JP7444098B2 - Slag outflow determination method, converter operating method, molten steel manufacturing method, slag outflow determination device, converter operating equipment, and molten steel manufacturing equipment - Google Patents

Description

The present invention provides a slag outflow determination technique for detecting the mixing of slag into the discharge stream (molten steel flow) discharged from the converter when molten steel is discharged from the converter into a ladle, and the use of the technique. This technology is related to converter operation and molten steel production.
Here, in this specification, machine learning includes deep learning.

The molten steel refined in the converter is discharged from the molten steel port of the converter toward the ladle by tilting the converter. If the timing to finish tapping the steel into the ladle (stopping the tapping) is late, the slag in the converter is discharged from the tap port and flows into the ladle. If a large amount of slag flows into the ladle, impurities such as P in the slag return to the molten steel, affecting the processing time of the next step and the auxiliary materials used. Furthermore, since the alloy to be introduced is captured in the slag, it also causes a decrease in alloy yield. On the other hand, if the steel is tapped into the ladle too quickly, a large amount of steel will remain in the converter, which will also lead to a decrease in yield. Therefore, it is important to determine whether or not slag is mixed in the discharge stream discharged from the converter.

Conventionally, when a skilled person observes the color of the discharge stream discharged from a converter and determines that slag is mixed in the discharge stream, he or she stops tapping the steel. However, there were problems such as the need for skill in making judgments and the limited number of locations where the exhaust flow could be clearly seen.

As a conventional technique for solving this problem, for example, there is a determination method using an infrared method (see Patent Document 1).

The infrared method is a method in which an infrared camera images the discharge stream (molten steel flow) discharged from a converter, and it is determined whether slag is mixed in based on the captured image. Specifically, this method is based on the energy difference between the radiant energy of the molten steel and the radiant energy of the slag in the discharge stream measured by the infrared camera, that is, the discharge stream from the converter is photographed with an infrared camera. This method determines whether slag is mixed into the discharge stream based on the difference in brightness between slag and molten steel. If it is determined that slag has been mixed in, the discharge of molten steel (steel tapping) from the converter is stopped.

Patent No. 5444692

Slag outflow determination using the infrared method is either a method that determines slag based on a fixed threshold value mainly based on differences in radiant energy (appears as a difference in brightness), or a method that determines slag based on a change in the average value of energy measurements up to a determination point (threshold value). A method of determining is adopted. However, no matter which method is adopted, the threshold value for determination must be set with a certain degree of leeway in order to avoid erroneous determinations or to accommodate differences between steel types. Therefore, there is a problem in that the determination timing is delayed accordingly.

In addition, in order to suppress the influence of dust associated with the discharge of molten steel and to make stable judgments, this infrared method uses a discharge stream that is less susceptible to the influence of dust associated with discharge of molten steel and in which mixed slag is dispersed. It is necessary to make a determination based on an infrared image taken at the lower position (for example, around the bottom third). However, if the discharge stream branches or is narrow at the bottom of the discharged tap stream, the proportion of slag pixels included in the infrared image decreases, and there are cases where slag contamination is not detected. The inventor discovered that there is a problem. Another problem is that there is a delay in detecting slag contamination because the determination is made at the lower part of the discharge stream.

Furthermore, in the infrared method, many settings such as a plurality of initial values and threshold values (for example, 14 setting values) are required for determination. Therefore, it is necessary to adjust the plurality of set values described above according to the operating environment (steel type, furnace condition, etc.) that changes over time. However, since this adjustment of the set value is a phenomenon that cannot be determined by calculation, it is necessary to adjust the set value through a large amount of testing work.

The present invention was made with attention to the above points, and aims to improve the accuracy of determining slag contamination while reducing the maintenance load by using captured infrared images during converter operation. .

The inventor used an infrared camera to photograph the tap flow (discharge flow) flowing out of the tap port of a converter, and separated and labeled the obtained infrared images into images with slag mixed in and images with no slag mixed in. It was found that slag contamination can be determined with high accuracy by using a trained model obtained by training the training data using a CNN (convolutional neural network).

