Machine learning prediction method for impact fatigue life of metal in additive manufacturing

Abstract

The invention relates to an additive manufacturing metal impact fatigue life-oriented machine learning prediction method, and belongs to the field of machine learning and metal materials. According to the invention, a set of machine learning model capable of obtaining complex fatigue and impact fatigue performance through a manufacturing process and / or material microstructure characteristics and / or material basic mechanical experiments is established, and a rapid prediction method of the machine learning model is realized, and simulation of an additive manufacturing material considering a scale effect is combined; a data set of each test result of the material can be effectively expanded by utilizing a simulation method which is relatively easy to implement, more effective data is provided compared with a traditional machine learning method, and the model quality is favorably improved; according to the machine learning rapid prediction method for the additive manufacturing metal impact fatigue life, a traditional impact fatigue life prediction method is combined with productivity tools such as artificial intelligence machine learning, and the problem that the impact fatigue life of an additive manufacturing advanced material is difficult to efficiently predict through a traditional method is solved.

Description

Technical Field

[0001] The present invention relates to a machine learning prediction method for metal impact fatigue life in additive manufacturing, which is a method for quickly predicting impact fatigue life by using machine learning based on experimental data and simulation expansion data, and belongs to the field of machine learning and metal materials. Background Art

[0002] Impact fatigue is one of the harsh working conditions faced by vehicles in many fields of national economy and people's livelihood, such as aerospace, rail transit and ships. Therefore, structures and materials that bear such loads need to undergo impact fatigue tests or simulations to guide the design work. Materials and structures made by advanced manufacturing processes such as additive manufacturing often have higher specific performance than traditional manufacturing, and are widely used in fields such as aerospace that pursue structural efficiency. However, structures made by these advanced manufacturing processes often have large performance dispersion, among which fatigue, impact fatigue and other performances that are significantly affected by the microstructure, defects, surface quality and residual stress of additive manufacturing materials have a higher dispersion level. Both experimental and simulation verification require huge test costs and test time, which restricts the further application of advanced manufacturing processes such as additive manufacturing that can improve structural efficiency.

[0003] Advanced manufacturing processes represented by additive manufacturing materials generally have more adjustable manufacturing process parameters, such as powder bed melting, direct energy deposition, material extrusion and other process routes. Powder bed melting methods also include laser selective melting, electron beam melting, selective laser sintering and other technologies. Among them, the laser selective melting process with the highest manufacturing quality includes important process parameters such as powder particle size, powder bulk, laser power, laser spot shape and size, melting channel width and thickness, laser scanning speed, substrate preheating temperature, etc. In addition, there are post-processing methods and their process parameters, and machining related process parameters. These process parameters do not have clear optimal parameters, and often need to select different parameter combinations according to the structure and required performance. This leads to many traditional performance prediction methods that cannot effectively predict the performance of advanced manufacturing materials and can only rely on a large number of experiments. Fatigue, especially impact fatigue experiments, are highly dispersed, have long experimental cycles and high costs. Simulation calculations are also limited by the balance between precision and efficiency. Therefore, the current research and development of impact fatigue performance is relatively limited. Structures that bear impact fatigue loads often use high safety redundancy or overly conservative life design to ensure safety, resulting in a large waste of structural efficiency. To this end, the present invention discloses a machine learning prediction method for the impact fatigue life of additively manufactured metals to achieve rapid prediction of the impact fatigue life of additively manufactured and other advanced manufacturing metal structures, so as to facilitate fatigue research in additive manufacturing and promote the improvement of impact fatigue performance research efficiency and structural efficiency. Summary of the invention

[0004] The purpose of this invention is to propose a specific method based on machine learning using experimental and simulation data, to develop a machine learning model that effectively identifies the microstructure and static and dynamic properties of additively manufactured materials and then learns their relationship with fatigue performance, thereby forming an effective method for quickly acquiring complex mechanical properties through simple experimental tests, and to provide predictive evaluation methods and research ideas for the impact fatigue performance of additively manufactured materials in multiple fields.

