Martensitic heat-resistant steel creep state and residual life prediction method, system and storage medium based on multi-task deep learning

Abstract

The application discloses a martensitic heat-resistant steel creep state and residual life prediction method and system based on multi-task deep learning and a storage medium, and the method comprises the following steps: acquiring microstructure images of a plurality of creep stages of a sample steel, preprocessing and establishing stage labels and life consumption ratio labels to construct a sample set; dividing a training set, a verification set and a test set according to sample sources, and ensuring that the same sample does not cross the sets; constructing and training a multi-task deep learning model comprising a shared feature extraction network, a classification branch and a regression branch, and adopting a joint loss function for optimization; and predicting the creep stage and continuous life consumption ratio by using the model after training, and determining the residual life. The application realizes the promotion from qualitative rating to quantitative characterization of the creep damage, reduces the artificial subjectivity, and improves the efficiency and consistency.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of material failure analysis and computer vision, and in particular to a method, system, and storage medium for predicting the creep state and remaining life of martensitic heat-resistant steel based on multi-task deep learning. Background Technology

[0002] Martensitic heat-resistant steels, represented by P91 steel, are widely used in high-temperature pressure-bearing components such as main steam pipes and reheater headers in ultra-supercritical units. Under long-term high-temperature and high-pressure service conditions, the material will undergo creep damage. The tempered martensite laths gradually widen, break, and become blurred, eventually evolving into equiaxed subgrains, coarsening of M23C6 carbides, subgrain formation and growth, Laves phase precipitation, and final recrystallization. Accurately determining the stage of creep damage and quantitatively predicting the remaining service life are key technical bases for power plants to formulate maintenance strategies.

[0003] Currently, the industry mainly relies on professionals to perform on-site metallographic imaging or take samples from tubes and observe the metallographic structure in the laboratory, classifying the creep aging level by comparing it with standard atlases. This method has the following main problems: it is highly subjective, and different technicians may have different judgments on the same metallographic image; it is inefficient, as manual analysis of a single image is time-consuming; it can usually only provide qualitative or semi-quantitative aging levels, and it is difficult to provide specific quantitative indicators such as the percentage of remaining life or hours; and it has high labor costs and difficulty in ensuring consistency when handling large-scale testing tasks.

[0004] In recent years, some studies have attempted to use convolutional neural networks to classify the aging of metallographic images of heat-resistant steel, but certain limitations remain: First, most of these studies use classification frameworks to output discrete aging levels, which are essentially qualitative assessments and cannot provide direct quantitative basis for precise maintenance decisions based on remaining service life; second, existing regression methods are mostly based on macroscopic mechanical performance parameters (such as hardness and impact energy), and have not established a direct mapping relationship between microstructure and lifespan; third, existing image analysis methods do not fully consider the independence of sample levels when dividing data, which may lead to different images of the same sample being scattered in the training and testing sets, resulting in an overestimation of model performance and generalization performance. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method, system, and storage medium for predicting the creep state and remaining life of martensitic heat-resistant steel based on multi-task deep learning, which can automatically, objectively, and quantitatively predict the creep state and remaining life of steel.

[0006] Technical Solution: To achieve the above objectives, the method for predicting the creep state and remaining life of martensitic heat-resistant steel based on multi-task deep learning, as described in this invention, includes the following steps:

[0007] S1. Obtain metallographic images of the sample steel at multiple different creep stages, preprocess each metallographic image, and establish creep stage category labels and lifetime consumption ratio labels for it to construct a sample set.

[0008] S2. Based on the sample source information corresponding to the images in the sample set, the sample set is divided into a training set, a validation set, and a test set. Among them, images from the same sample source only appear in one of the training set, validation set, and test set datasets.

[0009] S3. Construct a multi-task deep learning model, which includes a shared feature extraction network and classification and regression branches connected to it; the shared feature extraction network is used to extract micro-tissue features related to creep damage, the classification branch is used to output the probability distribution of creep stages, and the regression branch is used to output the continuous lifetime loss ratio prediction value.

[0010] S4. Use the sample set to train, validate and test the multi-task deep learning model. During training, a joint loss function is used, which includes a classification loss term and a regression loss term.

