JP2023050291A - Learning model generation method, program and air ratio estimation device - Google Patents

Abstract

To estimate an air ratio with high accuracy only by imaging a flame image in a furnace from an observation window.SOLUTION: A learning model generation method comprises: for each air ratio, acquiring a plurality of flame image data with different combinations of a flame imaging angle and a distance to a flame as teaching data, where the flame image data is obtained by imaging a flame in a furnace comprising a burner for burning premixed gas in which fuel gas and air are mixed in advance through an observation window from the outside of the furnace; and using the teaching data, generating a learning model that inputs the flame image data imaged through the observation window and outputs an air ratio.SELECTED DRAWING: Figure 1

Description

The present invention relates to a learning model generation method, program, and air ratio estimation device.

A heating furnace used in a factory or the like uses a burner in which a plurality of flames are arranged. This type of burner is also called a line burner or a ribbon burner due to the characteristic that flames are arranged in a line.
At present, the state of the flame in the heating furnace is manually adjusted by manually adjusting the air ratio of the gas in which fuel gas and air are premixed (hereinafter referred to as "premixed gas"). Specifically, the worker observes the flame inside the furnace through an observation window provided in the heating furnace, and adjusts the air ratio based on his/her sense and know-how.

Japanese Patent Application Laid-Open No. 2021-83558

However, there are individual differences in the determination of the air ratio by the operator, and there is a limit to the accuracy of determination. Skill transfer is also a problem. Furthermore, in a factory or the like, hundreds of line burners are sometimes used, requiring a lot of time and effort.

SUMMARY OF THE INVENTION An object of the present invention is to make it possible to estimate an air ratio with high accuracy simply by taking a flame image in a heating furnace through an observation window.

The invention according to claim 1 is flame image data obtained by imaging a flame in a furnace equipped with a burner that burns a premixed gas in which fuel gas and air are premixed through an observation window from outside the furnace. Regarding the air ratio, a plurality of flame image data with different combinations of flame imaging angles and distances to the flame are acquired as teacher data, and the flame image data captured through the observation window is input using the teacher data, A learning model generation method for generating a learning model having an air ratio as an output.
The invention according to claim 2 is the learning model generation method according to claim 1, wherein the teacher data includes a plurality of the flame image data having different photographing magnifications.
In the invention according to claim 3, a flame in a furnace equipped with a burner that burns a premixed gas in which fuel gas and air are premixed is imaged from outside the furnace through an observation window to obtain flame image data, and each With respect to the air ratio of , the obtained flame image data is input to a learning model that has been trained using a plurality of flame image data with different combinations of the flame imaging angle and the distance to the flame as teacher data, and the corresponding It is a program that causes a computer to execute processing for outputting an air ratio.
The invention according to claim 4 is the program according to claim 3, wherein the flame image data is hand-held by an on-site worker.
The invention according to claim 5 is an acquisition unit that acquires flame image data obtained by imaging a flame in a furnace equipped with a burner that burns a premixed gas in which fuel gas and air are premixed through an observation window from outside the furnace. Then, for each air ratio, the obtained flame image data is input to a learning model that has been trained using a plurality of flame image data with different combinations of the flame imaging angle and the distance to the flame as teacher data. , and an output unit for outputting the corresponding air ratio.

According to the first aspect of the invention, the air ratio can be estimated with high accuracy simply by taking a flame image in the heating furnace through the observation window.
According to the second aspect of the invention, the accuracy of estimating the air ratio can be further improved.
According to the third aspect of the invention, the air ratio can be estimated with high accuracy simply by taking a flame image in the heating furnace through the observation window.
According to the fourth aspect of the invention, it is possible to perform imaging at various sites.
According to the fifth aspect of the invention, the air ratio can be estimated with high accuracy simply by taking a flame image in the heating furnace through the observation window.

