Furnace state variable estimator, estimator model generation device, and program and method therefor

Abstract

Furnace state quantity estimation device (1), comprising: a feature quantity extraction unit (21) which extracts feature quantities (F) from a captured image (V) of a furnace (7); a feature quantity selecting unit (22) which selects one or more of the feature quantities (F) from the extracted feature quantities (F); a learning data generation unit (23) which generates learning data (D) by associating the selected feature quantities (F) with a state quantity (S) of the furnace (7) corresponding to a combustion state displayed in the image (V); an estimation model generation unit (2) which generates an estimation model (M) for estimating the combustion state from the image (V) of the furnace (7) using the learning data (D); and an estimated state quantity calculation unit (5) which acquires the captured image (V) of the furnace (7) and then estimates the state quantity (S) corresponding to the combustion state displayed in the acquired image (V) using the estimation model (M), wherein the feature variables (F) are selected based on a contribution (C) to the state variable (S) to represent an estimation target.

Description

Technical field of the invention

[0001] The present invention relates to a technique for monitoring a combustion state based on image information obtained by photographing a furnace of a combustion furnace. Background of the invention

[0002] For example, Patent Documents 1 and 2 each disclose a method for determining a combustion state based on image information of a combustion furnace. Specifically, Patent Document 1 discloses a method for detecting an abnormal combustion state based on a surface temperature distribution of a burner flame obtained from a combustion image (image information). Further, Patent Document 2 discloses that, in the combustion furnace, an image taken with a television camera (image pickup device) installed in an immediately upper part of a primary combustion region is subjected to image processing, and a concentration fluctuation of carbon monoxide in a secondary combustion region is predicted by fuzzy inference means based on a brightness change amount of an image obtained by the image processing.In the primary combustion zone, fuel is burned with a flame. A method for monitoring the combustion state by visually observing the flame shape from an image captured by a TV camera is also proposed (see patent document 1).

[0003] Patent Document 3 discloses a control device that learns a method for generating a model input to achieve a model output target value using a model used to simulate the characteristics of a controlled system and generates an operation signal according to the learning results. The measurement signal of the controlled system determines the model output target value for achieving the operation target value.The control device includes a model for predicting the value of a measurement signal generated when an operation signal is input to a controlled system, a function for learning a method for generating a model input so that the model output can reach a target value, a function for determining an operation signal input to the controlled system according to the results of the learning, a database for storing the limit value of a preset measurement signal, an external input interface for acquiring the measurement signal of the controlled system, a measurement signal database for storing the value of the acquired signal, and a function for determining the initial value of the model output target value using the result of calculating at least one of the average value, the maximum value, and the minimum value of the stored measurement signal and the limit value of the measurement signal.

[0004] Patent Document 4 discloses that in a combustion device such as a waste incinerator, the nature of the waste varies depending on the season, and the combustion state judgment to be performed by an initial neural network may not be consistent with the actual combustion state. Therefore, neural network learning is performed while the combustion device is operating by replacing part of the teacher data of the initial neural network and the operation data of the combustion device with the teacher data prepared through online integration by an online neural network generator. That is, online neural network learning is performed, and a new neural network is generated. Patent literature Patent specification 1: JP H4-143515 A Patent Document 2: JP 2001-4116 A Patent Document 3: CN 101379447 A Patent Document 4: JP H10-38243 A Summary of the inventionTechnical problem

[0005] Since Patent Document 1 obtains the surface temperature distribution of a flame and thus requires that a flame be displayed in the image information, it is difficult to apply the technique disclosed in Patent Document 1 to a situation where a flame is not displayed in the image information due to influences of an exhaust gas during combustion, an installation position of a furnace camera, and the like. On the other hand, since Patent Document 2 makes an inference based on the brightness change amount of the image obtained by image processing, a flame may not be displayed in the image information, restrictions on the type of combustion furnace, an arrangement position of a camera, and the like are less, and thus the method disclosed in Patent Document 2 is considered more suitable.However, a rule for performing control must be generated by fuzzy inference, such as collecting and analyzing determination methods and the like, based on the experience of the person skilled in the art. First of all, since the combustion state changes due to various factors such as the type of combustion targets (fuels) (waste, coal, etc.), combustion environments (air amount, furnace temperature, etc.), and burner shape, it is difficult to appropriately determine the combustion state in a situation that the person skilled in the art has never experienced, and it is impossible to deny the possible existence of a determination criterion that has not been recognized even by the person skilled in the art.

[0006] A method for determining the combustion state by measuring, for example, state variables such as unburned combustible content in ash, NOx concentration, CO concentration, and the like during operation of the incinerator is also considered. However, in a case where time is required to measure the state variables (such as unburned combustible content in ash), it is difficult to use the result of combustion state determination based on a measurement result for operation control in real time.

[0007] In view of the foregoing, an object of at least one embodiment of the present invention is to provide a furnace state quantity estimation apparatus that quickly and accurately estimates a state quantity that can determine a combustion state in a furnace based on furnace image information. Solution to the problem

[0008] (1) A furnace state quantity estimation device according to at least one embodiment of the present invention includes a feature quantity extraction unit that extracts feature quantities from a captured image of a furnace; a feature quantity selection unit that selects one or more of the feature quantities from the extracted feature quantities; a learning data generation unit that generates learning data by associating the selected feature quantities with a state quantity of the furnace corresponding to a combustion state indicated in the image; an estimation model generation unit that generates an estimation model for estimating the combustion state from the image of the furnace using the learning data; and an estimated state quantity calculation unit that acquires the captured image of the furnace and then estimates the state quantity corresponding to the combustion state indicated in the acquired image using the estimation model.The feature variables are selected based on a contribution to the state variable to represent an estimation objective.

[0009] With the above configuration (1), although the state quantity such as an unburned combustible content in ash, a NOx concentration, a CO concentration, or the like changes according to the combustion state of the furnace, the state quantity (estimated state quantity) generated in the combustion state can be estimated from image information obtained by capturing in the furnace during combustion using the estimation model generated by machine learning the learning data, and the image information in the furnace during combustion is associated with the state quantity (a measured value, an estimated value, or the like) when the image information is captured. Thus, the state quantity can be quickly and accurately estimated from the furnace image information.

