Titanium alloy microstructure prediction method based on deep learning

Abstract

The invention discloses a titanium alloy microscopic structure prediction method based on deep learning. The method comprises the following steps: S1, constructing a microscopic structure sample data expansion module based on deep learning and a cross-working-condition microscopic structure prediction module based on transfer learning; s2, carrying out acquisition and data preprocessing on microscopic structure sample data; s3, in a microscopic structure sample data expansion module based on deep learning, constructing a data expansion model, and expanding the sample data under the new working condition according to the sample data under the existing working condition; s4, dividing the sample data under the existing working condition and the expanded sample data under the new working condition into a source domain data set and a target domain data set in a transfer learning-based cross-working-condition microstructure prediction module, pre-training the source domain data set, and then performing parameter fine tuning on the target domain data to obtain a transfer learning model; and S5, predicting the microscopic structure sample data under the new working condition by using the transfer learning model.

Description

Technical Field

[0001] The present invention relates to the field of titanium alloy microstructure prediction and deep learning technology, and in particular to a titanium alloy microstructure prediction method based on deep learning. Background Art

[0002] Titanium alloys are widely used in aerospace and other industries due to their high strength, low density, excellent high temperature performance and corrosion resistance. The key to improving the performance of titanium alloys is to form a specific microstructure during hot working, so exploring the correlation between hot working process parameters and microstructure evolution has become a research hotspot.

[0003] During the hot working of titanium alloys, the main process parameters are deformation temperature, solution temperature and strain. The deformation temperature refers to the temperature of the titanium alloy during thermal deformation; the solution temperature refers to the temperature of the titanium alloy during heat treatment; and the strain refers to the degree of deformation of the titanium alloy caused by thermal deformation. In order to study the correlation between the above process parameters and the evolution of microstructures, a set of fixed deformation temperature and solution temperature are set in the materials experiment, and the experimental conditions that adjust the strain range are called working conditions. For example, Figure 1 Each rectangular box in represents a working condition (i.e., a set of experimental conditions that produce related data), and different working conditions are called cross-working conditions (i.e., different experimental conditions).

[0004] Due to the large number of different working conditions, the experimental process is complex and time-consuming, making it difficult to obtain complete data on microstructure samples under all working conditions. Therefore, a prediction model between hot working process parameters and microstructure characteristic parameters is established to infer the microstructure characteristic parameters under cross-working conditions, which helps guide process optimization and improve the performance of titanium alloys. The establishment of the prediction model is mainly divided into three parts: sample preparation, data extraction, and construction of the prediction model. Figure 2 As shown in the figure, the first step is sample preparation to obtain the corresponding microstructure data of titanium alloy under different working conditions, that is, using the TC18 titanium alloy double cone sample with lamellar structure, high-temperature compression experiments are carried out at different temperatures to obtain samples with different deformation degrees; the second step is data extraction to obtain hot working process parameters and the corresponding organizational characteristic parameters. The process is divided into two parts. One part uses Deform software to extract the strain generated in different parts of the same sample after the hot working process, and together with the deformation temperature and solution temperature obtained in the previous step, it constitutes the hot working process parameters; the other part uses ImageJ software to extract the organizational characteristic parameters of the same sample after the material characterization experiment; the third step is to build a model to establish a prediction model between hot working process parameters and microstructure characteristic parameters through the above data, that is, to study the evolution law of microstructure under cross-working conditions and guide process optimization.

[0005] There are two main parts to achieve the prediction of titanium alloy microstructure under cross-working conditions. First, the expansion of sample data is used to improve the adaptability and generalization ability of the model to diversified working conditions; second, the prediction of microstructure under cross-working conditions is used to improve the prediction accuracy of the model for diversified working conditions. In terms of expanding sample data, small sample learning methods have attracted widespread attention due to their excellent performance under incomplete data conditions and have become an important means to solve this problem. Small sample learning methods mainly include methods based on model fine-tuning, methods based on data enhancement, methods based on metric learning, and methods based on meta-learning. Among them, model fine-tuning, metric learning, and meta-learning methods are difficult to cope with the complex characteristics of microstructures and task adaptability problems, while methods based on data enhancement expand small sample data sets by generating new samples that are similar but not exactly the same as real data, which can significantly improve the generalization ability and robustness of the model. At present, there are studies on applying data augmentation methods to few-shot learning. For example, the article [He Xijun, Ma Shan, Wu Yuying, et al. E-commerce product sales prediction based on multi-dimensional indicator fusion under small samples [J]. Computer Engineering and Applications, 2019, 55(15): 177-184.] adopts ensemble learning method combined with multi-dimensional indicator fusion technology to collect data from multiple dimensions. Through technical means such as entropy method, the indicators are effectively fused and weighted to reduce the correlation between indicators and extract more critical information for sales prediction. In order to enhance the generalization ability of the model, the article [LiuY, Lee J, Park M, et al. Learning to propagate labels: Transductive propagation network for few-shot learning [J]. arxiv preprint arxiv: 1805.10002, 2018.] designs a conductive propagation network. By learning the propagation path of the label, the label information can be effectively transmitted between samples, thereby improving the accuracy of classification. The article [Zhang R, Che T, Ghahramani Z, et al. Metagan: Anadversarial approach to few-shot learning [J]. Advances inneural information processing systems, 2018, 31.] proposes an adversarial small-shot learning method based on generative adversarial network technology. By generating samples that are similar to but not exactly the same as real data, the discriminator is forced to learn a more robust feature representation, which significantly improves the robustness and prediction performance of the model in small-shot learning. The above method uses the data expansion strategy of small-shot learning technology to significantly enhance the performance and accuracy of the model in the prediction field.Although small sample learning technology has made progress in data expansion, traditional deep learning models often fail to effectively capture the characteristics of new working conditions when using expanded data for prediction, resulting in poor prediction results. Therefore, it is very important to use expanded microstructure sample data to complete the prediction of microstructure under cross-working conditions.