FIG. 4 is a diagram illustrating an example of a learning system used to generate a model that learns the relationship between a flame image and an air ratio in a furnace; FIG. 4 is a diagram for explaining the positional relationship between a camera and flames when acquiring flame image data; (A) shows the shooting distance L and the angle θ at which the flame is imaged, and (B) shows a specific example of the angle θ. It is a figure explaining an example of the flame image given to a machine-learning apparatus. (A) to (J) are examples of flame images corresponding to each air ratio. It is a figure explaining an example of the hardware constitutions of a machine learning device. It is a figure explaining an example of a deep learning model. FIG. 3 is a diagram illustrating a configuration example of an estimation system for estimating an air ratio in a furnace from a flame image captured through an observation window of the heating furnace; It is a figure explaining an example of the hardware constitutions of an air ratio estimation apparatus. FIG. 4 is a diagram illustrating a specific example of teacher data used to generate a learning model; FIG. FIG. 5 is a diagram for explaining the accuracy of an air ratio estimation result using a learning model trained using captured flame images; (A) shows the transition of the correct answer rate of the air ratio obtained by giving the flame image used as teacher data to the learning model, and (B) is obtained by giving the flame image used as teacher data to the learning model. (C) shows the transition of the correct answer rate of the air ratio obtained by giving the flame image as test data to the learning model, and (D) shows the transition of the flame as test data. The transition of the error of the air ratio obtained by giving the image to the learning model is shown. It is a figure explaining an example of the hardware constitutions of the air-ratio estimation apparatus integrated with a camera. FIG. 10 is a diagram illustrating another configuration example of an estimation system for estimating the air ratio in the furnace from a flame image captured through the observation window of the heating furnace;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings.
<Embodiment>
<Learning system configuration>
FIG. 1 is a diagram illustrating an example of a learning system 1A used to generate a model (hereinafter referred to as a "learning model") that has learned the relationship between a flame image and the air ratio in the furnace.
A learning model generated by the learning system 1A shown in FIG. 1 is used for estimating the air ratio in the heating furnace 10 used in the factory.

The structure of the heating furnace 10 shown in FIG. 1 is simplified for explanation, and does not necessarily match the actual structure of the heating furnace 10 . Moreover, the structure of the heating furnace 10 shown in FIG. 1 expresses the heating furnace 10 transparently.
A ribbon burner 20, a sensor 30, and a conveyor 40 are arranged in the heating furnace 10 shown in FIG.

The ribbon burner 20 has a structure in which one or more nozzle rows are formed in a hollow tube (that is, pipe). For example, a plurality of nozzles are formed in one row in one nozzle row. With this structure, the ribbon burner 20 forms an array of flames arranged along the axial direction of the hollow tube.
The ribbon burner 20 is supplied with a premixed gas in which fuel gas and air are premixed. The flow rates of the fuel gas and air forming the premixed gas are adjusted by the opening degrees of the fuel gas valve 20A and the air valve 20B. A worker manually adjusts the valve.

In the case of this embodiment, for example, LPG (=Liquefied Petroleum Gas), city gas, or low-calorie gas is used as the fuel gas.
LPG is, for example, liquefied petroleum gas whose main components are propane and butane. City gas is, for example, natural gas whose main component is methane. A low-calorie gas is, for example, a gas containing hydrogen and carbon monoxide as main components.

By the way, the ratio of the fuel gas to the air (that is, the air ratio) in the heating furnace 10 does not necessarily match the mixing ratio of the fuel gas to the air introduced as the premixed gas. The causes include, for example, aging deterioration of the valves 20A and 20B, clogging of nozzles, and inflow of air from gaps and openings existing in the furnace wall 10A.

Therefore, a sensor 30 is arranged in the heating furnace 10 used in the learning system 1A to accurately measure the air ratio in the furnace.
In the case of this embodiment, the sensor 30 uses, for example, an oxygen concentration meter or a gas chromatograph.
Since the sensor 30 is used for linking the flame image and the air ratio, it may be removed from the heating furnace 10 after acquiring the flame image for learning.

A conveyor 40 shown in FIG. 1 conveys an object to be heated, which is to be baked, in the direction of the arrow. In the case of FIG. 1, the object to be heated moves horizontally above the tip of the flame. However, FIG. 1 is only an example, and the object to be heated may move below the flame.
The conveyor 40 shown in FIG. 1 conveys the object to be heated in the direction of the arrow, but the object to be heated may be conveyed in the direction normal to the paper surface. When conveying the object to be heated in the normal direction of the paper, a plurality of ribbon burners 20 shown in FIG. 1 are arranged in the normal direction of the paper.

A part of the furnace wall 10A of the heating furnace 10 is provided with a window (hereinafter referred to as an "observation window") 10B for observing the flame inside the furnace.
At least one observation window 10B is provided corresponding to each ribbon burner 20 . The observation window 10B is made of heat-resistant glass, for example.
The camera 50 is an imaging device used for imaging flames in the furnace from the observation window 10B. In the case of this embodiment, the camera 50 uses a single focus lens.