[0011] S5. Input the metallographic image of the steel to be predicted into the trained multi-task deep learning model to obtain the probability distribution of creep stage and the predicted value of life consumption ratio, and determine the creep stage and remaining life of the steel to be predicted.

[0012] Through the above steps, this invention utilizes a multi-task deep learning model to automatically extract microstructural features related to creep damage from metallographic images, simultaneously achieving qualitative classification of the creep stage of the steel under test and quantitative prediction of its lifespan consumption ratio. The sample set grouping strategy effectively avoids the same sample being placed in the training, validation, and test sets. This grouping method is mainly used to prevent similar or enhanced images from the same sample source from simultaneously entering the training and test sets, reducing the risk of inflated test results due to data leakage. The use of a joint loss function makes the classification and regression tasks mutually constrained, helping to improve the accuracy of regression prediction. The overall method reduces the subjectivity of manual rating, providing both qualitative and quantitative detection, and improving detection efficiency and result consistency.

[0013] Preferably, in step S1, the multiple different creep stages include at least five levels; the lifetime consumption ratio label is determined based on the ratio of the current creep time to the total creep fracture time.

[0014] By defining multiple discrete creep stage categories and continuous lifetime consumption ratio labels based on time ratios, a clear and unambiguous supervision signal with physical meaning is provided for model training, which is conducive to the model learning a continuous mapping relationship from microstructure morphology to damage degree.

[0015] Preferably, in step S2, the sample source information is the sample number, which is extracted from the image file name or image metadata; after the sample set is divided, the images in the training set are subjected to data augmentation processing.

[0016] Dividing the sample set based on the sample number ensures that different images collected from the same sample are not scattered across different datasets (training set, validation set, and test set); data augmentation is performed only on the training set, which expands the sample diversity while ensuring the independence of the validation set and test set and the authenticity of the evaluation results.

[0017] Preferably, the shared feature extraction network is a convolutional neural network, a visual Transformer network, or a combination thereof.

[0018] Preferably, the shared feature extraction network uses the EfficientNetV2-S network and initializes its parameters using pre-trained weights.

[0019] Preferably, the EfficientNetV2-S network outputs a feature vector after global average pooling; the classification branch has the following structure: linear layer, batch normalization layer, ReLU activation layer, Dropout layer, and linear layer; the regression branch has the following structure: linear layer, batch normalization layer, ReLU activation layer, Dropout layer, linear layer, and Sigmoid activation layer.

[0020] Preferably, in step S4, the classification loss term is the labeled smoothed cross-entropy loss, and the regression loss term is the mean squared error loss; during training, the AdamW optimizer, learning rate decay strategy, and early stopping mechanism are used.

[0021] Labeled smoothing of cross-entropy loss can effectively suppress overfitting, while mean squared error loss is suitable for continuous value regression. The AdamW optimizer, combined with learning rate decay and early stopping mechanisms, can synergistically improve training stability, generalization performance and efficiency, prevent overfitting and avoid invalid computation.

[0022] The creep state and remaining life prediction system for martensitic heat-resistant steel based on multi-task deep learning described in this invention includes:

[0023] The image acquisition module is used to acquire images of the metallographic structure of the steel to be predicted.

[0024] The image preprocessing module is used to perform size adjustment, pixel value normalization, and standardization on metallographic images;

[0025] The model invocation module is used to invoke the multi-task deep learning model trained by the method according to any one of claims 1 to 7;

[0026] The inference module is used to input the preprocessed metallographic images into the multi-task deep learning model and obtain the probability distribution of the creep stage and the predicted value of the lifetime consumption ratio.

[0027] The post-processing module is used to determine the creep stage of the steel to be predicted based on the creep stage probability distribution, and to calculate the remaining life of the steel to be predicted based on the life consumption ratio prediction value.

[0028] The display module is used to show the prediction results of the creep stage and the remaining lifetime.

[0029] Preferably, an interpretability analysis module is also included. The interpretability analysis module is used to calculate the gradient of the output score corresponding to the predicted category of the classification branch or the lifetime consumption ratio of the regression branch as the target output, and to generate an attention heatmap by using gradient global average pooling, feature map weighted summation, ReLU and normalization processing. The attention heatmap is then overlaid with the original metallographic image for display.