[0010] (2) In some embodiments according to the above configuration (1), the estimation model generation unit generates the estimation model by performing machine learning on the learning data.

[0011] With the above configuration (2), the learning data can be generated by associating the image information in the furnace during combustion with the state quantity when the image information is acquired, and the estimation model can be generated by performing machine learning from the generated learning data.

[0012] (3) In some embodiments according to the above configuration (1) or (2), the contribution represents an indicator indicating an amount of correlation to the state quantity.

[0013] With the above configuration (3), the contribution represents the indicator showing the amount of correlation to the state quantity, and a difference in the image can be highlighted by selecting the feature quantities based on the contribution.

[0014] (4) In some embodiments according to the above configuration (1) or (3), the feature quantity selecting unit calculates a regression equation for calculating the state quantity from each of the feature quantities, and sets as the contribution a contribution ratio of each of the feature quantities in the calculated regression equation.

[0015] With the above configuration (4), the contribution can be adjusted based on the contribution ratio of each of the feature quantities in the regression equation.

[0016] (5) In some embodiments according to any one of the above configurations (1) to (4), the feature size extraction unit extracts the feature sizes based on brightness information obtained from the image in the past.

[0017] With the above configuration (5), the brightness information represents information obtained from the image without detecting a flame itself. The image information can be obtained even in a case where the flame cannot be captured due to exhaust gas, because an image pickup device is installed in the upper part of the furnace. Thus, the image pickup device (furnace camera) for capturing the furnace can be easily installed.

[0018] (6) In some embodiments according to any one of the above configurations (1) to (5), the state quantity is a state quantity regarding a waste material or an exhaust gas generated in the furnace during combustion; the learning data generation unit generates the learning data by associating a measurement value with the past image information, the measurement value being measured at a measurement point of the state quantity after a lapse of a predetermined time from a time point when the image constituting the basis of the past image information is captured.

[0019] With the above configuration (6), the learning data is generated taking into account a time delay until the waste or exhaust gas generated in the combustion state indicated by the past image information reaches the state variable measurement point. Such a time delay may vary considerably depending on the type of combustion furnace or the position of the measurement point. Thus, an estimation model with high estimation accuracy can be generated by generating the learning data taking into account the time delay.

[0020] (7) In some embodiments according to the above configuration (1) or (6), the estimation model generation unit generates a plurality of estimation models using a plurality of machine learning methods.

[0021] With the above configuration (7), the state quantity can be estimated by each of the plurality of estimation models each generated based on the plurality of machine learning methods (algorithms). For example, the estimation model with high estimation accuracy can be selected from the plurality of estimation models by comparing an estimation result by each of the plurality of machine learning methods with the state quantity (such as the measured value) that configures the learning data, and the estimation model suitable for, for example, conditions (such as an image size and the number of learning data) of the learning data or the like and the type of a combustion furnace can be selected.

[0022] (8) In some embodiments according to any one of the above configurations (1) to (7), the furnace state quantity estimation device further includes a relearning determining unit that determines regeneration of the estimation model by relearning if a differential between the estimated state quantity and a measured value of the state quantity exceeds a predetermined threshold.

[0023] With the above configuration (8), it is determined that regeneration of the estimation model through retraining is required if a decrease in estimation accuracy is found. If the estimation model is regenerated through retraining and the new estimation model is reacquired according to the determination, estimation can be continued with an appropriate estimation accuracy. Thus, the state variable can be accurately estimated from the input image information while following, for example, a change in the operating environment of the combustion furnace.

[0024] (9) An estimation model generating device according to at least one embodiment of the present invention includes an estimation model generating unit that performs machine learning for learning data, wherein past image information is associated with a state quantity, and generates an estimation model for estimating an estimated state quantity from input image information obtained by capturing a furnace, wherein the past image information represents a feature quantity obtained based on a captured image of the furnace and obtained based on the image in the past, wherein the state quantity corresponds to a combustion state indicated in the image in the past.

[0025] With the above configuration (9), similar to the above configuration (2), the learning data can be generated by associating the image information in the furnace during combustion with the state variable when the image information is acquired, and the estimation model can be generated by performing machine learning on the generated learning data. With the estimation model, the state variable can be quickly estimated from the image information (input image information) in the furnace during combustion.

[0026] In addition, with the above configuration (9), since the state quantity is quickly estimated from the image information (input image information) in the furnace during combustion, an optimal combustion state can be maintained by real-time operation control in which, if the state quantities such as the unburned combustible content in ash, the NOx concentration, the CO concentration, and the like increase; a user is notified of the increase by an alarm, or an operation to decrease these state quantities is automatically performed.

[0027] (10) A furnace state quantity estimation program according to at least one embodiment of the present invention is a furnace state quantity estimation program for causing a computer to perform a feature quantity extraction step for extracting feature quantities from a captured image of a furnace, a feature quantity selection step for selecting one or more of the feature quantities from the extracted feature quantities, a learning data generation step for generating learning data by associating the selected feature quantities with a state quantity of the furnace corresponding to a combustion state displayed in the image, an estimation model generation step for generating an estimation model for estimating the combustion state from the image of the furnace using the learning data, and an estimated state quantity calculation step for acquiring the captured image of the furnace,and then estimating the state variable corresponding to the combustion state displayed in the acquired image using the estimation model.

[0028] The feature variables are selected based on a contribution to the state variable to represent an estimation objective.

[0029] With the above configuration (10), the same effect as with the above configuration (1) can be achieved.

[0030] (11) In some embodiments according to the above configuration (10), the estimation model generation step generates the estimation model by performing machine learning on the learning data.

[0031] With the above configuration (11), the same effect as with the above configuration (2) can be achieved.

[0032] (12) A furnace state quantity estimation method according to at least one embodiment of the present invention includes a feature quantity extraction step for extracting feature quantities from a captured image of a furnace; a feature quantity selection step for selecting one or more of the feature quantities from the extracted feature quantities; a learning data generation step for generating learning data by associating the selected feature quantities with a state quantity of the furnace corresponding to a combustion state displayed in the image; an estimation model generation step for generating an estimation model for estimating the combustion state from the image of the furnace using the learning data;and an estimated state quantity calculation step of acquiring the captured image of the furnace, and then estimating the state quantity corresponding to the combustion state displayed in the acquired image using the estimation model.;

[0033] The feature variables are selected based on a contribution to the state variable to represent an estimation objective.

