A neural network method for identifying mechanical properties of alloy layer materials based on genetic algorithm

Abstract

The present invention discloses a neural network alloy layer material mechanical property identification method based on a genetic algorithm, comprising the following steps: 1: selecting reasonable test parameters according to the microstructure and macrostructure of the bearing alloy layer material; 2: testing the bearing alloy layer using an indentation test to obtain an indentation force-indentation depth curve; 3: selecting a material constitutive model; 4: creating a neural network; 5: constructing a network training database using Abaqus numerical analysis software based on Latin hypercube theory; 6: training the neural network using the database constructed by Abaqus, wherein the neural network can construct a nonlinear mapping relationship between the material constitutive parameters and the indentation force-indentation depth curve through training; and 7: minimizing the least squares error between the indentation force-indentation depth curves obtained by the test and calculated by the neural network through a genetic algorithm, thereby obtaining the mechanical properties of the bearing alloy layer material.

Description

Technical Field

[0001] The present invention relates to the field of mechanical property analysis of metal coatings, and in particular to a neural network alloy layer material mechanical property identification method based on a genetic algorithm. Background Art

[0002] Sliding bearings are key supporting components of important mechanical equipment such as internal combustion engines, compressors, and generators. Accurate analysis of their performance is of great significance to improving the reliability of the entire machine. Sliding bearings are composed of bearing shells, which are multi-layer thin-walled structures. Figure 1 As shown in the figure, the main functional structure is the alloy layer. Obtaining the mechanical properties of the bearing alloy layer material is the prerequisite for accurate analysis of the bearing.

[0003] However, because the thickness of the alloy layer (tens to hundreds of microns) is very thin compared to the thickness of the substrate (more than ten millimeters), its material properties cannot be obtained using traditional uniaxial tensile testing. If the alloy layer material is sampled separately, it is difficult to ensure that the sample preparation process is completely consistent with large-scale production, which will cause deviations in mechanical properties. In addition, separate sample preparation greatly increases the testing cycle and cost. Therefore, as one of the most suitable methods for measuring the mechanical properties of coatings, indentation technology is widely used. Indentation technology obtains an indentation force-indentation depth curve through indentation testing and uses analytical methods to extract the mechanical properties of the tested material from the indentation force-indentation depth curve.

[0004] Currently, methods for extracting material mechanical properties from indentation force-depth curves are mostly divided into analytical and numerical methods. Analytical methods are fast but have low accuracy, while numerical methods offer high accuracy but suffer from slow calculation speed, uneven sampling point distribution, and sensitivity to initial values. Currently, no research has fully addressed these issues.

[0005] Therefore, it is of great significance to develop a neural network alloy layer material mechanical property identification method based on genetic algorithm, which can evenly obtain the sampling points of the numerical method and solve the initial value sensitivity problem of the numerical method, thereby quickly and accurately obtaining the macroscopic mechanical properties of the bearing alloy layer material. Summary of the Invention

[0006] In order to overcome the shortcomings of the existing technology, the present invention provides a neural network alloy layer material mechanical property identification method based on genetic algorithm, which aims to solve the problems of difficult measurement of bearing alloy layer material properties, uneven distribution of sampling points in existing indentation test analysis technology, slow analysis speed and initial value sensitivity of analysis.

[0007] In order to achieve the above-mentioned purpose of the invention, the technical solutions adopted to solve the technical problems are as follows:

[0008] A neural network alloy layer material mechanical property identification method based on genetic algorithm includes the following steps:

[0009] Step 1: Select reasonable test parameters based on the microstructure and macrostructure of the bearing alloy layer material;

[0010] Step 2: Use an indentation test to test the bearing alloy layer and obtain an indentation force-indentation depth curve;

[0011] Step 3: Select the material constitutive model;

[0012] Step 4: Create a neural network;

[0013] Step 5: Based on the Latin hypercube theory, Abaqus numerical analysis software is used to build a network training database;

[0014] Step 6: Use the database built by Abaqus to train the neural network. Through training, the neural network can construct a nonlinear mapping relationship between the material constitutive parameters and the indentation force-indentation depth curve;

[0015] Step 7: Through the genetic algorithm, the least square error between the indentation force-indentation depth curves obtained by the test and the neural network calculation is minimized, thereby obtaining the mechanical properties of the bearing alloy layer material.

