Method and system for predicting stress endurance time of cold-drawn GH4169 alloy bar

Abstract

The invention relates to a method for predicting the stress duration of a cold-drawn GH4169 alloy bar. The method comprises the following steps: S1, collecting original components, structural characteristics and text data of a heat treatment process class of the alloy bar and image data of a metallographic microstructure class after heat treatment; s2, setting model parameters, reading the CSV file data in the step S1 for data processing, and dividing into a training set, a verification set and a test set; s3, constructing a physical prior multi-modal neural network model comprising a text data branch and an image data branch based on a ResNet-18 architecture, and performing feature fusion to output a life prediction value; s4, based on a Huber regression loss function, model parameters are trained and updated through an Adam optimizer, and model performance is evaluated by using a loss loss curve graph, MAE and an Rindex; s5, a training process and a prediction result are visualized; according to the method, microstructure multi-modal data and deep learning are fused for the first time, high-precision prediction of the stress endurance life is achieved, and a reliable basis is provided for fastener raw material quality control.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of stress-rupture time prediction of high-temperature alloy, and relates to a stress-rupture time prediction method and system for cold-drawn GH4169 alloy rod. BACKGROUND

[0002] As the "heart" of aviation equipment, an aero-engine is a core symbol of national high-end equipment manufacturing capacity and comprehensive national strength. Under extreme working conditions (high temperature >750℃, stress >1000MPa, high-frequency vibration) of the engine, nickel-based high-temperature alloy fasteners, with their irreplaceable performance advantages, have become the core components of high-pressure rotor coupling bolts, turbine disc fasteners, and key load-bearing parts of the transmission system of military / civil aviation engines. A large number of such fasteners are needed for a single advanced aero-engine, and their failure will directly lead to structural chain damage. With the accelerated development of China's large aircraft industry, the demand for high-end high-temperature alloy fasteners in the aerospace field has surged.

[0003] Among them, GH4169 alloy is a Ni-Cr-Fe-based precipitation-hardened wrought high-temperature alloy with excellent high-temperature strength, good fatigue resistance and corrosion resistance, and is one of the main materials for manufacturing aerospace fasteners. The connecting structure is prone to stress-rupture failure under high-altitude complex environment, which poses a great threat to the safety of the whole machine. The GH4169 alloy rod for fasteners is mostly in a cold-drawn state, which is formed through processes such as forging, rolling, and cold drawing. The microstructure and stress-rupture performance of the rod have an important influence on the performance of the fastener product, so the fastener manufacturer needs to quickly evaluate the microstructure and performance of the raw material when purchasing GH4169 alloy rod for fasteners to avoid producing unqualified samples. Although the national standard (GB / T 14994-2008) requires that the stress-rupture performance of GH4169 cold-drawn rod be tested at 650℃, 690MPa for 23h or more, and the elongation after fracture be 4%, the standard requirement is too broad, and there is still a gap in ensuring the stress-rupture performance of the fastener product. Therefore, it is necessary to more accurately predict the stress-rupture performance of GH4169 cold-drawn rod to ensure the performance of the fastener product. SUMMARY

[0004] Therefore, the present application provides a stress-rupture time prediction method and system for cold-drawn GH4169 alloy rod to solve the problem that the prior art only relies on loose national standards for post-inspection, cannot quickly and accurately predict the stress-rupture life of GH4169 cold-drawn rod under corresponding test conditions based on its microstructure, and leads to insufficient guarantee of the qualification rate of fasteners.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] A stress-rupture time prediction method of cold-drawn GH4169 alloy rod, comprising the following steps:

[0007] S1, data collection: collect text data of original components, organizational characteristics and heat treatment process of cold-drawn GH4169 alloy rod and image data of microstructure after heat treatment, arrange the text data into a table and save it as a CSV file, and collect the pictures in the image data into a Data folder. While arranging the text data, pay attention to the one-to-one correspondence between the data in the CSV file and the pictures in the image data;

[0008] S2, initialization and data preparation: set model parameters and read CSV file data in step S1 for data processing, i.e. divide the training set, validation set and test set according to the ratio of 14:3:3;

[0009] S3, model construction: based on ResNet-18 architecture, construct a physical prior multi-modal neural network model including text data branch and image data branch and perform feature fusion to output life prediction value, wherein the physical constraint is added to the image data branch part of the physical prior multi-modal neural network model to make the image feature part prediction conform to the real physical law;

[0010] The physical constraint empirical formula is:

[0011]

[0012] In the formula, represents the final output value or result, and represents the target label Duration (endurance life), represents the “ content(after)” feature, and represents the delta content, represents the optimal value when it reaches its maximum value, represents the possible maximum output value, used to control the “steepness” or “width” of the curve;

[0013] S4, model training and evaluation: based on Huber regression loss function, update the model parameters through Adam optimizer, and evaluate the model performance by using loss loss curve, MAE and R² index;

[0014] S5, result visualization: visualize the training process and prediction results.

