Method and system for predicting fatigue life of surface strengthening material
By training the fatigue life of surface-strengthening materials using a neural network model and utilizing features such as residual stress and hardness, the problem of inaccurate fatigue life prediction in existing technologies has been solved, achieving efficient and low-cost fatigue life prediction.
Patent Information
- Application Number
- CN202411121864.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2024-11-26
AI Technical Summary
Existing technologies struggle to accurately predict the fatigue life of surface-strengthened metallic materials, impacting guidance for fatigue-resistant design.
By using a neural network model, multiple sets of fatigue data of surface-strengthening materials are obtained, and features such as residual stress, hardness, and surface roughness are used to train the neural network model to predict fatigue life.
It achieves high-precision, low-cost fatigue life prediction, guiding fatigue-resistant design.
Smart Images

Figure CN119026473A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fatigue life prediction of surface reinforcement materials, and more specifically to a method and system for predicting fatigue life of surface reinforcement materials. Background Art
[0002] In major engineering fields such as aerospace and marine equipment, there are higher requirements for the fatigue life of key components. Surface strengthening technology has been widely used to improve the fatigue life of materials. The gradient structure metal materials after surface strengthening have significantly improved their fatigue life through the coupling of multiple mechanisms. It has always been a challenge to accurately predict the fatigue life of surface-strengthened metal materials, explore the key factors to improve fatigue life, and guide anti-fatigue design accordingly. Summary of the invention
[0003] The purpose of the present invention is to provide a method for predicting the fatigue life of a surface reinforcement material, so as to accurately predict the fatigue life of the surface reinforcement material and provide guidance for fatigue resistance design.
[0004] Based on the above purpose, the present invention provides a method for predicting the fatigue life of a surface reinforcement material, comprising:
[0005] Obtain multiple groups of sample fatigue data of surface-strengthened materials and the fatigue life corresponding to each group of sample fatigue data, wherein each group of sample fatigue data includes the maximum stress, frequency, stress ratio of the fatigue test, as well as the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, the surface roughness, the residual stress at multiple representative depths of residual stress, and the hardness at multiple representative depths of hardness;
[0006] Get the best neural network model;
[0007] Taking the sample fatigue data as input and the fatigue life corresponding to the sample fatigue data as output, the optimal neural network model is trained using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model;
[0008] Fatigue data of the surface reinforcement material to be predicted is obtained, and the fatigue data of the surface reinforcement material to be predicted is input into a trained neural network model to obtain the fatigue life of the surface reinforcement material to be predicted.
[0009] Furthermore, the optimal neural network model includes a first fully connected layer, a second fully connected layer and a third fully connected layer, the first fully connected layer is used to output a comprehensive evaluation score of residual stress according to the residual stress at multiple representative depths of residual stress, the second fully connected layer is used to output a comprehensive evaluation score of hardness according to the hardness at multiple representative depths of hardness, and the third fully connected layer is used to output fatigue life according to the comprehensive evaluation score of residual stress, the comprehensive evaluation score of hardness, maximum stress, frequency, stress ratio of fatigue test, tensile strength of material, depth of plastic deformation layer after surface strengthening, and surface roughness.
[0010] Furthermore, the optimal neural network model is obtained, specifically including:
[0011] Obtain multiple different neural network models, wherein the number of neurons in each layer of each neural network model is different;
[0012] For each neural network model, a five-fold cross validation is performed on the neural network model using each group of sample fatigue data and its corresponding fatigue life to obtain the cross validation error of the neural network model;
[0013] The neural network model with the smallest cross-validation error is selected as the best neural network model.
[0014] Furthermore, the optimal neural network model is trained using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model, which specifically includes:
[0015] Select a part of each group of sample fatigue data and its corresponding fatigue life as a training set;
[0016] For each group of sample fatigue data in the training set, the sample fatigue data is input into the optimal neural network model to obtain the predicted value of fatigue life output by the optimal neural network model; the loss is calculated according to the predicted value of fatigue life and the fatigue life corresponding to the group of sample fatigue data; the gradient is calculated according to the loss; the gradient is back-propagated to each layer of the optimal neural network model to update the parameters of the optimal neural network model;
[0017] After the preset conditions are met, the update is stopped to obtain a trained neural network model.