[0030] The present invention provides a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0031] Beneficial Effects: By constructing and training a multi-task deep learning model that includes a shared feature extraction network, classification branches, and regression branches, the model can simultaneously and automatically predict the creep damage stage category and the continuous lifespan ratio from the microstructure image of the steel under test, achieving both qualitative and quantitative assessments. Data leakage is avoided through grouped data partitioning based on sample origin information, improving the reliability of the model's generalization ability assessment. The overall technical solution is objective, efficient, and consistent, providing strong support for lifespan assessment and maintenance decisions for high-temperature pressure-bearing components. Attached Figure Description

[0032] Figure 1 This is an overall flowchart of the present invention;

[0033] Figure 2 This is a diagram of the network structure of a multi-task deep learning model.

[0034] Figure 3 The training and validation loss curves are used during the training process;

[0035] Figure 4 To train and validate the accuracy curve during the training process;

[0036] Figure 5 Metallographic example image of creep after 1000 hours;

[0037] Figure 6 Example of a GradCAM heatmap for 1000 hours of creep;

[0038] Figure 7 This is the confusion matrix of the multi-task deep learning model of this invention on the test set;

[0039] Figure 8 This is the confusion matrix for the comparison model's test set. Detailed Implementation

[0040] The technical solution of the present invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0041] As shown in the figure, the method for predicting the creep state and remaining life of martensitic heat-resistant steel based on multi-task deep learning according to the present invention includes:

[0042] Step S1: Acquire metallographic images and construct a sample set

[0043] Metallographic images of the sample steel at multiple different creep stages were acquired. Each metallographic image was preprocessed, and creep stage category labels and lifetime consumption ratio labels were established to construct a sample set.

[0044] P91 steel specimens in service condition or subjected to accelerated creep in the laboratory can be used, with creep times controlled to 200 hours, 400 hours, 600 hours, 800 hours, and at break (approximately 1000 hours). Metallographic specimens should be prepared from at least three different locations at each stage.

[0045] The standard metallographic sample preparation process was adopted: cutting → mounting → grinding and polishing (up to 2000 grit sandpaper) → mechanical polishing → chemical etching (usually using ferric chloride hydrochloric acid alcohol solution). A total of 163 images were taken at 500x magnification using an optical microscope, covering five stages. After offline data augmentation, approximately 1000 images were obtained, and the training set, validation set, and test set were divided into 70%:15%:15%.

[0046] Image preprocessing operations include: resizing all images to a uniform size; normalizing pixel values; and then standardizing the image by subtracting the mean from each pixel and dividing by the standard deviation.

[0047] The creep stage category labels are divided into five levels according to creep time: 200h corresponds to category 0, 400h corresponds to category 1, 600h corresponds to category 2, 800h corresponds to category 3, and 1000h (tear-off) corresponds to category 4.

[0048] The formula for calculating the lifespan consumption ratio label is:

[0049] Lifetime attrition ratio = Current creep time / Total creep rupture time

[0050] The total creep rupture time was determined through accelerated creep rupture experiments in the laboratory. Using 1000 hours as the total creep rupture time, the lifetime attrition ratios were calculated to be 0.2 for 200 hours, 0.4 for 400 hours, 0.6 for 600 hours, 0.8 for 800 hours, and 1.0 for 1000 hours. The filename, stage category label, and lifetime attrition ratio label for each image were stored in a CSV file to complete the construction of the sample set.

[0051] Step S2: Grouping and partitioning the dataset at the sample level

[0052] Based on the sample source information corresponding to the images in the sample set, the sample set is divided into a training set, a validation set, and a test set. Images from the same sample source appear only in one of the training set, validation set, and test set.

[0053] In this embodiment, the image file name format is "creep time\sample number.tif", for example, "200h\0001.tif". The sample source information is the sample number, which is extracted from each file name (0001, 0002...0033) as the grouping basis. Two independent GroupShuffleSplit operations are used: the first grouping is based on the sample number, dividing all samples into a training set (70% of the total samples) and a temporary set (30%); the second grouping is based on the sample number again on the temporary set, dividing it into a validation set and a test set (15% each). After the division, the ratio of the number of images in the training set, validation set, and test set is 70:15:15.