Furthermore, the inventor used Grad-CAM (gradient load class activation mapping) as the basis for layer prediction by CNN, using the target infrared image as an image of the entire discharge flow from the tap hole to the hot water surface of the ladle. ) technology, it was found that the basis for classification judgments extends to the upper part of the tap flow. They also discovered that by expanding the target image area from the conventional lower area to also include the upper area, it is possible to generate a model with even greater accuracy.

In order to solve the problem, one aspect of the present invention is to image the discharge flow discharged from the tapping port of the converter toward the ladle with an infrared camera, and based on the captured image, A slag inclusion determination method for determining whether slag is included in a machine, the method includes an image acquisition step of continuously acquiring images of the discharge flow taken by the infrared camera, and a step of continuously acquiring images of the discharge flow taken by the infrared camera; The method further comprises a determination step of determining the presence or absence of slag contamination from the captured image obtained in the image acquisition step using a learned model that recognizes the presence or absence of slag contamination. .

Further, one aspect of the present invention is to image the discharge flow discharged from the tapping port of the converter toward the ladle with an infrared camera, and to determine whether or not slag is mixed into the discharge flow based on the captured image. A slag contamination determination device for determining slag contamination, comprising an image acquisition unit that acquires an image of the discharge flow taken by the infrared camera, and a machine learning model using the infrared image of the discharge flow. The present invention further comprises a determination unit that determines the presence or absence of slag contamination from the captured image acquired by the image acquisition unit using a trained model that recognizes the presence or absence of slag.
In addition, it is preferable that the captured image is an image that includes a range from the tap hole position to the position directly above the ladle in the discharge stream.

According to the aspect of the present invention, during converter operation, it is possible to accurately determine slag contamination while suppressing maintenance load using captured infrared images.

BRIEF DESCRIPTION OF THE DRAWINGS It is a schematic diagram which shows the converter operation equipment based on embodiment based on this invention. It is a diagram showing an example of the configuration of a slag outflow determination device. FIG. 2 is a diagram illustrating a configuration example of a trained model generation section. FIG. 3 is a diagram illustrating an example of captured images that constitute unique data. FIG. 2 is a diagram illustrating the configuration of a learned model updating section. FIG. 3 is a flow diagram illustrating an example of processing for determining an image area.

Next, embodiments of the present invention will be described with reference to the drawings.
When the refining of molten steel in the converter is completed, the ladle is moved to the lower part of the converter, and as shown in FIG. Steel is tapped from the steel mouth 3a toward the ladle 4.

(composition)
In this embodiment, a slag stopper 10 is installed near the tapping port 3a of the converter 3. The slag stopper 10 includes a rotatable arm 11, a stopper 12 made of cast iron attached to the tip of the arm 11, and a hydraulic cylinder 13 that drives the arm 11. The hydraulic cylinder 13 is controlled by a slag stopper control device 16, and the stopper portion 12 closes the tapping port 3a to stop discharging the molten steel 1 from the converter 3.
Here, the discharge stream 1A discharged from the tapping port 3a is usually called a "tapping stream."
Moreover, in this embodiment, an infrared camera 15 and a slag outflow determination device 20 are provided.

<Infrared camera 15>
The infrared camera 15 is a camera that images the discharge stream 1A discharged from the tapping port 3a toward the ladle 4.

In the conventional infrared method, as described above, the lower three of the discharge stream 1A in the range from the tapping port 3a to the height position of the upper end surface 4a of the ladle 4 (the position directly above the ladle 4) The image was taken of an area close to one-tenth of that of the previous one.

On the other hand, when we visualized the basis for judgment using Grad-CAM technology for the trained model 50 that was machine learned using CNN, we found a new finding that the upper region of the exhaust flow 1A is given importance. Obtained.

Therefore, in the present embodiment, the imaging range of the infrared camera 15 is set to a position directly above the ladle 4 that receives the discharge flow 1A from the tapping port 3a position in the discharge flow 1A (the height of the upper end 4a of the ladle 4). ) is set to be an area that includes the range up to The imaging range and the hot water level in the ladle 4 may also be included. This is because the condition of the hot water level changes depending on whether or not slag is mixed in.