[0015] A stress-rupture time prediction system for cold-drawn GH4169 alloy rod, comprising:

[0016] An acquisition module is configured to acquire original data of cold-drawn GH4169 alloy bar, including composition, microstructure characteristics, and heat treatment process structure data, and image data after heat treatment;

[0017] A model construction module is configured to construct a physical prior multi-modal neural network model including an image branch and a structure data branch based on a ResNet-18 architecture, and perform feature fusion to output a life prediction value.

[0018] A model training and evaluation module is configured to train and update model parameters based on a Huber regression loss function through an Adam optimizer, and evaluate model performance by using a loss loss curve, MAE, and R² index.

[0019] A visualization module is configured to visualize the training process and prediction results.

[0020] The present application has the following advantages:

[0021] 1. The stress-rupture life prediction method of the cold-drawn GH4169 alloy bar disclosed in the present application is configured to perform solid solution treatment (960℃-1020℃ / 0.5-3h, oil cooling) and double aging treatment (720℃ / 8h→620℃ / 8h, oil cooling) on the bar; the grain size and carbide content are counted by a metallographic microscope, the δ phase content, size are quantitatively analyzed by a scanning electron microscope (SEM) and classified according to the morphology (granular / short rod / long needle); the stress-rupture experiment (650℃ / 690MPa constant load for 23h, then increase the load by 35MPa every 8h until fracture) is performed; the composition, microstructure parameters, process data and images (≥350) of the bar are integrated, the image features and fully connected layer coding table data are extracted based on ResNet-18, a multi-modal deep learning model is constructed, and the training set, prediction set and validation set are divided according to 14:3:3; the model is used to predict the life and compared with the experimental results for verification. The method first integrates multi-modal microstructure data and deep learning to realize high-precision prediction of stress-rupture life, and provides a reliable basis for fastener raw material quality control.

[0022] 2. The stress rupture time prediction method for cold-drawn GH4169 alloy rod disclosed in the application is used for the prediction of the stress rupture life of GH4169 alloy in the form of text data and image data, wherein the image feature extraction process is integrated with physical constraints, so that multiple forms of pictures can be normalized, the model has better data preprocessing and alignment performance compared with other same type models, can accept multiple forms of pictures, and the traditional method depends on a single data source. The multi-modal fusion breaks through the limitation of a single data source, so that the model can capture the complex and nonlinear correlation between different modal data, thereby having a more comprehensive understanding of the material performance. The addition of physical constraints increases the interpretability and robustness of the model on the basis of improving the performance of the model, which makes the prediction of the model not only statistically related, but also physically reasonable, enhances the trustworthiness and practicality of the model, and in the context of material data scarcity, the physical constraints can make up for the lack of data quantity and reduce the excessive dependence of the model on pure data mode, so that the model can have better generalization ability under limited data.

[0023] Other advantages, objects, and features of the application will be in part apparent and in part pointed out hereinafter in the specification, and it is to be understood that both the foregoing description and the appended claims are exemplary and explanatory thereof, and are intended to be as broad as possible. As will be seen from the specification, the objects and advantages of the application can be obtained by means which are presently preferred. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to make the objects, technical solutions and advantages of the application clearer, the preferred detailed description of the application will be combined with the drawings, in which:

[0025] Figure 1 The flowchart of the stress rupture time prediction method for cold-drawn GH4169 alloy rod of the application is shown in the figure.

[0026] Figure 2 The main step flowchart of step S1 of the application is shown in the figure.

[0027] Figure 3 The metallographic and scanning graphs of the sample after heat treatment in step S1 of the embodiment are shown in the figures, wherein Figure 3 (a) is the metallographic graph of the sample after heat treatment, Figure 3 (b) is the scanning graph of the sample after heat treatment.

[0028] Figure 4 The training and verification loss graph of step S4 of the embodiment is shown in the figure.

