High-temperature alloy creep / thermal exposure condition prediction method based on machine learning

Abstract

The invention relates to a machine learning-based high-temperature alloy creep / thermal exposure condition prediction method, which is used for realizing accurate prediction of high-temperature alloy creep / thermal exposure conditions by utilizing a fusion neural network constructed in machine learning based on a tissue image-creep / thermal exposure time-creep / thermal exposure tissue parameter data set. According to the method, organization parameters applied to prediction are accurately screened out through a large amount of research experience summary of material experts and a machine learning correlation analysis method, image data are effectively expanded through a certain method, and reference significance is provided for some complex and small-data-volume machine learning problems in the field of material science. According to the prediction method disclosed by the invention, the service conditions of the high-temperature alloy used for different parts of the turbine blade are comprehensively, quickly and effectively evaluated, so that the high-temperature alloy used for the turbine blade is reasonably designed, the harm is reduced, the cost is reduced, and industrial large-scale production, popularization and application are facilitated.

Description

Technical Field

[0001] The present invention belongs to the field of material science and engineering technology, and specifically relates to a method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning. Background Art

[0002] Due to its excellent high-temperature comprehensive performance, high-temperature alloys have been widely studied and applied in the past decades, especially in the manufacture of hot-end high-temperature components such as aircraft engines and industrial gas turbine hot-end turbine blades. One of the outstanding characteristics of high-temperature alloys is that they can maintain a long service life under extreme temperature and stress conditions. However, with the extension of service time, the creep behavior and thermal exposure effect of high-temperature alloys gradually emerge, becoming a key factor affecting their performance and life.

[0003] Creep refers to the irreversible deformation of metal materials over time under high temperature stress, and creep damage is one of the most common failure modes of high-temperature alloys during service. Due to the uneven distribution of temperature and stress in different parts of the turbine blades, the degree of creep damage varies. Accurately predicting the creep behavior of high-temperature alloys under complex service conditions or the organizational changes caused by thermal exposure is crucial to optimizing material selection and design. It can not only reduce R&D costs, but also significantly improve the service performance and safety of the alloy.

[0004] In order to solve this problem, a method is needed to accurately predict the creep conditions of high-temperature alloys, so as to provide guidance for the rational design and material selection of high-temperature alloy turbine blades, reduce R&D costs and cycles, and promote the widespread application of high-temperature alloys.

[0005] Thermal exposure refers to the microstructural changes and performance degradation caused by long-term exposure of high-temperature alloys to high-temperature environments. When high-temperature components such as turbine blades are in high-temperature service environments for a long time, creep may accelerate. Therefore, accurate prediction of high-temperature alloys under different thermal exposure conditions can provide a theoretical basis for the rational use of materials and the extension of service life.

[0006] At present, most of the research on thermal exposure of high-temperature alloys focuses on the laws of microstructure evolution and the impact on performance, and there are relatively few studies on the prediction of thermal exposure conditions. There are still many shortcomings in the prediction methods for creep of high-temperature alloys. Traditional methods rely on a large amount of experimental data, and the experiments are costly and time-consuming. At the same time, the existing methods are easily affected by human factors, there are large deviations in data processing, and the prediction accuracy is low. There is little public data on high-temperature alloys, and high-throughput experiments are not only costly, but also time-consuming. Existing studies often only focus on creep rupture life, but do not accurately predict the evolution of specific creep temperatures, stresses, and thermal exposure conditions.

[0007] For example, Fu et al. constructed a prediction model for creep conditions of high-temperature alloys through artificial neural networks, but this method relies on a large amount of experimental data and is difficult to solve the problems of data shortage and experimental costs. Xu et al. built an artificial neural network model by combining manually extracted organizational parameters and image processing technology. However, this method ignores some key information during data processing, resulting in limited prediction accuracy.

[0008] Through the search of existing published patents, the following similar published patents were found:

[0009] 1. The invention patents with publication numbers CN115035963A and CN105628511A, although using machine learning or correction methods to predict creep life, still fail to provide accurate predictions of creep and thermal exposure conditions.

[0010] 2. The invention patent with publication number CN113376015A discloses a method for rapid characterization and analysis of the microstructural evolution of nickel-based single crystal high-temperature alloys. The method integrates high-throughput experiments, high-throughput characterization and high-throughput statistical technology to efficiently establish a quantitative relationship between the microstructural evolution of nickel-based single crystal alloys during high-temperature creep. Its high-throughput method will lead to high costs and high energy consumption. The training set, validation set and test set of the adopted U-Net network structure have a large amount of input and output data. Moreover, the U-Net network is a convolutional neural network, which has the disadvantages of too much redundancy and the inability to achieve both receptive field and positioning accuracy.

