Method for identifying aging of martensitic structure of high-chromium heat-resistant steel

Abstract

A high-chromium heat-resistant steel martensitic structure aging identification method comprises the following steps that a martensitic structure image is obtained and preprocessed; a trained VGGNet-MetStr model is utilized to identify the preprocessed martensite structure image to obtain an identification result, a network structure of the VGGNet-MetStr model comprises two parts, one part comprises an input module and a plurality of feature extraction modules, each feature extraction module is composed of two convolutional layers and a pooling layer, and the other part comprises a plurality of feature extraction modules; the other part is composed of two full-connection layers; the identification method comprises the following steps: inputting a preprocessed martensitic structure image into feature extraction modules through an input module, and processing the martensitic structure image through a plurality of feature extraction modules to obtain a feature image; according to the martensite structure aging identification method, the accuracy of martensite structure aging is improved, a better convergence effect is achieved, and the calculated amount is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of martensite structure aging identification, and specifically relates to a method for identifying martensite structure aging of high-chromium heat-resistant steel. Background Art

[0002] At present, high-chromium martensitic heat-resistant steels (such as P91 and P92, etc.) are widely used in key pressure-bearing components such as main steam pipelines and reheater hot-section pipelines of ultra (super) critical units due to their excellent creep resistance and low cost. When the pressure-bearing components are in service under high-temperature and high-pressure conditions, with the increase of operation time, the internal structure of the material will creep, resulting in a decline in performance or a shortening of service life. Therefore, in order to ensure the safe and stable operation of the equipment, it is necessary to regularly conduct metallographic structure inspections on the in-use high-chromium martensitic heat-resistant steel. Among them, the evolution of the martensite structure has a major impact on the material performance. In the classification and evaluation of metallographic structure aging and damage, how to evaluate the martensite morphology is crucial.

[0003] In the prior art, the aging of the martensite structure is mainly analyzed by on-site film covering or pipe cutting and sampling for laboratory metallographic observation and expert analysis and determination. This method has strong subjectivity, high dependence on the professional technical level and practical experience of researchers. At the same time, the manual observation and analysis means have low reusability, relatively large experimental result errors, and low efficiency. Therefore, using the methods of deep learning and neural networks to identify the aging of the martensite structure has become the focus. For example, using the deep learning method, the problem of defect detection of materials through pictures is transformed into the semantic segmentation problem of pictures, and the semantic segmentation algorithm is used to detect defects in metal materials. However, the existing deep learning and neural network methods have low recognition efficiency and low accuracy for the aging of the martensite structure. Summary of the Invention

[0004] In order to solve the problems of low accuracy and low efficiency in identifying the aging of the martensite structure in the prior art, the present invention provides a method for identifying the aging of the martensite structure of high-chromium heat-resistant steel, which improves the accuracy of the martensite structure aging, has a better convergence effect, and reduces the calculation amount.

[0005] To achieve the above object, the specific solution adopted by the present invention is: A method for identifying the aging of the martensite structure of high-chromium heat-resistant steel, comprising the following steps:

[0006] Obtain a martensite structure image and perform preprocessing;

[0007] The preprocessed martensite tissue image is recognized using the trained VGGNet-MetStr model to obtain the recognition result. The network structure of the VGGNet-MetStr model consists of two parts. One part includes an input module and multiple feature extraction modules, and each feature extraction module is composed of two convolutional layers and one pooling layer. The other part consists of two fully connected layers. The recognition method includes:

[0008] The preprocessed martensite tissue image is input into the feature extraction module through the input module, and a feature image is obtained after being processed by multiple feature extraction modules;

[0009] Two fully connected layers are used to recognize the feature image and output the recognition result.

[0010] As an optimization scheme of the above-mentioned method for identifying the aging of martensite tissue in high-chromium heat-resistant steel: A metallurgical microscope is used to obtain the martensite tissue image.

[0011] As another optimization scheme of the above-mentioned method for identifying the aging of martensite tissue in high-chromium heat-resistant steel: The preprocessing of the martensite tissue image includes image size normalization, grayscale value normalization, and Gaussian smoothing.

[0012] As another optimization scheme of the above-mentioned method for identifying the aging of martensite tissue in high-chromium heat-resistant steel: The activation function of the VGGNet-MetStr model is the ELU function.

[0013] As another optimization scheme of the above-mentioned method for identifying the aging of martensite tissue in high-chromium heat-resistant steel: The ELU function is:

[0014]

[0015] where α is an adjustable parameter.