[0006] Transfer learning provides an effective adaptive strategy for microstructure prediction under cross-operating conditions. The goal of transfer learning is to apply the knowledge learned in one field or task to another different but related field or task to improve learning efficiency and performance, especially when the target task data is limited. Transfer learning mainly includes several methods such as instance weighting, parameter transfer, feature adaptation and domain adaptation. Among them, instance weighting method, parameter transfer method and feature adaptation method are difficult to accurately capture the structural change characteristics of titanium alloy under different temperature and strain conditions when dealing with the diversity and complex evolution of titanium alloy microstructure. The domain adaptation method can better capture the difference between the source domain and the target domain by dynamically adjusting the data distribution of the two domains, and is more suitable for prediction under cross-operating conditions. At present, there are models that apply domain adaptation to cross-operating condition prediction. For example, the article [Li Wei, Chen Jian, Tao Shanyong. Adaptive coupled periodic potential system stochastic resonance signal enhancement method [J]. Journal of Jilin University (Engineering Edition), 2021, 51(03): 1091-1096.] uses the conditional generative adversarial network model to enhance data diversity and solve the mode collapse problem, the article [Liu Fei, Chen Renwen, Xing Kailing, et al. Fast fault diagnosis algorithm for rolling bearings based on transfer learning and deep residual network [J]. Vibration and Shock, 2022, 41(03): 154-164.] proposes a model combining deep residual variational autoencoder with self-attention mechanism to extract and retain key feature information, the article [Zhang B, Li W, Li XL, et al. Intelligent fault diagnosis under varying working conditions based on domain adaptive convolutional neural networks [J]. Ieee Access, 2018, 6: 66367-66384.] proposes a convolutional neural network based on maximum mean difference to align feature distributions under different working conditions, and the article [Tang Z, Bo L, Liu X, et al. A semi-supervised transferable LSTM with feature evaluation for fault diagnosis of rotatingmachinery[J]. Applied Intelligence, 2022, 52(2): 1703-1717.] A semi-supervised transferable network model that combines entropy increase ratio and long short-term memory network (LSTM) is used to evaluate and select features that are helpful for prediction.Although transfer learning provides an effective strategy for cross-condition microstructure prediction, the existing domain adaptation algorithm still has the insufficiency of retaining the learning ability of the pre-trained model when processing complex data due to the coexistence of differences and similarities between different working conditions. Therefore, how to further optimize the domain adaptation algorithm to enhance its adaptability in complex environments and improve prediction accuracy remains an unresolved problem. Summary of the invention

[0007] The purpose of the present invention is to provide a titanium alloy microstructure prediction method based on deep learning, which performs feature analysis on sample data under existing working conditions, uses the feature analysis results to expand the model of sample data under new working conditions, and uses the data expansion results for subsequent prediction work. At the same time, in order to fully consider the differences and similarities between different working conditions during model prediction, the sample data set under the existing working conditions and the expanded sample data under the new working conditions are divided into a source domain data set and a target domain data set, the source domain data set is pre-trained, and then the parameters are fine-tuned on the target domain data set to obtain a transfer learning model, thereby achieving accurate prediction of microstructures under cross-working conditions.

[0008] The technical solution adopted by the present invention is:

[0009] A titanium alloy microstructure prediction method based on deep learning comprises the following steps:

[0010] S1: Construct a method framework including a microstructure sample data expansion module based on deep learning and a cross-operating condition microstructure prediction module based on transfer learning;

[0011] S2: Preprocessing the real data of the titanium alloy microstructure sample, including the steps of collection and normalization, obtaining the preprocessed data, and dividing it into a training set and a test set;

[0012] S3: Construct a data expansion model in the microstructure sample data expansion module based on deep learning, perform feature analysis on the sample data under the existing working conditions, use the generative adversarial network method to expand the data, and provide data guidance for the expanded data based on the feature analysis results, and finally obtain the sample data under the new working conditions;

[0013] S4: In the cross-working condition microstructure prediction module based on transfer learning, the sample data under the existing working condition and the expanded sample data under the new working condition are divided into a source domain data set and a target domain data set. The source domain data set is pre-trained using the transfer learning technology and a deep learning algorithm, and then the parameters are fine-tuned on the target domain data to obtain the transfer learning model.

[0014] S5: For the microstructure sample data under the new working conditions, the transfer learning model is used for prediction.

[0015] Furthermore, in step S3, a data expansion model is constructed in the microstructure sample data expansion module based on deep learning to complete sample data expansion under new working conditions, including the following steps:

[0016] 1-1) Construct an isolated forest anomaly detection algorithm (iForest) to identify abnormal samples in microscopic sample data. Original sample X = [x 1 ,x 2 ,…,x n ], the algorithm constructs a set of decision trees, each of which partitions the sample data by randomly selecting a feature and a random value of the feature. The path length h(x i ) is the number of splits from the root node to the leaf node in the tree. Since outliers are easy to be isolated, their path length h(x i ) are usually shorter.

[0017] The anomaly score s(x) of the isolation forest algorithm i ) is calculated as:

[0018]

[0019] Among them, h(x i ) is the sample x i The average path length in the tree, and E(h(x i )) is the expected path length of a sample in a dataset of the same size. For a sample size n, the expected path length E(h(x i The formula for ) is:

[0020]

[0021] H(i)=ln(i)+γ

[0022] Where H(i) is the harmonic number, γ is the Euler constant, and when s(x i )≈1, sample x 1 It is judged as an outlier point; when s(x i )≈1, sample x 1 Determined as an outlier;

[0023] 1-2) After abnormality detection processing is performed on the microstructure sample data, trend analysis is performed on it.

[0024] The range of data variation is determined by calculating the upper and lower limits of the data. The minimum and maximum values ​​of strain (∈) are set to ∈ min and ∈ max , and then perform constraint analysis of linear and nonlinear trends on the data.

[0025] The linear trend of strain and content (C) can be expressed as:

[0026]

[0027] Among them, C trend (∈) represents the linear trend of the content C as the strain ∈ changes, C start and C end are the starting point and end point content values ​​respectively.