[0018] Furthermore, the representative depth of residual stress is a depth within a depth range where the trend of residual stress variation with depth has obvious fluctuations, and the representative depth of hardness is a depth within a depth range where the trend of hardness variation with depth has obvious fluctuations.
[0019] Another aspect of the present invention provides a system for predicting fatigue life of a surface reinforcement material, comprising:
[0020] The first acquisition module is used to obtain multiple groups of sample fatigue data of surface-strengthened materials and the fatigue life corresponding to each group of sample fatigue data, wherein each group of sample fatigue data includes the maximum stress, frequency, stress ratio of the fatigue test, as well as the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, the surface roughness, the residual stress at multiple representative depths of residual stress, and the hardness at multiple representative depths of hardness;
[0021] The second acquisition module is used to obtain the best neural network model;
[0022] A training module, which is used to take the sample fatigue data as input and the fatigue life corresponding to the sample fatigue data as output, and train the optimal neural network model using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model;
[0023] Fatigue data of the surface reinforcement material to be predicted is obtained, and the fatigue data of the surface reinforcement material to be predicted is input into a trained neural network model to obtain the fatigue life of the surface reinforcement material to be predicted.
[0024] Furthermore, the optimal neural network model includes a first fully connected layer, a second fully connected layer and a third fully connected layer, the first fully connected layer is used to output a comprehensive evaluation score of residual stress according to the residual stress at multiple representative depths of residual stress, the second fully connected layer is used to output a comprehensive evaluation score of hardness according to the hardness at multiple representative depths of hardness, and the third fully connected layer is used to output fatigue life according to the comprehensive evaluation score of residual stress, the comprehensive evaluation score of hardness, maximum stress, frequency, stress ratio of fatigue test, tensile strength of material, depth of plastic deformation layer after surface strengthening, and surface roughness.
[0025] Furthermore, the optimal neural network model is obtained, specifically including:
[0026] Obtain multiple different neural network models, wherein the number of neurons in each layer of each neural network model is different;
[0027] For each neural network model, a five-fold cross validation is performed on the neural network model using each group of sample fatigue data and its corresponding fatigue life to obtain the cross validation error of the neural network model;
[0028] The neural network model with the smallest cross-validation error is selected as the best neural network model.
[0029] Furthermore, the optimal neural network model is trained using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model, which specifically includes:
[0030] Select a part of each group of sample fatigue data and its corresponding fatigue life as a training set;
[0031] For each group of sample fatigue data in the training set, the sample fatigue data is input into the optimal neural network model to obtain the predicted value of fatigue life output by the optimal neural network model; the loss is calculated according to the predicted value of fatigue life and the fatigue life corresponding to the group of sample fatigue data; the gradient is calculated according to the loss; the gradient is back-propagated to each layer of the optimal neural network model to update the parameters of the optimal neural network model;
[0032] After the preset conditions are met, the update is stopped to obtain a trained neural network model.
[0033] Furthermore, the representative depth of residual stress is a depth within a depth range where the trend of residual stress variation with depth has obvious fluctuations, and the representative depth of hardness is a depth within a depth range where the trend of hardness variation with depth has obvious fluctuations.
