Classification method for microstructure degradation of P91 steel in long-term high-temperature service

Abstract

The invention discloses a P91 steel long-term high-temperature service microstructure degradation classification method, which comprises the following steps of 1, acquiring metallographic images of P91 steel high-pressure boiler tubes at different aging stages, and constructing a data set; 2, the Vision Transform model is subjected to fine adjustment, and an optimal model is obtained; and step 3, inputting the preprocessed test set image into the optimal model, extracting global and local features through a Transform encoder, outputting aging stage probabilities through a classification head, and selecting a stage with the maximum probability as a classification result. According to the method, the metallographic image does not need to be manually observed, the features can be automatically extracted and classified through a deep learning model, personal errors are avoided, the evaluation efficiency is greatly improved, and batch automatic detection can be realized; according to the method, the microstructure change of the metal material in the aging process can be accurately captured, the prediction accuracy can reach 91.24% through inspection, and the prediction precision is high.

Description

Technical Field

[0001] This invention pertains to steel classification methods, specifically a classification method for the microstructural degradation of P91 steel during long-term high-temperature service. Background Technology

[0002] P91 steel, a modified 9Cr-1Mo ferritic heat-resistant steel, is widely used in high-temperature and high-pressure key components such as boilers and steam turbines due to its excellent high-temperature strength, oxidation resistance, and cost advantages. It has been in service in my country's energy sector for over ten years, with some pipelines exceeding 100,000 hours of operation. However, under long-term high-temperature and high-pressure environments, the microstructure of P91 steel gradually degrades, manifesting as M... 23 C6 carbide coarsening, MX carbonitride dissolution and Z phase transformation, Laves phase precipitation, and the recovery and recrystallization of tempered martensitic lath structure lead to a decrease in high-temperature creep strength and a hardness often below 180 HBW, which seriously threatens the safety of components and easily causes unplanned downtime and economic losses.

[0003] Current methods for assessing the remaining service life of P91 steel high-temperature components employ a three-tiered approach: Tier 1 relies on design data and visual inspection, which is highly subjective and prone to errors; Tier 2 incorporates non-destructive testing and stress analysis, but still fails to accurately capture microstructural changes; Tier 3, through destructive sampling and laboratory testing, can obtain accurate microscopic data, but is costly and time-consuming, making it difficult to routinely apply to in-service components. Furthermore, traditional microstructural assessment relies on manual observation of metallographic images, which is inefficient and easily influenced by experience, failing to meet the demands of large-scale, automated testing. Therefore, a more efficient and accurate method for classifying the microstructural degradation of P91 steel is urgently needed. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of existing technologies such as insufficient quality of P91 steel microstructure degradation classification datasets, low accuracy of models in capturing micro-features, and disconnect between aging status and remaining life assessment, the purpose of this invention is to provide a classification method for the microstructure degradation of P91 steel under long-term high-temperature service that accurately captures subtle changes in microstructure and integrates aging classification and life assessment.

[0005] Technical solution: The present invention provides a classification method for the microstructural degradation of P91 steel during long-term high-temperature service, comprising the following steps:

[0006] Step 1: Collect metallographic images of P91 steel high-pressure boiler tubes at different aging stages to construct a high-quality dataset;

[0007] Step 2: Based on a high-quality dataset, conduct systematic fine-tuning training on the Vision Transformer model. Through multiple rounds of parameter iteration and performance verification, select the optimal model that meets the preset indicators.

[0008] Step 3: Preprocess the actual obtained metallographic images of P91 steel, input the preprocessed images into the optimal model for prediction and discrimination, and the model outputs the corresponding aging state assessment conclusion.

[0009] Step 4: Based on the aging status assessment conclusions, the remaining life of P91 steel high-pressure boiler tubes is assessed using a pre-set calculation model.

[0010] Furthermore, in step one, the P91 steel high-pressure boiler tube is a P91 steel high-pressure boiler tube conforming to the ASME SA-335M-2023 standard. Its original state is normalized at 1040℃~1080℃ and tempered at 730℃~780℃, and it has a tempered martensitic microstructure.