[0005] The present invention achieves the above-mentioned purpose through the following technical scheme, which is a machine learning prediction method for metal impact fatigue life in additive manufacturing, and mainly includes the following steps:

[0006] S1: Select process parameters within the mature process range to perform additive manufacturing of experimental parts;

[0007] S2: Conduct microstructural characterization of additively manufactured test pieces, i.e., microstructural characterization of the surface and cross-section by optical microscope and electron microscope, and retain images and relevant microstructural dimension data; S3: Conduct basic mechanical experiments, including uniaxial tensile tests to determine the Young's modulus, yield limit, elongation after fracture, and damage characteristics of the material, and conduct Hopkinson bar uniaxial impact tests to determine its rate effect, or directly retain the electrical signal data in the tensile test and impact test without calculating the relevant material parameters;

[0008] S4: Conduct impact fatigue experiments to determine the number of impact fatigue cycles under specific energy or strain; S5: Use simulation methods that take into account microstructure effects, rate effects, and damage characteristics, combine basic mechanical experimental results to simulate impact fatigue experiments, and expand the required data sets to improve data quality;

[0009] S6: Build a machine learning model, perform data processing, training optimization, and model evaluation, and freeze the trained model;

[0010] S7: Through the trained model, the impact fatigue life of the corresponding material is directly and quickly predicted using the process parameters selected by S1, the microstructure characteristics obtained by S2, or the experimental data of S3, thereby updating the database and guiding the optimization of the manufacturing process.

[0011] Furthermore, in the step S2, the cross section is divided into a cut cross section after manufacturing and a fracture cross section after stretching or impact stretching, and the two types of cross sections are independent microstructural characteristic variables.

[0012] Furthermore, in the step S2, the microstructural features are obtained by calibrating the microstructural parameters as the parameters of the microstructural features using the images of the material's microscopic morphology OM or SEM, phase XRD or EDS, texture EBSD or dislocation interface TEM.

[0013] Furthermore, the microstructure parameter is one or more of average grain size, size distribution, aspect ratio, number of defects, and dislocation density.

[0014] Furthermore, in step S3, the basic mechanical experiments include experiments reflecting the material properties related to the elastic properties, plastic properties, fracture toughness, rate effect and volume effect of the material and the impact fatigue process.

[0015] Furthermore, in step S5, simulation is performed by using the finite element method considering the impact fatigue characteristics of additively manufactured metals, and the envelope of the data set is expanded by interpolating within the range of the experimental data or extrapolating outside the range; or the data set is expanded by directly interpolating the experimental data through statistical methods.

[0016] Furthermore, in step S6, the machine learning model is determined according to the actual data scale and the number of variables in the data, and a deep learning framework with multiple fully connected layers is selected. For situations requiring image recognition processing, a convolutional neural network model is selected.

[0017] Furthermore, in step S6, the specific implementation process of establishing, training, and debugging the machine learning model is as follows:

[0018] S61: preprocessing relevant data in model training;

[0019] S62: Build a machine learning model, including an input layer, several fully connected layers and BN layers, and an output layer, and set the relevant parameters in each layer, and specify the optimizer, learning rate, and loss function parameters;

[0020] S63: Select a training set and a test set from the data set. Both subsets contain all the main features of the original data set. The data volume of the training set and the test set is between 1:1 and 5:1, and is optimized and adjusted according to the training results.

[0021] S64: Model training using training set data;

[0022] S65: Test the accuracy and quality of the model through test set data, conduct visual analysis of key variables, and verify whether the model can accurately predict the impact fatigue performance under qualified manufacturing process parameters;

[0023] S66: Optimize and adjust model parameters according to the test results;

[0024] S67: Evaluate the trained model, freeze relevant parameters and save the model.

[0025] Furthermore, in the step S61, the method for preprocessing the relevant data, in addition to the process parameters, involves mainly image data, experimental electrical signal data, and simulation data, wherein the image data needs to be processed or image elements need to be extracted as microstructure feature parameters, the experimental electrical signal data needs to be enhanced, denoised, interpolated or devalued, and the simulation data needs to be extracted and its sampling rate needs to be unified with that of the experimental data.

[0026] Further, in step S62, the loss function is defined as or Where n is the total number of samples in the training set, f(i) is the predicted value of the i-th sample, and f′(i) is the labeled value of the i-th sample.

[0027] Furthermore, in step S64, the larger the data set size required for model training, the larger the single batch training amount, and the more training iteration cycles, the lower the optimizer learning rate.

[0028] Furthermore, in step S66, the model parameters to be optimized and adjusted include the neural network type, structure, number of neurons in a layer, kernel initialization, Dropout layer, and activation function selection.