By using a single-focus lens, the distance L from the camera 50 to the focal position (hereinafter referred to as "shooting distance") is substantially constant depending on the optical characteristics of the single-focus lens used. It should be noted that, within the range of the depth of field, it is possible to pick up a focused flame image even if the photographing distance L is different. When capturing flame images with different shooting distances L, another single focus lens with a different focusing distance is exchanged. In this embodiment, a plurality of single focal lenses with different shooting distances L are used to capture flame images.

A color camera capable of capturing a color image is used as the camera 50 used in this embodiment. However, the camera 50 is required to have a function of outputting flame image data (hereinafter referred to as "flame image data") without color correction. This is because even if the color-corrected flame image is learned, the versatility of the learning model 60A is lost.
Therefore, if the camera 50 has a color correction function, the flame image is taken with the color correction function turned off. In the case of this embodiment, an image including the entire flame is called a "flame image".

The flame image may be taken while the camera 50 is attached to a mounting base (not shown) or the like, or may be taken while the operator holds the camera 50 by hand.
In the case of this embodiment, the positional relationship between the flame and the camera 50 is recorded when capturing the flame image.
The positional relationship here includes the object distance L and the angle θ formed by the optical axis of the lens and the imaged flame (hereinafter referred to as “angle”).

FIG. 2 is a diagram for explaining the positional relationship between the camera 50 and the flame when acquiring the flame image data. (A) shows the shooting distance L and the angle θ at which the flame is imaged, and (B) shows a specific example of the angle θ.
As described above, the photographing distance L is determined according to the optical characteristics of the single focus lens used to capture the flame image.
In the present embodiment, the shooting distance L is recorded in association with the flame image, but if the flame image is captured using a plurality of single focus lenses for one air ratio, recording the shooting distance L is not necessarily required. do not have.

Further, in the present embodiment, flame images are captured at a plurality of angles θ for one air ratio, and the angles θ at the time of capturing are recorded. However, as with the photographing distance L, if flame images are captured at a plurality of angles θ for one air ratio, it is not always necessary to record the angles θ.
The circles shown in FIG. 2(B) indicate an example of the central position of the single focus lens when capturing the flame image.

In other words, each circle represents the imaging position of the camera 50 . The set of center positions used for capturing flame images may be common to multiple air ratios, but may also be different. For example, the set of center positions used to capture the flame image associated with the air ratio “0.6” and the set of center positions used to capture the flame image associated with the air ratio “1.0” are different. good too.

Note that the absolute values of the angles θ corresponding to a plurality of circles positioned on the same concentric circle are all the same.
In the present embodiment, for one angle θ, a plurality of directions, such as the position of the circle where the flame is imaged obliquely from the left, the position of the circle where the image of the flame is imaged obliquely from above, and the position of the circle where the flame is imaged obliquely from the right. Take a flame image from

Returning to the description of FIG.
In this embodiment, for one air ratio, flame images are captured at a plurality of photographing distances L×a plurality of angles θ. For example, for an air ratio of 0.6, 100 flame images with 10 shooting distances L and 10 angles θ for each shooting distance L (including cases where the magnitude of the angle is the same but the direction is different). is imaged.
Time data, for example, is used to associate the air ratio with the flame image.

The machine learning device 60 is provided with a plurality of flame images captured at a combination of a plurality of photographing distances L and a plurality of angles θ for one air ratio.
Flame images with different photographing distances L can be captured by exchanging the single focus lens attached to the camera 50 .
Specifically, for one air ratio, when flame images are captured at a plurality of angles θ with a single single-focal lens, flame images are again captured at a plurality of angles θ after replacing with another single-focus lens.
When the flame images are captured by a plurality of single-focal lenses having different shooting distances L, the air ratio is adjusted to another value and the flame images are repeatedly captured.

FIG. 3 is a diagram explaining an example of a flame image given to the machine learning device 60. As shown in FIG. (A) to (J) are examples of flame images corresponding to each air ratio.
Incidentally, the flame image shown in FIG. 3(A) is an example of an air ratio of 0.6, the flame image shown in FIG. 3(B) is an example of an air ratio of 0.7, and the flame image shown in FIG. An example of an air ratio of 0.75, the flame image shown in FIG. 3(D) is an example of an air ratio of 0.8, and the flame image shown in FIG. The flame image shown in FIG. 3F is an example with an air ratio of 0.9, the flame image shown in FIG. 3G is an example with an air ratio of 0.95, and the flame image shown in FIG. The flame image shown in FIG. 3(I) is an example of an air ratio of 1.05, and the flame image shown in FIG. 3(J) is an example of an air ratio of 1.1.