[0034] The method and system for predicting the fatigue life of surface reinforced materials of the present invention use residual stress, hardness, surface roughness, depth of plastic deformation layer, ultimate tensile strength of material, maximum stress of fatigue test, stress ratio and frequency as inputs and fatigue life as output to train a neural network model, so as to characterize the highly nonlinear relationship between fatigue life and the above characteristics through the trained neural network model, and then use the trained neural network model to predict the fatigue life of the surface reinforced material, which has the advantages of small error, low cost and high efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A flowchart of a method for predicting fatigue life of a surface reinforcement material according to an embodiment of the present invention;
[0036] Figure 2 A schematic diagram of the structure of a neural network model according to an embodiment of the present invention;
[0037] Figure 3 A structural block diagram of a system for predicting fatigue life of a surface reinforcement material according to another embodiment of the present invention;
[0038] Figure 4 A schematic diagram of a cross-validation result of a neural network for predicting TC4 according to an embodiment of the present invention;
[0039] Figure 5 A schematic diagram of loss curves of the best neural network model for predicting TC4 on a training set and a test set according to an embodiment of the present invention;
[0040] Figure 6It is a schematic diagram of regression results of the best neural network model for predicting TC4 on the training set and the test set according to an embodiment of the present invention. DETAILED DESCRIPTION
[0041] The preferred embodiments of the present invention are given below in conjunction with the accompanying drawings and described in detail.
[0042] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting fatigue life of a surface reinforcement material, which comprises the following steps:
[0043] S100: Obtain multiple groups of sample fatigue data of surface strengthened materials and the fatigue life corresponding to each group of sample fatigue data, wherein each group of sample fatigue data includes the maximum stress, frequency, stress ratio of the fatigue test, as well as the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, the surface roughness, the residual stress at multiple representative depths of residual stress, and the hardness at multiple representative depths of hardness.
[0044] The surface strengthening material can be any suitable material, such as titanium alloy, aluminum alloy, etc. Sample fatigue data is a parameter that has a greater impact on fatigue life. Sample fatigue data of surface strengthening materials can be obtained from relevant literature or by measuring the surface strengthening materials. For example, surface roughness, material tensile strength, depth of plastic deformation layer, residual stress at a representative depth of residual stress, and hardness at a representative depth of hardness can be obtained by measuring the surface strengthening material, and fatigue life can be obtained by performing fatigue tests on the surface strengthening material, and the maximum stress, frequency, and stress ratio of the fatigue test can also be obtained.
[0045] In some embodiments, the residual stress representative depth refers to the depth within the depth range where the residual stress variation trend with depth has obvious fluctuations, and the hardness representative depth refers to the depth within the depth range where the hardness variation trend with depth has obvious fluctuations. In order to fully capture the fluctuations of the residual stress and hardness variation trends with depth, more residual stress representative depths and hardness representative depths can be selected within the range where the residual stress and hardness variation trends with depth are more significant, and fewer residual stress representative depths can be selected within the range where the residual stress and hardness variation trends with depth are not so significant, and no residual stress representative depth is selected within the range where the residual stress and hardness variation trends with depth tend to be stable. Exemplarily, the residual stress of the Ti-6Al-4V alloy (TC4) after surface strengthening varies significantly in the depth range of 0 to 100 μm, and reaches a maximum value near 100 μm, then the fluctuation of the residual stress disappears beyond the depth of 600 μm and tends to the residual stress of the matrix, so ten depths of 0 μm, 10 μm, 25 μm, 50 μm, 100 μm, 200 μm, 300 μm, 400 μm, 500 μm and 600 μm can be selected as the representative depths of each residual stress; using the same method, ten depths of 0 μm, 5 μm, 10 μm, 15 μm, 30 μm, 50 μm, 80 μm, 150 μm, 300 μm and 500 μm are selected as the representative depths of each hardness, so as to accurately predict the fatigue life using the residual stress at the ten representative depths of residual stress and the hardness at the ten representative depths of hardness.
[0046] It is understandable that although ten representative depths of residual stress and ten representative depths of hardness are selected in the above example, those skilled in the art may also select other numbers of representative depths of residual stress and representative depths of hardness as needed. The more representative depths there are, the higher the prediction accuracy of fatigue life will be. If the number of representative depths is too small, the prediction result may be inaccurate.
[0047] In some embodiments, when the sample fatigue data is obtained from the literature, the literature may not give the residual stress value or hardness value at the required representative depth. In this case, the value at the required depth can be obtained by interpolation or other methods based on the value at the depth given in the literature. For example, the literature may only give the residual stress at depths of 0μm, 50μm, 100μm, 300μm, and 500μm. For the residual stress at 10μm, 25μm, 200μm, and 400μm, it can be obtained by interpolation calculation based on the residual stress at 0μm, 50μm, 100μm, 300μm, and 500μm. For the residual stress at 600μm, the residual stress at 500μm can be directly obtained as the residual stress at 600μm. The required representative depth of hardness can also be obtained in the same way, which will not be repeated here.
