Crack propagation parameter identification and remaining life prediction method based on neural network

Abstract

The present invention provides a method for identifying crack propagation parameters and predicting remaining life based on a neural network, comprising the following steps: S1, using a sensor to observe the crack propagation length of a component and its corresponding number of cycles; S2, selecting a traditional physics-based crack propagation rate model and determining initial material parameters; S3, scaling the crack length, number of cycles, and material parameters; S4, a neural network for predicting the number of cycles of a component through crack length; S5, constructing a loss function of physical information mixing according to the crack propagation rate model; S6, optimizing the material parameters in the neural network and the crack propagation rate model at the same time; S7, obtaining a crack propagation constant that meets the individual component and using the adjusted neural network to predict the remaining life of crack propagation. The present invention overcomes the problem that the traditional physical model is greatly affected by the individual differences of the component under a small amount of observation data, and can directly predict the remaining life of crack propagation of the monitored object.

Description

Technical Field

[0001] The present invention relates to the technical field of next generation information network, and in particular to a method for identifying fatigue crack extension material parameters and predicting remaining life based on neural network. Background Art

[0002] Fatigue cracks are cracks caused by stress concentration and stress cycles in materials or structures under cyclic loading. The expansion of fatigue cracks may lead to failure of materials or structures, causing accidents or compromising personnel safety. The residual life of fatigue crack expansion refers to the remaining life of crack expansion in the presence of fatigue cracks in materials or structures. By accurately predicting the residual life of fatigue crack expansion, it can help formulate reasonable repair and maintenance strategies, take repair, replacement or reinforcement measures in advance, and ensure that materials or structures are used within a safe range.

[0003] With the advancement of computer technology, data-driven methods have been used as a promising tool for fatigue crack growth life prediction because they can approximate the potential fatigue damage process without assuming any functional form. Compared with physics-based prediction methods, data-driven methods can more easily consider multiple factors that affect structural performance at the same time through data combination, such as different materials, load information, defect distribution information, etc. However, data-driven methods often require a large amount of labeled data to train the model to ensure its prediction accuracy. Due to cost or test conditions, a large amount of fatigue life test data is usually difficult to obtain. In addition, current data-driven methods are usually black box models, which have received many doubts in terms of interpretability and extrapolated performance prediction, especially that they may overfit observed data and produce prediction results that are contrary to physics. In recent years, methods that integrate physics into data-driven models and achieve physics and data fusion have received great attention. Physically induced neural networks can achieve high-precision predictions with a small amount of data by integrating physical models or empirical formulas into neural networks.

[0004] However, most of the existing physical induced neural network methods focus on fatigue life prediction, while almost no attention is paid to the prediction of fatigue crack propagation remaining life. When predicting the remaining life of fatigue crack propagation, only part of the observed information is available, so the amount of data will be less when solving this problem. Based on this small part of information, it may be difficult to build an accurate physical model. And affected by individual differences, it is difficult for the same material parameter to achieve good results in different monitoring objects. At the same time, predicting the remaining life based on some existing data is an extrapolation behavior, and the data-driven method may lead to prediction results with very large errors or deviations. The existing purely physical or data-driven methods are facing bottlenecks. It is urgent to develop a method that can integrate physics and data, which can not only consider the factors ignored by the physical model through data-driven, but also ensure the consistency of the final output with the physical laws, and enhance the interpretability and extrapolation performance of the model. Therefore, it is very necessary to develop corresponding crack propagation constant identification and remaining life prediction methods using physical information neural network technology for the problem of crack propagation remaining life prediction. Summary of the invention

[0005] The purpose of the present invention is to construct a method for identifying material parameters of fatigue crack propagation and predicting remaining life based on a neural network for cracked structural parts. Specifically, the present invention proposes a method for identifying material parameters of fatigue crack propagation and predicting remaining life based on a neural network, comprising the following steps:

[0006] S1 uses sensors to observe the crack extension length of the component and its corresponding number of cycles;

[0007] S2 selects the traditional physics-based crack growth rate model and determines the initial material parameters;

[0008] S3 performs parameter scaling on crack length, number of cycles, and material parameters;

[0009] A neural network for predicting the number of cycles of S4 components through crack length;

[0010] S5 constructs a loss function of physical information mixing according to the crack growth rate model;

[0011] S6 simultaneously optimizes material parameters in the neural network and crack growth rate models;

[0012] S7 obtains the crack growth constants that match the individual components and uses the adjusted neural network to predict the remaining life of crack growth.