[0011] Furthermore, in step one, P91 steel high-pressure boiler tubes are treated through accelerated thermal aging experiments in the laboratory to obtain samples with different aging times covering the entire life cycle, thus obtaining samples at different aging stages. The conditions for the accelerated thermal aging experiment are: aging temperature 790℃, and aging time gradients set to 0h, 1000h, 3000h, 5000h, 8000h, 10000h, 15000h, 20000h, 25000h, and 30000h, corresponding to 10 aging stages.

[0012] Further, in step one, the specific process of constructing a high-quality dataset is as follows: P91 steel high-pressure boiler tubes at different aging stages are sequentially cut, ground, polished, and etched; metallographic images are acquired using an optical microscope at 100-1000x magnification, with multiple images acquired for each aging stage, preferably 300 images; the acquired images are labeled according to aging time, and the labeled images are proportionally divided into training, validation, and test sets. The etching process uses a 10% hydrochloric acid-ferric chloride solution for 5-8 seconds, followed by rinsing with distilled water and drying.

[0013] Furthermore, the ratio of training set, validation set, and test set is 7:1.5:1.5; during the construction of the dataset, the collected metallographic images are quality-screened to remove images with scratches, stains, or blurred focus, ensuring that the clarity of a single image is ≥200dpi.

[0014] Furthermore, step two specifically includes the following steps:

[0015] Step 2.1: Load the pre-trained visual Transformer base model and reuse its general visual feature extraction capabilities. Replace the original model's 1000 general classification heads with 10 classification heads to match the classification requirements of the 10 aging stages during the operation of P91 steel. The pre-trained visual Transformer base model is the ViT-Base model.

[0016] Step 2.2, preprocess the dataset images: scale the images to 224×224 size, then divide them into 16×16 non-overlapping patches, flatten the patches, and generate patch embedding vectors with a dimension of 768 through linear projection;

[0017] Step 2.3: Concatenate the learnable "category embedding vector" with the patch embedding vector, and then add learnable positional encoding to preserve the spatial and semantic association information of the image to obtain the fused embedding sequence; the positional encoding is sine and cosine positional encoding, with the same dimension as the patch embedding vector;

[0018] Step 2.4 involves inputting the fused embedded sequence into the Transformer encoder for feature interaction and extraction. Through multiple rounds of parameter iteration training and combined with validation set performance evaluation, the optimal model that meets the preset indicators is selected. The preset indicators are validation set accuracy ≥ 95% and F1 score ≥ 0.94.

[0019] Furthermore, the specific settings for fine-tuning the training are as follows: the AdamW optimizer is used, and the initial learning rate is set to 10. -4 A cosine annealing learning rate scheduling strategy is used to achieve precise parameter updates; the number of training rounds is set to 50, and data augmentation operations are applied to the training set, including random rotation of -15° to 15°, horizontal flip with a probability of 0.5, and color jitter of 0.8 to 1.2 times to enhance the model's generalization ability; during the fine-tuning process, the validation set accuracy is used as the core optimization indicator, the model performance is monitored in real time, and finally the model weights corresponding to the best performance are saved.

[0020] Preferably, in step three, the Transformer encoder contains a 12-layer encoding structure, with 12 attention heads set in each layer, and each attention head has a feature dimension of 64; the hidden layer dimension of the feedforward neural network of the Transformer encoder is 3072, and the activation function is the GELU function.

[0021] Furthermore, in step three, the model calculates the probability of each aging stage using a preset formula, and outputs an aging status assessment conclusion based on the probability results. The preset formula is: in, For the input image The probability of belonging to the i-th aging stage, z i Let z be the i-th output value of the Transformer encoder output feature after linear transformation by the classification head. jThis is the j-th output value of the Transformer encoder after linear transformation of the features by the classification head, where j is the aging stage number, 1≤j≤10. When the probability value of a certain aging stage is ≥0.8, the stage is directly output as the evaluation conclusion. When the probability values ​​of all stages are <0.8, the two stages with the highest probabilities and their corresponding probability values ​​are output. Preprocessing includes image grayscale conversion, noise removal, contrast enhancement, and size normalization. Noise removal uses a median filtering algorithm with a filter window size of 3×3. Contrast enhancement uses adaptive histogram equalization with a contrast threshold of 2.0.