[0048] In some embodiments, multiple groups of sample fatigue data may also include fatigue data of materials that have not undergone surface strengthening, in which the depth of the plastic deformation layer is 0, so that the model can more clearly understand the impact of surface integrity changes caused by surface strengthening on fatigue life.
[0049] S200: Get the best neural network model.
[0050] like Figure 2 As shown, in some embodiments, the neural network model includes a first fully connected layer 10, a second fully connected layer 20 and a third fully connected layer 30. The first fully connected layer 10 is used to process the residual stress at multiple representative depths of residual stress input thereto, and output a comprehensive evaluation score of the residual stress. The second fully connected layer 20 is used to process the hardness at multiple representative depths of hardness input thereto, and output a comprehensive evaluation score of the hardness. The third fully connected layer 30 is used to process the comprehensive evaluation score of the residual stress, the comprehensive evaluation score of the hardness and the remaining features (i.e., maximum stress, frequency, stress ratio, material tensile strength, depth of the plastic deformation layer after surface strengthening, surface roughness) input thereto, and output fatigue life.
[0051] In some embodiments, step S200 specifically includes the following steps:
[0052] S210: Acquire multiple different neural network models, wherein the number of neurons in each layer of each neural network model is different;
[0053] S220: for each neural network model, using each group of sample fatigue data and its corresponding fatigue life, perform a five-fold cross validation on the neural network model to obtain a cross validation error of the neural network model;
[0054] S230: Select the neural network model with the smallest cross-validation error as the best neural network model.
[0055] In an exemplary embodiment, the number of neurons in the first fully connected layer 10 and the second fully connected layer 20 is set to be the same, which can be 5 or 8, and the number of neurons in the third fully connected layer 30 can be 5, 8, 10, 12, 14, and thus 10 neural network models can be obtained by combining them, namely, 5 neural network models when the number of neurons in the first and second fully connected layers is 5, the number of neurons in the third fully connected layer 30 is 5, 8, 10, 12, 14 respectively, and 5 neural network models when the number of neurons in the first and second fully connected layers is 8, the number of neurons in the third fully connected layer 30 is 5, 8, 10, 12, 14 respectively. Through the five-fold cross-validation error, it is obtained that the neural network model error is the smallest when the number of neurons in the first and second fully connected layers is 8 and the number of neurons in the third fully connected layer 30 is 12, which is the best neural network model.
[0056] S300: taking sample fatigue data as input and fatigue life corresponding to the sample fatigue data as output, and using each group of sample fatigue data and its corresponding fatigue life to train the best neural network model to obtain a trained neural network model.
[0057] When training the best neural network model, the neural network model will forward propagate each set of sample fatigue data, and its forward propagation formula is:
[0058] Z=X·W+B
[0059] A=f(Z)
[0060] Where W is the network weight matrix, X is the input matrix, B is the bias unit in the hidden layer, f is the activation function, Z is the result of the linear transformation of the data, and A is the output after processing by the activation function.
[0061] The activation function is selected as ReLU, which is calculated as follows:
[0062]
[0063] After the residual stress at each representative depth of residual stress of each group of sample fatigue data is input into the first fully connected layer 10, the first fully connected layer 10 will calculate them according to the above-mentioned forward propagation formula and output the comprehensive evaluation score of the residual stress. After the hardness at each representative depth of hardness is input into the second fully connected layer 20, the second fully connected layer 20 will calculate them according to the above-mentioned forward propagation formula and output the comprehensive evaluation score of the hardness. Then the comprehensive evaluation score of the residual stress, the comprehensive evaluation score and the remaining features will be input into the third fully connected layer 30 together. The third fully connected layer 30 will calculate them according to the above-mentioned forward propagation formula and output the predicted value of the fatigue life; then the loss between the predicted value of the fatigue life and the fatigue life (i.e., the true value) corresponding to the group of sample fatigue data can be calculated, and then the gradient is calculated based on the loss (usually with the goal of minimizing the loss), and the gradient is back-propagated to each layer of the neural network to update the parameters of the neural network model; then the above process is repeated and stopped after the preset conditions are met, thereby obtaining a trained neural network model.