[0016] Furthermore, in step 1, the specific test requirements are:

[0017] (1) The maximum indentation depth is less than 1 / 10 of the thickness of the bearing alloy layer;

[0018] (2) The maximum indentation depth is greater than 20 times the surface roughness Ra value of the bearing alloy layer;

[0019] (3) The ratio between the maximum indentation depth and the indenter radius is greater than 0.3;

[0020] (4) The contact area should contain as many material grains as possible.

[0021] Furthermore, in step 2, the specific experimental steps are:

[0022] Step 21: Clamp the sample stably;

[0023] Step 22: Adjust the indenter to the position just above the surface of the sample being tested;

[0024] Step 23: Set the pressing speed so that the pressing head presses in slowly;

[0025] Step 24: When the pressure reaches the set limit, the pressure head slowly withdraws.

[0026] Furthermore, in step 3, the specific method is:

[0027] Based on prior knowledge or subsequent needs, select a reasonable Hollomon constitutive model or Swift constitutive model.

[0028] Furthermore, in step 4, the specific method is:

[0029] In a multi-layer neural network, the weighted sum of the inputs to each layer is:

[0030]

[0031] The weighted sum of the input is processed through the activation function to get the output:

[0032]

[0033] in, Represents the output of the i-th neuron in the previous layer; and are the weight and bias corresponding to the j-th neuron in the l-th layer, and f() is the activation function.

[0034] Furthermore, in step 5, when sampling the material constitutive parameters, the Latin hypercube theory is used to improve the sampling performance. The specific steps are:

[0035] Step 51: Set the number of sampling points m;

[0036] Step 52: Generate m unit sampling points first. The generation method is:

[0037] s[i,j]=(i+rand()) / m

[0038] Among them, s[i,j] represents the value of the i-th sampling point in the j-th dimension, that is, the value of the i-th sampling point on different constitutive parameters, and rand() represents a random number from 0 to 1;

[0039] Step 53: Map the unit sampling points to the actual range to generate actual sampling points. Here, the Hollomon constitutive model is used as an example, but it is not limited to the Hollomon constitutive model:

[0040] K=K min +(K max -K min )*sK

[0041] n=n min +(n max -n min )*sn

[0042] Among them, K is the hardening coefficient of the Hollomon material constitutive model, n is the hardening index of the Hollomon material constitutive parameter, K min With Kmax is the minimum and maximum value of the value range of K, n min With n max are the minimum and maximum values ​​of the range of n, sK and sn are the values ​​of K and n at the unit sampling point respectively.

[0043] Furthermore, in step 6, the specific steps are:

[0044] Step 61: Training of Neural Network:

[0045] Use optimization algorithms to iteratively optimize the weights and biases in the neural network to minimize the error between prediction and reality;

[0046] The training database is divided into three parts: 70% is used to determine the weights and biases, 15% is used to test its accuracy, and 15% is used to verify its effectiveness;

[0047] Step 62: Packaging of Neural Network:

[0048] After the neural network is trained, it can reflect complex nonlinear mapping relationships;

[0049] The trained neural network is packaged into a MATLAB .m file and called when the target function is subsequently established.

[0050] Furthermore, in step 7, the specific method is:

[0051] Step 71: Create the objective function

[0052] The least square error between the predicted value simulated by the neural network and the corresponding data on the penetration force-depth curve obtained by the experiment is set as the objective function, specifically:

[0053]

[0054] Where F is the objective function, is the indentation force-indentation depth result predicted by the neural network, X is the vector composed of independent parameters of the material constitutive model; Y exp is the indentation force-indentation depth result obtained through the test;

[0055] Step 72: Optimize using genetic algorithms

[0056] (1) Initialization: generate m individuals to form the initial population;

[0057] (2) Calculate the fitness function for each individual;

[0058] (3) Select the best individuals based on the fitness function, keep the individuals with high fitness function, and discard the individuals with low fitness function;

[0059] (4) Performing crossover operations on the selected population to produce offspring;

[0060] (5) Perform mutation operations on the population with a certain probability;

[0061] (6) Conduct convergence judgment on the newly generated population and decide whether to terminate the algorithm;

[0062] (7) If the convergence requirement is not met, return to step (2) for iteration;

[0063] Through the genetic algorithm, X is used as the optimization variable and F is used as the objective function for iterative solution, so that the objective function converges to the minimum, thereby obtaining the actual material constitutive model parameter X.