As described above, in the present embodiment, the infrared camera 15 continuously images the discharge stream 1A located at least in the range directly above the ladle 4 from the tapping port 3a, and the captured images are transferred to the slag outflow determination device 20. I am sending it to

<Slag outflow determination device 20>
As shown in FIG. 2, the slag outflow determination device 20 includes an image acquisition section 20A, a determination section 20B, and a transmission section 20C.

[Image acquisition unit 20A]
The image acquisition unit 20A performs a process of continuously acquiring images captured by the infrared camera 15 at a predetermined sampling period. Alternatively, the infrared camera 15 may capture a moving image, and the image acquisition unit 20A may extract a still image from the moving image as the captured image.

Note that the captured image may be obtained by cutting out a part of the infrared image taken by the infrared camera 15, but if the captured image area is located at least in the range from the tap hole 3a position to the position directly above the ladle 4. The image is set so that the image captures the discharge flow 1A.

[Determination unit 20B]
The determination unit 20B uses the trained model 50 to perform a process of determining whether or not slag is mixed in from the captured image acquired by the image acquisition unit 20A.

The learned model 50 is a model that has been machine learned using infrared images of the discharge flow, and is a model that recognizes the presence or absence of slag contamination. The trained model 50 uses captured images as input data, and uses the determination (classification) of presence or absence of slag contamination as output data.
The determination unit 20B of this embodiment stores the acquired captured image and the determination result using the captured image in the storage unit 21.
Note that the storage unit 21 is composed of a database and other components, but does not need to be physically one medium.

[Transmitter 20C]
The transmitter 20C executes a process of transmitting the determination result of the determiner 20B. When the transmitter 20C of the present embodiment determines that slag is mixed through the processing of the determination unit 20B, it outputs a command to the slag stopper control device 16 to close the tapping port 3a. Upon receiving the closing command, the slag stopper control device 16 drives the hydraulic cylinder 13 to close the tapping port 3a.

Here, the process of stopping the discharge from the tapping port 3a is not limited to the process using the slag stopper. For example, the transmitter 20C outputs a notification command to a notification device such as a buzzer, and in response to the notification command, the operator stops the tilting of the furnace body of the converter 3, thereby stopping the discharge of the molten steel 1. Also good.

<Generation of trained model 50>
Generation of the above trained model 50 will be explained.
As shown in FIG. 3, the slag outflow determination device 20 includes a trained model generation section 22. The trained model generation unit 22 performs processing to generate a trained model 50.
The trained model generation section 22 includes a learning data storage section 22A, a first learning section 22B, a singular data acquisition section 22C, and a second learning section 22D.

[Learning data storage unit 22A]
The learning data storage unit 22A performs a process of storing learning data 51 (teacher data) in the storage unit 21, which is a captured image captured by the infrared camera 15 and labeled with whether or not slag is mixed.
The imaging range of this captured image is also an area including the discharge stream 1A located in the range directly above the ladle 4 from the tapping port 3a.

For example, regarding the initial learning data 51, the entire discharge stream 1A is continuously imaged with an infrared camera 15, and an expert determines whether or not slag is mixed in the captured images, and the images are divided into images with and without slag. to classify. In addition, a conventional infrared method is used to determine the discharge flow 1A, and the captured image before determining that slag is mixed in is labeled as "no slag mixed in", and the captured image after determining that slag is mixed in is labeled with slag mixed in. Label. Thereafter, data may be generated by learning by manually correcting the judgment results for data that is difficult to judge from before and after the start of slag flow until it becomes completely slag, or for unusual data such as branching.

The number of learning models used for learning is, for example, 3000.

[First learning section 22B]
The first learning unit 22B performs machine learning (deep learning) using the learning data 51 stored in the storage unit 21, and executes a process to obtain a first learning model 52 that recognizes the presence or absence of slag contamination.

For example, the first learning section 22B sequentially inputs the learning data 51 accumulated in the accumulation section 21 and learns the input image using a CNN (convolutional neural network), thereby using the captured image as input data and using the slug 2 A trained model 50 whose output data is the classification of presence/absence of mixture is generated.