[0064] The loss J between the predicted value and the true value of fatigue life is calculated as follows:
[0065]
[0066] in, is the predicted value of fatigue life, and y is the true value of fatigue life.
[0067] The gradient back propagation method of the third fully connected layer 30 is:
[0068] δ [l] =(δ [l+1] ·(W [l+1] ) T )⊙f′ [l] (z [l] )
[0069] Among them, δ [l] is the gradient of the lth layer of the third fully connected layer 30, δ [l+1] is the gradient of the l+1th layer of the third fully connected layer 30, W [l+1] is the parameter of the l+1th layer of the third fully connected layer 30, ⊙ represents the multiplication of the elements in the matrix one by one, and f′ [l] Represents the derivative of the activation function used in the lth layer of the neural network, where l is the number of layers in the neural network.
[0070] The parameters of the third fully connected layer 30 in the preprocessing neural network are updated according to the gradient, and the updating method is:
[0071]
[0072] Where α represents the learning rate, J is the loss, The calculation method is:
[0073]
[0074] For the gradients in the first and second fully connected layers, in the second back propagation (i.e., the third fully connected layer 30 propagates to the first and second fully connected layers), the gradients calculated by the third fully connected layer 30 are recorded as the δ matrix. The first two columns in δ represent the gradient values related to the dimensionality reduction results, that is, the gradients of the residual stress and hardness comprehensive evaluation indicators. The back propagation technology is used again using the values of the first two columns of δ to adjust the network weight parameters related to the residual stress and hardness comprehensive evaluation indicators, that is, the parameters of the first fully connected layer 10 and the second fully connected layer 20.
[0075] The gradient propagation mode in the first fully connected layer 10 and the second fully connected layer 20 is:
[0076] δ [l] =(δ [l+1] [:,n]·(W [l+1] ) T )⊙f′ [l] (Z [l] )
[0077] Where W represents the network weight parameter, δ [l+1] [:,n] is the nth column in the gradient matrix of the third fully connected layer 30, i.e., the gradient of the residual stress or hardness comprehensive evaluation score.
[0078] The above forward propagation and back propagation processes are iterated repeatedly to train the model.
[0079] In some embodiments, satisfying the preset condition may be that the number of iterations reaches a preset number or the loss of the neural network reaches a preset value.
[0080] In some embodiments, when training a neural network model, each group of sample fatigue data and its corresponding fatigue life can be divided into a training set and a test set according to a preset ratio (e.g., 8:2), and then the training set is used to train the neural network model to obtain a trained neural network model, and then the trained neural network model is tested using the test set to verify its performance. Specifically, the performance of the trained neural network model can be verified by the determination coefficient, and the calculation method of the determination coefficient is as follows:
[0081]
[0082] Where y is the true value, y pre is the model regression prediction value (i.e. the predicted value of fatigue life obtained after the test set is input into the trained neural network model), is the average of the true values.
[0083] S400: Obtain fatigue data of the surface reinforcement material to be predicted, and input the fatigue data of the surface reinforcement material to be predicted into a trained neural network model to obtain the fatigue life of the surface reinforcement material to be predicted.
[0084] After obtaining the trained neural network model, it can be used to predict fatigue life. The fatigue data of the surface strengthened material to be predicted (i.e., the maximum stress, frequency, stress ratio of the fatigue test (these parameters can use the fatigue parameters under actual working conditions), as well as the material tensile strength, the depth of the plastic deformation layer after surface strengthening, the surface roughness, the residual stress at multiple representative depths of residual stress, and the hardness at multiple representative depths of hardness) are input into the model, and the model can output the predicted value of fatigue life.