[0022] Furthermore, the classification method also includes a model performance verification step: the optimal model is verified using a test set, the precision, recall and overall accuracy of the model at each aging stage are statistically analyzed, and multiple metallographic images of P91 steel in field service that were not used in training are selected for blind testing to verify the model's suitability for practical applications.

[0023] Furthermore, the key computational models in step four include a microstructure feature quantification model, an aging stage-service time mapping model, and a remaining lifetime assessment model. The microstructure feature quantification model, based on deep features extracted from the optimal model, calculates the tempered martensite lath width W, carbide particle density ρ, and grain boundary clarity C, using the following formulas: Where k1 is the width calibration coefficient, k2 is the density calibration coefficient, k3 is the sharpness calibration coefficient, and F w For the model output eigenvector components related to the calculation of tempered martensite lath width, F ρ For the model output eigenvector components related to carbide particle density calculation, F c The model outputs eigenvector components related to grain boundary clarity calculation, where t is the accelerated aging time corresponding to the m-th aging stage. p This refers to the time the equipment has been in service on-site.

[0024] The aging stage-service time mapping model uses the Arrhenius formula to establish the conversion relationship between accelerated aging time and actual service time: Among them, t service For actual service time, t accel To accelerate the aging time, Q is the activation energy, taken as 280 kJ / mol, R is the gas constant 8.314 J / (mol·K), and T is... service T represents the actual service temperature. acce To accelerate the aging temperature;

[0025] The remaining useful life assessment model calculates the remaining useful life L based on a comparison of quantitative characteristics and a full life-cycle threshold. rem : in, The design lifecycle of P91 steel is defined, where W0 is the original strip width, ρ0 is the original strip width, and C0 is the carbide density.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following significant features:

[0027] 1. No manual observation of metallographic images is required. The model automatically extracts features and classifies them, avoiding human experience errors and greatly improving evaluation efficiency. It can realize large-scale automated batch inspection of microstructure of P91 steel components.

[0028] 2. Relying on the multi-head self-attention mechanism of VisionTransformer, it can accurately capture subtle microscopic changes such as coarsening of M23C6 carbides and MX phase transformation. The overall accuracy of the test set reaches 91.24%, and the classification accuracy of early and late aging stages (such as 10h and 600h) exceeds 95%, providing accurate data support for material performance evaluation with high classification accuracy.

[0029] 3. Low cost and wide applicability: No complex special testing equipment is required. It can be achieved based on existing optical microscopes and computers. It is suitable for routine service condition monitoring of P91 steel high-temperature components in the power and nuclear power fields. It can be extended to the microstructure degradation classification of other heat-resistant steels.

[0030] 4. Solid dataset support and outstanding model generalization ability: A high-quality dataset is built through accelerated aging experiments throughout the entire life cycle. After three levels of quality screening and scientific division, combined with data augmentation strategies, the model still achieves a blind test accuracy of over 92% under different microscopes and lighting conditions, making it suitable for complex on-site testing environments.

[0031] 5. Achieve integrated aging classification and life assessment: Innovatively construct a triple model of micro-feature quantification, time mapping and remaining life assessment, with an assessment error of ≤8%, solving the pain point of traditional technology "only classifying without life prediction", and providing full-chain data support for equipment maintenance planning and safe service early warning;

[0032] 6. Comprehensive feature capture dimensions: The dual mechanism of "category embedding + sine and cosine position encoding" preserves the spatial and semantic associations of the image. The 12-layer encoder and multi-head attention head work together to extract global and local micro features. Compared with traditional CNN models, the sensitivity of recognition of degenerate features such as slat widening and grain boundary blurring is improved by more than 20%. Attached Figure Description