[0029] Figure 5 The comparison graph of the prediction result and the experimental result in step S5 of the embodiment is shown in the figure. DETAILED DESCRIPTION

[0030] Following, the embodiments of the present application will be described in detail by specific examples. Other advantages and effects of the present application can be easily understood by those skilled in the art from the disclosure of this specification. The present application can also be implemented or applied by other different specific embodiments, and various modifications or changes can be made to the details in this specification based on different views and applications without departing from the spirit of the present application.

[0031] As Figure 1 indicated, the stress-rupture time prediction method of the cold-drawn GH4169 alloy bar comprises the following steps:

[0032] S1, data collection: collect text data of original components, microstructure and heat treatment process of cold-drawn GH4169 alloy bar and image data of microstructure after heat treatment, arrange the text data into a table and save it as a CSV file, and collect the pictures in the image data into a folder. The text data is arranged at the same time, and the data in the CSV file and the pictures in the image data are one-to-one corresponding;

[0033] Specifically, as Figure 2 indicated, the acquisition of the original data of the cold-drawn GH4169 alloy bar is specifically:

[0034] S11, the cold-drawn GH4169 alloy bar is subjected to solution treatment and double aging treatment; the temperature range of the solution treatment is 960-1020℃, and the gradient adjustment is performed at an interval of 10℃; the time range of the solution treatment is 0.5-3h, and the gradient adjustment is performed at an interval of 0.5h, and the solution treatment is followed by oil cooling; the double aging treatment process is to heat at 720℃ for 8h, to cool in the furnace at a rate of 50±10℃ / h to 620℃ and heat for 8h, and to oil cool to room temperature after the double aging treatment;

[0035] S12, metallographic preparation and corrosion of the heat-treated sample; after the heat treatment of step S11, the sample is subjected to metallographic preparation and corrosion, the heat-treated sample is cut, and after rough grinding→fine grinding→polishing, the sample can be corroded, and 3g of anhydrous copper chloride (CuCl2) powder is accurately weighed for preparing the corrosion solution. 10ml of 36%-38% concentrated hydrochloric acid of analytical purity is measured and mixed with CuCl2. Then, 5ml of anhydrous methanol (analytical pure) is added. A cotton ball soaked with the corrosion solution is gently and uniformly wiped on the polished, clean and dry observation surface of the GH4169 sample for 1-2 seconds.

[0036] S13, observe the microstructure by metallographic microscope, and count the grain size and carbide content; the microstructure of the well-prepared observation surface of the sample is analyzed by optical microscope (metallographic microscope). Under 500 times magnification, 3 to 5 representative fields are randomly selected for image acquisition to ensure that different areas of the sample are covered to obtain statistically meaningful data. The obtained high-definition metallographic photos are imported into professional image analysis software Image-Pro Plus 6.0 for processing. The area method is used to automatically calculate the average equivalent diameter of multiple grains. For the content of carbides, the threshold value is set for binary segmentation according to the significant gray / color difference between the carbides and the matrix, and the carbide area is accurately extracted and the percentage of the total area of the field is calculated. The statistical results of each field are averaged, and finally the average grain size and average carbide content of the sample are reported respectively.

[0037] S14, observe by scanning electron microscope phases, count their content, size and morphological distribution; the microstructure of the sample is observed under scanning electron microscope (SEM). In order to accurately characterize the distribution state of delta phase, 5000 times high magnification is used for observation. In order to obtain statistically representative data, 3 to 5 different fields are randomly selected on the observation surface of the sample to take high-definition secondary electron images (SEI), avoiding local segregation area. Then, the obtained electron microscope photos are processed by using professional image analysis software (such as Image-Pro Plus), and the gray threshold is set to accurately identify and extract delta phase contour, calculate the proportion of its area in the total area of the field, so as to quantitatively count the content of delta phase. The statistical results of each field will be averaged and the standard deviation will be calculated to finally determine the average content of delta phase and the data dispersion degree of the sample, to ensure the objectivity and accuracy of the results.

[0038] According to the morphology and distribution of delta phase, the classification is carried out; according to the morphology and distribution of delta phase, the classification is carried out: (1) delta phase is distributed in the grain and grain boundary, and is in the form of particle / short rod / long needle; (2) delta phase is distributed in the grain and grain boundary, and is in the form of particle; (3) delta phase is distributed in the grain boundary, and is in the form of short rod.