[0085] The method for predicting the fatigue life of a surface-reinforced material in an embodiment of the present invention uses residual stress, hardness, surface roughness, depth of the plastic deformation layer, ultimate tensile strength of the material, maximum stress of a fatigue test, stress ratio, and frequency as inputs, and fatigue life as output to train a neural network model, so as to characterize the highly nonlinear relationship between fatigue life and the above characteristics through the trained neural network model, and then use the trained neural network model to predict the fatigue life of the surface-reinforced material, which has the advantages of small error, low cost, and high efficiency.
[0086] like Figure 3 As shown, on the other hand, the present invention provides a prediction system for fatigue life of surface reinforced materials, which includes a first acquisition module 510, a second acquisition module 520, a training module 530 and a prediction module 540; the first acquisition module 510 is used to obtain multiple groups of sample fatigue data of surface reinforced materials and the fatigue life corresponding to each group of sample fatigue data, wherein each group of sample fatigue data includes maximum stress, frequency, stress ratio, material tensile strength, depth of plastic deformation layer after surface reinforcement, surface roughness, residual stress at multiple residual stress representative depths and hardness at multiple hardness representative depths; the second acquisition module 520 is used to obtain the best neural network model; the training module 530 is used to take sample fatigue data as input and the fatigue life corresponding to the sample fatigue data as output, and use each group of sample fatigue data and its corresponding fatigue life to train the best neural network model to obtain a trained neural network model; the prediction module 540 is used to obtain fatigue data of the surface reinforced material to be predicted, and input the fatigue data of the surface reinforced material to be predicted into the trained neural network model to obtain the fatigue life of the surface reinforced material to be predicted.
[0087] Each embodiment of the present invention is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
[0088] The fatigue life prediction system of the surface reinforcement material of the embodiment of the present invention uses residual stress, hardness, surface roughness, depth of plastic deformation layer, ultimate tensile strength of material, maximum stress of fatigue test, stress ratio and frequency as input, and fatigue life as output to train the neural network model, so as to characterize the highly nonlinear relationship between fatigue life and the above characteristics through the trained neural network model, and then use the trained neural network model to predict the fatigue life of the surface reinforcement material, which has the advantages of small error, low cost and high efficiency.
[0089] The effect of the prediction method of the present invention is explained below by taking TC4 as an example.
[0090] S100a: 87 groups of TC4 sample fatigue data were obtained from the literature, including the maximum stress, frequency, stress ratio of the fatigue test, the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, the residual stress at multiple representative depths of residual stress, the hardness at multiple representative depths of hardness, the surface roughness, and the fatigue life corresponding to each group of sample fatigue data.
[0091] The residual stress of surface-strengthened TC4 varies significantly in the depth range of 0 to 100μm and reaches a maximum value near 100μm. Therefore, in order to fully capture these fluctuations, the residual stress values at depths of 0μm, 10μm, 25μm, 50μm and 100μm are selected for interpolation. For the surface-strengthened TC4 specimen, the residual stress value fluctuation disappears beyond the depth of 600μm and tends to the residual stress of the specimen matrix. Therefore, in order to ensure that the residual stress characteristic information fully covers the depth range, additional interpolation must be performed, including the residual stress values at depths of 200μm, 300μm, 400μm, 500μm and 600μm.
[0092] Considering the different surface strengthening technologies and processes, the maximum residual stress depth of samples in some literature is around 400μm. In order to maintain the consistency of the number of features, the residual stress values at 500μm and 600μm depths are set equal to the value at 400μm. This correction makes the data set uniform and facilitates the training of the neural network model and comparative analysis.
[0093] Correspondingly, for the hardness feature, the hardness feature was obtained by interpolation at depths of 0 μm, 5 μm, 10 μm, 15 μm, 30 μm, 50 μm, 80 μm, 150 μm, 300 μm and 500 μm.