[0033] Figure 1 This is a flowchart of the present invention;

[0034] Figure 2 These are metallographic images of the etched parts of this invention;

[0035] Figure 3 This is a diagram illustrating the deep learning model training process of this invention;

[0036] Figure 4 This is the confusion matrix diagram of the model of this invention;

[0037] Figure 5 This is the confusion matrix diagram of the traditional CNN model. Detailed Implementation

[0038] like Figure 1 A classification method for microstructural degradation of P91 steel during long-term high-temperature service based on the Vision-Transformer model includes the following steps:

[0039] Step 1: Collect metallographic images of P91 steel high-pressure boiler tubes at different aging stages to construct a high-quality dataset.

[0040] 1.1 Samples and Aging Treatment: P91 steel high-pressure boiler tubes (outer diameter 133.00 mm, wall thickness 25.00 mm) produced by Baosteel were selected, conforming to ASME SA-335M-2023 standard. In their original state, they were normalized at 1050℃ and tempered at 750℃, exhibiting a tempered martensitic microstructure. The chemical composition (mass percentage) was determined by SPECTROMAXx spectrometer as follows: C: 0.09%, Si: 0.28%, Mn: 0.39%, P: 0.01%, S: 0.002%, Cr: 8.35%, Mo: 0.88%, V: 0.19%, Nb: 0.073%, with the remainder being Fe and unavoidable impurities. The samples were cut into 10mm×10mm×5mm pieces using a DK7732 wire EDM machine at a cutting speed of 80mm / min. After cutting, burrs were removed with 240# sandpaper, and the samples were ultrasonically cleaned with anhydrous ethanol for 10 minutes. The samples were then aged in a KSL-1200X-J high-temperature furnace at 790℃ (heating rate 5℃ / min, furnace cooling). Five parallel samples (strip width deviation ≤5%) were taken at each of the 10 corresponding aging times, covering the early, middle, and late aging stages of actual boiler operation. The accelerated thermal aging test conditions were: aging temperature 790℃, and aging time gradients of 0h, 1000h, 3000h, 5000h, 8000h, 10000h, 15000h, 20000h, 25000h, and 30000h, corresponding to 10 aging stages.

[0041] 1.2 Image Acquisition and Dataset: Samples were processed using a Struers Tegramin-20 semi-automatic polishing machine (wet grinding with 240#→1200# sandpaper, 3 min per stage, 15 N pressure; polishing with 3μm→1μm diamond polishing compound, 10 N pressure, 300 r / min speed). The samples were then etched with a 10% ferric chloride hydrochloride solution (10g ferric chloride + 10mL hydrochloric acid + 80mL distilled water) at 25℃ for 6 s, rinsed with distilled water for 10 s, and then dried with cold air. Images were acquired using a Nikon Model-Eclipse Ci-S microscope with a DS-Ri2 camera at magnifications of 100×, 200×, 500×, and 1000×, with 5 fields of view per sample. The image resolution was 2048×1536 pixels, with a sharpness ≥300 dpi. 300 images were acquired for each aging stage (60 images for each of the 5 parallel samples), for a total of 3000 images across 10 aging stages. Images were filtered using Matlab R2023a (images with scratches > 50 μm and stains > 0.1 mm² were removed, and images with a Tenengrad gradient value ≥ 200 were retained). The images were then labeled according to aging time (corresponding to 10 aging stages, labeled as Stage 1 to Stage 10). The labels were independently labeled by 3 material testing engineers, with a labeling consistency of ≥ 98%. The training set (2100 images), validation set (450 images), and test set (450 images) were then divided into a 7:1.5:1.5 ratio. The KS test was used to verify the consistency of the data distribution. The test statistic D = 0.07 < 0.15, indicating that the characteristics of each set are uniformly distributed. Figure 2 Metallographic images obtained under laboratory conditions.

[0042] Step 2: Based on the high-quality dataset, systematically fine-tune the Vision Transformer model. Through multiple rounds of parameter iteration and performance verification, select the optimal model that meets the preset indicators.