[0028] Next, the nonlinear trend analysis of strain and spheroidization rate (S) was performed, and the parabolic model was used to describe the relationship between the two:

[0029]

[0030] Among them, S trend (∈) represents the nonlinear trend of the spheroidization rate S as the strain ∈ changes. The constant p determines the curvature of the parabola, S min is the lower limit of spheroidization rate;

[0031] 1-3) Data augmentation is implemented by the Generative Adversarial Network (GAN) algorithm. GAN consists of a generator and a discriminator. The generator generates sample data G(z|v) through random noise z and conditional input v, and the discriminator classifies the real sample x and the generated sample data G(z|v);

[0032] 1-4) The generator is mainly divided into two parts: core calculation and output layer.

[0033] The core calculation part of the generator is implemented by the following formula:

[0034]

[0035] in, represents the output of the generator, that is, the generated samples or data, θ g is the parameter of the generator, z is random noise, v is the conditional input, and the generator generates target data through a fully connected layer and a nonlinear activation function.

[0036] The output layer of the generator adds a cropping layer and a distribution regularization layer. The cropping layer constrains the generated data to ensure that it is within a predetermined range:

[0037]

[0038] in, Represents the value of the data output by the generator after clipping, B i and T iThey are the lower and upper limits of the data respectively. The max() function returns the maximum value in a set of values, and the min() function returns the minimum value in a set of values.

[0039] In addition, the distribution regularization layer introduces a penalty term to prevent the generated data from being overly concentrated near the upper and lower limits:

[0040]

[0041] Among them, L penalty represents the penalty term of the distribution regularization layer, which is used to introduce a constraint in the generation process to avoid excessive concentration of generated data near the upper and lower limits, λ i To penalize weights, control the distribution of generated data;

[0042] 1-5) The discriminator classifies the real data x and the generated data G(z|v), and its loss function is:

[0043]

[0044] Where L D Represents the loss function of the discriminator, x~p data Indicates that x comes from the real data distribution p data The real sample, z~p z Denote z as coming from the generating data distribution p z The generated sample, D(x) is the probability prediction of the discriminator for the real sample x;

[0045] 1-6) By training the GAN model, the generator and the discriminator are trained adversarially to gradually optimize the quality of the generated samples. The generator can output the initial generated sample X'=[x' 1 ,x' 2 ,...,x' m ];

[0046] 1-7) Double trend constraints are applied to the initial generated samples, which are the linear trend constraints of strain (∈) and content C) and the nonlinear trend constraints of strain (∈) and spheroidization rate (S). Finally, the expanded sample X”=[x” under the new working condition after double trend constraints is obtained. 1 ,x” 2 ,...,x” m ].

[0047] Furthermore, in the described step S4, constructing a transfer learning model in the cross-operating condition microstructure prediction module based on transfer learning includes the following steps:

[0048] 2-1) Divide the original sample X and the expanded sample X' into a source domain dataset and a target domain dataset;

[0049] 2-2) The construction process of the transfer learning model is divided into two processes: pre-training and parameter fine-tuning;

[0050] 2-3) Construct a pre-training model, which is divided into three main parts: feature extraction, sequence information processing and fusion output;

[0051] 2-4) The feature extraction part is performed through multiple convolution operations. The input data is transformed through multiple layers of linear transformation to calculate a new feature map. Let the output of the convolution operation be F, then the output of each layer of convolution can be expressed as:

[0052] F (i) =σ(W (i) *X (i-1) +b (i) )

[0053] Among them, F (i) represents the convolution output of the i-th layer, W (i) and b (i) are the weight matrix and bias term of the i-th layer, X (i -1) , represents the output feature map of the i-1th layer, that is, the input data of the convolution operation of the current layer (i-th layer), * represents the convolution operation, and σ(·) is the activation function;

[0054] 2-5) The sequence information processing part relies on recursive calculation to capture the time-dependent characteristics of the data. For each time step of input, the hidden state is updated through a recursive function. For the hidden state p t , and its update rule is:

[0055] p t =φ(M*p t-1 +N*x t +c)

[0056] Among them, p t represents the hidden state at the current time t, M and N are weight matrices, p t-1 is the hidden state at the previous time t-1, x t is the input data at the current time t, c is the bias term, and φ(·) is the activation function. In order to further stabilize the training process, normalization is applied to the hidden state, and a random dropout mechanism is added to the last layer to prevent overfitting;

[0057] 2-6) The fusion output part uses the attention mechanism to perform weighted combination on the processed sequence features. Let the feature matrix be K and the attention weight be A, then the fused output Z can be expressed as:

[0058] Z=A·K

[0059] Among them, A is obtained by calculating the correlation in K;

[0060] 2-7) After multiple layers of fully connected operations, the fused feature input is finally mapped to the output space to obtain the prediction result The calculation formula is:

[0061]

[0062] Among them, W out and b out are the weight matrix and bias term of the fully connected layer respectively. Finally, the model uses the mean square error (MSE) as the loss function, which is defined as follows:

[0063]

[0064] in, represents the loss function MSE, is the model's predicted value y for the i-th sample i is the true value of the ith sample, and N is the total number of samples.

[0065] By minimizing the loss function, the model achieves effective prediction of the data and completes the pre-training process;

[0066] 2-8) Perform fine-tuning of transfer learning. Starting from the pre-trained model, adjust its parameters to adapt to the target domain data. Update the model parameters θ by minimizing a new loss function L(θ,X”). The goal of fine-tuning is to minimize the model loss function, that is:

[0067]

[0068] Among them, θ * It means finding the best parameters in the parameter space that minimize the loss function L(θ,X').

[0069] During fine-tuning, the layer freezing mechanism is used to freeze only the first few layers of the pre-trained model, keeping their parameters unchanged, and only updating the subsequent layers of the model. The parameter updates of the frozen layers are set to zero, that is:

[0070] Δθ freeze =0

[0071] Among them, Δθ freeze Indicates the update amount of the parameters in the frozen layer.