[0094] In addition to the surface-strengthened material data, the fatigue data of each group of samples also includes the original material data of the control group in the literature, that is, the fatigue life data of the material without surface reinforcement. The introduction of the fatigue life data of the metal material without surface reinforcement helps the model to more clearly understand the impact of the surface integrity change caused by surface reinforcement on the fatigue life. This is mainly reflected in the data of the material without surface reinforcement, where the residual stress and hardness values remain unchanged, and the depth of the plastic deformation layer is 0.
[0095] In addition, in order to speed up model training and reduce the possibility of falling into the local optimal solution during model parameter training, the collected data were normalized.
[0096] The normalization formula is:
[0097]
[0098] The above data are divided into a training set and a test set in the ratio of 8:2. In order to ensure that the training set information is not leaked, the training set and the test set are normalized respectively, and all the data in the same document are put into the training set or the test set, rather than putting part of the data into the training set and part of the data into the test set.
[0099] S200a: Use the training set to perform five-fold cross validation on different neural network models to obtain the best neural network model, such as Figure 4 The figure shows the cross-validation results. The cross-validation error MSE of the best neural network model is 0.1478. Its first fully connected layer 10 and the second fully connected layer 20 have 8 neurons, and the third fully connected layer 30 has 12 neurons.
[0100] S300a: Train the best neural network model using the training set to obtain a trained neural network model.
[0101] like Figure 5 The figure shows the loss curve of the best neural network on the training set and the test set. The loss value of the model drops rapidly in the early stage of training and gradually approaches the optimal point. When the loss fluctuates, the weight parameters of the surface model are close to the optimal solution. At this time, the training stops, indicating that the model has reached an appropriate training level.
[0102] S400a: After training 7000 times using all the training set data, the performance of the model is verified on the test set using the coefficient of determination; Figure 6 As shown, the prediction results are as follows:
[0103] The final R of the model on the test set 2 The value reached 0.91, showing excellent prediction performance; almost all data prediction results fell within the ±1.5 times error band.
[0104] The above is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. The above embodiment of the present invention can also be modified in various ways. That is, all simple, equivalent changes and modifications made according to the claims and the description of the present invention fall within the scope of protection of the claims of the present invention. The contents not described in detail in the present invention are all conventional technical contents.
Claims
1. A method for predicting fatigue life of surface reinforced materials, characterized in that: include: Obtain multiple groups of sample fatigue data of surface-strengthened materials and the fatigue life corresponding to each group of sample fatigue data, wherein each group of sample fatigue data includes the maximum stress, frequency, stress ratio of the fatigue test, as well as the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, the surface roughness, the residual stress at multiple representative depths of residual stress, and the hardness at multiple representative depths of hardness; Get the best neural network model; Taking the sample fatigue data as input and the fatigue life corresponding to the sample fatigue data as output, the optimal neural network model is trained using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model; Fatigue data of the surface reinforcement material to be predicted is obtained, and the fatigue data of the surface reinforcement material to be predicted is input into a trained neural network model to obtain the fatigue life of the surface reinforcement material to be predicted.
2. The method for predicting fatigue life of surface reinforcement materials according to claim 1, characterized in that: The optimal neural network model includes a first fully connected layer, a second fully connected layer and a third fully connected layer. The first fully connected layer is used to output a comprehensive evaluation score of residual stress according to the residual stress at multiple representative depths of residual stress. The second fully connected layer is used to output a comprehensive evaluation score of hardness according to the hardness at multiple representative depths of hardness. The third fully connected layer is used to output fatigue life according to the comprehensive evaluation score of residual stress, the comprehensive evaluation score of hardness, the maximum stress, frequency, stress ratio of fatigue test, and the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, and the surface roughness.
3. The method for predicting fatigue life of surface reinforcement materials according to claim 1, characterized in that: Get the best neural network model, including: Obtain multiple different neural network models, wherein the number of neurons in each layer of each neural network model is different; For each neural network model, a five-fold cross validation is performed on the neural network model using each group of sample fatigue data and its corresponding fatigue life to obtain the cross validation error of the neural network model; The neural network model with the smallest cross-validation error is selected as the best neural network model.
