A method and system for automatic rating of solidification structure of continuous casting billet

Abstract

The present invention discloses a method and system for automatic rating of solidification structure of continuous casting billets, which relates to the field of quality evaluation of continuous casting billets; by combining deep learning technology with machine vision technology, the segmentation of equiaxed crystals at the center of continuous casting billets and the calculation of equiaxed crystal ratio are realized, so that the accuracy and efficiency are higher when identifying the central equiaxed crystals; the interference of subjective factors during low-magnification rating is solved, the work intensity of workers is reduced, the efficiency and accuracy of rating are improved, and efficient automation of rating is realized. Since the boundaries of the solidification structure image are fuzzy and the gradients are complex, res‑unet is used to segment it, and the accurate segmentation of the solidification structure is achieved by fusing high-dimensional and low-dimensional features. The model has strong robustness and good universality, and can be deployed on computers and mobile terminals to complete the segmentation and evaluation of the solidification structure of continuous casting billets. By collecting the solidification structure and continuous casting process data, the internal quality of the billet can be effectively improved, with significant economic benefits.

Description

Technical Field

[0001] The invention relates to the field of continuous casting billet quality evaluation, and in particular to a method and system for automatic rating of continuous casting billet solidification structure. Background Art

[0002] In the steel production process, the internal quality of the continuous casting billet needs to be evaluated. At present, the evaluation method is: low-power corrosion operation is performed on the billet, and then the rating personnel subjectively judge its quality level by comparing the rating chart. This manual method has low accuracy and strong subjective will. Due to the different work experience of each worker, different evaluation standards are often generated, which leads to deviations in the judgment of the billet quality and cannot achieve the role that low-power inspection should play in the continuous casting billet production.

[0003] At present, machine vision technology has begun to be widely used in the field of defect detection, which can overcome the shortcomings of the above-mentioned manual methods. However, its application to the solidification structure and defects of continuous casting billets mainly focuses on the identification of central segregation and cracks, and the detection of the solidification structure of continuous casting billets is usually visual. This is because the gradient of the equiaxed crystal zone of the billet is complex and irregular, and the boundary between the equiaxed crystal zone and the columnar crystal zone is extremely fuzzy, which makes it difficult to identify and segment it. The solidification structure of the continuous casting billet includes edge equiaxed crystals, columnar crystals and central equiaxed crystals. In most cases, steel with good performance requires a large number of central equiaxed crystals, and overdeveloped columnar crystals will lead to serious low-magnification defects. The solidification structure of the billet has a great influence on the quality of the billet, and is widely present in the billet, which is an important part of the evaluation of the billet quality. Therefore, it is necessary to design a method and system for automatic rating of the solidification structure of continuous casting billets. Summary of the invention

[0004] In order to address the deficiencies in the prior art, the present invention provides a method and system for automatic rating of the solidification structure of a continuous casting billet. By combining deep learning technology with machine vision technology, the segmentation of the central equiaxed crystals of the continuous casting billet and the calculation of the equiaxed crystal ratio are achieved, thereby achieving higher accuracy and efficiency in identifying the central equiaxed crystals, which is convenient for users to use.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A method for automatic rating of solidification structure of continuous casting billet comprises the following steps:

[0007] S1: Send the low-magnification sample image of the continuous casting billet to the deep learning model for training;

[0008] S2: setting the casting process information of the continuous casting billet to be evaluated and selecting the low-magnification sample image of the continuous casting billet to be evaluated; the casting process information includes process parameters and dimensions;

[0009] S3: Use the trained deep learning model to segment the low-magnification sample image of the continuous casting billet to be evaluated from the background, and count the number of pixels of the low-magnification sample of the continuous casting billet as N1;

[0010] S4: Using the trained deep learning model, the central equiaxed crystal area of ​​the low-magnification sample image of the continuous casting billet to be evaluated is segmented, and the number of pixels in the central equiaxed crystal area is counted as N2;

[0011] S5: Calculate the equiaxed crystal ratio W of the continuous casting billet according to the number of pixels N1 of the low-magnification sample of the continuous casting billet and the number of pixels N2 of the central equiaxed crystal area (W = N2 / N1*100%);

[0012] S6: According to the ingot casting process information and equiaxed crystal ratio, the selected continuous casting ingot low-magnification sample image is rated, and the sample parameters and rating results are entered into the database.

