when extrapolating. In addition, recently there are also basic theoretical models based on material microstructure and defect evolution at home and abroad, which are mainly based on the dislocation evolution mechanism and the evolution of internal pores in the material. These models do not involve fitting parameters or adjustable parameters, and can reveal the creep failure mechanism of materials. However, because their theories and models are too complex and large, they have not yet been widely used.

[0003] In this case, neural networks can be used as a better auxiliary tool. Neural networks are also the core algorithm of the current development of artificial intelligence. Neural networks usually do not consider the failure mechanism of materials and establish a "parameter-free" model. A good fitting effect can be achieved through simple input and output. However, when using neural networks for fitting, it is easy to cause overfitting; and when extrapolating, it is very easy to produce results that violate the real physical meaning. Summary of the invention

[0004] In order to overcome the deficiencies in the above-mentioned prior art, the present invention provides a method for predicting the creep properties of metal structural materials based on a hard-constrained neural network model, which can solve the overfitting problem during prediction and obtain reasonable extrapolation results, thereby predicting the long-term creep properties of the material.

[0005] To achieve the above objectives, the present invention adopts the following technical solutions.

[0006] A method for predicting creep properties of metal structural materials based on a hard constraint neural network model, comprising:

[0007] S1, the establishment of the constraint conditions of the first derivative and the second derivative of the creep strength creep life curve. Here, the creep strength creep life curve refers to a curve whose abscissa is the creep rupture life or creep rupture time and whose ordinate is the creep rupture strength or creep rupture stress.

[0008] S2, combining the constraints of the first derivative and the second derivative, establishes a hard-constrained neural network model, that is, the establishment of the network structure and the derivation of the network structure, as well as the constrained loss function.

[0009] S3, setting the network structure, input parameters, output parameters, training methods, etc. of the hard-constrained neural network model, fitting the experimental data, obtaining the fitting results and prediction results, and comparing them with the experimental data.

[0010] S4, analyze the accuracy of the results obtained.

[0011] In step S1, the constraints of the first derivative and second derivative of the creep strength creep life curve are as follows:

[0012]

[0013] Where m is the negative value of the reciprocal of the first derivative of the creep strength-creep life curve, t R is the creep rupture time, σ is the creep rupture strength, and T is the absolute temperature.

[0014] In step S2, a hard constraint neural network model is established by combining the constraints of the first derivative and the second derivative. The hard constraints here are relative to the soft constraints. Soft constraints usually use the constraints to limit the final results of the neural network algorithm, and make restrictions and trade-offs; while hard constraints change the algorithm from both the process and the results by changing the code of the neural network algorithm, and finally obtain the results that meet the conditions.

[0015] In step S2, the constraints of the first derivative and the second derivative are combined to establish a hard-constrained neural network model, specifically: the constraints in step S1 are combined and encoded into the neural network algorithm to establish a hard-constrained neural network. The specific steps include:

[0016] S21, the first layer of the network structure; establish the input parameters and output parameters of the first layer of the network structure:

[0017]

[0018]

[0019] Where p is the input parameter, a 1 is the output parameter, p contains n input Each neuron has a corresponding weight W for each input parameter.1 , therefore, W 1 The size of the matrix is ​​n neuron ×n input , where n neuron is the number of neurons in the first layer. b 1 is the threshold of the neuron. The transition input function v 1 Input to transition function And get the output result a of the first layer 1 The superscript 1 indicates the first layer, and the subscripts k and i indicate the corresponding specific variables. is a scalar function. The output of the first layer will become the input variable of the next layer.

[0020] S22, the qth layer of the network structure; based on step S21, the qth layer of the network structure can be expressed as:

[0021]

[0022]

[0023] The superscript q indicates the qth layer, and the subscripts k and i indicate the corresponding specific variables. q-1 is the output of layer q-1, and is now the input variable of layer q. q is the output parameter. Similarly, each neuron for each input parameter has a corresponding weight W q and threshold b q . Transition input function v q Input to transition function And get the output result a of the qth layer q .