[0011] In summary, the existing technology has not considered the use of deep learning to extract features of creep / heat exposure structures of high-temperature alloys under different conditions, has not considered the role of convolutional neural networks and artificial neural networks combined with fusion neural networks on data sets, and has not considered how to accurately predict creep / heat exposure conditions with small data volumes. Summary of the invention

[0012] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for predicting creep / thermal exposure conditions of high-temperature alloys based on machine learning.

[0013] The present invention solves the technical problem by the following technical solutions:

[0014] A method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning, the method comprising the following steps:

[0015] S1. Collect creep / heat exposure tissue parameter data and creep / heat exposure tissue images of high-temperature alloys at different temperatures and times. Each tissue image corresponds to a set of data, including creep / heat exposure temperature, creep / heat exposure time and tissue parameter data.

[0016] S2, cropping the creep / heat exposure tissue images selected in S1 respectively to obtain small-sized unit RVE images, so that the images can represent the characteristics of the original images, and randomly rotating and flipping them, and the RVE units obtained from the same tissue image have the same tissue parameters, and the above RVE units and corresponding tissue parameters constitute a data set;

[0017] S3, using the Pearson characteristic correlation calculation method to select creep / heat exposure tissue parameters related to temperature and stress from the data set constructed in S2, using the corresponding data of the selected tissue parameters and the RVE unit as input feature data, and temperature and stress as output data, to build a fusion neural network;

[0018] S4. Training the fusion neural network constructed in S3, and verifying the obtained model on the data that have not participated in the training and testing, and predicting the creep temperature, creep stress or thermal exposure temperature corresponding to the tissue image in the data.

[0019] Moreover, the data in S1 at least include creep / heat exposure tissue images under 15 different test conditions; the high-temperature alloy in S1 is MarM247LC high-temperature alloy or DZ125 high-temperature alloy or equiaxed MarM247 high-temperature alloy; the tissue parameters in S1 include γ' volume fraction, γ channel width, γ' thickness, γ' average size, γ' film thickness and rafting degree.

[0020] Moreover, the tissue parameters of S1 and S3 are selected based on a large amount of previous research experience and are further screened through Pearson correlation analysis.

[0021] Moreover, the difference between the tissue parameter values ​​of each group of unit RVE in S2 and the tissue parameter values ​​of the original image is within 5%, and the RVE obtained for each group of original tissue images in S2 is 10 to 200, with a size of 600×600, and a total of 300 to 1300 groups of data are obtained.

[0022] Moreover, the fusion neural network model in S3 is a fusion structure of the convolutional neural network EfficientNetB7 and the artificial neural network.

[0023] Furthermore, the creep temperature, creep stress and predicted value of thermal exposure temperature corresponding to the creep / thermal exposure tissue image in S4 are R 2 All are above 0.9.

[0024] The advantages and beneficial effects of the present invention are:

[0025] The prediction method of the present invention can make a rapid and accurate prediction of the creep / thermal exposure conditions of a high-temperature alloy using a relatively small number of data sets within the time range of the creep / thermal exposure test conditions covered by the data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of the present invention;

[0027] Figure 2 It is a schematic diagram of a typical small-sized unit RVE cutout of the present invention;

[0028] Figure 3 This is a schematic diagram of the structure of the fusion neural network model built by the present invention. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0030] like Figure 1 As shown, a method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning, the steps of the method are:

[0031] S1. Collect creep / heat exposure tissue parameter data and creep / heat exposure tissue images of high-temperature alloys at different temperatures and times. Each tissue image corresponds to a set of data, including creep / heat exposure temperature, creep / heat exposure time and tissue parameter data.

[0032] S2. Crop the creep / heat-exposed tissue images selected in S1 to obtain small-sized unit RVE images, such as Figure 2 As shown, the image can represent the characteristics of the original image, and is randomly rotated and flipped, and the RVE units obtained from the same tissue image have the same tissue parameters, and the above RVE units and corresponding tissue parameters constitute a data set;

[0033] S3, using the Pearson characteristic correlation calculation method to select creep / heat exposure tissue parameters related to temperature and stress from the data set constructed in S2, using the selected tissue parameter corresponding data and RVE unit as input feature data, and temperature and stress as output data, to build a fusion neural network, such as Figure 3As shown, it includes two parts of structure; first, for the feature extraction of microscopic tissue images, the EfficientNet (model E) convolutional network method is used to flatten the extracted image features, and then the features are nonlinearly mapped, and three layers of fully connected layers with 512, 16 and 8 neurons are built. Each fully connected layer is followed by a batch normalization (BN) layer and an activation function ReLU, which improves the nonlinear mapping ability of the model; for each microscopic tissue image, there is a corresponding tissue parameter, which is used as data input, and two layers of fully connected layers with 16 and 8 neurons are built. The two parts are connected with a fully connected layer to generate a structure with 16 neurons, and then the mapping of creep / heat exposure conditions is completed with a structure with two layers of 8 and 8 neurons. This part can be considered as a BP artificial neural network structure;

[0034] S4. Training the fusion neural network constructed in S3, and verifying the obtained model on the data that have not participated in the training and testing, and predicting the creep temperature, creep stress or thermal exposure temperature corresponding to the tissue image in the data.