[0013] As a further solution of the present invention: in said S1, the crack expansion length a and its corresponding number of cycles n of the component during the service life are monitored and obtained, as well as other available crack expansion information of the structural material, such as the crack expansion threshold value, material fracture toughness, etc.

[0014] As a further solution of the present invention: in S2, the crack growth model rate model is in the following form:

[0015]

[0016] Among them, g(·) will include crack growth material parameters, such as the classic crack growth rate model Paris formula, which is expressed as:

[0017]

[0018] It contains material parameters C and m. ΔK is the amplitude of the stress intensity factor at the crack tip, which is closely related to the crack length a. ΔK of simple structural parts can be obtained through analytical expressions, and ΔK of complex structural parts can be solved by finite element method.

[0019] As a further solution of the present invention: in S3, the parameter scaling is as follows:

[0020]

[0021] Where a represents the crack length, a 0 represents the initial crack length, n represents the number of cycles, and l is the order of magnitude of the crack growth life (for example: 10 4 or 10 5 ), C and m are material parameters in the crack growth model, and [·] indicates rounding down.

[0022] As a further solution of the present invention: in S5, the hybrid loss function is as follows:

[0023] Loss = λ 1 Loss D +λ 2 Loss P

[0024] Among them, Loss D For data loss, Loss P is the physical loss of the crack growth rate model, λ 1 and λ 2 is the hyperparameter of the balanced loss term. D It is used to measure the error between the number of cycles corresponding to the crack length predicted by the neural network and the actual number of cycles observed. Its expression is as follows:

[0025]

[0026] Where MSELoss represents the root mean square loss function, a obs represents the observed crack length, NN(·) represents the neural network, and n pred represents the number of cycles predicted by the neural network, and nobs represents the number of cycles actually observed. P It is used to measure the gap between the crack growth rate extracted by the neural network and the physical model. Its expression is as follows:

[0027]

[0028] Among them, a p is the virtual crack length randomly sampled from the initial crack length to the critical crack length, and the sampling number is the neural network hyperparameter.

[0029] As a further solution of the present invention: in S6, during network training, the material parameters in the physical model are set as updateable parameters of the neural network, and are updated by the optimizer together with the weights and bias parameters of the neural network itself.

[0030] As a further solution of the present invention: in S7, the neural network adjusted by S6 can predict the remaining life by the following formula:

[0031] n Ref =NN(a c )-n t

[0032] Among them, a c is the critical crack length, n t is the number of cycles during observation, n Ref is the remaining life of crack growth. The material parameters in the crack growth model are obtained along with the update of the model, and finally, through the parameter scaling in S3, the reverse solution is obtained as the numerical scale of the raw material parameters.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The physical laws used in the present invention come from the crack growth rate model based on elastic mechanics and fracture mechanics. For cracked components, a physical information hybrid function is constructed to constrain the training of the neural network. This gives the neural network as a data-driven method stronger interpretability and reliability. The material parameters in the physical model are regarded as updateable parameters, and in the step-by-step training process, the material parameters that conform to the individual crack growth laws of the monitored object are found.

[0035] 2. The present invention adopts a physical induced neural network to predict the remaining life of crack growth. With a small amount of observation data at the initial stage of crack growth, it can overcome the problem that the physical model is greatly affected by individual differences and the data-driven model has poor extrapolation prediction ability for small sample data. While predicting the remaining life of crack growth, it can also identify the material parameters in the crack growth rate model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0037] Figure 1 The present invention provides a flow chart of a method for identifying fatigue crack propagation material parameters and predicting remaining life based on a neural network.

[0038] Figure 2 Schematic diagram of the MT specimen shape in Example 1.

[0039] Figure 3 This is a schematic diagram of MT data set division in Example 1.

[0040] Figure 4 It is the framework of the physically induced neural network in the present invention.

[0041] Figure 5 This is the crack propagation remaining life prediction result in Example 1 of the present invention.

[0042] Figure 6 This is the material parameter identification result of Example 1 of the present invention.

[0043] Figure 7 It is a schematic diagram of the shape of the CT specimen in Example 2.

[0044] Figure 8 This is the crack propagation remaining life prediction result in Example 2 of the present invention.