[0016] As another optimization scheme of the above-mentioned method for identifying the aging of martensite tissue in high-chromium heat-resistant steel: The calculation formula of the convolutional layer is:

[0017]

[0018] where w ij is the weight value, x m+i,n+j is the input value, b is the bias, and y mn is the feature value.

[0019] As another optimization scheme of the above-mentioned method for identifying the aging of martensite tissue in high-chromium heat-resistant steel: The sampling method of the pooling layer is uniform sampling, and the calculation formula is:

[0020]

[0021] where ymn is the output value of the S1×S2 region, is any input value within the S1×S2 region.

[0022] As another optimization scheme of the above method for identifying the aging of martensite structure in high-chromium heat-resistant steel: The fully connected layer uses a softmax classifier:

[0023]

[0024]

[0025] where k represents a certain classification, represents the value of this classification, p(y (i) =k|x (i) ,θ) represents the probability of the k-th class, and θ is the weight.

[0026] As another optimization scheme of the above method for identifying the aging of martensite structure in high-chromium heat-resistant steel: The loss function of the fully connected layer is the cross-entropy function:

[0027]

[0028] where 1{y (i) =j} means that if the true classification of sample i is j, the value is 1.

[0029] As another optimization scheme of the above method for identifying the aging of martensite structure in high-chromium heat-resistant steel: The training process of the VGGNet-MetStr model includes the following steps:

[0030] Collect martensite structure images and perform preprocessing to obtain a preprocessed data set;

[0031] Divide the preprocessed data set into a training set and a test set according to 20:3;

[0032] Input the training set into the VGGNet-MetStr model for training, and use the test set to verify the trained VGGNet-MetStr model to obtain a trained VGGNet-MetStr model.

[0033] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for identifying the aging of martensite structure in high-chromium heat-resistant steel. The VGGNet-MetStr model is used to identify the aging of martensite structure. The VGG-MetStr network model uses deeper layers and smaller convolution kernels to calculate feature images, can extract fine features in the images, has higher accuracy, and better convergence effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1It is a schematic flowchart of the identification method of the present invention;

[0035] Figure 2 It is a characteristic image of the martensite aging morphology extracted by the VGGNet-MetStr model;

[0036] Figure 3 It is a characteristic image of the normal martensite morphology extracted by the VGGNet-MetStr model;

[0037] Figure 4 It is a comparison of the original martensite tissue image, grayscale value normalization processing, and Gaussian smoothing processing;

[0038] Figure 5 It is the change process of the accuracy obtained during the training process of the VGGNet-MetStr model. Specific embodiments

[0039] The technical solution of the present invention will be further elaborated in detail below in conjunction with specific embodiments. For parts not detailedly recorded and disclosed in the following embodiments of the present invention, they should all be understood as the prior art known or should be known to those skilled in the art.

[0040] Embodiment 1

[0041] A method for identifying the aging of martensite structure in high-chromium heat-resistant steel, comprising the following steps:

[0042] Obtain a martensite tissue image and perform preprocessing. Specifically:

[0043] Sampling: Sampling from the pipeline to be detected, and the material specimen is a square specimen with a side length of 5-15 mm or a circular specimen of Ф(5~15 mm)×15 mm. In this embodiment, the specimen is a square specimen.

[0044] Mounting: In this embodiment, the hot pressing mounting method is used to mount the specimen. The specimen and phenolic resin particles are placed in a hot mounting machine and heated to 140°C, and the hot pressing is completed in the mounting machine through low temperature and pressure. During the operation, ensure that one plane of the water sample is located at the bottom of the mounting machine for later use.

[0045] Grinding: The purpose of grinding the specimen is to obtain a flat grinding surface for polishing preparation. The specimen is manually ground on sandpapers of different mesh sizes, starting from coarse to fine on metallographic sandpapers (the metallographic sandpaper numbers are 120 mesh, 240 mesh, 500 mesh, 800 mesh, 1200 mesh, and 1800 mesh in sequence). During the grinding process of the specimen, every time the sandpaper is changed, the specimen must be rotated 90 degrees perpendicular to the previous grinding direction and ground in one direction until the old grinding marks completely disappear and the new grinding marks are uniform. Before replacing the new sandpaper after each grinding, the specimen and the sandpaper must be rinsed with clean water to prevent the coarse sand grains from the previous process from being carried onto the fine sandpaper and forming deep scratches. Note that when grinding, do not apply excessive force and continuously cool with water to avoid excessive heating of the specimen surface and causing internal tissue changes.

[0046] Polishing: After the ground specimen is rinsed with water, it is polished. The purpose of polishing is to remove the uniform and fine scratches generated by fine grinding on the grinding surface of the specimen and make the inspection surface present a bright mirror surface.