Here, a convolutional neural network (CNN) extracts the image features by stacking multiple convolutional layers and pooling layers on the input image in multiple stages, and in the final stage, creates an output that categorizes the image. It is a neural network that

[Special data acquisition unit 22C]
The singular data acquisition unit 22C uses the learning data 51 for evaluation prepared in advance, uses the learning data 51 for evaluation as input data, and uses the first learning model 52 obtained by the first learning unit 22B. A process is performed in which the presence or absence of slag contamination is determined, and specific data 53, which is an infrared image incorrectly determined in the determination and which is an infrared image labeled with a label indicating the presence or absence of slag contamination, is selected and acquired. The acquired unique data 53 is stored in the storage unit 21.

The learning data 51 for evaluation is an infrared image prepared for evaluation, and is an infrared image with a label indicating whether or not slag is mixed.
A unique captured image that is likely to be misjudged using the conventional infrared method is used as a captured image for evaluation. For example, as shown in FIG. 4, an image of a branched discharge stream 1A or an image when the discharge stream 1A becomes narrow are examples of unique captured images. Such unique captured images are used as learning data 51 for evaluation.
Alternatively, the learning data 51 used in the first learning section 22B may be used as the learning data 51 for evaluation.

[Second learning section 22D]
The second learning unit 22D performs additional learning using CNN on the first learning model 52 obtained by the first learning unit 22B using the singular data 53 obtained by the singular data obtaining unit 22C, and adds The model after learning is stored in the storage unit 21 as a learned model 50.

In this way, the accuracy of the learned model 50 is improved by performing additional learning using the incorrectly determined learning data 51. Note that the first learning model 52 obtained by the first learning unit 22B may be used as the learned model 50.

<Learned model update unit 23>
This embodiment includes a learned model updating section 23.
The trained model updating unit 23 performs updating processing of the trained model 50 as appropriate in accordance with the accumulation of new learning data 51.

Note that, as described above, the unique data 53 acquired by the unique data acquisition unit 22C is stored in the storage unit 21.
As shown in FIG. 5, the learned model update section 23 includes a learning data update section 23A and a relearning section 23B.

[Learning data update unit 23A]
The learning data updating unit 23A generates learning data 51 from the captured image and the determination result determined by the determining unit 20B, stores the generated learning data 51 in the storage unit 21, and stores the generated learning data 51 in the storage unit 21. The current learning data 51 is updated. Note that if the number of learning data stored in the storage unit 21 is limited to a finite range due to capacity limitations, the storage timing may be adjusted as appropriate, for example, so that the number of learning data is within a preset number of data. Old learning data may be deleted one by one. Alternatively, data to be deleted may be determined by setting other criteria. As a result, the learning data is updated according to the condition of the converter 3 and the usage environment.

[Re-learning section 23B]
The relearning unit 23B uses the learning data 51 of the storage unit 21 updated by the learning data update unit 23A and the specific data 53 stored in the storage unit 21 to obtain a learning model that recognizes the presence or absence of slag contamination. , changes the learned model 50.

For example, the relearning unit 23B may be executed every predetermined period, for example, every three months or every six months, or the relearning unit 23B may be configured to perform new learning when a predetermined number of data (for example, 1/3 or more data) of the accumulated learning data 51 is executed. If the data 51 is obtained, a new learned model 50 is generated and the learned model 50 is updated.

The model generation process executed by the relearning unit 23B is basically the same as that of the first learning unit 22B, except that the unique data 53 is also included as part of the learning data from the beginning.

When the preset conditions are satisfied, the trained model updating unit 23 automatically obtains and updates a new trained model 50, thereby automating the updating of the trained model 50. Furthermore, by using the singular data 53 obtained in advance, the accuracy of the model is improved.

<Regarding the range of the image area acquired by the image acquisition unit 20A>
The area of the discharge stream 1A imaged by the infrared camera 15 is expanded compared to the conventional infrared method, and is set to include at least the range from the tap hole 3a to directly above the ladle 4, and the basis for image prediction by CNN is set. was investigated by Grad-CAM (Gradient Loaded Class Activation Mapping). As a result, it was found that the basis for classification judgments extends to the upper part of the tap flow. They also discovered that by changing the target image area from just the lower area to include the shape area, it is possible to generate and use a model with even greater accuracy.

That is, in this embodiment, the area of the captured image to be acquired in the image acquisition step is determined by steps as shown in FIG. 6 using the Grad-CAM technology.