[0034] With the above configuration (12), the same effect as with the above configuration (1) can be achieved.

[0035] (13) In some embodiments according to the above configuration (12), the estimation model generation step generates the estimation model by performing machine learning on the learning data.

[0036] With the above configuration (13), the same effect as with the above configuration (2) can be achieved. Beneficial effects

[0037] According to at least one embodiment of the present invention, a furnace state quantity estimation apparatus is provided which quickly and accurately estimates a state quantity capable of determining a combustion state in a furnace based on furnace image information. Short description of the drawings Fig. 1 is a schematic configuration diagram of a combustion furnace with an installed furnace state quantity estimating device according to an embodiment of the present invention. Fig. 2 is a functional block diagram of the furnace state quantity estimation apparatus according to an embodiment of the present invention. Fig. 3 is a flowchart showing a furnace state quantity estimation method according to an embodiment of the present invention. Fig. 4 is a flowchart showing an estimation model generation step (S1) according to an embodiment of the present invention. Fig. 5 is a flowchart showing a relearning determination step according to an embodiment of the present invention. Detailed description of the invention

[0038] In the following, embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, unless explicitly identified, dimensions, materials, shapes, relative positions, and the like of components described in the embodiments are intended to be interpreted as merely illustrative and are not intended to limit the scope of the present invention.

[0039] For example, an expression of a relative or absolute arrangement, such as "in a direction", "along a direction", "parallel", "orthogonal", "centered", "concentric" and "coaxial" should not be construed as indicating only the arrangement in a strictly literal sense, but should also include a state in which the arrangement is relatively displaced by a deviation, or by an angle or a distance, whereby the same function can be achieved.

[0040] For example, an expression of a same state, such as "the same", "the same" and "equal" should not be construed as indicating only the state in which the characteristic is strictly speaking the same, but should also include a state in which there may be a deviation or a difference that can still achieve the same function.

[0041] Furthermore, for example, an expression of a shape such as a rectangular shape or a cylindrical shape should not be construed as merely the geometrically strict shape, but should also include a shape having an unevenness or beveled corners within the range in which the same effect can be achieved.

[0042] On the other hand, an expression such as “comprise”, “include”, “have”, “contain” and “form” does not intend that other components are not included.

[0043] Fig. 1 is a schematic configuration diagram of a combustion furnace 7 with an installed furnace state quantity estimation device 1 according to an embodiment of the present invention. The combustion furnace 7 of the Fig. 1 is a stoker-type combustor that uses municipal waste, industrial waste, or the like as fuel Fg. The furnace state quantity estimator 1 estimates, for example, a state quantity S of an unburned combustible content in ash, a NOx concentration, a CO concentration, or the like in the combustor. The present invention will be described below by taking a combustor having a combustion region 8 (combustion chamber) in which waste is burned as an example of the combustion furnace 7. However, the present invention is not limited to the present embodiment. In some other embodiments, the furnace state quantity estimator 1 may be configured to estimate the state quantity S that changes according to a combustion state in the combustion furnace 7 including the combustion chamber in which the fuel Fg is burned.The combustion furnace is, for example, a gasification furnace in which coal is gasified in a boiler, IGCC. The combustor, the boiler, the gasification furnace, and the like collectively include an upper part 7c and side wall parts 7s, which will be described below. In a case where the combustion furnace 7 is the boiler or the gasification furnace, the combustion furnace 7 described below is replaced with the boiler or the gasification furnace as needed.

[0044] In the Fig. In the combustor shown in Figure 1, in the combustion furnace 7, the fuel Fg is fed into the furnace from a fuel supply port 71 through a fuel feeder 72, and then dried, combusted, and burned on a grate 73 ("stoker") in the combustion region 8 to be converted into ash (combustion ash), and the ash is discharged from the furnace through an ash discharge port 74. Furthermore, the combustion furnace 7 includes the combustion region 8, which is formed by a primary combustion region 81 and a secondary combustion region 82 within the furnace. The primary combustion region 81 is formed of a main combustion region 81a in which the fuel Fg actively burns with a flame on the grate 73, and a burning combustion region 82b in which the fuel Fg burns on the grate 73.The secondary combustion region 82 is a region where unburned combustible fuel content is burned above the grate 73. A fuel gas G of the fuel Fg is supplied into the furnace through gas supply lines 77. The fuel gas G is supplied to the primary combustion region 81 from the lower part of the grate 73 via a first gas flow rate adjusting valve 78a from a gas supply device 76 such as a blower, or it is supplied to the secondary combustion region 82 from the side parts of the combustion furnace 7 via a second gas flow rate adjusting valve 78b from the gas supply device 76. The representative example of the fuel gas G is air. However, the fuel gas G may be a combustible gas, and it may be generated by mixing it with, for example, an EGR gas (combustion exhaust gas) discharged from the primary combustion region 81 at a predetermined mixing ratio.On the other hand, an exhaust gas E generated by burning the fuel Fg is discharged from a stack 93 through an exhaust gas processing device 92 through an exhaust gas passage 91.

[0045] In addition, as in Fig. 1, an image pickup device 6 for recording the furnace is installed in the combustion furnace 7. The image pickup device 6 is, for example, a furnace camera that can record at least one of a moving image and a still image. In particular, the image pickup device 6 can be, for example, a camera such as a digital camera, a video camera, or an infrared camera that can detect a specific wavelength. In the Fig. 1, the image pickup device 6 is installed on the upper part 7c of the combustion furnace 7, which is above (the immediately upper part in Fig. 1) the primary combustion region 81, and is configured to capture a combustion state directly above it. However, the present invention is not limited to the present embodiment. In some other embodiments, the image pickup device 6 may be installed at a position other than the upper part 7c of the combustion furnace 7, such as each side wall part 7s of the combustion furnace 7. For example, the image pickup device 6 is installed in a portion of the side wall parts 7s in which the secondary combustion region 82 is formed, or in a portion further above the portion (a portion closer to the upper part 7c), and may be installed obliquely downward to capture a lower portion at which the primary combustion region 81 is arranged.The image pickup device 6 thus captures a furnace image V (combustion image) when the fuel Fg is combusted. The image V is stored (accumulated) in a storage device connected to the image pickup device 6. In the embodiment shown in . Fig. 1, the image V is configured to be stored in a storage device 1m in the furnace state quantity estimation device 1. However, in some other embodiments, the image V may be stored in, for example, a storage device provided as a separate part from the furnace state quantity estimation device 1, such as a storage device of an estimation model generation device 2b to be described below or a storage device of another device.