In the example shown in FIG. 3, ten air ratios are selected as learning targets, but the number of learning target air ratios is not limited to ten. For example, the number of air ratios may be less than ten or more than ten.
For one air ratio, a flame image specified by a combination of a plurality of photographing distances L and a plurality of angles θ is captured. You can take an image.

The machine learning device 60 (see FIG. 1) receives as input a plurality of flame images captured by changing the combination of the photographing distance L and the angle θ for one air ratio, and learns the relationship in which the air ratio is output. A model 60A is generated by machine learning.
In the present embodiment, a flame image and the air ratio measured when the flame image is captured are given to the machine learning device 60 as teacher data.

That is, the machine learning device 60 generates a learning model 60A that has learned the relationship between the flame image and the air ratio through supervised learning. Specifically, the machine learning device 60 learns the relationship between the flame features (color and shape) included in the flame image and the air ratio.
FIG. 4 is a diagram illustrating an example of the hardware configuration of the machine learning device 60. As shown in FIG. The machine learning device 60 is configured by, for example, a computer.

The machine learning device 60 shown in FIG. 4 includes a processor 61 that executes a program, a ROM (=Read Only Memory) 62 that stores BIOS (=Basic Input Output System), and a RAM (=Random Access Memory) that is used as a work area. 63, a hard disk drive 64, and an interface 65 used for inputting flame images and air ratios.

The processor 61 is, for example, a CPU (Central Processing Unit) or a GPU (Graphical Processing Unit), and executes a machine learning program.
The machine learning program here is a program for machine-learning the relationship between the input of the flame image and the output of the air ratio.
The machine learning program is stored in the hard disk device 64, which is an auxiliary storage device.

Based on the training data, the machine learning program adjusts the variables of each layer that make up the deep learning model, and when a flame image is given as an input, it learns so that the air ratio at the time of capturing the flame image is output. proceed.
FIG. 5 is a diagram illustrating an example of a deep learning model. FIG. 5 illustrates a convolutional neural network (CNN) as an example of a deep learning model.

A convolutional network consists of a convolutional layer for extracting features from a flame image, a pooling layer for extracting the average or maximum value of the extracted features, and a fully connected layer.
A convolutional network has a multi-layered structure in which unit structures composed of convolutional layers and pooling layers are connected in multiple stages, and a fully connected layer is placed in the final stage thereof.
In the example of FIG. 5, any value of the air ratio between 0 and 1.5 is output as the output of the fully connected layer.

<Estimation system>
FIG. 6 is a diagram illustrating a configuration example of an estimation system 1B for estimating the air ratio in the furnace from the flame image captured through the observation window 10B of the heating furnace 10. As shown in FIG.
In FIG. 6, parts corresponding to those in FIG. 1 are shown with reference numerals. In FIG. 6, the heating furnace 10 is also included in the estimation system 1B, but the estimation system 1B in a narrow sense is composed only of the air ratio estimation device 70. As shown in FIG.

In the estimation system 1B, an operator holds the camera 50A in hand and picks up a flame image (that is, a flame image) observed through the observation window 10B.
However, the optical characteristics of the camera 50A used to capture the flame image must be the same as the optical characteristics of the camera 50 (see FIG. 1) used to capture the flame image used to learn the learning model 60A. be done.

Specifically, the photographing distance L of the camera 50A is within the range of the photographing distance L used during learning, and the optical characteristics of the lens of the camera 50A are the same as those of the lens of the camera 50 (see FIG. 1) used during learning. It is required to be within the properties. The camera 50A that satisfies these optical characteristics is specified by, for example, the operator that generated the learning model 60A.

Also, when capturing a flame image, it is required to use a camera 50A that does not have a function for color correction of the captured flame image, or to turn off the color correction function that the camera 50A has.
For example, the camera 50A is installed with a dedicated application program (hereinafter referred to as "dedicated application") provided by the operator that generated the learning model 60A. When the dedicated application is installed, the worker can concentrate on capturing the flame image without worrying about the color correction settings and the like by activating the dedicated application and capturing the flame image.

When the dedicated application is activated and a flame image is captured, a flame image processed under the same conditions as when learning the learning model 60A is output.
However, if the imaging conditions are satisfied, the use of the dedicated application is not essential. That is, if the worker can capture a flame image using the camera 50A specified by the operator or the like who generated the learning model 60A while satisfying the imaging conditions such as the processing mode specified by the same operator, the dedicated application The use of is not mandatory.