[0072] The parameter updates of the layers that need fine-tuning are calculated by the gradient descent method:

[0073]

[0074] Among them, Δθtrainable represents the parameter adjustment amount of the layer (trainable layer) allowed to be updated, η is the learning rate, is the gradient of the loss function with respect to the trainable parameters. In each iteration, the optimizer uses these gradients to update the parameters:

[0075]

[0076] Among them, θ t represents the model parameters at the current time t, θ t+1 represents the new model parameters obtained after updating at the current time t, Represents the loss function for the current model parameter θ t gradient.

[0077] Finally, through continuous iterative optimization, a transfer learning model that can predict the microstructure under new working conditions is obtained.

[0078] The beneficial effects of adopting the above technical solution are:

[0079] The purpose of the present invention is to provide a titanium alloy microstructure prediction method based on deep learning, which includes a microstructure sample data expansion model based on deep learning and a cross-operating condition microstructure prediction module based on transfer learning. The microstructure sample data expansion model based on deep learning first performs feature analysis on the sample data under the existing working conditions, uses the generative adversarial network method to expand the data, and according to the feature analysis results, performs data guidance on the expanded data to achieve high-quality expansion of the sample data under the new working conditions and the merging of the data sets to obtain the final data set; the cross-operating condition microstructure prediction module based on transfer learning divides the final data set into a source domain data set and a target domain data set, uses a deep learning algorithm to pre-train the source domain data set, obtains the feature representation obtained by the pre-training model on the source domain data, obtains the migration model after fine-tuning the parameters, and uses the migration model to predict the microstructure under cross-operating conditions for the target domain data set.

[0080] The cross-operating condition microstructure prediction method proposed in this paper was applied to the microstructure sample data of TC18 titanium alloy. Through experimental analysis, the characteristic parameters of the microstructure, including the α-phase content and the α-phase spheroidization percentage, were predicted on the microstructure sample data under two different operating conditions. In the prediction of the α-phase content, the mean absolute error and the root mean square error were reduced by at least about 20% and at most about 32% compared with the mainstream method. In the prediction of the α-phase spheroidization percentage, the mean absolute error and the root mean square error were reduced by at least about 10% and at most about 38% compared with the mainstream method, both of which are better than the mainstream method, verifying the cross-operating condition prediction ability of the titanium alloy microstructure prediction method based on deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a schematic diagram of the working conditions in microstructure research;

[0082] Figure 2 Schematic diagram of the process of establishing a predictive model for titanium alloys;

[0083] Figure 3 This is a framework diagram of the titanium alloy microstructure prediction method based on deep learning;

[0084] Figure 4 This is a framework diagram of the microstructure sample data expansion module based on deep learning;

[0085] Figure 5 This is the framework diagram of the cross-working condition microstructure prediction module based on transfer learning;

[0086] Figure 6 This is a diagram of the titanium alloy sample preparation and characterization process;

[0087] Figure 7 This is a graph of the predicted results of the comparative experiment;

[0088] Figure 8 This is the predicted result of the ablation experiment. DETAILED DESCRIPTION

[0089] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0090] The present invention takes titanium alloy microstructure prediction as the background, deep learning as the carrier, and uses the microstructure sample data expansion module based on deep learning and the cross-working condition microstructure prediction module based on transfer learning as the main algorithm framework. Figure 3 As shown, the following steps are included:

[0091] S1: Construct a method framework including a microstructure sample data expansion module based on deep learning and a cross-operating condition microstructure prediction module based on transfer learning;

[0092] S2: Preprocessing the real data of the titanium alloy microstructure sample, including the steps of collection and normalization, obtaining the preprocessed data, and dividing it into a training set and a test set;

[0093] S3: construct a data expansion model in the microstructure sample data expansion module based on deep learning, perform feature analysis on the sample data under the existing working conditions, use the generative adversarial network method to expand the data, and provide data guidance for the expanded data based on the feature analysis results, and finally obtain the sample data under the new working conditions, including the following steps:

[0094] 1-1) Construct an isolated forest anomaly detection algorithm (iForest) to identify abnormal samples in microscopic sample data. Original sample X = [x1 ,x 2 ,…,x n ], the algorithm constructs a set of decision trees, each of which partitions the sample data by randomly selecting a feature and a random value of the feature. The path length h(x i ) is the number of splits from the root node to the leaf node in the tree. Since outliers are easy to be isolated, their path length h(x i ) are usually shorter.

[0095] The anomaly score s(x) of the isolation forest algorithm i ) is calculated as:

[0096]

[0097] Among them, h(x i ) is the sample x i The average path length in the tree, and E(h(x i )) is the expected path length of a sample in a dataset of the same size. For a sample size n, the expected path length E(h(x i The formula for ) is:

[0098]

[0099] H(i)=ln(i)+γ

[0100] Where H(i) is the harmonic number, γ is the Euler constant, and when s(x i )≈1, sample x 1 It is judged as an outlier point; when s(x i )≈1, sample x 1 Determined as an outlier;

[0101] 1-2) After abnormality detection processing is performed on the microstructure sample data, trend analysis is performed on it.

[0102] The range of data variation is determined by calculating the upper and lower limits of the data. The minimum and maximum values ​​of strain (∈) are set to ∈ min and ∈ max , and then perform constraint analysis of linear and nonlinear trends on the data.

[0103] The linear trend of strain and content (C) can be expressed as:

[0104]

[0105] Among them, C trend (∈) represents the linear trend of the content C as the strain ∈ changes, C start and C endare the starting point and end point content values ​​respectively.

[0106] Next, the nonlinear trend analysis of strain and spheroidization rate (S) was performed, and the parabolic model was used to describe the relationship between the two:

[0107]

[0108] Among them, S trend (∈) represents the nonlinear trend of the spheroidization rate S as the strain ∈ changes. The constant p determines the curvature of the parabola, S min is the lower limit of spheroidization rate;

[0109] 1-3) Data augmentation is implemented by the Generative Adversarial Network (GAN) algorithm. GAN consists of a generator and a discriminator. The generator generates sample data G(z|v) through random noise z and conditional input v, and the discriminator classifies the real sample x and the generated sample data G(z|v);

[0110] 1-4) The generator is mainly divided into two parts: core calculation and output layer.