[0046] In the following, the furnace state quantity estimation device 1 will be described with reference to Fig. 2 described. Fig. 2 is a functional block diagram of the furnace state quantity estimation device 1 according to an embodiment of the present invention. As shown in Fig. 2, the furnace state quantity estimation device 1 comprises an estimation model acquiring unit 3, an input image information acquiring unit 4 and an estimated state quantity calculating unit 5. In the Fig. 2, the furnace state quantity estimation device 1 further comprises an estimation model generation unit 2, and is configured to estimate the state quantity S by the above-described functional units (3 to 5) using an estimation model M generated by the estimation model generation unit 2.

[0047] Each of the functional units described above is described.

[0048] The furnace state quantity estimation device 1 includes a computer including a CPU (processor, not shown) and the storage device 1m serving as, for example, an external storage device or a memory such as a ROM or a RAM. Then, the CPU operates (for example, calculates the data) according to an instruction from a program (furnace state quantity estimation program, estimation model generation program) loaded into a memory (main storage device), thereby implementing each of the above-described functional units of the furnace state quantity estimation device 1. In other words, the above-described program constitutes software for causing the computer to implement each of the above-described functional units.

[0049] The estimation model generation unit 2 generates the estimation model M by performing machine learning on learning data D, in which past image information Ip is associated with the state quantity S corresponding to a combustion state indicated in the past image information Ip. The past image information Ip represents image information I obtained based on the captured image V of the furnace of the combustion furnace 7 (hereinafter simply referred to as the image V) and obtained based on the past image V. The learning data D is constituted by a plurality of data (learning configuration data), in which the past image information Ip (image information I) obtained based on the past image V is associated with an estimated value or a measured value Sr of the state quantity S when the past image V is captured.The image information I may represent the captured image V of the furnace itself, or it may represent a feature quantity F extractable from the image V, as described below. Furthermore, the image V may be a still image, or it may represent frames (images) that construct a moving image. Then, the estimation model generation unit 2 uses at least one known machine learning method (algorithm) to perform machine learning on the learning data D, thereby generating the estimation model M for calculating the state quantity S from the image information I, and the generated estimation model M is stored in the storage device 1m, such as the external storage device.

[0050] In the Fig. In the embodiment shown in FIG. 2, all the learning configuration data are generated by obtaining the state quantity S by sampling the unburned combustible content in ash while synchronizing with a shooting time of the past image V, or measuring the NOx concentration, the CO concentration, or the like by a sensor, and associating the obtained state quantity S with the past image information Ip obtained based on the past image V. The value of the unburned combustible content in ash is an indicator of combustion efficiency and can be obtained by a method (JISM8815) of heating ash sampled at a furnace outlet with an analyzer and measuring a weight change at that time. Generation of the estimation model M will be described in detail below.

[0051] Then, the estimation model acquiring unit 3, the input image information acquiring unit 4, and the estimated state quantity calculating unit 5, to be described below, are configured to input the image information I (input image information It to be described below) used to obtain an estimated value of the state quantity S (estimated state quantity Se) into the estimation model M generated by the estimation model generating unit 2, and estimate the state quantity S corresponding to the input image information It.

[0052] The estimation model acquisition unit 3 acquires the estimation model M generated by the estimation model generation unit 2. In the Fig. In the embodiment shown in FIG. 2, the estimated model acquisition unit 3 acquires the estimated model M by loading the estimated model M stored in the external storage device into the memory, such as RAM. However, the present invention is not limited to the present embodiment. In some other embodiments, the estimated model generation device 2b provided as a separate part from the furnace state quantity estimation device 1 may include the estimated model generation unit 2 (estimated model generation program). In this case, the estimated model acquisition unit 3 of the furnace state quantity estimation device 1 may acquire the estimated model M generated by the estimated model generation device 2b via, for example, a communication network or a portable storage medium, such as a USB memory.

[0053] The input image information acquisition unit 4 acquires the input image information It, which represents the image information I described above and is to be input into the estimation model M. The input image information It must be similar to the past image information Ip. In the Fig. In the embodiment shown in Figure 2, the input image information acquisition unit 4 is connected to the image pickup device 6 installed in the combustion furnace 7 and receives the image captured by the image pickup device 6 in real time. Furthermore, the input image information acquisition unit 4 extracts the feature quantity F (to be described below) from the image V.

[0054] The estimated state quantity calculation unit 5 calculates, using the estimation model M, the estimated state quantity Se corresponding to the combustion state indicated in the input image information It acquired by the input image information acquisition unit 4. In other words, the estimated state quantity calculation unit 5 calculates the input image information It according to the estimation model M and outputs the estimated state quantity Se as a result of the calculation. Fig. In the embodiment shown in FIG. 2, the estimated state quantity Se is displayed on a display device such as a display. In this case, the display device may display one or more estimated state quantities Se calculated in the past and stored in the storage device 1m or the like, together with the estimated state quantity Se calculated at that time. Thus, the time shift of the estimated state quantities Se can be displayed. Consequently, the operator or the like of the combustion furnace 7 can quantitatively grasp the state quantity S corresponding to the combustion state displayed in the image information I by checking the estimated state quantities Se displayed on the display.