First, in a learning model generation step S10, a learning model that recognizes the presence or absence of slag contamination is generated by CNN learning using the learning data 51 in which an infrared image is labeled as indicating the presence or absence of slag contamination. For example, the learning data 51 is generated using 1000 pieces of learning data 51. The generation process may be performed, for example, by a process similar to that of the first learning unit 22B.

Next, in a convolutional layer influence evaluation step S20, an influence coefficient that a change in each pixel of a plurality of convolutional layers of the generated learning model has on the slug 2 determination output is calculated.

Next, in input evaluation image creation step S30, the infrared image captured by the infrared camera 15 is used as an input image using the influence coefficient obtained in step S20 and the plurality of convolution layers used for calculating the influence coefficient. The influence of each pixel of the input image on the slug 2 determination output is evaluated.

Then, from the input evaluation image created in the input evaluation image creation step S30, an area of the infrared image that has a high influence on the slug 2 determination output is extracted, and acquired in the image acquisition step so as to include the extracted area. Determine the area of the captured image to be captured.

Here, Grad-CAM is an abbreviation for Gradient-weighted Class Activation Mapping, and is a technology that visualizes important pixels (pixels with large gradients) by weighting gradients with respect to predicted values. Grad-CAM processing is performed, for example, as follows. Input the input image in forward direction and extract the convolutional layer and classification results. Next, error backpropagation is performed based on the classification results to form the gradient of the convolutional layer. Next, take the GAP of the gradient of the convolutional layer, take the sum of the weights of the convolutional layer using the GAP as a weight, resize the original image size, obtain a Grad-CAM image, and visualize the basis of the image.

(Modified example)
Here, the generation of the trained model 50 is not limited to the above method.
For example, by adding a new layer using fine tuning technology to the final stage of a known model consisting of a neural network that has already been trained on general phenomena, and relearning for the actual judgment, this embodiment You may also generate a trained model 50 to be employed.

For example, only the final four layers (Flatten, Dense, Dropout, and Dense) are added to an existing stable model and relearning is performed. Fine tuning replaces the final output layer of the trained model 50 and changes the parameters of the part near the input layer and the output layer. This makes it possible to adapt to new tasks by making it possible to adapt to new tasks that are not included in existing learning.

Here, the Flatten layer is a layer for one-dimensionalizing data for fully connected layer input. The Dense (fully connected) layer is a layer that combines image data from which feature parts have been extracted into one and outputs a value (feature variable) converted by an activation function. The Dropout layer is a layer that prevents overfitting by randomly cutting off some of the connections between the fully connected layer and the output layer.

(Other operations)
The slag outflow determination device 20 of this embodiment can accurately determine whether slag is mixed in by determining whether or not slag is mixed in by using the trained model 50 trained by CNN on continuously captured infrared images. It becomes possible.

As a result, in the operation of the converter 3, it becomes possible to suppress the amount of slag 2 mixed into the ladle 4, and it becomes possible to improve the alloy yield.

In addition, when the basis for judgment was investigated using Grad-CAM technology, based on new knowledge that the basis for classification judgment extends to the upper part of the tapping stream, in this embodiment, when determining slag contamination based on infrared images. In addition, the area in which the discharge stream 1A of the infrared image (captured image) used for determination is shown was made wider than before to include the area from the tap hole 3a to the top of the ladle 4. This further improves the determination accuracy of the trained model 50. Here, the finding that the learning model trained on CNN places importance on the upper part of the steel tapping flow is a new finding.

In the infrared method, a determination area is provided at the bottom of the tap flow in order to determine where the slag 2 has fallen and the proportion of slag 2 has increased, leading to a delay in determination. In contrast, in the present embodiment, the judgment is based on the point immediately after the steel comes out of the tapping port 3a (the upper part of the tapping flow), and it was confirmed that the contamination of slag could be determined at an early stage, just like an expert. ing.

Further, when generating the trained model 50, additional learning is performed using a specific image (specific data 53) that is likely to be incorrectly determined using the conventional infrared method, thereby further improving the determination accuracy of the trained model 50.