[0054] After the sample set is divided, data augmentation is performed only on the images in the training set. Data augmentation methods include random cropping, horizontal flipping, small-angle rotation (±5°), brightness / contrast adjustment (±30%), and slight Gaussian noise. Five augmented copies are generated for each original training image, and these copies, along with the original image, are used as part of the training set. The validation and test sets are not augmented; the original images are used directly for evaluation after preprocessing as described in step S1.

[0055] Step S3: Construct a multi-task deep learning model

[0056] Construct a multi-task deep learning model, which includes a shared feature extraction network and classification and regression branches connected to the shared feature extraction network.

[0057] A shared feature extraction network is used to extract microscopic tissue features related to creep damage from the input image. In this embodiment, the shared feature extraction network adopts the EfficientNetV2-S network in convolutional neural networks, and the parameters are initialized by loading weights pre-trained on the ImageNet dataset. The input image size is 384×384×3 channels. The pre-trained EfficientNetV2-S is used as the shared backbone network, its original output layer for ImageNet classification is removed, and the convolutional feature extraction part and global average pooling layer are retained to obtain a 1280-dimensional feature vector, which is then input into the classification branch and regression branch respectively.

[0058] The classification branch is structured as follows: a fully connected layer (1280-dimensional → 512-dimensional), a batch normalization layer (BatchNorm1d), a ReLU activation layer, a dropout layer (dropout probability of 0.4), and a fully connected layer (512-dimensional → 5-dimensional). The last fully connected layer outputs the logits of the five creep stage categories, which can then be used to obtain the probability distribution of each creep stage via the Softmax function.

[0059] The regression branch consists of the following layers: a fully connected layer (1280-dimensional → 256-dimensional), a batch normalization layer (BatchNorm1d), a ReLU activation layer, a Dropout layer (with a dropout probability of 0.3), another fully connected layer (256-dimensional → 1-dimensional), and a Sigmoid activation layer. The Sigmoid activation layer outputs a floating-point number between 0 and 1, which represents the lifetime consumption ratio predicted by the model (the proportion of lifetimes consumed).

[0060] Step S4: Model Training, Validation, and Testing

[0061] The multi-task deep learning model is trained, validated, and tested using a sample set. During training, a joint loss function is used, which includes a classification loss term and a regression loss term.

[0062] The joint loss function of a multi-task deep learning network is defined as:

[0063] ,

[0064] Among them, classification loss Using cross-entropy loss and adding a label smoothing factor ε=0.05, its expression is:

[0065] ,

[0066] In the formula, The true label for the i-th stage (only the correct category is 1, and the rest are 0). Let be the probability of the i-th stage predicted by the model.

[0067] Regression loss The mean squared error loss is expressed as follows:

[0068] ,

[0069] In the formula, This represents the lifetime attrition ratio predicted by the model. This represents the actual lifetime attrition ratio. The two loss components are directly added together, with each having a weighting factor of 1.0.

[0070] Optimizer: AdamW, initial learning rate lr=0.0001, weight decay weight_decay=5e-4.

[0071] Learning rate scheduling: Cosine AnnealingLR is used, with a total of 60 epochs and a minimum learning rate of 1e-7.

[0072] Early stopping strategy: Monitor the total_loss on the validation set. If the decrease in validation loss is less than 0.001 for 15 consecutive epochs, terminate training and backtrack to save the model weights at the point with the lowest validation loss as the optimal model weight file.

[0073] Step S5: Creep State Prediction and Remaining Life Determination

[0074] The metallographic image of the steel to be predicted is input into the trained multi-task deep learning model to obtain the probability distribution of creep stage and the predicted value of lifetime consumption ratio, and based on this, the creep stage and remaining life of the steel to be predicted are determined.

[0075] Specifically, the saved optimal model weight file is loaded, and the image to be predicted, after preprocessing as described in step S1, is fed into the model. After forward inference, the model outputs the probabilities of the five creep stages in the classification branch and the predicted lifetime consumption ratio p in the regression branch. The creep stage is selected as the category with the highest probability (i.e., the creep stage of the steel being tested). The predicted lifetime consumption ratio p is a continuous value between 0 and 1.