[0013] The low-magnification sample image of the continuous casting billet is obtained by subjecting the continuous casting billet to hot acid corrosion and using a digital camera to collect the image of the continuous casting billet after corrosion;

[0014] The low-magnification sample is a continuous casting slab corrosion sample with a magnification of no more than 10;

[0015] The deep learning model training includes the following steps:

[0016] S1.1: Collect low-magnification sample images of continuous casting billets, annotate low-magnification sample images and equiaxed crystal area images, and create training set Strain, test set Stest and validation set Sval;

[0017] S1.2: Perform data augmentation operations on the prepared dataset;

[0018] S1.3: Resize the dataset, and send it to the deep learning model after its image size is set to (H, W, C), where H represents the height of the resized image, W represents the width of the resized image, and C represents the number of channels of the resized image;

[0019] S1.4: Select binary_crossentropy in TensorFlow as the loss function and iou as the evaluation criterion to train the model until the specified number of training times is reached or the loss function is lower than the set value;

[0020] S1.5: Test the trained model and save the model whose rating index iou is greater than the set threshold. Otherwise, repeat step S1.4 to continue training.

[0021] The rating indicator iou is set to 0.9.

[0022] The data enhancement operation is performed by selecting one or more of the horizontal flipping, vertical flipping and translation operations of the low-magnification sample image of the continuous casting slab;

[0023] The deep learning model in step S3 is an improved unet model, which uses a unet model that has been downsampled four times, and adds a BN layer to it, and the activation function is selected as relu;

[0024] The deep learning model in step S4 is a res-unet model, which uses a unet model that has been downsampled six times, and adds a BN layer and a residual block of resnets, and the activation function is relu.

[0025] On the other hand, a system for automatic rating of solidification structure of continuous casting billets is provided, which is used to implement the method for automatic rating of solidification structure of continuous casting billets, including a recommendation scheme module, a sample processing module, an industry standard module, a historical data query module, a rating modification module for changing and supplementing historical rating data, and a prediction output module for displaying the continuous casting billet processing results in the form of pictures.

[0026] The recommendation scheme module: the user inputs the size and steel type of the continuous casting billet to be rated, and the acid immersion corrosion conditions corresponding to the continuous casting billet are recommended to achieve a quantitative corrosion effect.

[0027] The sample processing module is composed of selecting sample parameters, rating results and situation analysis; in the sample parameter part, the low-magnification corrosion photo of the continuous casting billet to be evaluated is selected and the corresponding sample parameters are input, including billet size, steel type, casting speed, furnace, furnace number, rater, furnace number, continuous casting machine, superheat, and electric stirring current information; after the system completes the rating of the input photos, the rating result is displayed in the rating result part, and the evaluation result and sample parameters are stored in the database;

[0028] The standard module: analyzes the characteristics of the equiaxed crystals in the standard and displays the equiaxed crystal pictures;

[0029] The historical data query module selects the start and end time, casting speed, steel type, continuous casting machine, and equiaxed crystal ratio to screen the historically rated steel types, and compares the process parameters to optimize the continuous casting process.

[0030] Beneficial technical effects

[0031] 1. The present invention provides a method and system for automatic rating of solidification structure of continuous casting billets, which solves the interference of subjective factors during low-magnification rating, reduces the work intensity of workers, improves rating efficiency and accuracy, and realizes efficient automation of rating.

[0032] 2. Since the boundaries of coagulated tissue images are fuzzy and the gradients are complex, the present invention uses res-unet to segment them, and achieves accurate segmentation of coagulated tissue by fusing high-dimensional and low-dimensional features.

[0033] 3. The model constructed by the present invention is highly robust and universal, and can be deployed on computers and mobile terminals to complete the segmentation and evaluation of the solidification structure of the continuous casting billet.

[0034] 4. The present invention can effectively improve the internal quality of the ingot by collecting the solidification structure and continuous casting process data, and has significant economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flow chart of a method for automatic rating of solidification structure of continuous casting slabs provided by an embodiment of the present invention;

[0036] Figure 2 A data enhancement effect diagram provided by an embodiment of the present invention;

[0037] in, Figure 2 (a) is a schematic diagram of the original low-magnification image. Figure 2 (b) is a schematic diagram of a horizontally flipped image. Figure 2 (c) is a schematic diagram of a vertically flipped image. Figure 2 (d) is a schematic diagram of the translated image;

[0038] Figure 3 A schematic diagram of a model rating function iou provided in an embodiment of the present invention;