[0039] S16, the component data, heat treatment process data, initial δ phase content / size, grain size, δ phase content / size after heat treatment of the cold-drawn GH4169 alloy bar in steps S11-S15 are summarized into a table and saved as a CSV file, and the CSV file is normalized and coded. The metallographic and scanning images of the metallographic microstructure of the cold-drawn GH4169 alloy bar in steps S13 and S14 are collected into the Data folder, and the pixel values of the Data folder images are normalized.

[0040] S17, the GH4169 alloy bar after heat treatment is subjected to a stress rupture test; the GH4169 high-temperature alloy bar after solid solution + aging standard heat treatment is processed into a smooth cylindrical stress rupture specimen with a diameter of φ3mm according to GB / T 2039. The surface of the specimen is precisely polished (Ra ≤ 0.3μm) to ensure that there are no machining defects. The specimen is placed in a radiant heating furnace, and the temperature is controlled by a K-type thermocouple in a closed loop, with a target temperature of 650℃ and a temperature control accuracy of ±1℃. A constant tensile stress of 690MPa is applied in the initial stage of the experiment, and maintained for 23 hours. After 23 hours, enter the stepwise loading stage: every 8 hours, the axial tensile stress is automatically increased by 35MPa according to the preset program. This process is repeated until the specimen breaks. The device automatically captures the precise time at the moment of rupture and records the final stress rupture time. The true value after the experiment is obtained through step S17, which is convenient for comparison with the model prediction value to verify the prediction effect of the model.

[0041] In machine learning, data normalization and encoding in CSV table files in text data are important steps in data preprocessing, which helps to improve the performance and convergence speed of the model. The normalization of the Data folder images refers to scaling the pixel values of the images to a specific range, usually [0, 1]. This can speed up the model training, improve the convergence and stability of the model.

[0042] S2, initialization and data preparation: set the model parameters and read the CSV file data in step S1 for data processing, i.e. divide the training set (245 groups), prediction set (52 groups) and validation set (53 groups) in the ratio of 14:3:3;

[0043] S3, model construction: based on the ResNet-18 architecture, a physical prior multi-modal neural network model including text data branch and image data branch is constructed, and feature fusion is performed to output the life prediction value;

[0044] Specifically, the table in the text data branch of the physical prior multi-modal neural network model receives the components (such as Al, Ti content), initial δ phase content / size, grain size, heat treatment process parameters, δ phase content / size after heat treatment The structural data such as phase content and size are encoded into tabular high-dimensional feature vectors through a fully connected layer. The image data branch utilizes a pre-trained ResNet-18 convolutional neural network (CNN) to extract features from the metallographic and scanning electron microscope images. After global average pooling, the image high-dimensional feature vectors are reduced in dimension. The image feature vectors and tabular feature vectors are spliced or fused and input into the subsequent fully connected layer. Finally, the life prediction value of the material is output.

[0045] A convolutional neural network (CNN), also known as ConvNet, is a deep learning model specifically designed for processing images and similar grid-structured data. It is a broad category in the field of deep learning. ResNet (Residual Network) is a special type (or architecture) of convolutional neural network. ResNet-18 specifically refers to a model in the ResNet series with 18 layers. ResNet-18 is a CNN, a specific architecture of deep convolutional neural networks.

[0046] The physical prior multi-modal neural network model is a dual-branch multi-modal deep learning model that combines physical prior regularization. The image data branch part adds physical constraints to make the image feature part prediction conform to the real physical law. The physical constraint empirical formula is:

[0047]

[0048] In the formula, represents the final output value or result, and represents the target label Duration (endurance life), represents the “ content(after)” feature, which represents the delta content, represents the optimal value when it reaches its maximum value , represents the possible maximum output value. Used to control the “steepness” or “width” of the curve. A larger value will make the curve narrower and sharper, which means that the performance is very sensitive to changes in the phase content, and a slight deviation from the optimal value will cause a sharp decline in performance. A smaller value will make the curve wider and flatter, meaning that the performance is not as sensitive to changes in the phase content, and the performance fluctuation is not large within a certain range.

[0049] The image feature extraction process based on physical constraints is as follows:​​​

[0050] S31, raw_v_delta value is obtained when data is loaded, and passed to the training function through DataLoader.