[0035] Example 1

[0036] The present invention adopts the high temperature alloy MarM247LC for prediction, and a creep condition prediction method of the MarM247LC high temperature alloy specifically comprises the following steps:

[0037] S1. Collect creep microstructure parameter data and creep microstructure images of high-temperature alloys at different temperatures and times. Each creep microstructure image corresponds to a set of creep data, including creep temperature, creep time and microstructure parameter data. The creep conditions and creep time of MarM247LC high-temperature alloy are shown in Table 1 below.

[0038] Table 1 Creep data used for training and validation

[0039]

[0040] S2, crop the creep tissue images selected in S1 to obtain 600×600 unit RVE images, so that the images can represent the characteristics of the original images, and randomly rotate and flip them, and the RVE units obtained from the same tissue image have the same creep tissue parameters. These RVE units and corresponding creep tissue parameters constitute a data set of 1600 sets of data;

[0041] S3. Using the Pearson characteristic correlation calculation method, the data set constructed in S2 is used to select creep tissue parameters related to creep temperature and stress as γ channel width and rafting degree. The selected tissue parameter corresponding data and RVE unit are used as input feature data, and creep temperature and stress are used as output data to build a fusion neural network.

[0042] S4, training the fusion neural network constructed in S3, and verifying the obtained model on the data not involved in the training and testing, and predicting the creep temperature and stress corresponding to the creep structure image in the data. The comparison between the prediction results and the true values ​​is shown in Table 2.

[0043] Table 2 Comparison of data prediction results and experimental values

[0044]

[0045] R of the predicted values ​​in the table 2 They are all above 0.9, proving that the constructed fusion neural network model E has good prediction performance.

[0046] Example 2

[0047] The present invention adopts high temperature alloy DZ125 for prediction, and a method for predicting creep conditions of DZ125 high temperature alloy specifically comprises the following steps:

[0048] S1. Collect creep microstructure parameter data and creep microstructure images of DZ125 high temperature alloy at different temperatures and times. Each creep microstructure image corresponds to a set of creep data, including creep temperature, creep time and microstructure parameter data. The creep conditions and creep time of DZ125 high temperature alloy are shown in Table 3 below.

[0049] Table 3 Creep data used for training and validation

[0050]

[0051]

[0052] S2, crop the creep tissue images selected in S1 to obtain 600×600 unit RVE images, so that the images can represent the characteristics of the original images, and randomly rotate and flip them, and the RVE units obtained from the same tissue image have the same creep tissue parameters. These RVE units and corresponding creep tissue parameters constitute a data set of 1260 sets of data;

[0053] S3. Using the Pearson characteristic correlation calculation method, the creep tissue parameters related to creep temperature and stress are selected as γ' phase thickness and γ' volume fraction from the data set constructed in S2. The corresponding data of the selected tissue parameters and the RVE unit are used as input feature data, and the creep temperature and stress are used as output data to build a fusion neural network.

[0054] S4, training the fusion neural network constructed in S3, and verifying the obtained model on the data not involved in the training and testing, and predicting the creep temperature and stress corresponding to the creep structure image in the data. The comparison between the prediction results and the true values ​​is shown in Table 4.

[0055] Table 4 Comparison of data prediction results and experimental values

[0056]

[0057] R of the predicted values ​​in the table 2 They are all above 0.9, proving that the constructed fusion neural network model E has good prediction performance.

[0058] Example 3

[0059] The present invention adopts high temperature alloy equiaxed MarM247 for prediction, and a method for predicting the thermal exposure conditions of the equiaxed MarM247 high temperature alloy specifically comprises the following steps:

[0060] S1. Collect thermal exposure tissue parameter data and thermal exposure tissue images of high-temperature alloys at different temperatures and times. Each thermal exposure tissue image corresponds to a set of thermal exposure data, including thermal exposure temperature, thermal exposure time and tissue parameter data. The thermal exposure conditions and thermal exposure time of the equiaxed MarM247 high-temperature alloy are shown in Table 5 below.

[0061] Table 5 Thermal exposure data used for training and validation

[0062]

[0063] S2. Crop the heat-exposed tissue images selected in S1 to obtain 600×600 unit RVE images, so that the images can represent the characteristics of the original images, and randomly rotate and flip them, and the RVE units obtained from the same tissue image have the same heat-exposed tissue parameters. These RVE units and the corresponding heat-exposed tissue parameters constitute a data set of 380 sets of data;

[0064] S3. Using the Pearson characteristic correlation calculation method, the data set constructed in S2 is used to select the thermal exposure tissue parameters related to temperature (and stress) as γ' membrane thickness and γ' average size. The selected tissue parameter corresponding data and RVE unit are used as input feature data, and the thermal exposure temperature is used as output data to build a fusion neural network.