In the example of FIG. 6, a flame image captured at an arbitrary shooting distance L and an arbitrary angle θ is supplied from the camera 50A to the air ratio estimating device 70. In the example of FIG. Incidentally, the photographing distance L and the angle θ may be within the range used when learning the learning model 60A.
The air ratio estimation device 70 inputs the flame image captured by the camera 50A to the learning model 60A, and outputs the air ratio output from the learning model 60A as an estimated value to the smartphone 80 used by the worker.

In FIG. 6, the smartphone 80 is exemplified as the notification destination of the estimated air ratio, but the information terminal serving as the notification destination of the air ratio is not limited to the smartphone. For example, it may be a desktop computer, a notebook computer, a tablet computer, or a wearable computer such as smart glasses or a smart watch.

In the case of this embodiment, the operator who has confirmed the notification of the air ratio manually operates either or both of the valves 20A and 20B to adjust the air ratio in the heating furnace 10 to match the target value. . That the air ratio in the heating furnace 10 matches the target value is confirmed by the value of the air ratio estimated from the flame image taken after the adjustment.

In the case of this embodiment, the air ratio estimation device 70 is implemented as, for example, a cloud server. In this operation mode, the flame image captured by the camera 50A is uploaded to the air ratio estimating device 70 via the camera 50A, a computer operated by the operator, or the like.
In this case, the air ratio estimating device 70 is operated by the operator or the like that generated the learning model 60A, and the air ratio estimated from the flame image is provided from the air ratio estimating device 70 to the smartphone 80 as a service.

FIG. 7 is a diagram illustrating an example of the hardware configuration of the air ratio estimation device 70. As shown in FIG. The air ratio estimating device 70 is configured by, for example, a computer.
The air ratio estimating device 70 shown in FIG. 7 includes a processor 71 for executing programs, a ROM 72 for storing BIOS, a RAM 73 for use as a work area, a hard disk device 74, and an interface for inputting and outputting flame images and air ratios. It consists of 75.

The processor 71 is, for example, a CPU or GPU, and uses the learning model 60A to estimate the current air ratio inside the heating furnace 10 from the flame image captured by the camera 50A. Therefore, the processor 71 functions as a flame image acquisition section 71A that acquires flame image data and an air ratio output section 71B that outputs the estimated air ratio. Here, the flame image acquisition section 71A is an example of an acquisition section, and the air ratio output section 71B is an example of an output section.

When the flame image data captured by the camera 50A is acquired, the flame image acquisition unit 71A inputs the acquired flame image data to the learning model 60A.
Air ratio output unit 71B transmits the air ratio output from learning model 60A to smartphone 80 .
In the case of FIG. 7, the learning model 60A is stored in a hard disk device 74, which is an auxiliary storage device.

In this embodiment, interface 75 is a communication interface.
The interface 75 receives the flame image from the heating furnace 10 side and transmits the estimated air ratio to the smartphone 80 .
When receiving the semiconductor memory storing the flame image, the flame image stored in the semiconductor memory is read out through the interface 75 . Also, when storing the air ratio in the semiconductor memory, writing of the air ratio is executed through the interface 75 .

<Specific example>
FIG. 8 is a diagram illustrating a specific example of teacher data used to generate a learning model.
The flame images used as training data are obtained by setting the air ratio in the heating furnace 10 (see FIG. 1) to "0.6", "0.7", "0.75", "0.8", "0. 85", "0.9", "0.95", "1.0", "1.05", and "1.1".
Incidentally, the air ratio of "0.6" was realized by supplying the premixed gas amount to the heating furnace 10 at 10-21 [NL/min]. Other air ratios are as shown in the chart of FIG.

In the case of FIG. 8, 522 flame images with different combinations of the photographing distance L and the angle θ were captured for the air ratio of "0.6". The air ratios "0.7" to "1.0" are as shown in the chart of FIG. In the case of FIG. 8, a total of 3220 flame images are captured.
FIG. 9 is a diagram for explaining the accuracy of the result of estimating the air ratio using the learning model 60A (see FIG. 6) trained using captured flame images.

(A) shows the transition of the correct answer rate of the air ratio obtained by giving the flame image used as teacher data to the learning model 60A, and (B) gives the flame image used as teacher data to the learning model 60A. (C) shows the transition of the correct answer rate of the air ratio obtained by giving the flame image as test data to the learning model 60A, and (D) shows the test data Fig. 6 shows the transition of the error in the air ratio obtained by giving the flame image as , to the learning model 60A.