[0055] With the above configuration, although the state quantity S, such as the unburned combustible content in ash, the NOx concentration, the CO concentration, or the like, changes according to the combustion state of the furnace, the state quantity S (estimated state quantity Se) generated in the combustion state can be estimated from the image information I obtained by capturing in the furnace during combustion using the estimation model M generated by machine learning the learning data D, and the image information I in the furnace during combustion is associated with the state quantity S (the measured value Sr, the estimated value, or the like) when the image information I is captured. Thus, the state quantity S can be quickly and accurately estimated from the furnace image information I.

[0056] In the following, some embodiments according to the estimation model generation unit 2 are described.

[0057] In some embodiments, the estimation model generation unit 2 comprises, as in Fig. 2, a feature quantity extraction unit 21, a learning data generation unit 23, and a machine learning execution unit 24. Each of the functional units described above will be described.

[0058] The feature quantity extraction unit 21 extracts at least one feature quantity F from the past image V. The feature quantity F is an indicator that can capture the feature of the image V. In other words, the feature quantity F is an indicator that can quantify a difference from another image V. The feature quantity F may represent, for example, pieces of information regarding a shape, a size, a color, a contrast density, a brightness, a temperature (temperature distribution), a wavelength (wavelength distribution), and the like that appear in the image V or are obtained based on the image V, or a change amount of at least a piece of information of these pieces of information.In a case where the feature quantity F represents the above-described change quantity or the like, the feature quantity F may be a feature obtained by, for example, a CHLAC (cubic higher order local auto-correlation) feature quantity, CNN (convolution neural network), or AE (auto-encoder). The CHLAC feature quantity has the advantage of being able to compactly express a spatiotemporal change occurring in the entire image V. The CNN has the advantage of being able to effectively reduce and acquire a feature of the entire image V by obtaining the feature through convolution processing or pooling processing. The AE has the advantage of being able to effectively reduce and acquire the feature of the entire image V by obtaining the feature during encoding processing.

[0059] In the Fig. 2, the feature quantity extraction unit 21 extracts the feature quantities F based on brightness information obtained from the past image V. Specifically, the feature quantities F extracted based on the brightness information may represent values ​​of brightness (brightness values) obtained from the image V itself, or they may represent a static value such as an average value, a peak value, or the like of the brightness values. Alternatively, the above-described feature quantities F may represent one of the area, shape, or size of a portion having a brightness of a predetermined value or larger. Alternatively, the above-described feature quantities F may represent amounts of change in the brightness values.The brightness information may include information that can be converted from the brightness values, such as the temperature or the temperature distribution obtained by converting the brightness values. The feature quantity extraction unit 21 extracts one or the plurality of feature quantities F of these feature quantities F from the past image V. The brightness information is information obtained from the image V without detecting a flame itself. It is possible to obtain the image information I even in a case where the flame cannot be captured due to the exhaust gas E because the image pickup device 6 is installed in the upper part of the furnace. Thus, the image pickup device 6 (furnace camera) for capturing the furnace can also be easily installed.

[0060] The learning data generation unit 23 generates the learning data D by associating the state quantity S with at least one feature quantity F extracted by the feature quantity extraction unit 21. That is, one of the plurality of feature quantities F each extracted from the plurality of past images V corresponds to the above-described past image information Ip associated with each of the state quantities S. The learning data generation unit 23 may perform any preprocessing such as averaging processing or the like on the extracted feature quantities F and associate the state quantities S with the preprocessed feature quantities F.

[0061] The machine learning execution unit 24 executes machine learning based on the learning data D generated by the learning data generation unit 23. The machine learning execution unit 24 can execute machine learning by one of known machine learning methods (algorithms), such as a neutral network and a generalized linear model.

[0062] With the above configuration, the learning data D in which the state quantity S is associated with the feature quantity F of the past image V is generated, and machine learning is performed using the learning data D. Thus, by performing machine learning focusing on the relationship between the feature quantity F and the state quantity S, the estimation model M with high estimation accuracy can be generated.

[0063] Furthermore, in a case where the above-described feature quantity extraction unit 21 extracts the plurality of feature quantities F according to some embodiments, as shown in Fig. 2, the estimation model generation unit 2 may further comprise a feature quantity selection unit 22 which selects the N (N indicates an integer not less than one) feature quantities F from the plurality of feature quantities F extracted by the feature quantity extraction unit 21 based on contributions C to the state quantities S. In this case, the above-described learning data generation unit 23 generates the learning data D by associating the state quantities S with the N feature quantities F selected by the feature quantity selection unit 22. In the Fig. 2, the feature size selection unit 22 automatically selects (narrows down) the top N feature sizes F having the largest contributions C based on the preset number of N. However, the present invention is not limited to the present embodiment. In some other embodiments, the feature sizes F and the contributions C automatically selected by the feature size selection unit 22 may be displayed to be seen by the operator or the like, thereby causing the operator to perform a selection operation (select, deselect) by, for example, clicking a checkbox. In this case, the feature size selection unit 22 selects (acquires) the feature sizes F by receiving the result of the selection operation by the operator.In some other embodiments, the feature quantity selection unit 22 may select the feature quantities F by displaying the list of the feature quantities F and their contributions C, and receiving the N feature quantities F selected by the operator through the selection operation.

[0064] The contributions C described above each represent an indicator indicating the magnitude of a correlation with a corresponding one of the state variables S. Accordingly, the state variable S changes more significantly according to a change in the feature variable F when the feature variable F has the larger contribution C. Assuming that the state variable S also changes according to the change in the combustion state-indicating image V (feature variable F), it can be considered that the contribution C represents an indicator indicating whether the feature variable F can detect the change in the state variable S. Accordingly, the feature variable F with the larger contribution C highlights a difference in the image V.

[0065] In some embodiments, each of the contributions C may represent a contribution ratio in a regression analysis. In particular, the same state variable S corresponding to the one past image V is associated with each of the plurality of (n) feature variables F extracted from the image V, thereby generating a plurality of (n) data sets that differ from each other only in the feature variables F (the aggregation of the n data sets = {(F1, S a ), (F2, S a ), ..., (F n , S a )}, where n is an integer and S aa value of the state quantity S, such as the unburned combustible content in ash). Aggregation is obtained for all of the plurality of learning configuration data configuring the learning data D. Then, regression analysis is performed on the aggregations made from the plurality of data sets having the same feature quantity F, such as F1, obtained from the plurality of learning configuration data, thereby obtaining a regression equation for calculating the state quantity S from the feature quantity F. Then, the contribution ratio of the regression equation can be calculated by dividing a variance of a predicted value of the state quantity S calculated from the regression equation with the above-described data sets by a variance of an actual value.