[0065] S4, training the fusion neural network constructed in S3, and verifying the obtained model on the data not involved in the training and testing, and predicting the heat exposure temperature corresponding to the heat exposure tissue image in the data. The comparison between the prediction result and the true value is shown in Table 6.

[0066] Table 6 Comparison of data prediction results and experimental values

[0067]

[0068]

[0069] R of the predicted values ​​in the table 2 They are all above 0.9, proving that the constructed fusion neural network model E has good prediction performance.

[0070] In the above scheme, the purpose of the present invention is to propose a prediction method for creep and thermal exposure conditions of high-temperature alloys based on machine learning, wherein the prediction value R of the high-temperature alloy creep or thermal exposure condition prediction method is 2 All of them are above 0.9, with a small error range, which shows that this method can make fast and accurate predictions of creep and thermal exposure conditions of high-temperature alloys within the test conditions and time range covered by the data set.

[0071] The data set in the present invention ranges from 15 to 120 groups of data, among which the tissue parameter data are screened based on a large amount of research experience summary of materials experts and Pearson correlation analysis method; a limited number of creep / heat exposure tissue images are sheared, rotated and flipped to obtain small-sized representative unit RVEs. This method is used to expand the image data, which has a strong reference significance for some complex machine learning problems in the field of materials science with less data conditions.

[0072] The method for predicting creep / heat exposure conditions of high-temperature alloys described in the present invention creatively builds a fusion neural network model using a convolutional neural network and an artificial neural network, thereby realizing the joint input of microstructure images, creep / heat exposure time, and tissue parameters, greatly reducing the errors caused by human bias in the prior art during the extraction of microstructure images.

[0073] In summary, the method for predicting creep / thermal exposure conditions of high-temperature alloys described in the present invention can accurately predict the creep and thermal exposure conditions of high-temperature alloys for turbine blades, and can comprehensively, quickly and effectively evaluate the service conditions of high-temperature alloys used in different parts of turbine blades, thereby rationalizing the design of high-temperature alloys for turbine blades, reducing hazards, reducing costs, and facilitating large-scale industrial production and promotion of use.

[0074] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning, characterized in that: The steps of the method are: S1. Collect creep / heat exposure tissue parameter data and creep / heat exposure tissue images of high-temperature alloys at different temperatures and times. Each tissue image corresponds to a set of data, including creep / heat exposure temperature, creep / heat exposure time and tissue parameter data. S2, cropping the creep / heat exposure tissue images selected in S1 respectively to obtain small-sized unit RVE images, so that the images can represent the characteristics of the original images, and randomly rotating and flipping them, and the RVE units obtained from the same tissue image have the same tissue parameters, and the above RVE units and corresponding tissue parameters constitute a data set; S3, using the Pearson characteristic correlation calculation method to select creep / heat exposure tissue parameters related to temperature and stress from the data set constructed in S2, using the corresponding data of the selected tissue parameters and the RVE unit as input feature data, and temperature and stress as output data, to build a fusion neural network; S4. Training the fusion neural network constructed in S3, and verifying the obtained model on the data that have not participated in the training and testing, and predicting the creep temperature, creep stress or thermal exposure temperature corresponding to the tissue image in the data.

2. The method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning according to claim 1, characterized in that: The data in S1 at least include creep / heat exposure tissue images under 15 different test conditions; the high-temperature alloy in S1 is MarM247LC high-temperature alloy or DZ125 high-temperature alloy or equiaxed MarM247 high-temperature alloy; the tissue parameters in S1 include γ' volume fraction, γ channel width, γ' thickness, γ' average size, γ' film thickness and rafting degree.

3. The method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning according to claim 1, characterized in that: The tissue parameters of S1 and S3 were selected based on a large amount of previous research experience and further screened through Pearson correlation analysis.

4. The method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning according to claim 1, characterized in that: The difference between the tissue parameter values ​​of each group of unit RVE in S2 and the tissue parameter values ​​of the original image is within 5%. The number of RVEs obtained for each group of original tissue images in S2 is 10 to 200, with a size of 600×600, and a total of 300 to 1300 groups of data are obtained.

5. The method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning according to claim 1, characterized in that: The fusion neural network model in S3 is a fusion structure of the convolutional neural network EfficientNetB7 and the artificial neural network.

6. The method for predicting creep / heat exposure conditions of high-temperature alloys based on machine learning according to claim 1, characterized in that: The creep temperature, creep stress and predicted value of thermal exposure temperature corresponding to the creep / thermal exposure tissue image in S4 are R 2 All are above 0.9.

Images