[0064] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0065] 1. The present invention adopts Latin hypercube sampling theory, which can improve the sampling uniformity of material constitutive parameters;

[0066] 2. The present invention uses ANN as a proxy model to replace the finite element calculation in the optimization process, greatly improving the inversion efficiency;

[0067] 3. The present invention adopts a genetic algorithm in the optimization process to solve the problem of initial value sensitivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without inventive work. In the drawings:

[0069] Figure 1 This is a structural diagram of a bearing in the prior art;

[0070] Figure 2 This is a flow chart of a method for identifying mechanical properties of alloy layer materials using a neural network based on a genetic algorithm according to the present invention;

[0071] Figure 3 A neural network structure diagram of the present invention;

[0072] Figure 4 is a scatter plot of the Latin hypercube sampling results in the present invention;

[0073] Figure 5 Abaqus finite element model diagram in the present invention;

[0074] Figure 6This is a flow chart of the neural network training method of the present invention;

[0075] Figure 7 This is a flow chart of the genetic algorithm in the present invention;

[0076] Figure 8 This is a graph showing the indentation depth-indentation force curve of the bearing alloy layer material of the present invention;

[0077] Figure 9 This is a stress-strain curve diagram of the bearing alloy layer material in the present invention. DETAILED DESCRIPTION

[0078] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0079] like Figure 2 As shown, this embodiment discloses a neural network alloy layer material mechanical property identification method based on genetic algorithm, comprising the following steps:

[0080] Step 1: Select reasonable test parameters based on the microstructure and macrostructure of the bearing alloy layer material;

[0081] Furthermore, in step 1, the specific test requirements are:

[0082] (1) The maximum indentation depth is less than 1 / 10 of the thickness of the bearing alloy layer;

[0083] (2) The maximum indentation depth is greater than 20 times the surface roughness Ra value of the bearing alloy layer;

[0084] (3) The ratio between the maximum indentation depth and the indenter radius is greater than 0.3;

[0085] (4) The contact area should contain as many material grains as possible.

[0086] Step 2: Use an indentation test to test the bearing alloy layer and obtain an indentation force-indentation depth curve;

[0087] Furthermore, in step 2, the specific experimental steps are:

[0088] Step 21: Clamp the sample stably;

[0089] Step 22: Adjust the indenter to the position just above the surface of the sample being tested;

[0090] Step 23: Set the pressing speed so that the pressing head presses in slowly;

[0091] Step 24: When the pressure reaches the set limit, the pressure head slowly withdraws.

[0092] Step 3: Select the material constitutive model;

[0093] Furthermore, in step 3, the specific method is:

[0094] Based on prior knowledge or subsequent needs, select a reasonable Hollomon constitutive model or Swift constitutive model.

[0095] Step 4: Create a neural network;

[0096] Furthermore, in step 4, the specific method is:

[0097] In a multi-layer neural network, the weighted sum of the inputs to each layer is:

[0098]

[0099] The weighted sum of the input is processed through the activation function to get the output:

[0100]

[0101] in, Represents the output of the i-th neuron in the previous layer; and are the weight and bias corresponding to the j-th neuron in the l-th layer, and f() is the activation function.

[0102] Matlab provides the fitnet function to build a multi-layer fitting neural network. Using X as the input parameter matrix and Y as the output parameter matrix, a three-layer multi-layer neural network with 20, 40, and 20 neurons in each layer is created. The structure of the neural network is as follows. Figure 3 As shown:

[0103] net=fitnet(X,Y,[20,40,20]);

[0104] The neural network input is the material constitutive independent parameter vector, and the output is the sampling information vector on the indentation depth-indentation force curve.

[0105] Step 5: Based on the Latin hypercube theory, Abaqus numerical analysis software is used to build a network training database;

[0106] Furthermore, in step 5, according to the material constitutive model selected in step 3, the Latin hypercube theory is used to obtain sampling points of independent parameters of the material constitutive model. The specific implementation method of the Latin hypercube is:

[0107] Taking the Hollomon constitutive model as an example, but not limited to the Hollomon constitutive model:

[0108] Step 51: Set the number of sampling points m;

[0109] Step 52: Generate m unit sampling points first. The generation method is:

[0110] s[i,j]=(i+rand()) / m

[0111] Among them, s[i,j] represents the value of the i-th sampling point in the j-th dimension, that is, the value of the i-th sampling point on different constitutive parameters, and rand() represents a random number from 0 to 1;

[0112] Step 53: Map the unit sampling points to the actual range to generate actual sampling points. Here, the Hollomon constitutive model is used as an example, but it is not limited to the Hollomon constitutive model:

[0113] K=K min +(K max -K min )*sK

[0114] n=n min +(n max -n min )*sn

[0115] Among them, K is the hardening coefficient of the Hollomon material constitutive model, n is the hardening index of the Hollomon material constitutive parameter, K min With K max is the minimum and maximum value of the value range of K, n min With n max is the minimum and maximum value of the range of n, sK and sn are the values ​​of K and n at the unit sampling point respectively, and the Latin hypercube sampling result diagram is as follows Figure 4 shown.