[0111] The core calculation part of the generator is implemented by the following formula:

[0112]

[0113] in, represents the output of the generator, that is, the generated samples or data, θ g is the parameter of the generator, z is random noise, v is the conditional input, and the generator generates target data through a fully connected layer and a nonlinear activation function.

[0114] The output layer of the generator adds a cropping layer and a distribution regularization layer. The cropping layer constrains the generated data to ensure that it is within a predetermined range:

[0115]

[0116] in, Represents the value of the data output by the generator after clipping, B i and T i They are the lower and upper limits of the data respectively. The max() function returns the maximum value in a set of values, and the min() function returns the minimum value in a set of values.

[0117] In addition, the distribution regularization layer introduces a penalty term to prevent the generated data from being overly concentrated near the upper and lower limits:

[0118]

[0119] Among them, L penaltyrepresents the penalty term of the distribution regularization layer, which is used to introduce a constraint in the generation process to avoid excessive concentration of generated data near the upper and lower limits, λ i To penalize weights, control the distribution of generated data;

[0120] 1-5) The discriminator classifies the real data x and the generated data G(z|v), and its loss function is:

[0121]

[0122] Where L D Represents the loss function of the discriminator, x~p data Indicates that x comes from the real data distribution p data The real sample, z~p z Denote z as coming from the generating data distribution p z The generated sample, D(x) is the probability prediction of the discriminator for the real sample x;

[0123] 1-6) By training the GAN model, the generator and the discriminator are trained adversarially to gradually optimize the quality of the generated samples. The generator can output the initial generated sample X'=[x' 1 ,x' 2 ,...,x' m ];

[0124] 1-7) Double trend constraints are applied to the initial generated samples, which are the linear trend constraints of strain (∈) and content (C) and the nonlinear trend constraints of strain (∈) and spheroidization rate (S). Finally, the expanded sample X”=[x” 1 ,x” 2 ,...,x” m ].

[0125] S4: In the cross-working condition microstructure prediction module based on transfer learning, the sample data under the existing working condition and the expanded sample data under the new working condition are divided into a source domain data set and a target domain data set. The source domain data set is pre-trained using the transfer learning technology and a deep learning algorithm, and then the parameters are fine-tuned on the target domain data to obtain a transfer learning model, including the following steps:

[0126] 2-1) Divide the original sample X and the expanded sample X' into a source domain dataset and a target domain dataset;

[0127] 2-2) The construction process of the transfer learning model is divided into two processes: pre-training and parameter fine-tuning;

[0128] 2-3) Construct a pre-training model, which is divided into three main parts: feature extraction, sequence information processing and fusion output;

[0129] 2-4) The feature extraction part is performed through multiple convolution operations. The input data is transformed through multiple layers of linear transformation to calculate a new feature map. Let the output of the convolution operation be F, then the output of each layer of convolution can be expressed as:

[0130] F (i) =σ(W (i) *X (i-1) +b (i) )

[0131] Among them, F (i) represents the convolution output of the i-th layer, W (i) and b (i) are the weight matrix and bias term of the i-th layer, X (i -1) , represents the output feature map of the i-1th layer, that is, the input data of the convolution operation of the current layer (i-th layer), * represents the convolution operation, and σ(·) is the activation function;

[0132] 2-5) The sequence information processing part relies on recursive calculation to capture the time-dependent characteristics of the data. For each time step of input, the hidden state is updated through a recursive function. For the hidden state p t , and its update rule is:

[0133] p t =φ(M*p t-1 +N*x t +c)

[0134] Among them, p t represents the hidden state at the current time t, N and B are weight matrices, p t-1 is the hidden state at the previous time t-1, x t is the input data at the current time t, c is the bias term, and φ(·) is the activation function. In order to further stabilize the training process, normalization is applied to the hidden state, and a random dropout mechanism is added to the last layer to prevent overfitting;

[0135] 2-6) The fusion output part uses the attention mechanism to perform weighted combination on the processed sequence features. Let the feature matrix be K and the attention weight be A, then the fused output Z can be expressed as:

[0136] Z=A·K

[0137] Among them, A is obtained by calculating the correlation in K;

[0138] 2-7) After multiple layers of fully connected operations, the fused feature input is finally mapped to the output space to obtain the prediction result The calculation formula is:

[0139]

[0140] Among them, W out and b out are the weight matrix and bias term of the fully connected layer respectively. Finally, the model uses the mean square error (MSE) as the loss function, which is defined as follows:

[0141]

[0142] in, represents the loss function MSE, is the model's predicted value y for the i-th sample i is the true value of the ith sample, and N is the total number of samples.

[0143] By minimizing the loss function, the model achieves effective prediction of the data and completes the pre-training process;

[0144] 2-8) Perform fine-tuning of transfer learning. Starting from the pre-trained model, adjust its parameters to adapt to the target domain data. Update the model parameters θ by minimizing a new loss function L(θ,X”). The goal of fine-tuning is to minimize the model loss function, that is:

[0145]

[0146] Among them, θ * It means finding the best parameters in the parameter space that minimize the loss function L(θ,X').

[0147] During fine-tuning, the layer freezing mechanism is used to freeze only the first few layers of the pre-trained model, keeping their parameters unchanged, and only updating the subsequent layers of the model. The parameter updates of the frozen layers are set to zero, that is:

[0148] Δθ freeze =0

[0149] Among them, Δθ freeze Indicates the update amount of the parameters in the frozen layer.

[0150] The parameter updates of the layers that need fine-tuning are calculated by the gradient descent method:

[0151]

[0152] Among them, Δθ trainable represents the parameter adjustment amount of the layer (trainable layer) allowed to be updated, η is the learning rate, is the gradient of the loss function with respect to the trainable parameters. In each iteration, the optimizer uses these gradients to update the parameters:

[0153]

[0154] Among them, θ t represents the model parameters at the current time t, θ t+1 represents the new model parameters obtained after updating at the current time t, Represents the loss function for the current model parameter θ t gradient.

