Fatigue crack growth rate prediction method based on active learning and physical loss

Abstract

The invention discloses an active learning and physical loss-based fatigue crack growth rate prediction method, which comprises the following steps of: splicing and fusing a preprocessed stress intensity factor, a stress ratio, pre-strain and stress amplitude, and taking the fused characteristics as the input of a model; embedding the Jones model into a loss function of the neural network, constructing a physical information time sequence model based on the Jones model, setting a parameter set, selecting a most valuable training sample from a training set by using active learning, retraining the model by using the selected most valuable training sample, selecting an optimal model, and evaluating the optimal model by using a test set; fusing the actually measured stress intensity factor, stress ratio, pre-strain and stress amplitude of the to-be-predicted material, and inputting into the obtained prediction model to obtain a predicted value of the fatigue crack growth rate. By adopting the technical scheme of the invention, the fatigue crack growth rate of the material in the whole life period can be efficiently and accurately predicted at low cost.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of structural damage assessment and life prediction of pressure equipment, and particularly relates to a fatigue crack propagation rate prediction method based on active learning and physical loss. BACKGROUND

[0002] As the core device in the industrial fields of petroleum chemical industry, energy and power, aerospace, pharmaceuticals, etc., pressure equipment often operates under harsh conditions such as high temperature, high pressure, corrosive medium and cyclic load. Fatigue failure is one of the most common and destructive failure modes. Small initial defects or cracks gradually expand under cyclic load, which may eventually lead to sudden rupture of the equipment, causing catastrophic accidents, huge economic losses, environmental pollution and even casualties. Therefore, accurately predicting the fatigue crack propagation rate in the whole life stage is the key basis for assessing the remaining life of pressure equipment, developing scientific maintenance strategies and ensuring safe operation.

[0003] PIML provides a new machine learning paradigm that guides the training of machine learning models through physical knowledge or phenomenological constraints. Although the theoretical system of PIML has just been established, it has shown great potential and broad research prospects in the fields of science and engineering. By introducing physical knowledge as a constraint, the network framework for fatigue crack propagation rate prediction and evaluation using PIML can provide physically consistent life prediction. However, the existing technology lacks a physical information time series model solution for fatigue crack propagation rate, so there is an urgent need to develop a fatigue crack propagation rate prediction method based on active learning and physical loss based on active learning and physical loss. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a fatigue crack propagation rate prediction method based on active learning and physical loss, which can predict the fatigue crack propagation rate of materials in the whole life stage at low cost, high efficiency and accuracy, thereby improving the service life of pressure equipment and reducing safety hazards.

[0005] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0006] A fatigue crack propagation rate prediction method based on active learning and physical loss, comprising:

[0007] S1: obtaining the corresponding relationship between the stress intensity factor, stress ratio, pre-strain, stress amplitude and fatigue crack propagation rate of the material, and taking it as an original data set;

[0008] S2: preprocessing the original data set obtained in step S1, and dividing the preprocessed original data set into a training set, a validation set and a test set;

[0009] S3: splicing and fusing the processed time sequence features and non-time sequence features, taking the fused features as the input of the model; wherein the time sequence features include: stress intensity factor; and the non-time sequence features include: stress ratio, pre-strain, stress amplitude;

[0010] S4: embedding the Jones model into a loss function of a neural network, constructing a physical information time sequence model based on the Jones model, setting a parameter set, selecting the most valuable training samples from the training set by active learning, retraining the model by using the selected most valuable training samples and selecting the best model, and evaluating the best model by using a test set;

[0011] S5: splicing and fusing the actually measured stress intensity factor, stress ratio, pre-strain and stress amplitude of the material to be predicted, and then inputting them into the prediction model obtained in step S4 to obtain the predicted value of the fatigue crack propagation rate.

[0012] Preferably, in step S4, the time sequence model is a Recurrent Neural Network model, a Long Short-Term Memory model or a Gated Recurrent Unit model, and the three time sequence models are used as committee members of active learning, wherein the parameter set to be set includes: time step, time step length, learning rate, batch size, iteration number, neuron number, activation function, L2 regularization weight and random inactivation;

[0013] The process of training the model by using active learning is as follows:

[0014] The initial 5 labeled samples are selected from the training set to train all learners, then all training set samples are input into the committee, new samples are selected from the training set according to the degree of disagreement between the committees and added to the training set of the next iteration until all samples in the training set are selected. In the final stage, all committee members retrain all selected samples and select the optimal model according to the evaluation standard.