[0045] Fig. 9 This is the material parameter identification result in Example 2 of the present invention. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] Example 1

[0048] like Figure 1 As shown, the present invention provides a method for fatigue crack propagation material parameter identification and remaining life prediction based on neural network, comprising the following steps:

[0049] S1 uses sensors to observe the crack extension length of the component and its corresponding number of cycles;

[0050] There are 68 sets of crack growth test results of the existing central crack tensile specimens. Each specimen is regarded as an observation object for residual life prediction. The specimen shape is as follows: Figure 2 As shown. One group of experiments out of 68 groups of experiments was selected as a validation set to determine the hyperparameters of the neural network and the training process, and the remaining 67 experiments were used to test the effectiveness of the method proposed in the present invention. The remaining life prediction task essentially refers to predicting the remaining life based on a small amount of observational data at the early stage of crack propagation. In this implementation, it is assumed that the small amount of data at the early stage of each crack propagation experiment is 20% of the total data volume, and it is necessary to predict the subsequent crack propagation based on this part of the observational data and predict its remaining life. The data set division in the implementation is as follows Figure 3 shown.

[0051] S2 selects the traditional physics-based crack growth rate model and determines the initial material parameters;

[0052] Select a suitable crack growth rate model based on the existing crack growth data in the data set. Since this data set only provides the necessary load information and no more crack growth parameter information, the most classic Paris formula is selected as the model for subsequent physical information fusion. The expression of the Paris formula is as follows:

[0053]

[0054] It contains material parameters C and m. ΔK is the amplitude of the stress intensity factor at the crack tip. For MT specimens, its calculation formula is as follows:

[0055]

[0056] Among them, ΔF represents the load range of the specimen, ΔF=F max -F min; W represents the width of the specimen, B represents the thickness of the specimen; α is the shape parameter of the specimen, α = 2a / W. The initial material parameters C and m are obtained from the validation set experiment through the least squares linear fitting.

[0057] S3 performs parameter scaling on crack length, number of cycles, and material parameters;

[0058] The variables and parameters involved in the neural network calculation process are scaled as follows.

[0059]

[0060] Where a represents the crack length, a 0 represents the initial crack length, n represents the number of cycles, l represents the order of magnitude of the crack growth life, C and m represent the material parameters in the crack growth model, and [·] represents rounding down.

[0061] A neural network for predicting the number of cycles of S4 components through crack length;

[0062] The neural network in this implementation consists of three layers of perceptrons. The output of the neural network can be expressed as follows:

[0063] n=NN(a)=W (3) (f(W (2) (f(W (1) a+b (1) ))+b (2) ))+b (3) (4)

[0064] Where n is the output of the neural network, i.e., the number of fatigue cycles; a is the input of the neural network, i.e., the crack extension length; W (i) and b (i) , i=1,2,3 are the network layers where the parameters are located, representing the weight parameters and bias parameters of different layers respectively; f· represents the activation function; NN represents the neural network function. The network includes two hidden layers with 64 nodes each and a single-node output layer. Tanh is selected as the activation function.

[0065] S5 constructs a loss function of physical information mixing according to the crack growth rate model;

[0066] The physical information hybrid loss function constructed in combination with the crack growth rate model is as follows:

[0067] Loss = λ 1 Loss D +λ 2 Loss P (5)

[0068] Among them, Loss DFor data loss, Loss P is the physical loss of the crack growth rate model, λ 1 and λ 2 is the hyperparameter of the balanced loss term, which is λ in this implementation 1 =100 and λ 2 =1. Loss D It is used to measure the error between the number of cycles corresponding to the crack length predicted by the neural network and the actual number of cycles observed. Its expression is as follows:

[0069]

[0070] Where MSELoss represents the root mean square loss function, a obs represents the observed crack length, NN(·) represents the neural network, and n pred Represents the number of cycles predicted by the neural network, n obs Represents the number of cycles actually observed. P It is used to measure the gap between the crack growth rate extracted by the neural network and the physical model. Its expression is as follows:

[0071]

[0072] Among them, a p is the virtual crack length randomly sampled from the initial crack length to the critical crack length, and a total of 500 were sampled in this implementation.

[0073] S6 simultaneously optimizes material parameters in the neural network and crack growth rate models;

[0074] The deep learning algorithm is used to optimize the neural network. The optimization flow chart of the proposed physical induced neural network is as follows: Figure 4 shown.

[0075] S7 obtains the crack growth constants that match the individual components and uses the adjusted neural network to predict the remaining life of crack growth;

[0076] The neural network adjusted by S6 can predict the remaining life by the following formula:

[0077] n Ref =NN(a c )-n t (8)

[0078] Among them, a c is the critical crack length, n t is the number of cycles during observation, n Refis the remaining life of crack growth. The material parameters in the crack growth model are obtained along with the model update, and finally the numerical scale of the raw material parameters is reversely solved through parameter scaling in S3. The obtained crack growth remaining life prediction results are as follows: Figure 5 As shown in Figure 2, the effect of material parameter identification is as follows: Figure 6 The results show that the prediction error of the crack propagation remaining life of the method proposed in the present invention is within 1.5 times the discrete band, while the prediction error of the traditional physical model is within 2.5 times the discrete band; at the same time, the linear relationship represented by the material parameters obtained by neural network identification is more consistent with the crack propagation law of the individual specimen.