[0024] S23, derivation of the first derivative of the network structure, based on steps S21 and S22, the derivative of the input variable in the vector p is derived. The derivative of the transition function of the first layer can be expressed as:

[0025]

[0026] The derivative of the transition input function from the q-1th layer to the qth layer is expressed as:

[0027]

[0028] The derivative of the output result of the qth layer is expressed as:

[0029]

[0030] By combining the above formulas, we can directly obtain the relationship between the transition input function in a certain layer and the transition input function in the previous layer, that is, the first derivative:

[0031]

[0032] The superscript 1 indicates the first layer, the superscript q indicates the qth layer, and the subscripts k and i indicate the corresponding specific variables. Through the above formula, the transition input function and the derivative of the output result of a certain layer can be calculated from the corresponding data of the previous layer.

[0033] S24, derivation of the second derivative of the network structure; the derivation of the second derivative is similar to the first derivative. From the above formula, it can be obtained that the second derivative of the transition input function of the first layer will disappear.

[0034]

[0035] Taking the derivative of the first derivative in step S23, we can get

[0036]

[0037] Combining the above formula, we can get that the second derivative of the output result of the qth layer can be expressed as

[0038]

[0039] S25, constrained loss function; the impact of the first derivative and second derivative constraints on the error Δerr is added to the mean square error.

[0040]

[0041] Where Q is the last layer of the network structure, c1 and c2 are constants. is a logsig function. If the derivative is positive, this effect will be removed during the training of the network structure. Hard constraint neural network model structure parameter setting; When predicting creep rupture life, a two-layer network structure is used: the first layer is a hidden layer, containing 3-10 neurons, and using a logsig transition function; the second layer is an output layer, with only one output variable and one neuron, and using a linear transition function.

[0042] In step S2, the constraints of the first derivative and the second derivative are combined to establish a hard-constrained neural network model. Since there is currently no specific neural network code or software containing specific constraints, the present invention designs and proposes a new code to establish a hard-constrained neural network model by compiling the constraints into the neural network algorithm.

[0043] In step S2, a hard-constrained neural network model is established by combining the constraints of the first derivative and the second derivative. Different numbers of layers can be set here, and different numbers of neurons can also be set. Specifically, multiple hidden layers and the number of neurons in the hidden layers can be set.

[0044] In step S2, a hard constraint neural network model is established by combining the constraints of the first derivative and the second derivative. Here, the weight and threshold of the transition function are determined by the back propagation neural network algorithm.

[0045] In step S3, the structure, input and output parameters, training method, etc. of the hard constraint neural network model are set, the experimental data are fitted, the fitting results and prediction results are obtained, and compared with the experimental data. Specifically, it includes:

[0046] S31, setting the structure, input, output and other parameters of the hard-constrained neural network model; according to the hard-constrained neural network model established in step S2, setting different numbers of layers, numbers of neurons, input and output parameters, training methods, and training data classification.

[0047] S32, fitting the experimental data, obtaining fitting and extrapolation results, and comparing them with the experimental data. Through the hard-constrained neural network model established in step S2, using the network structure and related parameters set in step S31, combined with known experimental data, fitting and prediction are performed to obtain fitting and extrapolation structures, which are compared with experimental data and the results are plotted.

[0048] In step S4, the accuracy of the obtained results is analyzed. Specifically, the following are performed:

[0049] S41, verifying that the first derivative and second derivative of the result obtained by using the hard constraint neural network model should meet the requirements of the constraint conditions.

[0050] S42, verify that the model does not produce unreasonable prediction curves when extrapolated for a longer period of time.

[0051] S43, verify that there is no large deviation between the model results and the experimental results, and that 95% of the deviations between the model results and the experimental data are no greater than 2.5 times the standard deviation.

[0052] According to the analysis in step S4, it should be found that the hard-constrained neural network model can obtain relatively stable, reliable and accurate prediction results.

[0053] The advantages of the present invention are:

[0054] By adding hard constraints to the neural network algorithm, namely the constraints of the first derivative and second derivative of the creep strength creep life curve, the local overfitting caused by the neural network and the extrapolation results that do not conform to the real physical meaning are avoided. And combined with the verification method of the analysis results, the accuracy and reliability of the model are verified. Therefore, the high temperature creep performance of the material can be predicted more reliably.

[0055] The present invention will greatly promote the evaluation and prediction of long-term creep performance of high-temperature pressure-bearing metal structural materials used in thermal power, nuclear power and other fields, and has important significance for the research and development and application of high-temperature alloy materials. In addition, the present invention can also provide guiding reference and help for similar complex engineering problems in other fields.