[0015] Preferably, in step S4, the physical loss function used by the constructed physical information time sequence model is a MSE loss function nested according to the Jones model; the MSE loss function nested according to the Jones model is as follows:

[0016]

[0017] Loss=Loss D +λ2Loss D

[0018] Wherein, C v =8.8×10 -6 , mv = 0.9363, is the model output, is the true value, and λ2 is a hyperparameter that maintains the balance between physical constraints.

[0019] The application can fully consider the non-linear problem in the fatigue crack propagation rate prediction process, avoid the calculation error caused by the traditional linear fitting to the crack propagation life prediction, and can conveniently, quickly and accurately predict the fatigue crack propagation rate in the whole life stage under different loading conditions, thereby reducing the calculation cost and saving the research and test cost. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only a part of the embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0021] Figure 1 It is a flow chart of the fatigue crack propagation rate prediction method based on active learning and physical loss of the embodiment of the present application;

[0022] Figure 2 It is a comparison chart of the crack propagation rate and the test value predicted by the fatigue crack propagation rate prediction method based on active learning and physical loss of the physical information time sequence model based on active learning of the present application;

[0023] Figure 3 It is a comparison chart of the crack propagation rate and the test data predicted by the fatigue crack propagation rate prediction method based on active learning and physical loss of the physical information time sequence model based on active learning of the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0026] Embodiment 1:

[0027] As Figure 1As shown, the embodiment of the present application provides a fatigue crack propagation rate prediction method based on active learning and physical loss, comprising:

[0028] S1: using a fatigue crack propagation test, obtaining the corresponding relationship between the stress intensity factor, stress ratio, pre-strain, stress amplitude and fatigue crack propagation rate of the material, and taking it as an original data set;

[0029] S2: preprocessing the original data set obtained in step S1, and dividing the preprocessed original data set into a training set, a validation set and a test set;

[0030] S3: splicing and fusing the processed time sequence features (stress intensity factor) and non-time sequence features (stress ratio, pre-strain, stress amplitude), and taking the fused features as the input of the model;

[0031] S4: embedding the Jones model into the loss function of the neural network, constructing a physical information time sequence model based on the Jones model, setting a parameter set, selecting the most valuable training samples from the training set by active learning, wherein the committee members of the active learning are Recurrent Neural Network, Long Short-Term Memory and Gated Recurrent Unit, retraining the model with the selected most valuable training samples and selecting the best model, and evaluating the best model by using the test set;

[0032] S5: splicing and fusing the actually measured stress intensity factor, stress ratio, pre-strain and stress amplitude of the material to be predicted, and then inputting them into the prediction model obtained in step S4 to obtain the predicted value of the fatigue crack propagation rate.

[0033] As an embodiment of the present application, in step S1, a standard compact tension specimen is subjected to a fatigue test by a fatigue testing machine, and the stress intensity factor ΔK and the corresponding fatigue crack propagation rate da / dN data under different stress ratios R, pre-strains ε P and stress amplitudes ΔF are obtained as an original data set;

[0034] As an embodiment of the present application, in step S2, the original data set is specifically standardized and logarithmized, the stress ratio R, the pre-strain ε P and the stress amplitude ΔF are standardized, the stress intensity factor ΔK is transformed by , and the fatigue crack propagation rate is logarithmized to obtain lg(da / dN):

[0035] df[da / dN]=log(df[da / dN])

[0036] Wherein the standardization processing is:

[0037]

[0038] Wherein, μ is the mean, σ is the standard deviation, is the normalized feature, X is the input feature.

[0039] As an embodiment of the present application, in step S2, 12 fatigue crack propagation samples of the preprocessed original data set are randomly selected as the training set, 2 samples are selected as the validation set, and 4 samples are selected as the test set.