[0029] Beneficial effects of the present invention

[0030] 1) A machine learning model that can obtain complex fatigue and impact fatigue through manufacturing process and / or material microstructure characteristics and / or basic material mechanical experiments and realize its rapid prediction method is established, which can greatly reduce costs and improve efficiency compared with existing impact fatigue test research methods;

[0031] 2) Combined with the additive manufacturing material simulation that takes into account the scale effect, it is possible to effectively expand the data set of various material test results using a relatively easy-to-implement simulation method. Compared with traditional machine learning methods, it has more effective data and helps improve the model quality.

[0032] 3) A machine learning rapid prediction method for the impact fatigue life of additively manufactured metals was implemented, combining traditional impact fatigue life prediction methods with productivity tools such as artificial intelligence machine learning, solving the problem that traditional methods are difficult to efficiently predict the impact fatigue life of additively manufactured advanced materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of rapid prediction of impact fatigue life of additive manufacturing materials by machine learning;

[0034] Figure 2 Schematic diagram of the neural network structure of the fatigue prediction embodiment of the present invention;

[0035] Figure 3 It is the prediction of the extended data set model of a single experiment - fatigue experiment in an embodiment of the present invention;

[0036] Figure 4 It is the convergence situation under different model parameters in the embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention is described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0038] like Figure 1 As shown, the present invention achieves the above-mentioned purpose through the following technical scheme, which is a machine learning prediction method for metal impact fatigue life in additive manufacturing, and mainly includes the following steps:

[0039] S1: Select process parameters within the mature process range, such as using smaller spot size and melting channel width for fine structures, using higher substrate preheating temperature for materials that require grain control, and performing additive manufacturing of experimental parts;

[0040] S2: Conduct microstructural characterization on additively manufactured experimental parts, i.e., microstructural characterization on the surface and cross-section using optical microscopes and electron microscopes, and retain images and related microstructural dimension data; the cross-sections are divided into cut cross-sections after manufacturing and fracture cross-sections after stretching and impact stretching, and the two types of cross-sections are independent microstructural characteristic variables.

[0041] Figure 1 The microstructural features shown in the figure are images of the material's microscopic morphology OM or SEM, phase XRD or EDS, texture EBSD or dislocation interface TEM; the microstructural parameters are calibrated as parameters of the microstructural features. The microstructural parameters are one or more of the average grain size, size distribution, aspect ratio, number of defects, and dislocation density.

[0042] S3: Conduct basic mechanical experiments, including uniaxial tensile tests to determine the Young's modulus, yield limit, elongation after fracture, and damage characteristics of the material, and conduct Hopkinson bar uniaxial impact tests to determine its rate effect, or directly retain the electrical signal data in the tensile test and impact test without calculating the relevant material parameters; basic mechanical experiments include experiments that reflect the elastic properties, plastic properties, fracture toughness, rate effect and volume effect of the material and other material properties related to the impact fatigue process, such as three-point or four-point bending tests in statics, pendulum and drop hammer impact tests in impact experiments, but uniaxial tensile and Hopkinson bar tests are preferred.

[0043] S4: Conduct impact fatigue experiments to determine the number of impact fatigue cycles under specific energy or strain;

[0044] S5: Through simulation methods that take into account microstructural effects, rate effects, and damage characteristics, combined with basic mechanical experimental results, impact fatigue experiments are simulated to expand the required data set to improve data quality; through finite element and other methods that consider the impact fatigue characteristics of additively manufactured metals, simulations are performed, interpolation is performed within the range of experimental data or partial extrapolation is performed outside the range to expand the envelope of the data set; or statistical methods are directly used to interpolate experimental data to expand the data set to improve data acquisition efficiency.

[0045] S6: Establish a machine learning model, perform data processing, training optimization and model evaluation, and freeze the trained model; the machine learning model is determined based on the actual data scale and the number of variables in the data, and a deep learning framework with multiple fully connected layers is selected. For situations where image recognition processing is required, a convolutional neural network model is selected.

[0046] The specific implementation process of establishing, training, and debugging the machine learning model is as follows:

[0047] S61: Preprocess the relevant data in the model training; the method for preprocessing the relevant data, in addition to the process parameters, mainly involves image data, experimental electrical signal data, and simulation data, wherein the image data needs to be processed or the image elements need to be extracted as microstructure feature parameters, the experimental electrical signal data needs to be enhanced, denoised, interpolated or devalued, and the simulation data needs to be extracted and unified with the sampling rate and other relevant indicators of the experimental data.