[0039] Figure 4 The basic framework of the deep learning model unet provided for the embodiment of the present invention;

[0040] Figure 5 A residual block added to unet provided in an embodiment of the present invention;

[0041] Figure 6 A segmentation effect diagram provided by an embodiment of the present invention;

[0042] in, Figure 6 (a) is the original image of the low-magnification sample of the continuous casting billet. Figure 6 (b) is a manually labeled image. Figure 6 (c) is the model prediction graph;

[0043] Figure 7 A diagram of a sample sheet processing module provided in an embodiment of the present invention;

[0044] Figure 8 A diagram showing the equiaxed crystal region division results provided by an embodiment of the present invention;

[0045] Fig. 9 A schematic diagram of the structure of a system for automatic rating of solidification structure of continuous casting billets provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] To better illustrate the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0047] This embodiment proposes a method for automatic rating of solidification structure of continuous casting billets, such as Figure 1 As shown, the following steps are included:

[0048] S1: Send the low-magnification sample image of the continuous casting billet to the deep learning model for training, including the following steps:

[0049] In step S1, the deep learning model training includes the following steps:

[0050] S1.1: Collect low-magnification sample images of continuous casting billets, manually annotate low-magnification sample images and equiaxed crystal region images, and create training set S train , test set S test And the validation set S val ;

[0051] S1.2: Perform data enhancement on the prepared data set. In deep learning, it is generally required that the number of samples is sufficient. The more samples there are, the better the trained model will be and the stronger the generalization ability of the model will be. However, since it is difficult to obtain high-quality low-magnification images of continuous casting billets and there are few samples, data enhancement is performed on the low-magnification images of continuous casting billets to improve the quality of the data set.

[0052] In this embodiment, the data enhancement operation used includes horizontal flipping, vertical flipping, and translation operations on the low-magnification sample image of the continuous casting billet, and the enhancement effect is as follows: Figure 2 (a)- Figure 2 (d)

[0053] S1.3: Resize the dataset. Experiments show that the best segmentation effect is achieved when the image size is set to (1536, 1536, 3). Then, the dataset is fed into the deep learning model.

[0054] S1.4: Select binary crossentropy in TensorFlow as the loss function and iou as the evaluation function to train the model until the specified number of training times is reached or the loss function is lower than the set value;

[0055] The binary crossentropy formula is:

[0056]

[0057] where y i is the target true value, y i ∈(0,1); is the model prediction value, Loss is a function that measures the difference between the predicted value and the true value.

[0058] iou is defined as Figure 3 As shown, the formula is:

[0059]

[0060] Among them, A is the target area, B is the model prediction area, and iou is used to measure the overlap between the model segmented image and the real image in image segmentation.

[0061] S1.5: Test the trained model. If the test result is good, save the model. Otherwise, repeat step S14 to continue training.

[0062] S2: Input the parameters of the steel type to be evaluated and select the low-magnification sample image of the continuous casting slab to be evaluated on the system interface;

[0063] S3: Use the trained deep learning model to complete the segmentation of the low-magnification sample image of the continuous casting billet and the background, and count the number of pixels of the low-magnification sample of the continuous casting billet as N1;

[0064] In step S3, the deep learning model is an improved unet model, in which a four-time downsampled unet model is selected, a BN layer is added, and the activation function is selected as relu. The structure of the improved unet model is as follows: Figure 4 As shown, the specific steps are:

[0065] S3.1 performs two convolution operations on the image with a size of (1536, 1536, 3), sends it to the BN layer for normalization, and then sends it to the pooling layer for pooling to obtain a feature map with a size of (768, 768, 32);

[0066] S3.2 performs two convolution operations on (768, 768, 32), sends it to the BN layer for normalization, and then sends it to the pooling layer for pooling.

[0067] S3.3 performs an upsampling operation on the obtained (384, 384, 64) feature map, and then concatenates the obtained (768, 768, 32) feature map with the corresponding encoder feature map, and then performs two convolution operations and normalization;

[0068] S3.4 performs an upsampling operation on the obtained (768, 768, 32) feature map, and then concatenates the obtained (1536, 1536, 16) feature map with the corresponding encoder feature map, and then performs two convolution operations and sends it to the convolution layer with a convolution kernel of (1, 1). Finally, after activation by the activation function sigmoid, a prediction map of the low-magnification sample of the continuous casting billet is obtained, and its size is (1536, 1536, 1). The sigmoid function formula is as follows:

[0069]

[0070] This function adds a nonlinear function to the model that only has the linear function of convolution, and constrains the result to the range of 0-1, which enhances the fitting ability of the model and simplifies the model calculation.