In the case of FIG. 9, some of the 3220 flame images are used as training data. The flame images used as test data are the rest of the 3,220 flame images that were not used as teacher data.
The horizontal axis of FIGS. 9A to 9D is the number of repetitions (Epoch).
The vertical axis of FIGS. 9A and 9C is the percentage of correct answers, which is calculated by the following equation.
Correct answer rate (Accuracy) = (TP + TN) / (TP + FP + TN + FN)
The vertical axes in FIGS. 9B and 9D are errors.

TP means true positive when the estimated value is correct and the actual value is also correct.
TN means a case where the estimated value is incorrect but the actual answer is incorrect, that is, True Negative.
FP means the case where the estimated value is correct but the actual answer is incorrect, that is, false positive.
FN means the case where the estimated value is incorrect but actually correct, that is, False Negative.

FIGS. 9A to 9D show graphs of changes in results when learning by four types of deep learning models is repeated.
The four types of deep learning models are a simple convolutional neural network (Simple CNN), a 16-layer deep convolutional neural network (VGG16), a 19-layer deep convolutional neural network (VGG19), and a 50-layer deep convolutional neural network (VGG19). It is a convolutional neural network (ResNet50).

As shown in FIGS. 9A and 9B, the estimated value of the air ratio when the flame image used as the teacher data is input to the learning model 60A is obtained by learning regardless of which deep learning model is used. There was a tendency that the more the test was repeated, the higher the correct answer rate and the less the error.
On the other hand, as shown in FIGS. 9(C) and 9(D), when flame images were given as test data, in VGG16 and VGG19, the more the learning was repeated, the higher the rate of correct answers and the fewer the errors. It was found that results equivalent to the data were obtained. This means that it is possible to estimate the air ratio with high precision in practice.

On the other hand, in the case of Simple CNN and ResNet50, the correct answer rate was slightly low even after repeated learning, and the error did not decrease.
As far as the results of this time are concerned, it was found that the adoption of VGG16 or VGG19 is desirable for generating the learning model 60A.

<Summary>
Using the learning system 1A (see FIG. 1) described in the present embodiment, a learning model 60A is generated by learning flame images taken at various combinations of the shooting distance L and the angle θ for one air ratio. By doing so, it is possible to generate the learning model 60A with high accuracy in estimating the air ratio.
Therefore, an on-site worker can capture a flame image from an arbitrary position (in other words, capture a flame image at an arbitrary shooting distance L and angle θ) with a high accuracy of 0.05 increments. An air ratio estimation result can be obtained.

As a result, the efficiency of the air ratio measurement work is improved, and the air ratios of many heating furnaces 10 can be measured with high accuracy in a short period of time.
In addition, by using the air ratio estimated using the learning model 60A, the air ratio in the heating furnace 10 can be accurately determined compared to the conventional method in which a skilled worker observes and judges the state of the flame. can grasp.
Therefore, even when the air ratio of the premixed gas is adjusted by manually adjusting the valves 20A and 20B, the result of the adjustment can be confirmed only by imaging the flame image. Moreover, as a result of being able to adjust the air ratio in the heating furnace 10 with high accuracy, the energy saving of the heating furnace 10 can also be achieved.
In addition, the method of using the air ratio estimated using the learning model 60A does not depend on the skill of the operator, so the problem of skill transfer can be resolved.

<Other embodiments>
(1) Although the embodiments of the present invention have been described above, the technical scope of the present invention is not limited to the scope described in the above-described embodiments. It is clear from the scope of claims that the technical scope of the present invention includes various modifications and improvements to the above-described embodiment.

(2) For example, in the above embodiment, the air ratio estimating device 70 (see FIG. 6) is realized as a cloud server, but the air ratio estimating device 70 is the heating furnace 10 (see FIG. 6) may be implemented as a computer operated by an operator who manages the . In this case, the computer operated by the operator is installed with a program for estimating the air ratio from the flame image using the learning model 60A.

The air ratio estimating device 70 here may be implemented as a computer integrated with the camera 50A that captures the flame image. For example, it may be realized as a smartphone 80 or a tablet computer.
FIG. 10 is a diagram illustrating an example of the hardware configuration of a camera-integrated air ratio estimation device 70A. In FIG. 10, the parts corresponding to those in FIGS. 6 and 7 are indicated by the reference numerals.

In the air ratio estimating device 70A shown in FIG. 10, a memory card 74A with a large storage capacity is used in place of the hard disk device 74 (see FIG. 7). A learning model 60A is stored in the memory card 74A.
A touch panel 76 and a camera 50A are mounted on the air ratio estimation device 70A shown in FIG.
The touch panel 76 is composed of, for example, a liquid crystal display and a capacitive touch sensor.
Of course, the camera 50A is required to have optical characteristics equivalent to those of the camera 50 (see FIG. 1) used when generating the learning model 60A.