4. The method for predicting fatigue life of surface reinforcement materials according to claim 1, characterized in that: The optimal neural network model is trained using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model, which specifically includes: Select a part of each group of sample fatigue data and its corresponding fatigue life as a training set; For each group of sample fatigue data in the training set, the sample fatigue data is input into the optimal neural network model to obtain the predicted value of fatigue life output by the optimal neural network model; the loss is calculated according to the predicted value of fatigue life and the fatigue life corresponding to the group of sample fatigue data; the gradient is calculated according to the loss; the gradient is back-propagated to each layer of the optimal neural network model to update the parameters of the optimal neural network model; After the preset conditions are met, the update is stopped to obtain a trained neural network model.
5. The method for predicting fatigue life of surface reinforcement material according to claim 1, characterized in that: The representative depth of residual stress is the depth within the depth range where the trend of residual stress changing with depth has obvious fluctuations, and the representative depth of hardness is the depth within the depth range where the trend of hardness changing with depth has obvious fluctuations.
6. A system for predicting fatigue life of surface reinforcement materials, characterized in that: include: The first acquisition module is used to obtain multiple groups of sample fatigue data of surface-strengthened materials and the fatigue life corresponding to each group of sample fatigue data, wherein each group of sample fatigue data includes the maximum stress, frequency, stress ratio of the fatigue test, as well as the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, the surface roughness, the residual stress at multiple representative depths of residual stress, and the hardness at multiple representative depths of hardness; The second acquisition module is used to obtain the best neural network model; A training module, which is used to take the sample fatigue data as input and the fatigue life corresponding to the sample fatigue data as output, and train the optimal neural network model using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model; Fatigue data of the surface reinforcement material to be predicted is obtained, and the fatigue data of the surface reinforcement material to be predicted is input into a trained neural network model to obtain the fatigue life of the surface reinforcement material to be predicted.
7. The system for predicting fatigue life of surface reinforcement materials according to claim 6, characterized in that: The optimal neural network model includes a first fully connected layer, a second fully connected layer and a third fully connected layer. The first fully connected layer is used to output a comprehensive evaluation score of residual stress according to the residual stress at multiple representative depths of residual stress. The second fully connected layer is used to output a comprehensive evaluation score of hardness according to the hardness at multiple representative depths of hardness. The third fully connected layer is used to output fatigue life according to the comprehensive evaluation score of residual stress, the comprehensive evaluation score of hardness, the maximum stress, frequency, stress ratio of fatigue test, and the tensile strength of the material, the depth of the plastic deformation layer after surface strengthening, and the surface roughness.
8. The system for predicting fatigue life of surface reinforcement materials according to claim 6, characterized in that: Get the best neural network model, including: Obtain multiple different neural network models, wherein the number of neurons in each layer of each neural network model is different; For each neural network model, a five-fold cross validation is performed on the neural network model using each group of sample fatigue data and its corresponding fatigue life to obtain the cross validation error of the neural network model; The neural network model with the smallest cross-validation error is selected as the best neural network model.
9. The system for predicting fatigue life of surface reinforcement materials according to claim 6, characterized in that: The optimal neural network model is trained using each group of sample fatigue data and its corresponding fatigue life to obtain a trained neural network model, which specifically includes: Select a part of each group of sample fatigue data and its corresponding fatigue life as a training set; For each group of sample fatigue data in the training set, the sample fatigue data is input into the optimal neural network model to obtain the predicted value of fatigue life output by the optimal neural network model; the loss is calculated according to the predicted value of fatigue life and the fatigue life corresponding to the group of sample fatigue data; the gradient is calculated according to the loss; the gradient is back-propagated to each layer of the optimal neural network model to update the parameters of the optimal neural network model; After the preset conditions are met, the update is stopped to obtain a trained neural network model.
10. The system for predicting fatigue life of surface reinforcement materials according to claim 6, characterized in that: The representative depth of residual stress is the depth within the depth range where the trend of residual stress changing with depth has obvious fluctuations, and the representative depth of hardness is the depth within the depth range where the trend of hardness changing with depth has obvious fluctuations.
Citation Information
Cited By
Fatigue life prediction method considering surface roughness
CN121051940A