[0043] 2.1 Model fine-tuning parameters: Based on the Ubuntu 20.04LTS system, it was run on an NVIDIA RTX 3090 graphics card (24GB VRAM), Intel Core i9-12900K CPU, and 64GB RAM device; using the PyTorch 1.13.1 framework, the ViT-Base model pre-trained on the ImageNet-21k dataset was loaded, and the original 1000-class classification head was replaced with a 10-class classification head (corresponding to 10 aging stages).

[0044] 2.2 Input Image Preprocessing: The 2048×1536 pixel image was scaled to 224×224 pixels by bilinear interpolation (black background fill at the edges ≤5%), segmented into 16×16 non-overlapping patches (196 in total), and flattened to generate a 768-dimensional embedding vector through a 256×768 linear projection layer;

[0045] 2.3 Add sine and cosine position encoding: Among them, PE pos,2i Encoding for even-numbered dimension positions, where i is the dimension index, PE pos,2i+1 Encoding for odd-numbered dimension positions.

[0046] By concatenating the 768-dimensional category embedding vectors, a 197×768 fusion sequence is obtained.

[0047] 2.4 Training parameters: Batch size 32 (16 is too slow, 64 is insufficient in GPU memory), initial learning rate 10. -4 (10) -3 Fluctuations, 10 -5 (Slow convergence), AdamW optimizer (weight decay 0.01, β1=0.9, β2=0.999, ε=10) -8 Cosine annealing learning rate scheduling (50 rounds per cycle, minimum 10) -6 The training set was subjected to a -15° to 15° rotation, a horizontal flip with a probability of 0.5, and a 0.8 to 1.2x color jitter enhancement (reducing generalization error by 4.2%). Cross-entropy loss was used. Where Loss is the loss function, N is the number of samples, and y n,i Let x be the true label value of the nth sample belonging to the i-th class. n Let be the feature vector of the nth sample.

[0048] The training and monitoring metrics reached their optimal state in the 38th round (validation set accuracy 91.24%, F1=0.94), and a 340MB weight file was saved.

[0049] Step 3: Preprocess the actual obtained metallographic images of P91 steel, input the preprocessed images into the optimal model for prediction and discrimination, and the model outputs the corresponding aging state assessment conclusion.

[0050] Classification Results: Microstructure Degradation Classification: Test set images were preprocessed (weighted average grayscale: Gray=0.299×R+0.587×G+0.114×B; 3×3 median filtering for noise reduction; adaptive histogram equalization, threshold 2.0; normalized to 224×224) and then input into the optimal model. Features were extracted through a 12-layer Transformer encoder (12 attention heads per layer, feature dimension 64; 3072-dimensional hidden layers in a feedforward neural network, GELU activation), and the classification head output the probability: in, For the input image The probability of belonging to the i-th aging stage, z i Let z be the i-th output value of the Transformer encoder output feature after linear transformation by the classification head. j This is the j-th output value of the Transformer encoder after the feature is linearly transformed by the classification head, where j is the aging stage number, 1≤j≤10; when the probability value of a certain aging stage is ≥0.8, the stage is directly output as the evaluation conclusion; when the probability values ​​of all stages are <0.8, the two stages with the highest probabilities and their corresponding probability values ​​are output.

[0051] The criteria were "outputting a single stage if the probability is ≥0.8, otherwise outputting the first two stages," and 92% of the samples met the single-stage condition. The overall accuracy of the test set was 91.24%. The accuracy for the early aging stage (e.g., stage 1) was 96.7%, precision 97.4%, and F1 = 0.970; the accuracy for the late aging stage (e.g., stage 10) was 95.3% and F1 = 0.950; and the accuracy for some intermediate stages (e.g., stages 4 and 5) ranged from 86.5% to 88.2%. The confusion matrix showed that there were 3 cases of cross-classification between adjacent intermediate stages (e.g., stages 4 and 5) (slat width difference <0.2μm), while the classification of other stages was good.