[0056] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present invention, and together with the description, are used to explain the principles of the present invention. In the drawings:

[0058] Figure 1 The present invention is a flow chart of a method for predicting creep properties of metal structural materials based on a hard constraint neural network model.

[0059] Figure 2 This is a graph showing the creep strength curves of Sanicro 25 austenitic steel at six different temperatures of 600-800°C, which are fitted and predicted using a hard-constrained neural network model in an embodiment of the present invention.

[0060] Figure 3 : is the relationship between the m value and the stress in the embodiment of the present invention, and the m value is the negative value of the reciprocal of the first derivative.

[0061] Figure 4 : is the relationship between the secondary derivative and the creep life in the embodiment of the present invention.

[0062] Figure 5 This is a prediction result diagram of the result obtained by using the hard-constrained neural network model in an embodiment of the present invention extrapolated to 1,000,000 hours.

[0063] Figure 6 The figure shows the linear regression comparison analysis between the predicted results and the experimental data. DETAILED DESCRIPTION

[0064] Now, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments in the present invention, not all of the embodiments. For ordinary technicians in this field, other drawings can also be obtained based on these drawings without creative work. The embodiments can be implemented in a variety of ways and should not be understood as being limited to the examples described herein; on the contrary, these embodiments are provided to make the present invention more comprehensive and complete, and to fully convey the example embodiments to technicians in this field.

[0065] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, those skilled in the art will appreciate that the technical solution of the present invention can be practiced without one or more of the specific details, or other methods, components, steps, etc. may be adopted. In other cases, known methods, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0066] The present invention provides a method for predicting the creep performance of metal structural materials based on a hard-constrained neural network model, which can solve the overfitting problem during prediction and obtain reasonable extrapolation results, thereby predicting the long-term creep performance of materials. Figure 1 As shown, Figure 1 The schematic flow chart of the method is shown in FIG. As can be seen from the flow chart, the method mainly includes steps S1 to S5. The present invention will be further described below with reference to the accompanying drawings and the specific implementation of each step in the specific embodiment.

[0067] Example

[0068] In this embodiment, the long-term creep life is predicted by taking Sanicro 25 austenitic stainless steel as an example.

[0069] It should be stated that the creep test data of Sanicro 25 austenitic stainless steel comes from the online literature database. The experimental data can come from open source databases, published literature, actual production data of enterprises, experimental data from laboratories of universities, research institutes, etc. The source of the experimental data is not protected by this patent.

[0070] Combined with Figure 1 As shown, the operation steps of this embodiment are as follows:

[0071] S1, the establishment of the constraints of the first derivative and second derivative of the creep strength creep life curve. The constraints of the first derivative and second derivative are as follows:

[0072]

[0073] Where m is the negative value of the reciprocal of the first derivative of the creep curve, t R is the creep rupture time, σ is the creep rupture strength, and T is the absolute temperature.

[0074] In S2, the constraints of the first derivative and the second derivative are combined to establish a hard-constrained neural network model. That is, by compiling the constraints in S1 into the neural network algorithm, a hard-constrained neural network model is established. Specifically, it includes:

[0075] S21, the first layer of the network structure; establish the input parameters and output parameters of the first layer of the network structure:

[0076]

[0077]

[0078] Where p is the input parameter, a 1 For output parameters, p contains n input Each neuron has a corresponding weight W for each input parameter. 1 , therefore, W 1 The size of the matrix is ​​n neuron ×n input , where n neuron is the number of neurons in the first layer. b 1 is the threshold of the neuron. The transition input function v 1 Input to transition function And get the output result a of the first layer 1 The superscript 1 indicates the first layer, and the subscripts k and i indicate the corresponding specific variables. is a scalar function. The output of the first layer will become the input variable of the next layer.

[0079] S22, the qth layer of the network structure; based on step S21, the qth layer of the network structure can be expressed as:

[0080]

[0081]

[0082] The superscript q indicates the qth layer, and the subscripts k and i indicate the corresponding specific variables. q-1 is the output of layer q-1, and is now the input variable of layer q. q is the output parameter. Similarly, each neuron for each input parameter has a corresponding weight W q and threshold b q . Transition input function v q Input to transition function And get the output result a of the qth layer q .