[0066] In another embodiment, the contributions C may be calculated by obtaining, from the plurality of learning configuration data, the data sets made of the state quantity S (actual value) and the plurality of feature quantities F extracted from the one piece of past image information Ip, by a multiple regression analysis, a regression equation (S=Σ (b i × F i ) + c), which obtains the state variable S from the plurality of feature variables F (i = 1, 2, 3, ..., n, and n is an integer not less than two), and, for example, calculates the rate based on the magnitude of a coefficient (coefficient b i ) is calculated in the regression equation.

[0067] With the above configuration, the estimation accuracy of the state variable S can be improved by the estimation model M by constraining the feature variables F based on the contributions C.

[0068] However, the present invention is not limited to the above-described embodiment. In some other embodiments, the furnace state quantity estimation device 1 (estimation model generation unit 2) may not include the feature quantity extraction unit 21, the feature quantity selection unit 22, and the learning data generation unit 23. In this case, when the learning data D generated by the estimation model generation unit 2 is input to the furnace state quantity estimation device 1, the machine learning execution unit 24 performs machine learning using the input learning data D itself. In this case, the learning data D may be generated by another device provided as a separate part from the furnace state quantity estimation device 1, or may be generated manually.

[0069] Furthermore, in some embodiments, the estimation model generation device 2 generates the plurality of estimation models M using a plurality of machine learning methods. Known machine learning methods (algorithms) can be arbitrarily adopted. Estimation models M of, for example, the neutral network, the generalized linear model, and the like can be generated, respectively.

[0070] With the above configuration, the state quantity S can be estimated by each of the plurality of estimation models M each generated based on the plurality of machine learning methods (algorithms). For example, the estimation model M with high estimation accuracy can be selected from the plurality of estimation models M by comparing an estimation result by each of the plurality of machine learning methods with the state quantity S (such as the measurement value Sr) configuring the learning data D, and the estimation model M suitable for, for example, conditions (such as an image size and the number of learning data) of the learning data D or the like and the type of a combustion furnace 7 can be selected.

[0071] In the following, some embodiments according to a generation of the learning data D (learning data generation unit 23) are described.

[0072] In a case where the above-described state quantity S represents the state quantity S with respect to the exhaust gas E generated in the furnace during combustion or a waste material such as combustion ash contained in the exhaust gas E, the above-described learning data generation unit 23 generates, in some embodiments, as shown in Fig. 2, the learning data D is obtained by associating the measured value Sr with the above-described past image information Ip. The measured value Sr is measured at a measuring point of the state quantity S after a lapse of a predetermined time from a time when the image V to be the basis of the past image information Ip is captured. The exhaust gas E generated by burning the fuel Fg flows from the combustion region 8 to the exhaust passage 91 (see FIG. Fig. 1). However, if the distance from the image pickup point (position) indicated in the image V to the measurement point of the state quantity S is large, there is a time delay (predetermined time) until the exhaust gas E generated in the combustion state indicated in the image V reaches the measurement point of the state quantity S. Thus, the image V is associated with the measured value Sr of the state quantity S taking the time delay into account. For example, the time delay in the combustor is larger than the time delay in a pulverized coal firing boiler that pulverizes the fuel Fg into small particles and burns it.

[0073] For example, the learning data generation unit 23 may perform control to measure the state quantity S, such as the NOx concentration or the CO concentration, after the lapse of the predetermined time from the acquisition time of the image V, and associate the image V with the measured value Sr of the state quantity S. In the case of the unburned combustible content in ash, the learning data generation unit 23 may perform sampling after the lapse of the predetermined time from the acquisition time of the image V, and then associate the image V with an analysis result (measured value Sr) of the sampling.Alternatively, the learning data generation unit 23 may refer to time information such as a shooting time of the images V and a measurement time of each of the state quantities S from the set of the images V and the set of the measurement values ​​Sr of the state quantities S that have already been stored, search for the measurement value Sr having a measurement time that coincides with a time after the lapse of the predetermined time that has been checked in advance from the shooting time of the image V, and associate the images V with the searched measurement values ​​Sr.

[0074] With the above configuration, the learning data D is generated taking into account the time delay until the exhaust gas E generated in the combustion state indicated by the past image information Ip reaches the measurement point of the state variable S. Such a time delay may undoubtedly vary depending on the type of a combustion furnace or the position of the measurement point. Thus, the estimation model M with high estimation accuracy can be generated by generating the learning data D taking into account the time delay.

[0075] Furthermore, in some embodiments, the above-described learning data generation unit 23 may generate the learning data D according to a parameter regarding generation of the preset learning data D. Specifically, the parameter may represent at least one of the number of pixels (image size) that affect image quality, the number of learning configuration data to be included in the learning data D, contents of preprocessing for the image V, or the like. It is possible to appropriately adjust an estimation accuracy by the estimation model M by adjusting the parameter.

[0076] Furthermore, in some embodiments, the furnace state quantity estimation device 1 comprises, as shown in Fig. 2, further includes a relearning determination unit 13 that determines regeneration of the estimation model M by relearning if a differential between the estimated state quantity Se and the measured value Sr of the state quantity S exceeds a predetermined threshold. The measured value Sr of the state quantity S used for determination of relearning is measured at an arbitrary time point, such as a periodic timing, which is longer than an estimation interval of the furnace state quantity estimator 1. Then, the relearning determination unit 13 calculates the differential between the estimated state quantity Se and the measured value Sr of the state quantity S when the measured value Sr of the state quantity S is input to the furnace state quantity estimator 1, and determines the necessity of relearning based on the threshold.For example, the differential described above may represent an absolute value of a result of dividing the estimated state variable Se by the measured value Sr of the state variable S. Any calculation method may be used as long as it is possible to calculate the differential. In the case shown in . Fig. In the embodiment shown in Figure 2, if the relearning determination unit 13 determines relearning, the estimation model generation unit 2 is notified of the determination. Triggered by the notification, the estimation model generation unit 2 regenerates the estimation model M.