[0052] Step 4: Based on the aging status assessment conclusion, the remaining life of P91 steel high-pressure boiler tubes is assessed using a preset calculation model.

[0053] Remaining life assessment: Taking a sample at a certain mid-term aging stage (corresponding to 300 hours of accelerated aging) as an example, it is known that: =0.85, Fw=120, Fρ=85, Fc=90, t m =300h, t p =15000h, calibration coefficient k1=0.82,

[0054] k2=1.15, k3=0.97. Quantification of microscopic features:

[0055] W=(0.82×120) / (0.85×3000.3+0.15×5.8)≈1.59μm,

[0056] ρ=1.15×85×exp(-0.0002×15000)≈4.89 pieces / μm 2 ,

[0057] C=0.97×90 / 1+0.001×15000=21.83.

[0058] Service time mapping (Arrhenius formula):

[0059] t service=300×exp(280000 / 8.314×(1 / 813-1 / 1063))≈45200h

[0060] (T) service =540℃=813K,T accel =790℃=1063K).

[0061] Remaining life calculation:

[0062] L rem =100000×(1-(1.59 / 0.8+4.89 / 8.2+(1-21.83 / 30)) / 3)≈42800h, with an error of 7.2% compared to the actual field measurement.

[0063] The test set performance of the method in this embodiment is as follows: overall accuracy 91.24%; accuracy 96.7%, precision 97.4%, F1=0.970 in the early aging stage (stage 1: 0h); accuracy 95.3%, F1=0.950 in the late aging stage (stage 10: 30000h); accuracy 88.2%, F1=0.887 in the mid-stage (stage 4: 5000h); and accuracy 86.5%, F1=0.878 in stage 5: 8000h. The confusion matrix shows that only 3 cases were cross-classified between stages 4 and 5 (slat width difference <0.2μm), and the classification accuracy for the remaining stages was ≥90%. Figure 4 .

[0064] Comparative Example

[0065] Using the exact same dataset as this invention—that is, with identical data collection sources, annotation rules, and partitioning ratios—a comparative experiment was conducted using the ResNet-50 model. Specifically, a ResNet-50 pre-trained model with general image feature extraction capabilities was used. The dataset images underwent the same preprocessing operations as this invention, including resizing to 224×224 and normalization. The last fully connected layer of the model was replaced with a fully connected layer adapted to the 10 aging stage classifications in this experiment. The optimizer was set to SGD, with an initial learning rate of 0.01 and a decay strategy of multiplying the learning rate by 0.1 every 30 epochs. The model was trained for 100 epochs using the cross-entropy loss function and parameters were updated via backpropagation. Performance was evaluated using a test set after training.

[0066] The results showed that the overall accuracy was 70.8%, which was 20.44% lower than that of the present invention. The accuracy of early stage 1 was 82.1% (14.6% lower), the accuracy of late stage 10 was 79.3% (16.0% lower), the accuracy of mid-stage 4 was 62.5% (25.7% lower), and the accuracy of stage 5 was 59.8% (26.7% lower). The confusion matrix showed that there were 28 cases of misclassification in mid-stage 4-6, and there were more than 10 cases of misclassification in stages 2 and 3, and stages 7 and 8. The ability to capture subtle microscopic changes was significantly insufficient, as shown in Figure 5.

Claims

1. A classification method for the microstructural degradation of P91 steel during long-term high-temperature service, characterized in that, Includes the following steps: Step 1: Collect metallographic images of P91 steel high-pressure boiler tubes at different aging stages to construct a high-quality dataset; Step 2: Based on the high-quality dataset, systematically fine-tune the Vision Transformer model. Through multiple rounds of parameter iteration and performance verification, select the optimal model that meets the preset indicators. Step 3: Preprocess the actual obtained metallographic images of P91 steel, input the preprocessed images into the optimal model for prediction and discrimination, and the model outputs the corresponding aging state assessment conclusion. Step 4: Based on the aging status assessment conclusion, the remaining life of P91 steel high-pressure boiler tubes is assessed using a preset calculation model.

2. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 1, characterized in that: In step one, the P91 steel high-pressure boiler tube is a P91 steel high-pressure boiler tube conforming to the ASME SA-335M-2023 standard. Its original state is normalized at 1040℃~1080℃ and tempered at 730℃~780℃, and it has a tempered martensitic microstructure.

3. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 1, characterized in that: In step one, P91 steel high-pressure boiler tubes are treated by laboratory accelerated thermal aging experiments to obtain samples of different aging times covering the entire life cycle, so as to obtain samples of different aging stages.

4. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 1, characterized in that: In step one, the specific process of constructing a high-quality dataset is as follows: P91 steel high-pressure boiler tubes at different aging stages are sequentially cut, ground, polished, and corroded. Metallographic images were acquired using an optical microscope at magnifications of 100 to 1000x, with multiple images acquired at each aging stage. The acquired images were labeled according to the aging time, and the labeled images were divided into training, validation, and test sets according to the proportions.

5. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 4, characterized in that: The ratio of the training set, validation set, and test set is 7:1.5:1.

5. During the construction of the dataset, the collected metallographic images are quality-screened to remove images with scratches, stains, or blurred focus, ensuring that the clarity of each image is ≥200dpi.

6. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 1, characterized in that: Step two includes loading and modifying the pre-trained visual Transformer base model, preprocessing the dataset images, generating fused embedding sequences, and inputting them into the Transformer encoder for training and selection of the optimal model.

7. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 6, characterized in that: The fine-tuning training uses the AdamW optimizer, sets the initial learning rate, learning rate scheduling strategy, and training epochs, applies data augmentation operations, and monitors and saves the optimal model weights with validation set accuracy as the core indicator.

8. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 1, characterized in that: In step three, the model calculates the probability of each aging stage using a preset formula, and outputs an aging status assessment conclusion based on the probability results. The preset formula is: in, For the input image The probability of belonging to the i-th aging stage, z i Let z be the i-th output value of the Transformer encoder output feature after linear transformation by the classification head. j This is the j-th output value of the Transformer encoder output feature after linear transformation by the classification head, where j is the aging stage number, 1≤j≤10.

9. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 1, characterized in that: It also includes a model performance verification step: the optimal model is verified using a test set, the precision, recall and overall accuracy of the model at each aging stage are statistically analyzed, and multiple metallographic images of P91 steel in field service that were not used in training are selected for blind testing to verify the model's suitability for practical applications.

10. The classification method for microstructural degradation of P91 steel during long-term high-temperature service according to claim 1, characterized in that: The key computational models in step four include a microstructure feature quantification model, an aging stage-service time mapping model, and a remaining service life assessment model. The microstructure feature quantification model, based on deep features extracted from the optimal model, calculates the tempered martensite lath width W, carbide particle density ρ, and grain boundary clarity C, using the following formulas: Where k1 is the width calibration coefficient, k2 is the density calibration coefficient, k3 is the sharpness calibration coefficient, and F w For the model output eigenvector components related to the calculation of tempered martensite lath width, F ρ For the model output eigenvector components related to carbide particle density calculation, F c The model outputs eigenvector components related to grain boundary clarity calculation, where t is the accelerated aging time corresponding to the m-th aging stage. p This refers to the time the equipment has been in service on-site. The aging stage-service time mapping model uses the Arrhenius formula to establish the conversion relationship between accelerated aging time and actual service time: Among them, t service For actual service time, t accel To accelerate the aging time, Q is the activation energy, taken as 280 kJ / mol, R is the gas constant 8.314 J / (mol·K), and T is... service T represents the actual service temperature. accel To accelerate the aging temperature; The remaining life assessment model calculates the remaining life L based on a comparison of quantitative characteristics and a full life-cycle threshold. rem : in, The design lifecycle of P91 steel is defined, where W0 is the original strip width, ρ0 is the original strip width, and C0 is the carbide density.

Images