[0051] S32, the train_model_process function calls the model, and passes the image, table data, raw_v_delta and physical parameters into the model, and the total loss function is the weighted sum of the task loss (task_loss) and the physical loss (phys_loss): total_loss = task_loss + lambda_reg * phys_loss. So when calculating the gradient by backpropagation, the gradient of this total loss will flow through all the parameters of the entire neural network, including the image feature extraction part (convolutional layers of ResNet).

[0052] Among them, the build_resnet model internally contains CNN convolutional layers for extracting image features, and the model calculates the physical branch L_phys_normalized using raw_v_delta and physical parameters. The physical prediction value L_phys_normalized is calculated based on the physical formula using the input raw_v_delta and physical parameters. This physical prediction value has a direct physical relationship with the phase volume fraction.

[0053] In the train_model_process function, the physical loss phys_loss and the total loss total_loss are calculated, and the physical loss is added as a regularization term to the total loss. phys_loss measures the gap between the model's predicted value (L_model_normalized) and the physically formula predicted value (L_phys_normalized). Through phys_loss, the model is forced to make its predicted value L_model_normalized closer to L_phys_normalized given by the physical formula. This means that the output of the model must simultaneously meet the requirements of fitting the real label and the physical law between phase volume fraction and service life.

[0054] S33, backpropagation process: the gradient of the total loss is returned, guiding the learning of the entire network parameters including image feature extraction layers, so that its prediction meets the task requirements and physical constraints. In order to reduce the physical loss, the neural network extracts features from images and table data that are consistent with the change trend of the phase volume fraction. In particular, for the image part, the convolutional layers of ResNet extract features that are consistent with the change trend of ​The visual information related to the phase morphology, distribution, quantity, etc. is associated with the raw_v_delta, which is related to the life-related features.

[0055] S4, Model training and evaluation:

[0056] The model training is based on Huber regression loss function, trained by Adam optimizer (learning rate 0.001), and the core is data iteration. At the beginning of each Epoch, the data in the training set is divided into several small batches (batch) according to the set batch_size and sampling strategy (here is weighted random sampling), and the sample size is increased through data augmentation strategy. The processed image and table data are forward propagated, the loss is calculated, and finally the back propagation and parameter update are performed. The cycle of "forward propagation -> loss calculation -> back propagation -> parameter update" is repeated in each training batch. Specifically:

[0057] S41, Gradient zeroing

[0058] At the beginning of each batch training, the optimizer clears the gradient calculated in the previous batch to prevent gradient accumulation.

[0059] S42, Forward propagation

[0060] Image data enters CNN (ResNet-18 part) to extract image features. Table data enters MLP (fully connected layer) to encode table features. Image features and table features are fused in the model, and the fused features are sent to the subsequent fully connected layer after splicing. The fused features are further processed by MLP to obtain the final neural network prediction value (L_model_normalized). At the same time, raw_v_deltas are input into the model as parameters to calculate the physical formula prediction value (L_phys_normalized). The model outputs the normalized prediction value.

[0061] S43, Loss calculation

[0062] The task loss between the neural network prediction and the true label, the physical loss between the neural network prediction and the physical formula prediction, and the total loss are calculated in turn.

[0063] S44, Back propagation

[0064] According to the total_loss and the calculation graph, the gradients of all trainable parameters (weights and biases) in the model are automatically calculated.

[0065] S45, Parameter update

[0066] The Adam optimizer uses the computed gradients combined with learning rate, momentum, weight decay, etc. to adjust the parameters of the model to minimize the total_loss.

[0067] Model evaluation is based on quantitative calculations of predicted loss, mean absolute error, and R² score, and the validation set performance is continuously monitored during the training process, and finally validated on an independent test set. The Loss plot provides a dynamic view of the training process, MAE provides a direct numerical value of the average error, and R² evaluates the overall fitting ability of the model from the perspective of variance explanation. Combined with them, a comprehensive and in-depth understanding of the performance of the model can be obtained.

[0068] S5, result visualization: visualize the training process and prediction results. Visualize the training process by displaying the evaluation indicators of each round of training rounds, and finally generate the loss curve of the training results and the scatter plot of the true value and the predicted value, and then visualize the prediction results. Loss curve: show the change of training and validation loss with Epoch. Two R² scatter plots: show the true value and predicted value of the validation set prediction results of the best validation Epoch and the final test set prediction results, respectively, with R² score.

[0069] Figure 3 The metallographic and scanning images after heat treatment of the example step S1 are shown in Figure 3 (a) is the metallographic image of the sample after heat treatment, Figure 3 (b) is the scanning image of the sample after heat treatment; for the model, the two images mainly show the evolution of δ phase, the size difference before and after heat treatment, the volume fraction, etc.