[0076] The remaining service life is calculated based on the predicted service life ratio. The remaining service life ratio can be calculated as 1-p. If the total creep rupture time T under the corresponding test conditions is known (determined by tensile testing), then the remaining test time under those test conditions can be calculated as (1-p)×T. For example, if the model outputs p = 0.78 and T = 1000 hours, then the remaining hours = (1 - 0.78)×1000 = 0.22×1000 = 220 hours. It should be noted that the above remaining hours are the remaining test time under laboratory accelerated creep test conditions. To convert them to the remaining operating hours under actual service conditions, extrapolation should be performed using the material's time-temperature parameter master curve, such as the Larson-Miller curve.

[0077] It should be further noted that although this embodiment uses P91 steel as an example for illustration, those skilled in the art will understand that the method of the present invention is also applicable to other heat-resistant steel materials with similar creep damage microstructure evolution laws, and only requires retraining or fine-tuning the model using the creep test data of the corresponding material.

[0078] The present invention relates to a system for predicting the creep state and remaining life of martensitic heat-resistant steel based on multi-task deep learning. The system includes an image acquisition module, an image preprocessing module, a model calling module, an inference module, a post-processing module, and a display module.

[0079] The image acquisition module is used to acquire metallographic images of the steel to be predicted. It can be connected to a microscope camera or read stored image files.

[0080] The image preprocessing module is used to resize the acquired images to a uniform value, normalize and standardize the pixel values ​​to adapt to the input requirements of the model.

[0081] The model calling module is used to load the multi-task deep learning model and its corresponding optimal model weights trained according to the method described in Embodiment 1.

[0082] The inference module is used to input the preprocessed image into the loaded multi-task deep learning model, perform forward inference, and obtain the probability distribution of the creep stage and the predicted lifetime consumption ratio.

[0083] The post-processing module is used to determine the creep stage of the steel to be predicted (taking the stage with the highest probability) based on the creep stage probability distribution, and to calculate the remaining life of the steel to be predicted based on the life consumption ratio prediction value. When the total creep fracture time T under the corresponding test conditions is input, the post-processing module can also calculate the remaining test time under the test conditions.

[0084] The display module is used to display the creep stage prediction results and remaining lifetime (which can be the remaining lifetime percentage or the remaining test time).

[0085] The prediction system also includes an interpretability analysis module. This module uses the output score corresponding to the predicted category in the model's classification branch or the lifetime consumption ratio output value in the regression branch as the target output. It calculates the gradient of this gradient with respect to the last convolutional feature map of the shared feature extraction network, and generates an attention heatmap through gradient global average pooling, feature map weighted summation, ReLU, and normalization. This attention heatmap is then overlaid on the original metallographic image for display. The heatmap allows users to visually observe the microscopic tissue regions that the model focuses on during stage identification or lifetime consumption ratio prediction. In this embodiment, the attention heatmap is generated using the GradCAM method.

[0086] This embodiment provides a computer-readable storage medium for storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform any implementation of the above-described method.

[0087] To better illustrate the present invention, a specific example is provided below:

[0088] S1. Data Acquisition Details

[0089] A batch of P91 steel samples were subjected to accelerated creep tests at 650℃ and 94.3MPa. Samples were taken after furnace shutdown at 200h, 400h, 600h, 800h, and 1000h. In this embodiment, the total creep rupture time was 999h, but for ease of explanation, it was approximated as 1000h. Therefore, the lifetime consumption ratios corresponding to 200h, 400h, 600h, 800h, and 1000h were 0.2, 0.4, 0.6, 0.8, and 1.0, respectively. A 10mm × 10mm × 5mm metallographic block was cut axially from each sample, mounted, polished, and then etched with ferric chloride alcohol solution for 30 seconds. Using an optical microscope at 500x magnification, approximately 30 images were acquired for each stage, totaling 163 original images for the five stages.