[0048] S62: Construct a machine learning model, including an input layer, several fully connected layers and BN layers, and an output layer, and set the relevant parameters in each layer, and specify the optimizer, learning rate, and loss function parameters; the number of neural network layers and the number of neurons in each layer should match the size and complexity of the data set to avoid overfitting or underfitting, and appropriate regularization parameters should be introduced to prevent overfitting. The loss function is defined as or Where n is the total number of samples in the training set, f(i) is the predicted value of the i-th sample, and f′(i) is the labeled value of the i-th sample.

[0049] S63: Select a training set and a test set from the data set. Both subsets contain all the main features of the original data set. The data volume of the training set and the test set is between 1:1 and 5:1, and is optimized and adjusted according to the training results.

[0050] S64: Perform model training using training set data; the larger the data set size, the larger the single batch training amount, and the more training iteration cycles, the lower the optimizer learning rate.

[0051] S65: Test the accuracy and quality of the model through test set data, conduct visual analysis of key variables, and verify whether the model can accurately predict the impact fatigue performance under qualified manufacturing process parameters;

[0052] S66: Optimize and adjust model parameters according to the test results; the optimized and adjusted model parameters include neural network type, structure, number of neurons in a layer, kernel initialization, Dropout layer, activation function selection, etc.

[0053] S67: Evaluate the trained model, freeze relevant parameters and save the model.

[0054] S7: Through the trained model, the impact fatigue life of the corresponding material is directly and quickly predicted using the process parameters selected in S1, the microstructure characteristics obtained in S2, or the experimental data in S3, and the database is updated and the manufacturing process optimization is guided. The trained model depends on the content and quality of the input and output data of its training data set, and continuously maintains and updates the impact fatigue database through verification iteration to improve the model quality; through the trained model, the impact fatigue life of the corresponding material is directly and quickly predicted using process parameters, microstructure characteristics, or basic mechanical experimental data, and the database is used to promote the optimization of additive manufacturing process.

[0055] The present invention aims at the problem of impact fatigue energy input value of an additively manufactured metal material with specific material parameters under fixed fatigue cycles, and combines a set of expanded data sets for machine learning modeling and training. The specific implementation of this embodiment is as follows:

[0056] To establish a neural network model, this embodiment uses multiple layers of the most basic fully connected layers to build it, starting from the input layer, through multiple layers of fully connected layers and BN layers, and finally connecting to the output layer. The neural network structure diagram of the embodiment is shown in Figure 2 shown.

[0057] The additive manufacturing material used in this embodiment has the following input parameters: Young's modulus, yield limit, elongation after fracture, impact toughness, strain hardening coefficient, microstructure characteristic parameters, and the output parameter is the impact energy under a fixed number of fatigue cycles. Among the selected material parameters, the two with the most significant changes (elongation after fracture and microstructure characteristic parameters) are selected and visualized in Figure 3 As shown in the left picture.

[0058] In this embodiment, RELU is selected as the activation function of each fully connected layer, and a first-order regularization term with a parameter of 0.0001 is introduced. The "random_normal" kernel is initialized in the first layer, and the optimizer is SGD, and its learning rate is set to 10 -6 ,The loss function and evaluation function in model training are selected as MSE, and the number of model iterations is set to 200.

[0059] After model training and debugging, the predicted output parameters of this embodiment for the two independent variables in the test set within the corresponding original input data set range are as follows: Figure 3 As shown in the figure on the right, it can be seen that the results predicted by the model are basically accurate and have high feasibility.