[0070] Figure 4 The training and validation loss plot of the example step S4 is shown in the figure; the training and validation loss curve is used to monitor the learning process of the model, to judge whether there is overfitting or underfitting, and whether the physical constraints are effective. The closer the loss curve is to 0, the better the model performance, which intuitively shows the learning effect and fitting trend of the model on the training data and independent validation data.

[0071] Figure 5 The comparison chart of prediction results and experimental results of the example step S5 is shown in the figure. The validation set R² scatter plot shows the scatter distribution between the predicted value and the actual true value of the model on the validation set, and marks the corresponding R² score. It is a direct tool for evaluating the goodness of fit of the regression model. It directly shows the consistency between the predicted value of the model and the actual true value. The higher the R² value, the stronger the ability of the model to explain the variance of the data, and the better the prediction effect.

[0072] Finally, it is to be explained that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions, and all should be covered in the scope of the claims of the present application.

Claims

1. A method for predicting the stress persistence time of cold-drawn GH4169 alloy bars, characterized in that: The following steps are involved: S1. Data collection: Collect text data on the original composition, microstructure characteristics, and heat treatment process of cold-drawn GH4169 alloy bars, and image data on the metallographic microstructure after heat treatment. Organize the text data into tables and save them as CSV files. Collect the images in the image data into the Data folder. When organizing the text data, ensure that the data in the CSV file corresponds to the images in the image data. S2, initialization and data preparation: set model parameters and read the CSV file data in step S1 for data processing, that is, divide the training set, validation set, and test set into a ratio of 14:3:3; S3. Model Construction: Based on the ResNet-18 architecture, a physical priori multimodal neural network model is constructed, including a text data branch and an image data branch. Feature fusion is performed to output the lifespan prediction value. Physical constraints are added to the image data branch of the physical priori multimodal neural network model to ensure that the prediction of the image feature part conforms to the laws of real physics. The empirical formula for physical constraints is: Where, Indicates the final output value or result, representing the target tag Duration (persistent lifespan), represent" content(after)” feature, indicating delta content, express Reach its maximum value The optimal value, represents the maximum possible output value, Used to control the "steepness" or "width" of the curve; S4. Model training and evaluation: Based on the Huber regression loss function, the Adam optimizer is used to train and update the model parameters. The model performance is evaluated using the loss curve, MAE, and R² indicators. S5. Result visualization: visualize the training process and prediction results.

2. The stress persistence time prediction method according to claim 1, wherein: The composition and heat treatment process data in step S1 include solution temperature / time, aging temperature / time; process data include endurance test conditions, Phase morphology classification; image data after heat treatment includes metallographic microscope images and SEM scanning electron microscope images, with ≥350 pieces of collected data.

3. The stress persistence time prediction method according to claim 2, wherein: The acquisition of raw data of cold-drawn GH4169 alloy bar in step S1 is specifically as follows: S11, performing solution treatment and double aging treatment on the cold drawn GH4169 alloy bar; S12, performing metallographic preparation and etching on the heat-treated sample; S13. Observe the microstructure through a metallographic microscope and calculate the grain size and carbide content; S14. Observation by scanning electron microscopy phase, and count its content, size and morphological distribution; S15. The composition data, heat treatment process data, initial δ phase content / size, grain size, and δ phase content / size after heat treatment of the cold-drawn GH4169 alloy bar in steps S11 to S15 are summarized and organized into a table and saved as a CSV file. The CSV file is normalized and encoded. The image files of the metallographic images and scanned images of the metallographic microstructure of the cold-drawn GH4169 alloy bar in steps S13 and S14 are uniformly collected into a Data folder, and the pixel values ​​of the images in the Data folder are normalized.

4. The stress persistence time prediction method according to claim 1, wherein: The model parameters in step S2 include hyperparameters such as image size, learning rate, and batch size.

5. The stress persistence time prediction method according to claim 4, wherein: In step S3, the table of the text data branch in the physical prior multimodal neural network model receives components, initial Phase content / size, grain size, heat treatment process parameters, after heat treatment Structural data such as phase content and size are encoded into tabular high-dimensional feature vectors through the fully connected layer. The image data branch uses the pre-trained ResNet-18 convolutional neural network (CNN) to extract features from metallographic and scanning electron microscope images. After global average pooling, the dimension is reduced to a high-dimensional image feature vector. The image feature vector and the tabular feature vector are spliced ​​or fused and input into the subsequent fully connected layer to finally output the material life prediction value.