[0071] S4: Use the trained deep learning model to segment the central equiaxed crystal area of ​​the low-magnification sample image of the continuous casting billet. The effect is shown in the figure below. Figure 6 (a)- Figure 6 (c) shows that the number of pixels in the central equiaxed crystal area is counted as N2;

[0072] The deep learning model is a res-unet model, which uses a unet model that has been downsampled six times, adds a BN layer to it, and replaces each pooling layer and the first two convolutional layers of the upsampling layer in unet with a residual block of resnets. The activation function is relu, and the residual block is as follows: Figure 5 As shown in the figure, the input feature map x first undergoes a convolution operation, is activated by the relu function, and then convolved to obtain the feature map F(x), which is then added to x and activated by the relu function to obtain the output feature map, which is sent to the pooling layer or upsampling layer as input. The residual block acts as a differential amplifier, which can effectively solve the model degradation problem that occurs as the model depth increases, and enhance the robustness, generalization and learning ability of the model.

[0073] S5: According to the number of pixels of the low-magnification sample of the continuous casting billet obtained and the number of pixels of the central equiaxed crystal region thereof, the present invention refers to the quality evaluation standard "GB / T 24178-2009, Low-magnification evaluation method for solidification structure of continuous casting billet", and calculates the equiaxed crystal ratio W (W=N2 / N1*100%) of the continuous casting billet;

[0074] S6: According to the steel grade parameters and equiaxed crystal ratio, the selected continuous casting billet low-magnification sample image is comprehensively rated, and the rating results are displayed on the sample processing module, such as Figure 7 The division result of the equiaxed crystal area is displayed in the system's prediction output - equiaxed crystal module, as shown in Figure 8As shown, in this module, users can freely adjust the low-magnification image size to facilitate the observation of the equiaxed crystal area segmentation effect. The sample parameters and rating results are entered into the database. If the user has any questions about the rating results, the data can be modified in the rating modification module.

[0075] In this embodiment, 36 low-magnification sample images of continuous casting billets are used as training sets, 4 as test sets, and 4 as verification sets are sent to the deep learning model for training. The segmentation effect of the equiaxed crystal zone on the test set reaches an iou of more than 92%.

[0076] On the other hand, this embodiment provides a system for automatic rating of solidification structure of continuous casting billets, which is used to implement the method for automatic rating of solidification structure of continuous casting billets, such as Fig. 9 As shown, it includes a recommendation module, a sample processing module, an industry standard module, a historical data query module, a rating modification module for changing and supplementing historical rating data, and a prediction output module for displaying the continuous casting billet processing results in the form of pictures.

[0077] The recommendation module: the user inputs the size and steel type of the continuous casting billet to be rated, and the system will recommend the corresponding low-magnification corrosion conditions for the continuous casting billet to achieve a quantitative corrosion effect;

[0078] The sample processing module: This module consists of sample parameters, rating results and situation analysis. In the sample parameter section, the user needs to select the low-magnification corrosion photo of the continuous casting billet to be evaluated and input the corresponding sample parameters such as billet size, steel type, casting speed, furnace, furnace number, rater, furnace number, continuous casting machine, superheat, and electric stirring current information. After the system completes the rating of the input photos, the rating results will be displayed in the rating result section of the module, and the results will be displayed in the situation analysis section. At the same time, the corresponding results and sample parameters will be stored in the database;

[0079] The standard module: analyzes the characteristics of the equiaxed crystals in the standard and displays the equiaxed crystal pictures;

[0080] The historical data query module: In this module, the user selects the start and end time, casting speed, steel type, continuous casting machine, and equiaxed crystal ratio options to screen the historically rated steel types, compare their process parameters, and optimize the continuous casting process.