[0060] In this embodiment, the relevant parameters in the model are adjusted, and the convergence of the model training process under different parameter combinations is as follows: Figure 4 As shown, the left and right figures are respectively the relationship diagrams of the loss function and the MSE error function with the number of training iterations. It can be seen that different model parameter selections will affect the convergence speed and accuracy of the results, but the appropriately selected parameters have relatively high quality. It should be pointed out that according to the needs of implementation, the steps described in this embodiment can be split and combined into new steps to achieve the purpose of the present invention. It is easy for those skilled in the art to understand that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A machine learning prediction method for metal impact fatigue life in additive manufacturing, characterized in that: And mainly includes the following steps: S1: Select process parameters within the mature process range to perform additive manufacturing of experimental parts; S2: Conduct microstructural characterization of additively manufactured experimental parts, i.e., microstructural characterization of the surface and cross-section using optical microscopes and electron microscopes, and retain images and related microstructural dimension data; S3: Conduct basic mechanical experiments, including uniaxial tensile tests to determine the Young's modulus, yield limit, elongation after fracture, and damage characteristics of materials, and conduct Hopkinson bar uniaxial impact tests to determine its rate effect, or directly retain the electrical signal data in the tensile test and impact test without measuring the relevant material parameters; S4: Conduct impact fatigue experiments to determine the number of impact fatigue cycles under specific energy or strain; S5: Through the simulation method that takes into account the microstructure effect, rate effect and damage characteristics, the impact fatigue experiment is simulated in combination with the basic mechanical experimental results, and the required data set is expanded to improve the data quality; S6: Build a machine learning model, perform data processing, training optimization, and model evaluation, and freeze the trained model; S7: Through the trained model, the impact fatigue life of the corresponding material is directly and quickly predicted using the process parameters selected by S1, the microstructure characteristics obtained by S2, or the experimental data of S3, thereby updating the database and guiding the optimization of the manufacturing process.

2. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 1, characterized in that: In step S2, the cross section is divided into a cut cross section after manufacturing and a broken cross section after experiment, and the two types of cross sections are independent microstructure characteristic variables.

3. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 1, characterized in that: In step S2, the microstructural features are images of the material's microscopic morphology OM or SEM, phase XRD or EDS, texture EBSD or dislocation interface TEM; and the microstructural parameters are calibrated as parameters of typical microstructural features of grain size and defect distribution.

4. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 3, characterized in that: The microstructure parameters are one or more of average grain size, size distribution, aspect ratio, number of defects, and dislocation density.

5. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 1, characterized in that: In step S3, the basic mechanical experiments include experiments reflecting the material's elastic properties, plastic properties, fracture toughness, rate effect and volume effect and the material properties related to the impact fatigue process.

6. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 1, characterized in that: In step S5, simulation is performed by using the finite element method considering the impact fatigue characteristics of additively manufactured metals, and the envelope of the data set is expanded by interpolating within the range of the experimental data or partially extrapolating outside the range; or the data set is expanded by directly interpolating the experimental data using a statistical method.

7. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 1, characterized in that: In step S6, the machine learning model is determined according to the actual data scale and the number of variables in the data, and a deep learning framework with multiple fully connected layers is selected. For situations where image recognition processing is required, a convolutional neural network model is selected.

8. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 1, characterized in that: In step S6, the specific implementation process of establishing, training, and debugging the machine learning model is as follows: S61: preprocessing relevant data in model training; S62: Build a machine learning model, including an input layer, several fully connected layers and BN layers, and an output layer, and set the relevant parameters in each layer, and specify the optimizer, learning rate, and loss function parameters; S63: Select a training set and a test set from the data set. Both subsets contain all the main features of the original data set. The data volume of the training set and the test set is between 1:1 and 5:1, and is optimized and adjusted according to the training results. S64: Model training using training set data; S65: Test the accuracy and quality of the model through test set data, conduct visual analysis of key variables, and verify whether the model can accurately predict the impact fatigue performance under qualified manufacturing process parameters; S66: Optimize and adjust model parameters according to the test results; S67: Evaluate the trained model, freeze relevant parameters and save the model.

9. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 8, characterized in that: In the step S61, the method for preprocessing the relevant data, in addition to the process parameters, the data involved are mainly image data, experimental electrical signal data, and simulation data, among which the image data needs to be processed or the image elements need to be extracted as microstructure feature parameters, the experimental electrical signal data needs to be enhanced, denoised, interpolated or devalued, and the simulation data needs to be extracted and its sampling rate needs to be unified with that of the experimental data.

10. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 8, characterized in that: In step S62, the loss function is defined as or Where n is the total number of samples in the training set, f(i) is the predicted value of the i-th sample, and f ′ (i) is the label value of the i-th sample.

11. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 8, characterized in that: In step S64, the training set data for model training requires that the larger the data set size is, the larger the single batch training amount is, and the more training iteration cycles are, the lower the optimizer learning rate is.

12. The machine learning prediction method for metal impact fatigue life in additive manufacturing according to claim 8, characterized in that: In step S66, the model parameters to be optimized and adjusted include the neural network type, structure, number of neurons in a layer, kernel initialization, Dropout layer, and activation function selection.

Images