The system also includes a system that automatically updates the trained model 50 by appropriately regenerating the trained model 50 using newly acquired learning data 51 and pre-obtained singular data 53 (fixed learning data). This makes it possible to suppress deterioration in the determination accuracy of the trained model 50 even if the steel type or operating conditions change over time.

In other words, the infrared method does not require the adjustment of many setting values, and automatically calculates the current operating conditions using CNN using the learning data 51 (teacher data) that is updated as appropriate. Since the detected parameters are automatically learned, the maintenance load is reduced.

(others)
The present disclosure can also take the following configuration.
(1) A slag inclusion determination method that images the discharge stream discharged from the tapping port of the converter toward the ladle with an infrared camera, and determines whether or not slag is mixed in the discharge stream based on the captured image. The method includes an image acquisition step of continuously acquiring images of the discharge stream taken by the infrared camera, and a machine learning model using the infrared images of the discharge stream to determine whether or not the slag is mixed. and a determination step of determining the presence or absence of slag contamination from the captured image acquired in the image acquisition step using the trained model for recognition.

This slag outflow determination method, for example, uses an infrared camera to image the discharge stream discharged from the tapping port of the converter toward the ladle, and based on the captured image, determines whether or not slag has been mixed into the discharge stream. A slag contamination determination device for determining slag contamination, comprising an image acquisition unit that acquires an image of the discharge flow taken by the infrared camera, and a machine learning model using the infrared image of the discharge flow. and a determination section that determines whether or not slag is mixed in from the captured image acquired by the image acquisition section using a trained model that recognizes the presence or absence of slag.
According to this configuration, during converter operation, it is possible to accurately determine slag contamination while suppressing the maintenance load using the captured infrared image.

(2) The infrared image used for the learning and the captured image used for determining slag contamination are images that include the range from the tap hole position to the position directly above the ladle in the discharge flow.
For example, the captured image acquired by the image acquisition unit is an image that includes a range from the tap hole position in the discharge stream to a position directly above the ladle.

According to this configuration, the upstream discharge flow position, which is not used in the conventional infrared method, is also included in the captured image, thereby improving the determination accuracy using the trained model.
Furthermore, by including the upstream discharge flow position in the captured image area, the determination speed is improved compared to the conventional infrared method.

(3) Learning data that labels infrared images captured by the infrared camera as to whether or not slag is present is stored in the storage section, and machine learning is performed using the learning data stored in the storage section to detect slag contamination. A first learning model that recognizes the presence or absence of slag is determined, and the first learning model determined above is used to detect the presence or absence of slag in the infrared image prepared for evaluation and labeled with the presence or absence of slag contamination. The presence or absence of slag is determined, and unique data consisting of infrared images that are incorrectly determined in the determination and labeled with the presence or absence of slag contamination is acquired, and the acquired data is applied to the first learning model obtained above. Additional learning is performed using the unique data, and the learned model after the additional learning is defined as the learned model.

This method includes, for example, a trained model generation unit that generates the trained model, and the trained model generation unit generates training data that is labeled with a label indicating whether or not slag is present in an image taken by the infrared camera. a learning data storage unit that stores the learning data in the storage unit; a first learning unit that performs machine learning using the learning data stored in the storage unit to obtain a first learning model that recognizes the presence or absence of slag contamination; For the infrared image prepared for evaluation and labeled with the presence or absence of slag contamination, use the first learning model obtained in the first learning section to determine the presence or absence of slag contamination, and make the determination. A singular data acquisition unit that selects singular data that is an infrared image that is incorrectly determined in the above and is labeled as indicating whether or not slag is mixed; and a first learning model obtained by the first learning unit; A second learning section that performs additional learning using the singular data acquired by the singular data acquisition section and updates the model after the additional learning as the learned model.
According to this configuration, it is possible to generate a trained model with high accuracy.

(4) The acquired specific data is stored in the storage section in advance, and the captured image whose presence or absence of slag has been determined in the determination step is stored in the storage section 21 as learning data and accumulated in the storage section. The trained model that recognizes the presence or absence of slag contamination is retrained using the updated learning data in the storage section and the unique data stored in the storage section.