[0090] S2. Sample numbering, extraction, and grouping

[0091] The image file names are in the format "200h\0001.tif", "400h\0001.tif", ..., "800h\0001.tif". A total of 163 metallographic images of P91 steel (representing five creep stages) were generated. To expand the training samples, the original images were first divided into training, validation, and test sets according to the sample number. Subsequently, data augmentation was performed only on the training set images, while the validation and test sets underwent only resizing, normalization, and standardization. Each image was randomly cropped, horizontally flipped, rotated at a small angle (±5°), its brightness / contrast adjusted (±30%), and slightly modified with Gaussian noise to generate five augmented copies, totaling 978 images along with the original images, forming the augmented dataset. To avoid data leakage when dividing the training, validation, and test sets, the original P91 steel metallographic images were first divided into training, validation, and test sets based on the sample number, ensuring that all images corresponding to the same sample number entered only the same dataset. Subsequently, data augmentation is performed only on the training set images. Augmentation methods include random cropping, horizontal flipping, small-angle rotation, brightness / contrast adjustment, and slight Gaussian noise. The validation and test sets are not randomly augmented; they are only resized, normalized, and standardized. This method avoids the original or augmented images of the same sample appearing simultaneously in the training, validation, and test sets, thereby improving the reliability of the model's generalization performance evaluation.

[0092] S3, Model Training and Testing

[0093] Using the PyTorch framework, the model was trained for 60 epochs. An early stop was triggered on the 31st epoch (to verify that the loss decreased by less than 0.001 for 15 consecutive epochs). At this point, the saved model achieved 95.5% accuracy on the test set. The software framework described above is merely one implementation method in this embodiment and does not constitute a limitation on the scope of protection of this invention.

[0094] S4, Prediction Example

[0095] Running the Streamlit application and uploading the same image, the system outputs: Predicted creep stage: 800h; Lifetime consumption: 82.93%; Remaining lifetime under this test condition: 170.7 hours; Probabilities of each stage: 200h: 4.19%, 400h: 0.22%, 600h: 13.97%, 800h: 78.74%, 1000h: 2.89%. It should be noted that the "170.7 hours" is the remaining time under accelerated testing conditions of 650℃ / 94.3MPa. To obtain the remaining operating hours under service conditions of 550℃, extrapolation and conversion using the Larson-Miller master curve of the material are required.

[0096] Comparative Example

[0097] To verify the performance advantages of the multi-task model provided in this embodiment, a comparative experiment was conducted using the classic ResNet50 classification network under the exact same dataset and partitioning method. ResNet50 only performs creep stage classification (5 classes) and does not include a lifetime attrition ratio regression branch. Therefore, it can only output the discrete creep stage category corresponding to the metallographic image under test and cannot directly provide the continuous lifetime attrition ratio and remaining lifetime prediction results. During ResNet50 training, the same optimizer (AdamW), learning rate scheduling strategy (cosine annealing), and early stopping condition as in this embodiment were used. Figure 8 The confusion matrix of the ResNet50 contrastive model on the test set is shown. Figure 8 The ResNet50 model correctly identified 131 images out of 156 images in the test set, achieving an accuracy of 83.97% (approximately 84.0%). The 200h stage showed good results, with all 30 samples correctly classified. In the 1000h stage, 33 out of 36 samples were correctly identified, while 3 were misclassified as 800h. However, the ResNet50 model exhibited significant misclassification in the intermediate creep stages: in the 400h stage, 22 out of 30 samples were correctly identified, with 6 misclassified as 600h and 2 as 200h; in the 600h stage, only 20 out of 30 samples were correctly identified, with 10 misclassified as 200h; and in the 800h stage, 26 out of 30 samples were correctly identified, with 4 misclassified as 1000h. The above results indicate that single-classification networks still have certain limitations in distinguishing adjacent or intermediate creep damage stages, especially when the microscopic tissue morphology evolution is continuous and the stage boundaries are not obvious, making stage confusion easy. In contrast, such as Figure 7 The multi-task deep learning model described in this embodiment achieves an accuracy of 95.5% on the same test set, which is approximately 11.5 percentage points higher than the ResNet50 comparison model. Furthermore, in addition to outputting the creep stage category, this model can also simultaneously output the continuous lifetime attrition ratio through a regression branch and further calculate the remaining lifetime. Therefore, this model not only improves the accuracy of creep stage identification but also extends from discrete stage classification to quantitative assessment of continuous lifetime, making it more suitable for evaluating the creep state and predicting the remaining lifetime of martensitic heat-resistant steel.