[0040] As an embodiment of the present application, in step S3, at each time step, the time series feature (stress intensity factor ΔK) and the non-time series feature (stress ratio R, pre-strain ε P , stress amplitude ΔF) are spliced into a longer feature vector, and then input into the time series model;

[0041] As an embodiment of the present application, in step S4, the time series model used is the Recurrent Neural Network model, the Long Short-Term Memory model, and the Gated Recurrent Unit model, which are used as committee members of active learning, and the parameter set to be set is: time step, time step length, learning rate, batch size, iteration number, neuron number, activation function, L2 regularization weight, and random inactivation;

[0042] The process of training the model by active learning is as follows:

[0043] The initial 5 labeled samples are selected from the training set to train all learners, then all training set samples are input into the committee, new samples are selected from the training set according to the degree of disagreement between the committees and added to the training set of the next iteration, until all samples in the training set are selected. In the final stage, all committee members retrain all selected samples and select the optimal model according to the evaluation standard.

[0044] As an embodiment of the present application, in step S4, the physical information time series model uses a MSE loss function nested according to the Jones model as the physical loss function. Wherein, the MSE loss function nested according to the Jones model is:

[0045]

[0046] Loss=Loss D +λ2Loss D

[0047] wherein C v = 8.8 x 10 -6 , m v = 0.9363, is the model output, is the true value, and λ2 is a hyperparameter that maintains the balance between physical constraints.

[0048] As an embodiment of the present application, in step S5, the regression coefficient R 2 and the mean square error MSE are used as evaluation indexes of the prediction accuracy of the model to determine whether the prediction model meets the requirements. The model determination coefficient R 2 is:

[0049]

[0050] wherein n is the number of data groups in the test set, y j is the crack propagation rate of the jth data in the test set, is the crack propagation rate predicted by the model for the jth data in the test set, is the average value of all crack propagation rates in the test set.

[0051] The mean square error MSE is:

[0052]

[0053] wherein n is the number of data groups in the test set, y i is the crack propagation rate of the ith data in the test set, is the crack propagation rate predicted by the model for the ith data in the test set, is the square of the error between the predicted value and the actual value of the ith sample.

[0054] Embodiment 2:

[0055] The present application provides a fatigue crack propagation rate prediction method based on active learning and physical loss, comprising:

[0056] Step S1: using a fatigue crack propagation test, the corresponding relationship between the stress intensity factor, stress ratio, pre-strain, load amplitude and fatigue crack propagation rate of the material is obtained, and is used as the original data set. The data from crack initiation to accelerated crack propagation under each loading condition is taken as a group of data, and a total of 18 groups of data constitute the original data set.

[0057] Step S2: the original data set obtained in step S1 is preprocessed, and the preprocessed original data set is divided into a training set, a validation set and a test set. Among them, 12 groups are used as the training set, 2 groups are used as the validation set, and 4 groups are used as the test set.

[0058] Step S3: splice and fuse the processed time sequence features (stress intensity factor) and non-time sequence features (stress ratio, pre-strain, load amplitude), and take the fused features as the input of the model;

[0059] Step S4: embed the Jones model into the loss function of the neural network, construct a physical information time sequence model based on the Jones model, set a parameter set, select the most valuable training samples from the training set by active learning, wherein the committee members of the active learning are Recurrent Neural Network, Long Short-Term Memory, Gated Recurrent Unit, retrain the model with the selected most valuable training samples and select the best model, and evaluate the best model by using the test set;

[0060] The set parameter set includes physical loss weight (λ2), active learning cycle number (n), time step (timesteps), step (stride), learning rate (learning rate), batch size (batch size), iteration number (epoch), neuron number (N), activation function, and random inactivation (dropout)

[0061] The modified Jones model is:

[0062]

[0063] Wherein, C v = 8.8 x 10 -6 , m v = 0.9363; ΔK th is the fatigue crack propagation threshold value, K IC is the fracture toughness

[0064] The expression of the loss function embedded in the neural network is specifically:

[0065]

[0066] Wherein, is the model output, is the true value, and λ2 is a hyperparameter for maintaining the balance between physical constraints.

[0067] The uncertainty index is:

[0068]

[0069] Wherein, is the average prediction value, and y j is the prediction value of each model.

[0070] The stability index is:

[0071] Stability Score = mean_val_loss + std_val_loss

[0072] where mean_val_loss is the average loss of the validation set, and std_val_loss is the standard deviation of the loss of the validation set.