(3) In the above-described embodiment, the operator who confirmed the air ratio estimated by the air ratio estimating device 70 adjusts the valve 20A and the like, but the adjustment of the valve 20A and the like may be automated.
FIG. 11 is a diagram illustrating a configuration example of another estimation system 1C for estimating the air ratio in the furnace from the flame image captured through the observation window 10B of the heating furnace 10. As shown in FIG. In FIG. 11, parts corresponding to those in FIG. 6 are shown with reference numerals.

In the estimation system 1</b>C shown in FIG. 11 , the air ratio estimated by the air ratio estimation device 70 is output to the flow rate adjustment device 90 . The flow rate adjusting device 90 adjusts the opening of both or one of the valves 20A and 20B so that the estimated air ratio matches the target value. In this case, the operator can match the air ratio in the heating furnace 10 to the target value simply by imaging the flame in the heating furnace 10 through the observation window 10B.

(4) In the above-described embodiment, the heating furnace 10 using the ribbon burner 20 (see FIG. 1) was exemplified, but the heating furnace 10 using a pian line burner that forms a wide curtain-like flame or a heat-resistant metal It can also be used in the heating furnace 10 using a metal knit burner in which fibers are woven into a knit. The ribbon burner 20, the pian line burner, and the metal knit burner are examples, and the heating furnace 10 using other types of burners may be used.

(5) In the above-described embodiment, the camera 50 ( 1) and the camera 50A (see FIG. 6), a zoom lens with a variable photographic distance L may be used.
When a zoom lens is used, it is possible to change the imaging magnification, and the flame in the heating furnace 10 can be imaged from a position away from the observation window 10B. In this case, teacher data used for learning the learning model 60A can include a plurality of flame images with different shooting magnifications.

(6) In the above-described embodiment, the case where the learning system 1A (see FIG. 1) and the estimation system 1B (see FIG. 6) are used in a factory has been described. Can be used in facilities.

DESCRIPTION OF SYMBOLS 1A... Learning system, 1B, 1C... Estimation system, 10... Heating furnace, 10A... Furnace wall, 10B... Observation window, 20... Ribbon burner, 20A, 20B... Valve, 30... Sensor, 40... Conveyor, 50, 50A... Camera 60 Machine learning device 60A Learning model 70, 70A Air ratio estimating device 80 Smartphone 90 Flow rate adjusting device

A heating furnace used in a factory or the like uses a burner in which a plurality of flames are arranged. This type of burner is also called a line burner due to the characteristic that the flames are arranged in a line.
At present, the state of the flame in the heating furnace is manually adjusted by manually adjusting the air ratio of the gas in which fuel gas and air are premixed (hereinafter referred to as "premixed gas"). Specifically, the worker observes the flame inside the furnace through an observation window provided in the heating furnace, and adjusts the air ratio based on his/her sense and know-how.

The invention according to claim 1 is flame image data obtained by imaging a flame in a furnace equipped with a nozzle array for burning premixed gas, which is a premixed gas of fuel gas and air, from a point outside the furnace through an observation window, , for each air ratio, a plurality of flame image data having different combinations of the flame imaging angle and the distance to the flame are acquired as teacher data, and the flame image data captured through the observation window using the teacher data. is an input and an air ratio is an output.
The invention according to claim 2 is the learning model generation method according to claim 1, wherein the teacher data includes a plurality of the flame image data having different photographing magnifications.
In the invention according to claim 3, the flame in the furnace equipped with a plurality of rows of burners that burns a premixed gas in which fuel gas and air are premixed is captured by an operator through an observation window from a point outside the furnace. The acquired flame image data is acquired, and for each air ratio, a plurality of flame image data with different combinations of the angle at which the flame is imaged and the distance to the flame are used as teacher data to train the learning model. It is a program that causes a computer to execute a process of inputting flame image data and outputting the corresponding air ratio.
In the invention according to claim 4 , the flame in the furnace equipped with a plurality of rows of burners that burns premixed gas in which fuel gas and air are premixed is captured by an operator each time through an observation window from a point outside the furnace. A learning model that is trained using a plurality of flame image data with different combinations of flame imaging angle and distance to the flame for each air ratio as teacher data, and an output unit for inputting the acquired flame image data and outputting the corresponding air ratio.