[0047] Etching: Picric acid hydrochloric acid alcohol solution is selected as the etchant. After the specimen is etched, it is rinsed with clean water, wiped with an alcohol cotton ball, and dried. Then, a metallographic microscope is used to obtain the martensite tissue image.

[0048] The obtained martensite tissue image is subjected to image size normalization processing, gray value normalization processing, and Gaussian smoothing processing to obtain the preprocessed martensite tissue image.

[0049] The trained VGGNet-MetStr model is used to identify the preprocessed martensite tissue image to obtain the recognition result. The network structure of the VGGNet-MetStr model includes two parts. One part includes an input module and multiple feature extraction modules. Each feature extraction module consists of two convolutional layers and one pooling layer. The other part consists of two fully connected layers. In this embodiment, the number of feature extraction modules is 5. Each convolutional layer uses a 3×3 dpi convolutional kernel with a stride of 1 for convolution, deepening the depth of the network structure, having good generalization ability and good transfer learning ability. Each pooling layer uses a 2×2 dpi pooling kernel with a stride of 1 for pooling calculation. The network structure of the VGGNet-MetStr model is shown in Table 1, which can extract fine features in the martensite tissue image, improve the accuracy and efficiency of martensite tissue aging recognition, and obtain a feature image with a deeper layer and a smaller convolutional kernel, having a better convergence effect.

[0050] Table 1 Network structure of the VGGNet-MetStr model

[0051] The recognition method includes:

[0052] The preprocessed martensite tissue image is input into the feature extraction module through the input module, and the feature image is obtained after being processed by multiple feature extraction modules. In this embodiment, the preprocessed martensite tissue image is processed by 5 feature extraction modules. Among them, the convolutional layer performs convolutional calculation, and the specific calculation formula is as follows:

[0053]

[0054] Among them, w ij is the weight value, x m+i,n+j is the input value, b is the bias, and y mn is the feature value.

[0055] The pooling layer is used to reduce the scale of the image after convolution by the convolutional layer. The sampling methods of the pooling layer include uniform sampling and max pooling. In this embodiment, the pooling layer adopts uniform sampling, and the average value of the input value x within S1×S2 is taken as the output value y of this area. The specific calculation formula is:

[0056]

[0057] Among them, y mn is the output value of the S1×S2 area, is any input value within the S1×S2 area.

[0058] Two fully connected layers are used to identify the feature image and output the recognition result, Figure 2 is the feature image of the martensite aging morphology extracted by the VGGNet-MetStr model; Figure 3 is the feature image of the normal martensite morphology extracted by the VGGNet-MetStr model. Specifically, the fully connected layer is a shallow perceptron, which plays a classification role. The softmax classifier is used to judge the sample type by calculating the probability, and can complete the classification of multiple categories. The specific calculation formula is as follows:

[0059]

[0060]

[0061] Among them, k represents a certain classification, represents the value of this classification, p(y (i) =k|x (i) ,θ) represents the probability of the kth class, and θ is the weight.

[0062] In this embodiment, the activation function of the VGGNet-MetStr model is the ELU function, which solves the problem of gradient disappearance. The specific calculation formula is as follows:

[0063]

[0064] Among them, α is an adjustable parameter. The ELU function combines the Sigmoid function and the Relu function, effectively solving the situation where the Relu function is not activated and the gradient disappears when it is negative, making it more robust and having a faster convergence speed.

[0065] The loss function of the fully connected layer is the cross-entropy function:

[0066]

[0067] Among them, 1{y (i) =j} means that if the true classification of sample i is j, the value is 1.

[0068] The training process of the VGGNet-MetStr model includes the following steps:

[0069] Collect martensite tissue images and perform preprocessing to obtain a preprocessed data set; specifically:

[0070] Sampling: Use high-chromium heat-resistant steels T / P91 and T / P92 to make specimens. The material specimens are square specimens with a side length of 5 - 15 mm or circular specimens with Ф(5 - 15 mm)×15 mm. In this embodiment, the specimens are square specimens.

[0071] After the square specimens are all inlaid, ground, polished, etched, rinsed with clean water, wiped with alcohol cotton balls, and dried, use a metallurgical microscope to obtain a martensite tissue image data set, and perform image size normalization, gray value normalization, and Gaussian smoothing processing on the obtained martensite tissue image data set to obtain a preprocessed data set. Figure 4 It is a comparison of the gray value normalization and Gaussian smoothing processing of the martensite tissue image.