[0073] Regression coefficient R 2 is:

[0074]

[0075] where n is the number of data groups in the test set; y j is the crack propagation rate of the jth data in the test set; is the crack propagation rate predicted by the model for the jth data in the test set; is the average value of all crack propagation rates in the test set;

[0076] The mean square error MSE is:

[0077]

[0078] where n is the number of data groups in the test set; y i is the crack propagation rate of the ith data in the test set; is the crack propagation rate predicted by the model for the ith data in the test set; is the square of the error between the predicted value and the actual value of the ith sample.

[0079] The final set of training parameters is set as the physical loss weight λ2 = 0.1, the number of active learning loops n = 5, the time step timesteps = 4, the step learning rate = 0.001 and using Adam optimizer to dynamically adjust the learning rate, batch size = 16, epoch = 20000, number of hidden layer neurons N1 = 32, N2 = 16, activation function selection relu function, dropout = 0.3.

[0080] Step S5: The actually measured stress intensity factor, stress ratio, pre-strain, and load amplitude of the material to be predicted are first spliced and fused, and then input into the prediction model obtained in step S4 to obtain the predicted value of the fatigue crack propagation rate.

[0081] where the sample used for the fatigue test to be predicted is a standard compact tension specimen in ASTM E647-15 standard.

[0082] The performance of the optimal model on the test set is the regression coefficient R 2 = 0.9873, and the mean square error MSE = 0.0317. The comparison between the predicted fatigue crack propagation rate of the model and the test value is shown in FIG. 2, and the comparison between the predicted fatigue crack propagation rate curve of the model and the test value is shown in FIG. 3. Figure 2 Figure 3

[0083] The method described in the embodiments of the present application can fully consider the non-linear problem in the process of predicting the fatigue crack propagation rate of the material. Under different stress ratios, pre-strains, and stress amplitudes, the crack propagation rate of the sample in the full life stage can be predicted in a portable, efficient, and accurate manner, avoiding the calculation error of the fatigue crack propagation rate caused by the traditional linear fitting, reducing the limitations of the application range of the existing fatigue crack propagation rate theoretical equation, and reducing the calculation cost and the test cost.

[0084] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.​​

Claims

1. A fatigue crack propagation rate prediction method based on active learning and physical loss, characterized in that, include: S1: Obtain the correspondence between the stress intensity factor, stress ratio, prestrain, stress amplitude and fatigue crack propagation rate of the material, and use it as the original dataset; S2: Preprocess the original dataset obtained in step S1, and divide the preprocessed original dataset into training set, validation set and test set; S3: The processed temporal and non-temporal features are concatenated and fused, and the fused features are used as the input to the model; among them, the temporal features include: stress intensity factor; the non-temporal features include: stress ratio, pre-strain, and stress amplitude; S4: Embed the Jones model into the loss function of the neural network, construct a physical information time series model based on the Jones model, set the parameter set, use active learning to select the most valuable training samples from the training set, retrain the model using the selected most valuable training samples and select the best model, and use the test set to evaluate the best model. S5: First, the actual measured stress intensity factor, stress ratio, pre-strain, and stress amplitude of the material to be predicted are spliced ​​and fused together, and then input into the prediction model obtained in step S4 to obtain the predicted value of fatigue crack propagation rate.

2. The fatigue crack propagation rate prediction method based on active learning and physical loss according to claim 1, characterized in that, In step S4, the time series model is a Recurrent Neural Network model, a Long Short-Term Memory model, and a Gated Recurrent Unit model. These three time series models are used as members of the active learning committee. The parameter set that needs to be set is: time step, time step size, learning rate, batch size, number of iterations, number of neurons, activation function, L2 regularization weight, and random deactivation. The process of training this model using active learning is as follows: The learning machines are trained using an initial set of 5 labeled samples selected from the training set. Subsequently, all training set samples are fed into a committee. Based on the degree of disagreement among committee members, new samples are selected from the training set to be added to the training set for the next iteration, until all training set samples have been selected. In the final stage, all committee members retrain the learning machines on all selected samples and select the optimal model based on evaluation criteria.

3. The fatigue crack propagation rate prediction method based on active learning and physical loss according to claim 2, characterized in that, In step S4, the physical loss function used in the constructed physical information time series model is the MSE loss function nested based on the Jones model; the MSE loss function nested based on the Jones model is: Loss=Loss D +λ2Loss D Among them, C v =8.8×10 -6 m v =0.9363, For model output, λ is the true value, and λ2 is a hyperparameter that maintains the balance between physical constraints.

Images