[0155] Finally, through continuous iterative optimization, a transfer learning model that can predict the microstructure under new working conditions is obtained.

[0156] S5: For the microstructure sample data under the new working conditions, the transfer learning model is used for prediction.

[0157] Based on the above steps, the present invention effectively solves the problem that the preparation experiment of titanium alloy microstructure samples takes a long time and is difficult, and it is difficult to obtain microstructure sample data under all working conditions, resulting in the traditional technology having little effect on the prediction of microstructure under new working conditions. A titanium alloy microstructure prediction method based on deep learning is proposed. The method performs feature analysis on the microstructure sample data set under existing working conditions, and uses the feature analysis results to assist in constructing a new working condition sample data expansion model. The expanded new working condition sample data is used for subsequent prediction work. At the same time, in order to fully consider the differences and similarities between different working conditions during model prediction, the sample data set under the existing working conditions and the expanded new working condition sample data set are divided into source domain data and target domain data set, the source domain data set is pre-trained, and the parameters are fine-tuned on the target domain data set to obtain a transfer learning model, and the transfer learning model is used to perform microstructure prediction under cross-working conditions, thereby achieving accurate prediction of microstructure under cross-working conditions.

[0158] Experimental verification of the titanium alloy microstructure prediction method based on deep learning of the present invention:

[0159] Test environment

[0160] Windows 11 system, Intel Core i5-1035G7 CPU@1.20GHz and NVIDIA GeForce RTX2080Ti GPU, the development environment is Python 3.9.5, the deep learning framework is PyTorch 2.1.0, and the machine learning method uses the API provided by Scikit-learn1.3.0.

[0161] Test verification

[0162] Experimental results and analysis on TC18 titanium alloy microstructure sample data

[0163] (1) Dataset description

[0164] In order to verify the effectiveness of the method proposed in this paper, the experiment uses the microstructure sample data of TC18 titanium alloy. The data acquisition process is divided into two parts: sample preparation and data extraction. The sample preparation process includes the following steps: First, the material is cut into a cylindrical sample of φ55×50mm, such as Figure 6 (a), and processed into a double-conical specimen; then subjected to thermal deformation, the obtained specimen is as shown Figure 6 As shown in (b), after the deformation is completed, a thin slice 10 mm away from the core of the sample is taken along the axial direction. Figure 6 (c), followed by a heat treatment experiment; finally, the surface of the sample was mechanically polished and etched. Figure 6 As shown in (d), the sample was characterized by Zeiss Sigma 300 scanning electron microscope. Some of the characterization results are shown in Figure 6 As shown in (e), this paper studies the characteristic parameters of the α phase indicated in the figure. Data extraction process The data extraction process uses Deform software and Image J software to extract the sample after thermal deformation and the sample after characterization to obtain the final data.

[0165] The TC18 titanium alloy microstructure sample data contains the experimental results of TC18 titanium alloy under 10 different working conditions, which are divided into hot working process parameters and organizational characteristic parameters. Among them, the hot forming process parameters include deformation temperature, solution temperature and strain, and the organizational characteristic parameters include α phase content and α phase spheroidization percentage. Some of the data are shown in Table 1. The experiment selected sample data under 8 working conditions as the training set, and sample data under 2 working conditions as the test set, and selected hot forming process parameters as input variables, and organizational characteristic parameters as output variables to be tested.

[0166] Table 1 Partial data of TC18 titanium alloy microstructure specimen

[0167]

[0168]

[0169] (2) Evaluation indicators

[0170] In order to evaluate the prediction performance of the model, the root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators, which are defined as follows:

[0171]

[0172] Among them, y i , and Represents the true value of the target variable, the mean of the true value, and the model prediction. RMSE measures the average error between the model prediction and the true value, and is more sensitive to the prediction error. The smaller the index, the better the model performance. MAE measures the average absolute error between the prediction and the true value. The smaller the index, the better the model performance.

[0173] (3) Comparative experiment

[0174] In order to verify the effectiveness of the titanium alloy microstructure prediction method based on deep learning, a comparative experiment was conducted on the model. The method in this paper is denoted as GAN-ACTL, and two different transfer learning methods are selected for comparison with GAN-ACTL. Although the selected transfer learning methods come from different fields, these models are representative in processing complex data distribution and cross-domain migration, and can provide reference for cross-condition prediction of microstructure. A brief description of each model is as follows:

[0175] Spatiotemporal Transfer Network (TL-TCN-BIGRU): It integrates transfer learning and deep learning, transfers the knowledge of source domain data to the target domain through transfer learning, uses the k-center clustering algorithm to analyze data features, and combines the temporal convolutional network and bidirectional gated recurrent unit to extract multi-scale and nonlinear features of time series. Hybrid Deep Learning Model (TraTCN-LSTM-BiGR U): Based on the transfer learning theory, it uses the mean heat map method to select relevant features, divides the source domain and target domain data by relative entropy, combines the temporal convolutional network to extract spatial features, and the long short-term memory network to extract temporal features, and finally uses the bidirectional gated recurrent unit for feature fusion.

[0176] Table 2 Comparison of prediction results of TL-TCN-BIGRU, TraTCN-LSTM-BiGRU and GAN-ACTL

[0177]

[0178] According to the experimental results, the GAN-ACTL model has significantly improved in all indicators compared with the TL-TCN-BIGRU model and the TraTCN-LSTM-BiGRU model. In the prediction of α-phase content, the RMSE and MAE of the GAN-ACTL model were reduced by 20.59% to 32.04%, respectively, showing higher prediction accuracy. In the prediction of α-phase spheroidization percentage, the RMSE and MAE of the GAN-ACTL model were 10.77% to 38.0% lower than other models, showing more stable prediction ability. This is because the GAN-ACTL model introduces the idea of ​​microstructure sample data expansion and hierarchical freezing mechanism on the basis of the transfer learning model. On the one hand, it can expand more microstructure data under new working conditions and enrich the training data of the model, thereby improving the model's prediction ability for new samples; on the other hand, it can reduce the risk of overfitting of the model during training by freezing some hierarchical parameters, and improve the generalization ability on test data.