In this method, for example, the singular data acquired by the singular data acquisition unit is stored in advance in the storage unit, and learning data is further generated from the captured image and the determination result determined by the determination unit. a learning data updating unit that stores the learned data in the storage unit and updates the learning data stored in the storage unit, the learning data of the storage unit updated by the learning data updating unit, and the storage unit; A slag outflow determination device includes a relearning section that uses unique data stored in the retraining section to determine a learning model that recognizes the presence or absence of slag contamination, and updates the determined learning model as the learned model. .

According to this configuration, even if the conditions of the converter state are changed, the learned model is automatically updated according to the conditions. In addition, since the updated learning is performed including the peculiar data for which a correct judgment could not be made initially, highly accurate learning judgment performance is maintained.

(5) The above learning is performed using a convolutional neural network (CNN).
According to this configuration, it is possible to reliably generate a trained model.

(6) Recognize the presence or absence of slag contamination using learning data labeled with the presence or absence of slag contamination for an infrared image that includes at least the range from the tap hole position to the position directly above the ladle in the discharge flow. a learning model generation step that generates a learning model for the learning model generated in the learning model generation step; and a convolution layer impact evaluation that calculates an influence coefficient that changes in each pixel of a plurality of convolutional layers of the learning model generated in the learning model generation step has on the slug determination output. step, the above-mentioned influence coefficient, and the plurality of convolutional layers used to calculate the above-mentioned influence coefficient, the infrared image captured by the above-mentioned infrared camera is used as an input image, and each pixel of the input image is converted into a slag judgment output. an input evaluation image creation step for evaluating the influence on the slag determination output, from the input evaluation image created in the input evaluation image creation step, extracting an area of the infrared image that has a high influence on the slag determination output, The area of the captured image to be acquired in the image acquisition step is determined so as to include the area.
According to this configuration, it is possible to set an optimal imaging area range that can improve the determination accuracy.

(6) A converter operating method, which operates a converter using the slag outflow determination method described above.
According to this configuration, it is possible to provide a converter operating method with a high yield.
(7) A method for manufacturing molten steel, comprising the converter operating method described above.
According to this configuration, it is possible to provide a molten steel manufacturing method with a high yield.

(8) Operating equipment for a converter, comprising the slag outflow determination device described above.
According to this configuration, it is possible to provide operating equipment for a converter with a high yield.
(9) Molten steel manufacturing equipment comprising the slag outflow determination device described above.
According to this configuration, it is possible to provide molten steel manufacturing equipment with a high yield.

1 Molten steel 1A Discharge stream 2 Slag 3 Converter 3a Tap port 4 Ladle 4a Upper end of ladle (directly above position)
15 Infrared camera 20 Slag outflow determination device 20A Image acquisition unit (image acquisition step)
20B Judgment section (judgment step)
20C Transmission unit 21 Accumulation unit 22 Learned model generation unit 22A Learning data accumulation unit 22B First learning unit 22C Singular data acquisition unit 22D Second learning unit 23 Learned model update unit 23A Learning data update unit 23B Relearning unit 50 Trained model 51 Learning data 52 First learning model 53 Specific data

Claims (11)