[0077] With the above configuration, it is determined that regeneration of the estimation model M through retraining is required if a decrease in estimation accuracy is detected. If the estimation model M is regenerated through retraining and the new estimation model M is reacquired according to the determination, estimation can be continued with an appropriate estimation accuracy. Thus, the state variable S can be accurately estimated from the input image information It while following, for example, a change in an operating environment of the combustion furnace.

[0078] A furnace state quantity estimation method corresponding to a process (furnace state quantity estimation program) executed by the above-described furnace state quantity estimation device 1 will be described below with reference to Fig. 3 to 5. Fig. 3 is a flowchart showing the furnace state quantity estimation method according to an embodiment of the present invention. Fig. 4 is a flowchart showing an estimation model generation step (S1) according to an embodiment of the present invention. Fig. 5 is a flowchart showing a relearning determination step according to an embodiment of the present invention. The furnace state quantity estimation program causes the computer to perform the respective steps described below.

[0079] In some embodiments, the furnace state variable estimation method comprises, as in Fig. 3, an estimation model acquisition step (S2), an input image information acquisition step (S3), and an estimated state quantity calculation step (S4). In the present embodiment, the furnace state quantity estimation method according to some embodiments, as shown in Fig. 3 shows an estimation model generation step (S1) performed before the estimation model acquisition step (S2) described above.

[0080] The Fig. 3 shown furnace state quantity estimation method is carried out in the order of steps in Fig. 3 described.

[0081] In step S1 of Fig. 3, the estimation model generation step is performed. The estimation model generation step is a step of generating the above-described estimation model M by performing machine learning on the above-described learning data D. The estimation model generation step corresponds to the processing contents performed by the above-described estimation model generation unit 2. The details will be described below. A program (estimation model generation program) different from the furnace state quantity estimation program can cause the computer to perform the estimation model generation step.

[0082] In step S2, the estimation model acquisition step is performed. The estimation model acquisition step is a step for acquiring the above-described estimation model M. The estimation model acquisition step corresponds to the processing contents performed by the above-described estimation model acquisition unit 3, and thus its details are omitted.

[0083] In step S3, the input image information acquisition step is performed. The input image information acquisition step is a step of acquiring the above-described input image information It. The input image information acquisition step corresponds to the processing contents performed by the above-described input image information acquisition unit 4, and thus its details are omitted.

[0084] In step S4, the estimated state quantity calculation step is performed. The estimated state quantity calculation step is a step of calculating, using the estimated model M acquired by the estimated model acquisition step (S2), the estimated state quantity Se corresponding to the combustion state indicated in the input image information It acquired by the above-described input image information acquisition step (S3). The estimated state quantity calculation step corresponds to the processing contents performed by the above-described estimated state quantity calculation unit 5, and thus its details are omitted.

[0085] With the above configuration, the state variable S can be quickly and accurately estimated from the furnace image information I.

[0086] In some embodiments, the estimation model generation step (S1) comprises, as in Fig. 4, a feature size extraction step (S11), a learning data generation step (S13), and a machine learning execution step (S14). In the present embodiment, according to some embodiments, as shown in Fig. 4, the estimation model generation step (S1) further comprises a feature size selection step (S12) performed between the feature size extraction step (S11) and the learning data generation step (S13).

[0087] The Fig. The estimation model generation step (S1) shown in Figure 4 is carried out in the order of the steps in Fig. 4 described.

[0088] In step S11 of Fig. 4, the feature size extraction step (S11) is performed. The feature size extraction step is a step of extracting at least one feature size F from the past image V captured by the image pickup device 6. The feature size extraction step corresponds to the processing contents performed by the feature size extraction unit 21 described above, and thus its details are omitted.

[0089] In step S12 of Fig. 4, the feature size selection step is performed. The feature size selection step is a step of selecting, from the plurality of feature sizes, the N (N indicates an integer not less than one) feature sizes F based on contributions to the state quantity S in a case where the plurality of feature sizes F are extracted by the above-described feature size extraction step (S11). The feature size extraction step (S12) corresponds to the processing contents performed by the above-described feature size selection unit 22, and thus its details are omitted.

[0090] In step S13 of Fig. 4, the learning data generation step is performed. The learning data generation step is a step of generating the learning data D by associating the state quantity S with at least one feature quantity F representing the above-described past image information Ip. The learning data generation step corresponds to the processing contents performed by the above-described learning data generation unit 23, and thus its details are omitted.

[0091] In step S14, the machine learning execution step is performed. The machine learning execution step is a step of executing machine learning on the learning data D generated by the above-described learning data generation step (S13). The machine learning execution step corresponds to the processing contents performed by the above-described machine learning execution unit 24, and thus its details are omitted.

[0092] With the above configuration, the estimation model M with high estimation accuracy can be generated by performing machine learning focusing on the relationship between the feature variable F and the state variable S.

[0093] Furthermore, in some embodiments, the furnace state quantity estimation method includes, as in Fig. 5, further relearning determination steps (S51 to S55). The relearning determination steps (S51 to S55) correspond to the processing contents performed by the relearning determination unit 13 described above. In the Fig.In the embodiment shown in Figure 5, the measured value Sr of the state variable S is acquired in step S51 by performing a measurement at an arbitrary time. In step S52, the estimated state variable Se calculated by the estimated state variable calculation step (S4) is acquired. In step S53, the difference (differential) is calculated by dividing the estimated state variable Se by the measured value Sr of the state variable S.

[0094] Then, if it is determined in step S54 that the above-described difference exceeds a predetermined threshold, relearning is determined in step S55. In contrast, if it is determined that the above-described difference is not greater than the predetermined threshold, the relearning determination step is terminated without determining relearning.