[0179] Furthermore, in order to explore the contribution of the cross-condition prediction module in the GAN-ACTL model to the performance, the base model GAN-ACTL-SingleRD, which does not have the cross-condition prediction capability, was set as the control group and tested on the microstructure data of working conditions 1 and 2, respectively.

[0180] Table 3 Comparison of prediction results between GAN-ACTL-SingleRD and GAN-ACTL

[0181]

[0182] It can be seen from Table 3 that GAN-ACTL exhibits lower RMSE and MAE. Compared with the control group GAN-ACTL-SingleRD, the RMSE of the α-phase content on the microstructure data under two different working conditions is reduced by about 69.4% and 36.8%, the RMSE of the α-phase spheroidization percentage is reduced by 13.1% and 33.5%, the MAE of the α-phase content is reduced by 75% and 48.2%, and the MAE of the α-phase spheroidization percentage is reduced by 20.3% and 40.7%. This is because the base model GAN-ACTL-SingleRD cannot learn the microstructure data characteristics under new working conditions, while the GAN-ACTL with data expansion can effectively use the expanded data features to adjust the parameters of the pre-trained model, so that the model can more accurately capture the actual change trend of the data across working conditions, significantly reducing the prediction error, proving the effectiveness of the cross-condition microstructure prediction ability in the face of differential changes in microstructures.

[0183] In summary, the GAN-ACTL prediction method proposed in this paper is superior to other prediction models in microstructure prediction results under cross-working conditions. GAN-ACTL performs feature analysis on the microstructure sample data under existing working conditions, and uses the feature analysis results to guide the expansion of sample data under new working conditions, thereby achieving high-quality expansion of sample data under new working conditions and merging of data sets to obtain the final data set. At the same time, a cross-working condition microstructure prediction module based on transfer learning is constructed, and the final data set is divided into a source domain data set and a target domain data set. Deep learning is used to pre-train the source domain data set, and parameter fine-tuning is performed on the target domain data set to obtain a transfer learning model, so that the model prediction can fully consider the differences and similarities between different working conditions, thereby achieving accurate prediction of microstructure data under cross-working conditions.

[0184] In view of the long time and difficulty of preparing titanium alloy microstructure samples, it is difficult to obtain microstructure sample data under all working conditions, which leads to the problem that traditional technology has little effect on microstructure prediction under new working conditions. This paper proposes a titanium alloy microstructure prediction method based on deep learning. This aspect includes two parts. First, the sample data under existing working conditions are feature analyzed, and the data is expanded using the generative adversarial network method. According to the feature analysis results, the expanded data is guided by data to achieve high-quality expansion of sample data under new working conditions and merge the data sets to obtain the final data set; secondly, the final data set is divided into a source domain data set and a target domain data set, and the source domain data set is pre-trained using a deep learning algorithm to obtain the feature representation of the pre-trained model on the source domain data, and the migration model is obtained after fine-tuning the parameters; finally, the migration model is used to predict microstructures under cross-working conditions.

[0185] The above examples of the present invention are described in detail, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A titanium alloy microstructure prediction method based on deep learning, characterized in that: The steps include: S1: Construct a method framework including a microstructure sample data expansion module based on deep learning and a cross-operating condition microstructure prediction module based on transfer learning; S2: Preprocessing the real data of the titanium alloy microstructure sample, including the steps of collection and normalization, obtaining the preprocessed data, and dividing it into a training set and a test set; S3: Construct a data expansion model in the microstructure sample data expansion module based on deep learning, perform feature analysis on the sample data under the existing working conditions, use the generative adversarial network method to expand the data, and provide data guidance for the expanded data based on the feature analysis results, and finally obtain the sample data under the new working conditions; S4: In the cross-working condition microstructure prediction module based on transfer learning, the sample data under the existing working condition and the expanded sample data under the new working condition are divided into a source domain data set and a target domain data set. The source domain data set is pre-trained using the transfer learning technology and a deep learning algorithm, and then the parameters are fine-tuned on the target domain data to obtain the transfer learning model. S5: For the microstructure sample data under the new working conditions, the transfer learning model is used for prediction.