[0072] Divide the preprocessed data set into a training set and a test set according to 20:3.

[0073] Input the training set into the VGGNet-MetStr model for training. The learning feature of the training set samples is that the softmax classifier learns the optimal parameters to minimize the loss value of the loss function. The loss function uses the cross-entropy function, and the specific calculation formula is as follows:

[0074]

[0075] Among them, 1{y (i) =j} means that if the true classification of sample i is j, the value is 1.

[0076] In this embodiment, the hardware required for the VGG-MetStr model is an Intel i7-9750H CPU, an NVIDIA GTX1650 GPU, 8GB of memory, the Ubuntu18 operating system, the Pycharm environment, the TensorFlow and Keras platforms, and the Python language. The parameters of the VGG-MetStr model are randomly generated, and the experimental data results are visualized using TensorBoard. The images in the training set are used as inputs. After training, the accuracy acc and the loss function loss are obtained.

[0077] During the training process, the value of the loss function gradually decreases, the accuracy gradually increases, and the VGGNet-MetStr model gradually converges until the loss function is less than the allowed range. Figure 5 This is the change process of the accuracy obtained during the training process of the VGGNet-MetStr model. As can be seen from the figure, after about 600 iterations, the accuracy of the VGGNet-MetStr model starts to increase rapidly. After the number of iterations reaches 700, the accuracy reaches about 0.8. The training process is slightly long. When the number of iterations is around 4500 to 5000, the convergence effect is excellent and the accuracy is high.

[0078] The trained VGGNet-MetStr model is verified using the test set to obtain the trained VGGNet-MetStr model.

[0079] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the aging of martensite structure in high-chromium heat-resistant steel, characterized in that, It includes the following steps: Obtain a martensite microstructure image and perform preprocessing; Use the trained VGGNet-MetStr model to identify the preprocessed martensite microstructure image to obtain an identification result. The network structure of the VGGNet-MetStr model includes two parts. One part includes an input module and multiple feature extraction modules, and each feature extraction module consists of two convolutional layers and one pooling layer. The other part consists of two fully connected layers; The identification method includes: The preprocessed martensite microstructure image is input into the feature extraction module through the input module, and a feature image is obtained after being processed by multiple feature extraction modules; Use two fully connected layers to identify the feature image and output the identification result.

2. The method for identifying the aging of a martensite structure of a high-chromium heat-resistant steel according to claim 1, characterized in that: Use a metallurgical microscope to obtain a martensite microstructure image.

3. A method for identifying the aging of martensite structure of high-chromium heat-resistant steel according to claim 1, characterized in that: The preprocessing of the martensite microstructure image includes image size normalization processing, gray value normalization processing, and Gaussian smoothing processing.

4. The method for identifying the aging of martensite structure of high chromium heat-resistant steel according to claim 1, characterized in that: The activation function of the VGGNet-MetStr model is the ELU function.

5. The aging identification method of martensite structure of a high-chromium heat-resistant steel according to claim 4, characterized in that: The ELU function is: where α is an adjustable parameter.

6. A method for identifying the aging of a martensitic structure of a high-chromium heat-resistant steel as described in claim 1, characterized in that: The calculation formula of the convolutional layer is: Among them, w ij is the weight value, x m+i,n+j is the input value, b is the bias, and y mn is the eigenvalue.

7. A method for identifying the aging of martensite structure of high-chromium heat-resistant steel according to claim 1, characterized in that: The sampling method of the pooling layer is uniform sampling, and the calculation formula is: Among them, y mn is the output value of the S1×S2 region, is any input value within the S1×S2 region.

8. A method for identifying the aging of martensite structure of high chromium heat-resistant steel according to claim 1, characterized in that: The fully connected layer uses a softmax classifier: where k represents a certain classification, denotes the value of this classification, and p(y (i) = k|x (i) , θ) represents the probability of the k-th class, and θ is the weight.

9. A method for identifying the aging of martensite structure of high chromium heat-resistant steel according to claim 1, characterized in that: The loss function of the fully connected layer is the cross-entropy function: where l{y (i) = j} means that if the true classification of sample i is j, then the value is 1.

10. A method for identifying the aging of martensite structure of high chromium heat-resistant steel according to claim 1, characterized in that: The training process of the VGGNet-MetStr model includes the following steps: Collect martensite microstructure images and perform preprocessing to obtain a preprocessed dataset; Divide the preprocessed dataset into a training set and a test set at a ratio of 20:3; Input the training set into the VGGNet-MetStr model for training, and use the test set to verify the trained VGGNet-MetStr model to obtain a trained VGGNet-MetStr model.

Images