[0095] The present invention is not limited to the above-described embodiment, and also includes an embodiment obtained by modifying the above-described embodiment and an embodiment obtained by appropriately combining these embodiments. List of reference symbols 1 furnace state variable estimator 1m storage device 13 Relearning determination unit 2 Estimation model generation unit 21 Feature size extraction unit 22 Feature size selection unit 23 Learning data generation unit 24 Machine Learning Execution Unit 2b Estimation model generation device 3 Estimation model acquisition unit 4 Input image information acquisition unit 5 Estimated state variable calculation unit 6 Image recording device 7 incinerator 7c Upper part of incinerator 7s side wall part of incinerator 71 Fuel supply opening 72 Fuel input device 73 Grating (Stoker) 74 Ash ejection opening 76 Gas supply device 77 Gas supply line 78a First gas flow rate adjustment valve 78b Second gas flow rate adjustment valve 8 Burn area 81 Primary burn area 81a Main cremation area 82b Burning combustion area 82 Secondary combustion area 91 Exhaust passage 92 Exhaust gas processing device 93 stacks Fg fuel G Fuel gas E exhaust S state variable Sr measured value of state variable D Learning data V Image I Image information IP Past Image Information It input image information M estimation model Se Estimated state variable F feature size C Contribution a i coefficient b i coefficient

Claims

[1] Furnace state quantity estimation device (1), comprising: a feature quantity extraction unit (21) which extracts feature quantities (F) from a captured image (V) of a furnace (7); a feature quantity selecting unit (22) which selects one or more of the feature quantities (F) from the extracted feature quantities (F); a learning data generation unit (23) which generates learning data (D) by associating the selected feature quantities (F) with a state quantity (S) of the furnace (7) corresponding to a combustion state displayed in the image (V); an estimation model generation unit (2) which generates an estimation model (M) for estimating the combustion state from the image (V) of the furnace (7) using the learning data (D); and an estimated state quantity calculation unit (5) which acquires the captured image (V) of the furnace (7) and then estimates the state quantity (S) corresponding to the combustion state displayed in the acquired image (V) using the estimation model (M), wherein the feature variables (F) are selected based on a contribution (C) to the state variable (S) to represent an estimation target. [2] The furnace state quantity estimation device (1) according to claim 1, wherein the estimation model generation unit (2) generates the estimation model (M) by performing machine learning on the learning data (D). [3] The furnace state quantity estimating device (1) according to claim 1 or 2, wherein the contribution (C) is an indicator indicating an amount of correlation to the state quantity (S). [4] Furnace state quantity estimation device (1) according to one of claims 1 to 3, wherein the feature size selection unit (22) a regression equation for calculating the state variable (S) from each of the feature variables (F), and as the contribution (C) sets a contribution ratio of each of the feature variables (F) in the calculated regression equation. [5] The furnace state quantity estimation device (1) according to any one of claims 1 to 4, wherein the feature quantity extraction unit (21) extracts the feature quantities (F) based on brightness information obtained from the image (V) in the past. [6] Furnace state quantity estimation device (1) according to one of claims 1 to 5, wherein the state variable (S) represents a state variable with respect to a waste material or exhaust gas (E) generated in the furnace (7) during combustion, and wherein the learning data generation unit (23) generates the learning data (D) by associating a measurement value (Sr) measured at a measurement point of the state quantity (S) after a lapse of a predetermined time from the time at which the image (V) is captured. [7] The furnace state quantity estimation device (1) according to any one of claims 1 to 6, wherein the estimation model generation unit (2) generates a plurality of estimation models (M) using a plurality of machine learning methods. [8] Furnace state quantity estimation device (1) according to one of claims 1 to 7, further comprising a relearning determination unit (13) which determines a regeneration of the estimation model (M) by performing relearning if a differential between an estimated value of the state quantity (S) and a measured value (Sr) of the state quantity (S) exceeds a predetermined threshold. [9] An estimation model generating device (2b) comprising an estimation model generating unit (2) that performs machine learning of learning data (D) in which past image information (Ip) is associated with a state quantity (S), and generates an estimation model (M) for estimating an estimated state quantity (Se) from input image information (It) obtained by capturing a furnace (7), wherein the past image information (Ip) represents a feature quantity (F) obtained based on a captured image (V) of the furnace (7) and obtained based on the image (V) in the past, wherein the state quantity (S) corresponds to a combustion state indicated in the past image (V). [10] The estimation model generating device (2b) according to claim 9, wherein the feature quantity (F) is selected based on a contribution (C) to the state quantity (S) to represent an estimation target. [11] Furnace state quantity estimation program for causing a computer to perform the following: a feature size extraction step for extracting feature sizes (F) from a captured image (V) of a furnace (7); a feature quantity selecting step for selecting one or more of the feature quantities (F) from the extracted feature quantities (F); a learning data generation step for generating learning data (D) by associating the selected feature quantities (F) with a state quantity (S) of the furnace (7) corresponding to a combustion state displayed in the image (V); an estimation model generation step for generating an estimation model (M) using the learning data (D) to estimate the combustion state from the image (V) of the furnace (7); and an estimated state quantity calculation step for acquiring the captured image (V) of the furnace (7), and then estimating the state quantity (S) corresponding to the combustion state displayed in the acquired image (V) using the estimation model (M), wherein the feature variables (F) are selected based on a contribution to the state variable (S) to represent an estimation target. [12] The furnace state quantity estimation program according to claim 11, wherein the estimation model generating step generates the estimation model (M) by performing machine learning on the learning data (D). [13] Furnace state variable estimation method, comprising: a feature size extraction step for extracting feature sizes (F) from a captured image (V) of a furnace (7); a feature quantity selecting step for selecting one or more of the feature quantities (F) from the extracted feature quantities (F); a learning data generation step for generating learning data (D) by associating the selected feature quantities (F) with a state quantity (S) of the furnace (7) corresponding to a combustion state displayed in the image (V); an estimation model generation step for generating an estimation model (M) using the learning data (D) to estimate the combustion state from the image (V) of the furnace (7); and an estimated state quantity calculation step for acquiring the captured image (V) of the furnace (7), and then estimating the state quantity (S) corresponding to the combustion state displayed in the acquired image (V) using the estimation model (M), wherein the feature variables (F) are selected based on a contribution to the state variable (S) to represent an estimation target. [14] The furnace state quantity estimation method according to claim 13, wherein the estimation model generating step generates the estimation model (M) by performing machine learning on the learning data (D).

Images