Claims

1. A method for predicting the creep state and remaining life of martensitic heat-resistant steel based on multi-task deep learning, characterized in that, Includes the following steps: S1. Obtain metallographic images of the sample steel at multiple different creep stages, preprocess each metallographic image, and establish creep stage category labels and lifetime consumption ratio labels for it to construct a sample set. S2. Based on the sample source information corresponding to the images in the sample set, the sample set is divided into a training set, a validation set, and a test set. Among them, images from the same sample source only appear in one of the training set, validation set, and test set datasets. S3. Construct a multi-task deep learning model, which includes a shared feature extraction network and classification and regression branches connected to it. The shared feature extraction network is used to extract microstructure features related to creep damage. The classification branch is used to output the probability distribution of creep stages, and the regression branch is used to output the continuous lifetime loss ratio prediction. S4. Use the sample set to train, validate and test the multi-task deep learning model. During training, a joint loss function is used, which includes a classification loss term and a regression loss term. S5. Input the metallographic image of the steel to be predicted into the trained multi-task deep learning model to obtain the probability distribution of creep stage and the predicted value of life consumption ratio, and determine the creep stage and remaining life of the steel to be predicted.

2. The method according to claim 1, characterized in that: In step S1, the multiple different creep stages include at least five levels; the lifetime consumption ratio label is determined based on the ratio of the current creep time to the total creep fracture time.

3. The method according to claim 1, characterized in that: In step S2, the sample source information is the sample number, which is extracted from the image file name or image metadata; after the sample set is divided, the images in the training set are subjected to data augmentation processing.

4. The method according to claim 1, characterized in that: The shared feature extraction network is a convolutional neural network, a visual Transformer network, or a combination thereof.

5. The method according to claim 4, characterized in that: The shared feature extraction network uses the EfficientNetV2-S network and initializes its parameters using pre-trained weights.

6. The method according to claim 5, characterized in that: The EfficientNetV2-S network outputs a feature vector after global average pooling; the classification branch has the following structure: linear layer, batch normalization layer, ReLU activation layer, Dropout layer, and linear layer; the regression branch has the following structure: linear layer, batch normalization layer, ReLU activation layer, Dropout layer, linear layer, and Sigmoid activation layer.

7. The method according to claim 1, characterized in that: In step S4, the classification loss term is the labeled smooth cross-entropy loss, and the regression loss term is the mean squared error loss; during training, the AdamW optimizer, learning rate decay strategy, and early stopping mechanism are used.

8. A system for predicting the creep state and remaining life of martensitic heat-resistant steel based on multi-task deep learning, characterized in that, include: The image acquisition module is used to acquire images of the metallographic structure of the steel to be predicted. The image preprocessing module is used to perform size adjustment, pixel value normalization, and standardization on metallographic images; The model invocation module is used to invoke the multi-task deep learning model trained by the method according to any one of claims 1 to 7; The inference module is used to input the preprocessed metallographic images into the multi-task deep learning model and obtain the probability distribution of the creep stage and the predicted value of the lifetime consumption ratio. The post-processing module is used to determine the creep stage of the steel to be predicted based on the creep stage probability distribution, and to calculate the remaining life of the steel to be predicted based on the life consumption ratio prediction value. The display module is used to show the prediction results of the creep stage and the remaining lifetime.

9. The system according to claim 8, characterized in that, It also includes an interpretability analysis module, which uses the output score corresponding to the predicted category of the model's classification branch or the lifetime consumption ratio of the regression branch as the target output, calculates its gradient on the last convolutional feature map of the shared feature extraction network, and generates an attention heatmap through gradient global average pooling, feature map weighted summation, ReLU and normalization processing, and displays the attention heatmap superimposed on the original metallographic image.

10. A computer-readable storage medium for storing one or more programs, characterized in that: The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods according to claims 1 to 7.

Images