[0083] S23, derivation of the first derivative of the network structure, based on steps S21 and S22, the derivative of the input variable in the vector p is derived. The derivative of the transition function of the first layer can be expressed as:

[0084]

[0085] The derivative of the transition input function from the q-1th layer to the qth layer is expressed as:

[0086]

[0087] The derivative of the output result of the qth layer is expressed as:

[0088]

[0089] By combining the above formulas, we can directly obtain the relationship between the transition input function in a certain layer and the transition input function in the previous layer, that is, the first derivative:

[0090]

[0091] The superscript 1 indicates the first layer, the superscript q indicates the qth layer, and the subscripts k and i indicate the corresponding specific variables. Through the above formula, the transition input function and the derivative of the output result of a certain layer can be calculated from the corresponding data of the previous layer.

[0092] S24, derivation of the second derivative of the network structure; the derivation of the second derivative is similar to the first derivative. From the above formula, it can be obtained that the second derivative of the transition input function of the first layer will disappear.

[0093]

[0094] Taking the derivative of the first derivative in step S23, we can get

[0095]

[0096] Combining the above formula, we can get that the second derivative of the output result of the qth layer can be expressed as

[0097]

[0098] S25, constrained loss function; the impact of the first derivative and second derivative constraints on the error Δerr is added to the mean square error.

[0099]

[0100] Where Q is the last layer of the network structure, c1 and c2 are constants. is a logsig function. If the derivative is positive, this effect will be removed during the training of the network structure. Hard constraint neural network model structure parameter setting; When predicting creep rupture life, a two-layer network structure is used: the first layer is a hidden layer, containing 3-10 neurons, and using a logsig transition function; the second layer is an output layer, with only one output variable and one neuron, and using a linear transition function.

[0101] In step S2, different numbers of layers and different numbers of neurons can be set. Specifically, multiple hidden layers and the number of neurons in the hidden layers can be set.

[0102] In step S2, the weight and threshold of the transition function are determined by a back propagation neural network algorithm.

[0103] S3, set the structure of the hard-constrained neural network model, input and output parameters, training methods, etc., fit the experimental data, obtain the fitting results and prediction results, and compare them with the experimental data. Specifically include:

[0104] S31, according to the hard-constrained neural network model established in step S2, the structure, input and output parameters of the hard-constrained neural network model are set; the training method, and the training data classification.

[0105] In this embodiment, the neural network structure is set to 2-N-1, where N is the number of neurons in the hidden layer. In this embodiment, the number of neurons in the hidden layer is set to 3-6, and the output layer is the creep life.

[0106] The input parameters are test stress and test temperature, and the output parameter is creep rupture time, which refers to the time it takes for a material to rupture under given test temperature and test stress conditions.

[0107] In this embodiment, all experimental data are divided into training data, verification data and test data according to a ratio of 70:15:15.

[0108] The training method adopted in this embodiment is the back propagation neural network algorithm.

[0109] S32, fitting the experimental data, obtaining fitting and extrapolation results, and comparing them with the experimental data. Through the hard-constrained neural network model established in step S2, using the network structure and related parameters set in step S31, combined with known experimental data, fitting and prediction are performed to obtain fitting and extrapolation structures, which are compared with experimental data and the results are plotted.

[0110] In this embodiment, the long-term creep life is predicted by taking Sanicro 25 austenitic stainless steel as an example.

[0111] The input parameter test temperatures are 600°C, 650°C, 700°C, 725°C, 750°C and 800°C. Using the hard constraint neural network structure obtained in step S2, all experimental data are fitted and the predicted creep curve is plotted, as shown in Figure 2 The figure shows the use of hard constraint neural network model to fit and predict the creep strength of Sanicro 25 austenitic steel at six different temperatures of 600-800℃ and compare it with the experimental data. In the figure, "exp" is the experimental data at each test temperature, "pred" is the predicted creep curve drawn by fitting the experimental data at each test temperature, and "epol" is the extrapolated result curve based on the predicted creep curve at each test temperature. It can be seen from the figure that a more reasonable fitting and prediction result is obtained, and the occurrence of overfitting and unreasonable extrapolation results is avoided.