Claims

1. A method for automatic rating of solidification structure of continuous casting billet, characterized by: The following steps are involved: S1: Send the low-magnification sample image of the continuous casting billet to the deep learning model for training; S2: setting the casting process information of the continuous casting billet to be evaluated and selecting the low-magnification sample image of the continuous casting billet to be evaluated; the casting process information includes process parameters and dimensions; S3: Use the trained deep learning model to segment the low-magnification sample image of the continuous casting billet to be evaluated from the background, and count the number of pixels of the low-magnification sample of the continuous casting billet as N1; S4: Using the trained deep learning model, the central equiaxed crystal area of ​​the low-magnification sample image of the continuous casting billet to be evaluated is segmented, and the number of pixels in the central equiaxed crystal area is counted as N2; S5: Calculate the equiaxed crystal ratio W of the continuous casting billet according to the number of pixels N1 of the low-magnification sample of the continuous casting billet and the number of pixels N2 of the central equiaxed crystal area (W = N2 / N1*100%); S6: According to the ingot casting process information and equiaxed crystal ratio, the selected continuous casting ingot low-magnification sample image is rated, and the sample parameters and rating results are entered into the database.

2. The method for automatic rating of solidification structure of continuous casting billet according to claim 1 is characterized in that: The low-magnification sample image of the continuous casting billet is obtained by subjecting the continuous casting billet to hot acid corrosion, and using a digital camera to collect the image of the continuous casting billet after corrosion.

3. The method for automatic rating of solidification structure of continuous casting according to claim 2, characterized in that: The low-magnification sample is a continuous casting slab corrosion sample with a magnification of no more than 10.

4. The method for automatic rating of solidification structure of continuous casting billet according to claim 1, characterized in that: The deep learning model training includes the following steps: S1.1: Collect low-magnification sample images of continuous casting billets, annotate low-magnification sample images and equiaxed crystal area images, and create training set Strain, test set Stest and validation set Sval; S1.2: Perform data augmentation operations on the prepared dataset; S1.3: Resize the dataset, and send it to the deep learning model after its image size is set to (H, W, C), where H represents the height of the resized image, W represents the width of the resized image, and C represents the number of channels of the resized image; S1.4: Select binary_crossentropy in TensorFlow as the loss function and iou as the evaluation criterion to train the model until the specified number of training times is reached or the loss function is lower than the set value; S1.5: Test the trained model and save the model whose rating indicator iou is greater than the set threshold, otherwise repeat step S1.4 to continue training.

5. The method for automatic rating of solidification structure of continuous casting billet according to claim 4 is characterized in that: The rating indicator iou is set to 0.

9.

6. The method for automatic rating of solidification structure of continuous casting billet according to claim 4, characterized in that: The data enhancement operation adopts one or more of the horizontal flipping, vertical flipping and translation operations of the low-magnification sample image of the continuous casting billet.

7. The method for automatic rating of solidification structure of continuous casting billet according to claim 1, characterized in that: The deep learning model in step S3 is an improved unet model, which uses a unet model that has been downsampled four times, and a BN layer is added to it, and the activation function is selected as relu.

8. The method for automatic rating of solidification structure of continuous casting billet according to claim 1, characterized in that: The deep learning model in step S4 is a res-unet model, which uses a unet model that has been downsampled six times, and adds a BN layer and a residual block of resnets, and the activation function is relu.

9. A system for automatic rating of solidification structure of continuous casting billets, used to implement the method for automatic rating of solidification structure of continuous casting billets as claimed in claim 1, characterized in that: It includes a recommendation module, a sample processing module, an industry standard module, a historical data query module, a rating modification module for changing and supplementing historical rating data, and a prediction output module that displays the continuous casting billet processing results in the form of pictures.

10. The system for automatic rating of solidification structure of continuous casting billet according to claim 9, characterized in that: The recommendation module: the user inputs the size and steel type of the continuous casting billet to be rated, and the acid immersion corrosion conditions corresponding to the continuous casting billet are recommended to achieve a quantitative corrosion effect; The sample processing module is composed of selecting sample parameters, rating results and situation analysis; in the sample parameter part, the low-magnification corrosion photo of the continuous casting billet to be evaluated is selected and the corresponding sample parameters are input, including billet size, steel type, casting speed, furnace, furnace number, rater, furnace number, continuous casting machine, superheat, and electric stirring current information; after the system completes the rating of the input photos, the rating result is displayed in the rating result part, and the evaluation result and sample parameters are stored in the database; The standard module: analyzes the characteristics of the equiaxed crystals in the standard and displays the equiaxed crystal pictures; The historical data query module selects the start and end time, casting speed, steel type, continuous casting machine, and equiaxed crystal ratio to screen the historically rated steel types, and compares the process parameters to optimize the continuous casting process.

Images