A method for determining slag contamination, in which a discharge stream discharged from a tapping port of a converter toward a ladle is imaged with an infrared camera, and based on the captured image, it is determined whether or not slag is mixed in the discharge stream. ,
an image acquisition step of continuously acquiring images of the discharge flow taken by the infrared camera;
Judgment to determine the presence or absence of slag contamination from the captured image acquired in the image acquisition step using a trained model that is machine learned using the infrared image of the discharge flow and recognizes the presence or absence of slag contamination. step and
has
The infrared image used for the machine learning and the captured image used for determining slag contamination are images that include the range from the position of the tapping port in the discharge stream to the position directly above the ladle,
Learning data labeled with the presence or absence of slag contamination is stored in the storage unit for the infrared images taken by the above infrared camera,
Perform machine learning using the learning data accumulated in the storage unit to obtain a first learning model that recognizes the presence or absence of slag contamination,
The first learning model obtained above is used to determine the presence or absence of slag contamination for the infrared image prepared for evaluation and labeled with the presence or absence of slag contamination. Obtain specific data consisting of infrared images labeled with slag contamination or not.
Perform additional learning on the first learning model obtained above using the acquired singular data, and use the learning model after additional learning as the learned model,
The above-mentioned unique data includes image data of a branched discharge flow and image data of a narrowed discharge flow,
A slag outflow determination method characterized by the following. The specific data acquired above is stored in the storage section in advance,
storing the captured image whose presence or absence of slag has been determined in the determination step in the storage unit as learning data, and updating the learning data stored in the storage unit;
retraining the trained model that recognizes the presence or absence of slag contamination using the updated learning data of the storage unit and the unique data stored in the storage unit;
The slag outflow determination method according to claim 1 . 3. The slag outflow determination method according to claim 1 , wherein the machine learning is performed using a convolutional neural network (CNN). Learning to recognize the presence or absence of slag contamination using learning data labeled with slag contamination for infrared images that include at least the range from the position of the tapping port in the discharge stream to the position directly above the ladle. a learning model generation step for generating a model;
a convolutional layer influence evaluation step of calculating an influence coefficient that a change in each pixel of a plurality of convolutional layers of the learning model generated in the learning model generation step has on the slug determination output;
Using the above influence coefficient and the multiple convolutional layers used to calculate the above influence coefficient, the infrared image captured by the above infrared camera is used as an input image, and the influence of each pixel of the input image on the slag determination output is calculated. an input evaluation image creation step to be evaluated;
Equipped with
From the input evaluation image created in the input evaluation image creation step, a region of the infrared image that has a high influence on the slag determination output is extracted, and the captured image is acquired in the image acquisition step so as to include the extracted region. determine the area of
4. The slag outflow determination method according to claim 3 . A converter operating method, comprising operating a converter using the slag outflow determination method according to any one of claims 1 to 4 . A molten steel manufacturing method comprising the converter operating method according to claim 5 . A slag contamination determination device that images a discharge stream discharged from a tap port of a converter toward a ladle with an infrared camera, and determines whether or not slag is mixed in the discharge stream based on the captured image. ,
an image acquisition unit that acquires an image of the discharge flow taken by the infrared camera;
Determine the presence or absence of slag contamination from the captured image acquired by the image acquisition unit using a trained model that is machine-learned using infrared images of the discharge flow and recognizes the presence or absence of slag contamination. A determination section;
has
The infrared image used in the machine learning and the captured image acquired by the image acquisition unit are images including a range from the position of the tapping port in the discharge flow to the position directly above the ladle,
comprising a trained model generation unit that generates the trained model,
The trained model generation section above is
a learning data storage unit that stores learning data in which a label indicating presence or absence of slag is added to the captured image taken by the infrared camera;
a first learning unit that performs machine learning using the learning data accumulated in the storage unit to obtain a first learning model that recognizes the presence or absence of slag contamination;
For the infrared image prepared for evaluation and labeled with the presence or absence of slag contamination, use the first learning model obtained in the first learning section to determine the presence or absence of slag contamination, and make the determination. a unique data acquisition unit that selects unique data that is an infrared image that has been incorrectly determined in the infrared image with a label indicating whether or not slag is mixed;
A second step in which additional learning is performed on the first learning model obtained by the first learning section using the singular data acquired by the singular data acquisition section, and the model after the additional learning is updated as the learned model. Equipped with 2 learning sections,
The above-mentioned unique data includes image data of a branched discharge flow and image data of a narrowed discharge flow,
A slag outflow determination device characterized by: The specific data acquired by the specific data acquisition unit is stored in the storage unit in advance,
Furthermore,
A learning data update unit that generates learning data from the captured image and the determination result determined by the determination unit, stores the generated learning data in the storage unit, and updates the learning data stored in the storage unit. and,
A learning model that recognizes the presence or absence of slag contamination is obtained by using the learning data of the storage section updated by the learning data updating section and the singular data stored in the storage section, and the obtained learning model is used in the training described above. a relearning unit that updates the model as a completed model;
The slag outflow determination device according to claim 7 , comprising: The slag outflow determination device according to claim 7 or 8 , wherein the machine learning is performed using a convolutional neural network (CNN). Converter operating equipment comprising the slag outflow determination device according to any one of claims 7 to 9 . Molten steel manufacturing equipment, comprising the slag outflow determination device according to any one of claims 7 to 9 .

Images