[0112] S4, the accuracy of the results obtained by analysis. Specifically including:

[0113] S41, verify whether the first derivative and second derivative of the result obtained by using the hard constraint neural network model meet the requirements of the constraint conditions. Specifically, in this embodiment, the first derivative and second derivative of the result obtained by using the hard constraint neural network model are as follows: Figure 3 and Figure 4 shown. Figure 3 Shown is the relationship between the first derivative and stress. From right to left in the figure are the first derivatives of the "NN" curve at 600°C, 650°C, 700°C, 725°C, 750°C and 800°C. Figure 4 Shown is the relationship between the secondary derivative and the creep life. From bottom to top, the figure shows the secondary derivatives of the "NN" curve at 600°C, 650°C, 700°C, 725°C, 750°C and 800°C. Figure 3 and Figure 4 It can be seen that the results obtained by using the hard constraint neural network model to predict meet the requirements of the constraints.

[0114] S42, verifying that the model does not produce unreasonable prediction curves when extrapolated for a longer period of time. Specifically, in this embodiment, the prediction results obtained by using the hard constraint neural network model are extended to 1,000,000 hours. Figure 5 As shown in the figure, from top to bottom are the predicted and extrapolated curves of 600℃, 625℃, 650℃, 675℃, 700℃, 725℃, 750℃, 775℃, 800℃ and 825℃. Figure 5The results obtained using the hard-constrained neural network model are shown in the figure. The prediction results extended to 1,000,000 hours do not contain any bends or crosses that do not conform to the actual physical meaning.

[0115] S43, verify that there is no large deviation between the model results and the experimental results. Specifically, in this embodiment, the results obtained by the hard constraint neural network model are compared with the results obtained by the experimental data by linear regression analysis. The results are as follows: Figure 6 shown. Figure 6 The figure shows the linear regression comparison analysis between the predicted results and the experimental data. It can be seen from the figure that there is no significant deviation between the experimental data and the model results.

[0116] According to the analysis in step S4, it can be concluded that the hard-constrained neural network model established in this embodiment can obtain relatively stable, reliable and accurate prediction results.

[0117] The present invention has a significant effect on the evaluation, analysis and prediction of the long-term service performance of high-temperature metal structural materials in service. It has important guiding significance for the research of new materials currently in the development stage. It can predict the long-term creep performance of materials by combining short-term experimental data, saving a lot of time and economic costs. It also greatly promotes the research on extrapolation and prediction technology under long-term service conditions of high-temperature alloy materials.

[0118] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the disclosed invention herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0119] It should be understood that the present disclosure is not limited to the exact construction and results described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for predicting creep properties of metal structural materials based on a hard constraint neural network model, characterized in that: The following steps are involved: S1, establishment of constraint conditions for the first derivative and second derivative of the creep strength creep life curve; the creep strength creep life curve refers to a curve whose abscissa is creep rupture life or creep rupture time and whose ordinate is creep rupture strength or creep rupture stress; S2, combining the constraints of the first derivative and the second derivative of step S1, establishing a hard-constrained neural network model, including at least the establishment of a network structure, the derivation of the network structure, and a constrained loss function; wherein the hard constraint means: encoding the constraint into the neural network algorithm, changing the neural network algorithm in terms of both the process and the result by changing the code of the neural network algorithm, and finally obtaining a result that meets the conditions; S3, setting the network structure, input parameters, output parameters, and training method of the hard-constrained neural network model, fitting the experimental data, obtaining fitting results and prediction results, and comparing them with the experimental data; S4, accuracy of the results obtained by analysis; Wherein, in step S1, the constraints of the first derivative and the second derivative are as follows: Where m is the negative value of the reciprocal of the first derivative of the creep strength-creep life curve, t R is the creep rupture time, σ is the creep rupture strength, and T is the absolute temperature; Among them, in step S2, the step of establishing a hard constraint neural network model by combining the constraints of the first derivative and the second derivative of step S1 includes: S21, the first layer of the network structure; establish the input parameters and output parameters of the first layer of the network structure: Where p is the input parameter, a 1 is the output parameter, p contains n input input parameters; each neuron for each input parameter has a corresponding weight W 1 , therefore, W 1 The size of the matrix is ​​n neuron ×n input , where n neuron is the number of neurons in the first layer; b 1 is the threshold of the neuron; the transition input function v 1 Input to transition function And get the output result a of the first layer 1 ; The superscript 1 indicates the first layer, and the subscripts k and i indicate the corresponding specific variables; is a scalar function; the output of the first layer will become the input variable of the next layer; S22, the qth layer of the network structure; based on step S21, the network structure of the qth layer is expressed as: The superscript q indicates the qth layer, and the subscripts k and i indicate the corresponding specific variables; a q-1 is the output of layer q-1, and is now the input variable of layer q. q is the output parameter; similarly, each neuron of each input parameter has a corresponding weight W q and threshold b q ; Transition input function v q Input to transition function And get the output result a of the qth layer q ; S23, derivation of the first derivative of the network structure, based on steps S21 and S22, the derivative of the input variable in the derivation vector p; the derivative of the transition function of the first layer is expressed as: The derivative of the transition input function from the q-1th layer to the qth layer is expressed as: The derivative of the output result of the qth layer is expressed as: By combining the above formulas, we can directly obtain the relationship between the transition input function in a certain layer and the transition input function in the previous layer, that is, the first derivative: The superscript 1 indicates the first layer, the superscript q indicates the qth layer, and the subscripts k and i indicate the corresponding specific variables. Through the above formula, the transition input function and the derivative of the output result of the next layer can be calculated from the corresponding data of the previous layer. S24, derivation of the second derivative of the network structure; the second derivative of the transition input function of the first layer will disappear: Taking the first derivative of the transition input function of the qth layer in step S23, we can get The second derivative of the output result of the qth layer in step S23 is expressed as S25, constrained loss function: the effect of the first derivative and second derivative constraints on the error Δerr is added to the mean square error; Where Q is the last layer of the network structure, c1 and c2 are constants; is a logsig function; if the derivative is positive, this effect will be removed during the training of the network structure; The network structure setting of the hard-constrained neural network model: When predicting the creep rupture life, a two-layer network structure is used: the first layer is a hidden layer, which contains 3-10 neurons and uses a logsig transition function; the second layer is an output layer, which has only one output variable and one neuron, and uses a linear transition function.