2. The titanium alloy microstructure prediction method based on deep learning according to claim 1, characterized in that: In step S3, a data expansion model is constructed in the microstructure sample data expansion module based on deep learning to complete sample data expansion under new working conditions, including the following steps: 1-1) Construct an isolated forest anomaly detection algorithm (iForest) to identify abnormal samples in microscopic sample data. The original sample X = [x1, x2, ..., x n ], the algorithm constructs a set of decision trees, each tree divides the sample data by randomly selecting a feature and a random value of the feature, and the path length h(x i ) is the number of splits from the root node to the leaf node in the tree. Since the abnormal point is easy to be isolated, its path length h(x i ) is usually shorter, and the anomaly score s(x i ) is calculated as: Among them, h(x i ) is the sample x i The average path length in the tree, and E(h(x i )) is the expected path length of samples in a dataset of the same size. For a sample size n, the expected path length E(h(x i The formula for ) is: H(i)=ln(i)+γ Where H(i) is the harmonic number, γ is the Euler constant, and when s(x i )≈1, sample x1 is judged as an outlier; when s(x i )≈1, sample x1 is judged as an outlier; 1-2) After abnormal detection processing of microstructure sample data, trend analysis is performed on it; The range of data variation is determined by calculating the upper and lower limits of the data, and the minimum and maximum values ​​of the strain (∈) are set to ∈ min and ∈ max ,Then perform constraint analysis of linear and nonlinear trends on the data; The linear trend of strain and content (C) can be expressed as: Among them, C trend (∈) represents the linear trend of the content C as the strain ∈ changes, C start and C end are the content values ​​of the starting point and the end point, respectively; Next, the nonlinear trend analysis of strain and spheroidization rate (S) was performed, and the parabolic model was used to describe the relationship between the two: Among them, S trend (∈) represents the nonlinear trend of the spheroidization rate S as the strain ∈ changes. The constant p determines the curvature of the parabola, S min is the lower limit of spheroidization rate; 1-3) Data augmentation is implemented by the Generative Adversarial Network (GAN) algorithm. GAN consists of a generator and a discriminator. The generator generates sample data G(z|v) through random noise z and conditional input v, and the discriminator classifies the real sample x and the generated sample data G(z|v); 1-4) The generator is mainly divided into two parts: core calculation and output layer; The core calculation part of the generator is implemented by the following formula: in, represents the output of the generator, that is, the generated samples or data, θ g is the parameter of the generator, z is random noise, v is the conditional input, and the generator generates target data through a fully connected layer and a nonlinear activation function; The output layer of the generator adds a cropping layer and a distribution regularization layer. The cropping layer constrains the generated data to ensure that it is within a predetermined range: in, Represents the value of the data output by the generator after clipping, B i and T i They are the lower and upper limits of the data, respectively. The max() function returns the maximum value in a set of values, and the min() function returns the minimum value in a set of values. In addition, the distribution regularization layer introduces a penalty term to prevent the generated data from being overly concentrated near the upper and lower limits: Among them, L penalty represents the penalty term of the distribution regularization layer, which is used to introduce a constraint in the generation process to avoid excessive concentration of generated data near the upper and lower limits, λ i To penalize weights, control the distribution of generated data; 1-5) The discriminator classifies the real data x and the generated data G(z|v), and its loss function is: Where L D Represents the loss function of the discriminator, x~p data Indicates that x comes from the real data distribution p data The real sample, z~p z Denote z as coming from the generating data distribution p z The generated sample, D(x) is the probability prediction of the discriminator for the real sample x; 1-6) By training the GAN model, the generator and the discriminator are trained adversarially to gradually optimize the quality of the generated samples. The generator can output the initial generated sample X'=[x'1,x'2,...,x' m ]; 1-7) Double trend constraints are applied to the initial generated samples, which are the linear trend constraints of strain (∈) and content (C) and the nonlinear trend constraints of strain (∈) and spheroidization rate (S). Finally, the expanded sample X' under the new working condition after double trend constraints is obtained. ’ =[x'1 ’ ,x'2 ’ ,...,x ’ ' m ].

3. The titanium alloy microstructure prediction method based on deep learning according to claim 1, characterized in that: In the described step S4, constructing a transfer learning model in the cross-operating condition microstructure prediction module based on transfer learning includes the following steps: 2-1) The original sample X and the expanded sample X' ’ Divide into source domain dataset and target domain dataset; 2-2) The construction process of the transfer learning model is divided into two processes: pre-training and parameter fine-tuning; 2-3) Construct a pre-training model, which is divided into three main parts: feature extraction, sequence information processing and fusion output; 2-4) The feature extraction part is performed through multiple convolution operations. The input data undergoes multiple layers of linear transformation to calculate a new feature map. Let the output of the convolution operation be F, then the output of each layer of convolution can be expressed as: F (i) =σ(W (i) *X (i-1) +b (i) ) Among them, F (i) represents the convolution output of the i-th layer, W (i) and b (i) are the weight matrix and bias term of the i-th layer, X (i-1) , represents the output feature map of the i-1th layer, that is, the input data of the convolution operation of the current layer (i-th layer), * represents the convolution operation, and σ(·) is the activation function; 2-5) The sequence information processing part relies on recursive calculation to capture the time-dependent characteristics of the data. For each time step input, the hidden state is updated through a recursive function. For the hidden state p t , and its update rule is: p t =φ(M*p t-1 +N*x t +c) Among them, p t represents the hidden state at the current time t, N and B are weight matrices, p t-1 is the hidden state at the previous time t-1, x t is the input data at the current time t, c is the bias term, φ(·) is the activation function. In order to further stabilize the training process, the hidden state is normalized and a random dropout mechanism is added to the last layer to prevent overfitting. 2-6) The fusion output part uses the attention mechanism to perform weighted combination on the processed sequence features. Let the feature matrix be K and the attention weight be A. The fused output Z can be expressed as: Z=A·K Among them, A is obtained by calculating the correlation in K; 2-7) After multiple layers of full connection operations, the fused feature input is finally mapped to the output space to obtain the prediction result The calculation formula is: Among them, W out and b out are the weight matrix and bias term of the fully connected layer respectively. Finally, the model uses the mean square error (MSE) as the loss function, which is defined as follows: in, represents the loss function MSE, is the model's predicted value y for the i-th sample i is the true value of the i-th sample, and N is the total number of samples; By minimizing the loss function, the model achieves effective prediction of the data and completes the pre-training process; 2-8) Perform fine-tuning of transfer learning. Starting from the pre-trained model, adjust its parameters to adapt to the target domain data, and update the model parameters θ by minimizing a new loss function L(θ,X”). The goal of fine-tuning is to minimize the loss function of the model, that is: Among them, θ * It means finding the best parameters in the parameter space to minimize the loss function L(θ,X”); During fine-tuning, the layer freezing mechanism is used to freeze only the first few layers of the pre-trained model, keeping their parameters unchanged, and only the subsequent layers of the model are updated. The parameter updates of the frozen layers are set to zero, that is: Dth freeze =0 Among them, Δθ freeze Indicates the update amount of the parameters in the frozen layer; The parameter updates of the layers that need fine-tuning are calculated by the gradient descent method: Among them, Δθ trainable represents the parameter adjustment amount of the layer (trainable layer) allowed to be updated, η is the learning rate, are the gradients of the loss function with respect to the trainable parameters. At each iteration, the optimizer uses these gradients to update the parameters: Among them, θ t represents the model parameters at the current time t, θ t+1 represents the new model parameters obtained after updating at the current time t, Represents the loss function for the current model parameter θ t The gradient of Finally, through continuous iterative optimization, a transfer learning model that can predict the microstructure under new working conditions is obtained.

Images