[0116] Furthermore, the constitutive parameter sampling points obtained by the Latin hypercube theory are input into the Abaqus model, such as Figure 5 As shown in the figure; the model is used to simulate the indentation process and obtain the indentation depth-indentation force curve, and the sampling information vector corresponding to the neural network is extracted from the curve, thereby constructing the mapping relationship between the constitutive information and the indentation depth-indentation force curve.

[0117] Step 6: Use the database built by Abaqus to train the neural network. Through training, the neural network can construct a nonlinear mapping relationship between the material constitutive parameters and the indentation force-indentation depth curve;

[0118] Furthermore, in step 6, the specific steps are:

[0119] Step 61: Training of Neural Network:

[0120] In MATLAB, the Levenberg-Marquardt algorithm was used to iteratively optimize the weights and biases in the neural network to minimize the error between prediction and reality. Specifically, the training database was divided into three parts: 70% for determining weights and biases, 15% for testing their accuracy, and 15% for verifying their effectiveness.

[0121] [net,tr]=train(net,X,Y);

[0122] net.divideParam.trainRatio=70%

[0123] net.divideParam.testRatio=15%

[0124] net.divideParam.valRatio=15%

[0125] Among them, X is the input matrix, which corresponds to the matrix composed of the Latin hypercube sampling results; Y is the output matrix, which corresponds to the matrix composed of the Abaqus calculation results. The training process is as follows Figure 6 shown.

[0126] Step 62: Packaging of Neural Network:

[0127] After the neural network is trained, it can reflect complex nonlinear mapping relationships;

[0128] The trained neural network is packaged into a MATLAB .m file and called when the target function is subsequently established.

[0129] Step 7: Through the genetic algorithm, the least square error between the indentation force-indentation depth curves obtained by the test and the neural network calculation is minimized, thereby obtaining the mechanical properties of the bearing alloy layer material.

[0130] Furthermore, in step 7, the specific method is:

[0131] Step 71: Create the objective function

[0132] The least square error between the predicted value simulated by the neural network and the corresponding data on the penetration force-depth curve obtained by the experiment is set as the objective function, specifically:

[0133]

[0134] Where F is the objective function, is the indentation force-indentation depth result predicted by the neural network, X is the vector composed of independent parameters of the material constitutive model; Y exp is the indentation force-indentation depth result obtained through the test;

[0135] Step 72: Optimize using genetic algorithm, such as Figure 7 As shown, specifically;

[0136] (1) Initialization: generate m individuals to form the initial population;

[0137] (2) Calculate the fitness function for each individual;

[0138] (3) Select the best individuals based on the fitness function, keep the individuals with high fitness function, and discard the individuals with low fitness function;

[0139] (4) Performing a crossover operation on the selected population to produce offspring;

[0140] (5) Perform mutation operations on the population with a certain probability;

[0141] (6) Conduct convergence judgment on the newly generated population and decide whether to terminate the algorithm;

[0142] (7) If the convergence requirement is not met, return to step (2) for iteration;

[0143] Through the genetic algorithm, X is used as the optimization variable and F is used as the objective function for iterative solution, so that the objective function converges to the minimum, thereby obtaining the actual material constitutive model parameter X.

[0144] The indentation test results Figure 8 As an example, the bearing alloy layer indentation depth-indentation force curve shown in FIG. 1 is used to obtain the mechanical properties obtained by the present invention, that is, the stress-strain curve is as follows: Figure 9 shown.

[0145] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A neural network alloy layer material mechanical property identification method based on genetic algorithm, characterized in that: The following steps are involved: Step 1: Select reasonable test parameters based on the microstructure and macrostructure of the bearing alloy layer material; Step 2: Use an indentation test to test the bearing alloy layer and obtain an indentation force-indentation depth curve; Step 3: Select the material constitutive model; Step 4: Create a neural network; In step 4, the specific method is: In a multi-layer neural network, the weighted sum of the inputs to each layer is: The weighted sum of the input is processed through the activation function to get the output: in, Represents the output of the i-th neuron in the previous layer; and are the weight and bias corresponding to the j-th neuron in the l-th layer, and f() is the activation function; Step 5: Based on the Latin hypercube theory, Abaqus numerical analysis software is used to build a network training database; Step 6: Use the database built by Abaqus to train the neural network. Through training, the neural network can construct a nonlinear mapping relationship between the material constitutive parameters and the indentation force-indentation depth curve; Step 7: Through the genetic algorithm, the least square error between the indentation force-indentation depth curves obtained by the test and the neural network calculation is minimized, thereby obtaining the mechanical properties of the bearing alloy layer material.