2. The method for predicting creep properties of metal structural materials based on a hard constraint neural network model according to claim 1, characterized in that: The step S2 also includes setting different numbers of layers and different numbers of neurons, that is, setting multiple hidden layers and the number of neurons in each hidden layer.

3. The method for predicting creep properties of metal structural materials based on a hard constraint neural network model according to claim 2, characterized in that: In step S2, the weight and threshold of the transition function are determined by a back propagation neural network algorithm.

4. The method for predicting creep properties of metal structural materials based on a hard constraint neural network model according to claim 1, characterized in that: The step S3 specifically comprises the following steps: S31, setting the network structure, input parameters, output parameters, training method, and training data classification of the hard-constrained neural network model; S32, fitting the experimental data, obtaining fitting and extrapolation results, and comparing them with the experimental data: by building the hard-constrained neural network model established in step S2, using the network structure and parameters set in step S31, and combining known experimental data for fitting and prediction, obtaining fitting and extrapolation results, and comparing them with the experimental data.

5. The method for predicting creep properties of metal structural materials based on a hard constraint neural network model according to claim 4, characterized in that: The network structure set in step S31 is 2-N-1, where N is the number of neurons in the hidden layer, and the number of neurons in the hidden layer is set in the range of 3-6; The input parameters set in step S31 are test stress and test temperature, and the output parameter is creep rupture time; The creep rupture time refers to the time it takes for a material to rupture under given test temperature and test stress conditions; The training data classification set in step S31 is: dividing the experimental data into three groups of training data, verification data and test data according to a ratio of 70:15:15; The training method set in step S31 is a back propagation neural network algorithm.

6. The method for predicting creep properties of metal structural materials based on a hard constraint neural network model according to claim 1, characterized in that: Step S4 specifically includes the following steps: S41, verifying whether the first derivative and the second derivative of the fitting and extrapolation results in S3 meet the requirements of the constraint conditions; S42, verifying whether an unreasonable prediction curve will appear when the hard constraint neural network model is extrapolated for long-term creep performance; S43, verifying whether there is a large deviation between the prediction results of the hard-constrained neural network model and the experimental data, that is, satisfying that the deviation between 95% of the model results and the experimental data is no more than 2.5 times the standard deviation.

7. The method for predicting creep properties of metal structural materials based on a hard constraint neural network model according to claim 6, characterized in that: The S43 verifies whether there is a deviation between the prediction results of the hard constraint neural network model and the experimental data, and the verification method adopted is linear regression comparative analysis.

Images