6. The stress persistence time prediction method according to claim 5, wherein: The specific process of image feature extraction based on physical constraints in step S3 is as follows: S31. Get the original raw_v_delta value when loading data and pass it to the training function through DataLoader; S32, train_model_process function calls the model, passing the image, table data, raw_v_delta, and physical parameters into the model. The model contains a CNN convolution layer to extract image features. The model uses raw_v_delta and physical parameters to calculate the physical prediction value L_phys_normalized and establishes a direct physical relationship with the phase volume fraction through the physical branch. Calculate phys_loss and total_loss, and add the physical loss as a regularization term to the total loss: total_loss = task_loss + lambda_reg * phys_loss. When backpropagating to calculate the gradient, the gradient of this total loss flows through all parameters of the entire neural network, including the image feature extraction part. In the train_model_process function, the physical loss phys_loss and the total loss total_loss are calculated, and the physical loss is added to the total loss as a regularization term; phys_loss measures the gap between the model prediction value L_model_normalized and the physical formula prediction value L_phys_normalized; through phys_loss, the model is forced to make its prediction value L_model_normalized closer to the L_phys_normalized given by the physical formula; this means that the output of the model must simultaneously meet the requirements of fitting the real label and The physical laws between phase volume fraction and lifetime; S33, backpropagation process: The gradient of the total loss is returned to guide the learning of the entire network parameters, including the image feature extraction layer, so that its predictions meet both task requirements and physical constraints.

7. The stress persistence time prediction method according to claim 6, wherein: In step S4, the model training is based on the Huber regression loss function and is performed by the Adam optimizer (learning rate 0.001) training, the core of which is data iteration. At the beginning of each epoch, the data in the training set is divided into several small batches according to the set batch_size and sampling strategy. The number of samples is increased through data augmentation strategies. The processed image and table data are forward propagated to calculate the loss, and finally backpropagated and parameter updated. The "forward propagation -> loss calculation -> backpropagation -> parameter update" cycle is repeated for each training batch. Model evaluation is based on the quantitative calculation of prediction loss, mean absolute error, and R² score. The performance of the validation set is continuously monitored during training, and the final verification is performed on an independent test set.

8. The stress persistence time prediction method according to claim 7, wherein: The model training process in step S4 is as follows: S41, gradient clear At the beginning of each batch training, the optimizer clears the gradients calculated in the previous batch to prevent gradient accumulation; S42, forward propagation Image data enters the CNN to extract image features; table data enters the fully connected layer and is encoded as table features. The image features and table features are fused within the model, concatenated, and fed into the subsequent fully connected layer. The fused features are further passed through the fully connected layer to obtain the final neural network prediction value. At the same time, raw_v_deltas is passed into the model as a parameter to calculate the physical formula prediction value, and the model outputs the normalized prediction value. S43. Loss calculation Calculate the task loss between the neural network prediction and the true label, the physical loss between the neural network prediction and the physical formula prediction, and the total loss in turn; S44. Backpropagation Automatically calculate the gradients of all trainable parameters in the model based on total_loss and the computational graph; S45. Parameter update The Adam optimizer uses the calculated gradient combined with learning rate, momentum, weight decay, etc. to adjust the model parameters to minimize total_loss.

9. The stress persistence time prediction method according to claim 8, wherein: Step S5 visualizes the training process by displaying the evaluation indicators of each training round, and finally generates a loss curve graph of the training results and a scatter plot comparing the true value and the predicted value, thereby visualizing the prediction results.

10. A stress persistence time prediction system for cold-drawn GH4169 alloy bars, based on the stress persistence time prediction method according to any one of claims 1 to 9, characterized in that: include: The acquisition module is used to obtain the original data of the cold-drawn GH4169 alloy bar, including composition, microstructure characteristics, heat treatment process structure data, and image data after heat treatment; The model building module builds a physical priori multimodal neural network model based on the ResNet-18 architecture, including image branches and structural data branches, and performs feature fusion to output life prediction values; The model training and evaluation module uses the Adam optimizer to train and update model parameters based on the Huber regression loss function, and uses loss curves, MAE, and R² indicators to evaluate model performance. Visualization module, which visualizes the training process and prediction results.

Images