2. The method for identifying mechanical properties of alloy layer materials using a neural network based on a genetic algorithm according to claim 1, characterized in that: In step 1, the specific test requirements are: (1) The maximum indentation depth is less than 1 / 10 of the thickness of the bearing alloy layer; (2) The maximum indentation depth is greater than 20 times the surface roughness Ra value of the bearing alloy layer; (3) The ratio between the maximum indentation depth and the indenter radius is greater than 0.3; (4) The contact area should contain as many material grains as possible.

3. The method for identifying mechanical properties of alloy layer materials using a neural network based on a genetic algorithm according to claim 1, characterized in that: In step 2, the specific experimental steps are: Step 21: Clamp the sample stably; Step 22: Adjust the indenter to be directly above the surface of the sample to be tested; Step 23: Set the pressing speed so that the pressing head presses in slowly; Step 24: When the pressure reaches the set limit, the pressure head slowly withdraws.

4. The method for identifying mechanical properties of alloy layer materials using a neural network based on a genetic algorithm according to claim 1, wherein: In step 3, the specific method is: Based on prior knowledge or subsequent needs, select a reasonable Hollomon constitutive model or Swift constitutive model.

5. The method for identifying mechanical properties of alloy layer materials using a neural network based on a genetic algorithm according to claim 1, characterized in that: In step 5, when sampling the material constitutive parameters, the Latin hypercube theory is used to improve the sampling performance. The specific steps are: Step 51: Set the number of sampling points m; Step 52: Generate m unit sampling points first. The generation method is: s[i,j]=(i+rand()) / m Among them, s[i,j] represents the value of the i-th sampling point in the j-th dimension, that is, the value of the i-th sampling point on different constitutive parameters, and rand() represents a random number from 0 to 1; Step 53: Map the unit sampling points to the actual range to generate actual sampling points. Here, the Hollomon constitutive model is used as an example, but it is not limited to the Hollomon constitutive model: K=K min +(K max -K min )*sK n=n min +(n max -n min )*sn Among them, K is the hardening coefficient of the Hollomon material constitutive model, n is the hardening index of the Hollomon material constitutive parameter, K min With K max is the minimum and maximum value of the value range of K, n min With n max are the minimum and maximum values ​​of the range of n, sK and sn are the values ​​of K and n at the unit sampling point respectively.

6. The method for identifying mechanical properties of alloy layer materials using a neural network based on a genetic algorithm according to claim 1, characterized in that: In step 6, the specific steps are: Step 61: Training of Neural Network: Use optimization algorithms to iteratively optimize the weights and biases in the neural network to minimize the error between prediction and reality; The training database is divided into three parts: 70% is used to determine the weights and biases, 15% is used to test its accuracy, and 15% is used to verify its effectiveness; Step 62: Packaging of Neural Network: After the neural network is trained, it can reflect complex nonlinear mapping relationships; The trained neural network is packaged into a MATLAB .m file and called when the target function is subsequently established.

7. The method for identifying mechanical properties of alloy layer materials using a neural network based on a genetic algorithm according to claim 1, characterized in that: In step 7, the specific method is: Step 71: Create the objective function The least square error between the predicted value simulated by the neural network and the corresponding data on the penetration force-depth curve obtained by the experiment is set as the objective function, specifically: Where F is the objective function, is the indentation force-indentation depth result predicted by the neural network, X is the vector composed of independent parameters of the material constitutive model; Y exp is the indentation force-indentation depth result obtained through the test; Step 72: Optimize using genetic algorithms (1) Initialization: generate m individuals to form the initial population; (2) Calculate the fitness function for each individual; (3) Select the best individuals based on the fitness function, keep the individuals with high fitness function, and discard the individuals with low fitness function; (4) Performing a crossover operation on the selected population to produce offspring; (5) Perform mutation operations on the population with a certain probability; (6) Conduct convergence judgment on the newly generated population and decide whether to terminate the algorithm; (7) If the convergence requirement is not met, return to step (2) for iteration; Through the genetic algorithm, X is used as the optimization variable and F is used as the objective function for iterative solution, so that the objective function converges to the minimum, thereby obtaining the actual material constitutive model parameter X.

Images