[0079] Example 2

[0080] The above method can also be applied to different types of structural parts. First, a compact tensile specimen (CT specimen, such as Figure 7 The feasibility of the proposed method is verified on a crack growth data set (shown in Figure 2). This data set only provides crack growth data and necessary load information, so the Paris formula is also selected as the crack growth rate model.

[0081] In modification S2, the expression of the amplitude of the stress intensity factor at the crack tip is as follows:

[0082]

[0083] Wherein, α = a / W. The processes of S3-7 are similar to those of Example 1, but the corresponding parameters are adjusted to the parameters of the CT specimen data set. The trained neural network is used to predict the remaining life of crack propagation and identify material parameters. The results are as follows: Figure 8 and Fig. 9 The results show that compared with the traditional remaining life prediction method based on the Paris formula, the proposed method also shows good prediction ability in CT specimens. At the same time, the material parameters obtained by neural network optimization can better reflect the individual crack propagation behavior.

[0084] The effectiveness and versatility of the method of the present invention are verified by the MT experiment and the CT experiment data. Therefore, the present invention is not limited to the above specific implementations. A person skilled in the art can implement the present invention by adopting other specific implementations according to the disclosure of the embodiments and the drawings. Therefore, any design that adopts the design structure and ideas of the present invention and makes some simple changes or modifications falls within the scope of protection of the present invention.

Claims

1. A method for crack propagation parameter identification and remaining life prediction based on neural network, characterized in that: The fatigue crack propagation material parameter identification and remaining life prediction method based on neural network comprises the following steps: S1 uses sensors to monitor the crack extension length of the component and its corresponding number of cycles; S2 selects the traditional physics-based crack growth rate model and determines the initial material parameters; S3 performs parameter scaling on crack length, number of cycles, and material parameters; S4 builds a neural network to predict the number of cycles through crack length; S5 constructs a loss function of physical information mixing according to the crack growth rate model, wherein the loss function includes two parts: data loss and physical loss, and is balanced and adjusted through hyperparameters; wherein the data loss part uses a root mean square loss function (MSELoss) to evaluate the error between the number of cycles (npred) corresponding to the crack length predicted by the neural network and the number of cycles (nobs) actually observed, and the physical loss part is used to measure the difference between the crack growth rate extracted by the neural network and the physical model; S6 simultaneously optimizes the material parameters in the neural network and crack growth rate model for the proposed physically induced neural network; S7 obtains the crack growth constants that match the individual components and uses the adjusted neural network to predict the remaining life of crack growth.

2. The method for crack propagation parameter identification and remaining life prediction based on neural network according to claim 1, characterized in that: In the S1, crack extension test results of N groups of crack extension test pieces are used as observation objects, wherein one group of experimental data is used as a validation set to determine the hyperparameters of the neural network and the training process, and the remaining N-1 groups of experimental data are used to test the effectiveness of this method.

3. The method for crack propagation parameter identification and remaining life prediction based on neural network according to claim 1, characterized in that: In S2, for the case where the data set only provides necessary load information but no more crack extension parameter information, the Paris formula is selected as the model for subsequent physical information fusion; wherein the Paris formula includes material parameters, and the initial material parameters are obtained by linear fitting of the validation set experimental data using the least squares method.

4. The method for crack propagation parameter identification and remaining life prediction based on neural network according to claim 1, characterized in that: The S3 includes scaling the crack length, initial crack length, number of cycles and the order of magnitude l of the crack extension life.

5. The method for crack propagation parameter identification and remaining life prediction based on neural network according to claim 1, characterized in that: The neural network in S4 is composed of three layers of perceptrons, including two hidden layers each containing 64 nodes and an output layer with a single node.

6. The method for crack propagation parameter identification and remaining life prediction based on neural network according to claim 1, characterized in that: In S7, the neural network adjusted by S6 is used to predict the remaining life using a formula, which takes into account the critical crack length, the number of cycles during observation, and the remaining life of crack extension; at the same time, the material parameters in the crack extension model are obtained as the model is updated, and are reversely solved as the numerical scale of the raw material parameters through parameter scaling in S3.

Images