According to the first aspect of the invention, the air ratio can be estimated with high accuracy simply by taking a flame image in the heating furnace through the observation window.
According to the second aspect of the invention, the accuracy of estimating the air ratio can be further improved.
According to the third aspect of the invention, the air ratio can be estimated with high accuracy simply by taking a flame image in the heating furnace through the observation window.
According to the fourth aspect of the present invention , the air ratio can be estimated with high accuracy simply by capturing the image of the flame inside the heating furnace through the observation window.

The structure of the heating furnace 10 shown in FIG. 1 is simplified for explanation, and does not necessarily match the actual structure of the heating furnace 10 . Moreover, the structure of the heating furnace 10 shown in FIG. 1 expresses the heating furnace 10 transparently.
A line burner 20, a sensor 30, and a conveyor 40 are arranged in the heating furnace 10 shown in FIG.

The line burner 20 has a structure in which one or more nozzle rows are formed in a hollow tube (that is, pipe). For example, a plurality of nozzles are formed in one row in one nozzle row. With this structure, the line burner 20 forms an array of flames arranged along the axial direction of the hollow tube.
A premixed gas in which fuel gas and air are premixed is introduced into the line burner 20 . The flow rates of the fuel gas and air forming the premixed gas are adjusted by the opening degrees of the fuel gas valve 20A and the air valve 20B. A worker manually adjusts the valve.

A conveyor 40 shown in FIG. 1 conveys an object to be heated, which is to be baked, in the direction of the arrow. In the case of FIG. 1, the object to be heated moves horizontally above the tip of the flame. However, FIG. 1 is only an example, and the object to be heated may move below the flame.
The conveyor 40 shown in FIG. 1 conveys the object to be heated in the direction of the arrow, but the object to be heated may be conveyed in the direction normal to the paper surface. When the object to be heated is conveyed in the normal direction of the paper, a plurality of line burners 20 shown in FIG. 1 are arranged in the normal direction of the paper.

A part of the furnace wall 10A of the heating furnace 10 is provided with a window (hereinafter referred to as an "observation window") 10B for observing the flame inside the furnace.
At least one observation window 10B is provided corresponding to the line burner 20 . The observation window 10B is made of heat-resistant glass, for example.
The camera 50 is an imaging device used for imaging flames in the furnace from the observation window 10B. In the case of this embodiment, the camera 50 uses a single focus lens.

(4) In the above-described embodiment, the heating furnace 10 using the line burner 20 (see FIG. 1) was exemplified, but the heating furnace 10 using a pian line burner that forms a wide curtain-like flame or a heat-resistant metal It can also be used in the heating furnace 10 using a metal knit burner in which fibers are woven into a knit. The line burner 20, the pian line burner, and the metal knit burner are examples, and the heating furnace 10 using other types of burners may be used.

DESCRIPTION OF SYMBOLS 1A... Learning system, 1B, 1C... Estimation system, 10... Heating furnace, 10A... Furnace wall, 10B... Observation window, 20... Line burner, 20A, 20B... Valve, 30... Sensor, 40... Conveyor, 50, 50A... Camera 60 Machine learning device 60A Learning model 70, 70A Air ratio estimating device 80 Smartphone 90 Flow rate adjusting device

Claims (5)

Flame image data obtained by imaging a flame in a furnace equipped with a burner that burns a premixed gas in which fuel gas and air are premixed through an observation window from the outside of the furnace, and the angle at which the flame is imaged for each air ratio. and flame image data with different combinations of distances to the flame are acquired as training data,
Using the teacher data, generating a learning model that inputs the flame image data captured through the observation window and outputs the air ratio;
Learning model generation method. The training data includes a plurality of flame image data with different imaging magnifications;
The learning model generation method according to claim 1. Acquiring flame image data of a flame in a furnace equipped with a burner that burns a premixed gas in which fuel gas and air are mixed in advance through an observation window from outside the furnace,
For each air ratio, the obtained flame image data is input to a learning model that has been trained using a plurality of flame image data with different combinations of the flame imaging angle and the distance to the flame as teacher data, and correspondence is performed. A program that causes a computer to execute the process of outputting the air ratio. The flame image data is hand-held by an on-site worker.
4. A program according to claim 3. an acquisition unit that acquires flame image data of a flame in a furnace equipped with a burner that burns a premixed gas obtained by mixing fuel gas and air in advance through an observation window from outside the furnace;
For each air ratio, the obtained flame image data is input to a learning model that has been trained using a plurality of flame image data with different combinations of the flame imaging angle and the distance to the flame as teacher data, and correspondence is performed. an output unit that outputs an air ratio that
air ratio estimator.

Images