JP2021042948A - Combustion abnormality prediction device, combustion abnormality prediction program, combustion control system, combustion abnormality prediction method, abnormality prediction program generation method and abnormality avoidance means prediction program generation method - Google Patents

Abstract

To provide a combustion abnormality prediction device and a combustion abnormality prediction program capable of suitably predicting an abnormality by constructing a machine learning model (program) using an abnormality determination made by an operator as teacher data and a combustion control system including the combustion abnormality prediction device and the combustion abnormality prediction program.SOLUTION: A combustion abnormality prediction device predicts whether state amount related to a combustion furnace is abnormal or not by using an abnormality prediction program for predicting a combustion abnormality generated by an intellectual information processing technique by using learning data at least including the state amount related to the combustion furnace and teacher data including at least operator manual intervention data, and outputs a predicted result.SELECTED DRAWING: Figure 1

Description

The present invention relates to a combustion abnormality prediction device such as a waste incineration device, a combustion abnormality prediction program, and a combustion control system including the same, and for example, a combustion control system that predicts (infers) a combustion abnormality and performs combustion control.

Patent Documents 1 to 3 disclose various control methods using a machine learning model.
Patent Document 1 is an in-core state amount estimation device that estimates a state amount capable of determining a combustion state in a furnace based on image information in the furnace. This in-core state quantity estimation device corresponds to the past image information obtained based on the past image which is the image information obtained based on the image of the inside of the furnace and the combustion state indicated by the past image information. The estimation model created by machine learning of the training data associated with the state quantity is used. Input image information is input to this estimation model, and an estimated state quantity corresponding to the combustion state indicated by the input image information is calculated by this estimation model. As a result, the state quantities such as unburned content in ash, NOx concentration, and CO concentration change according to the combustion state in the furnace, but the image information at the time of combustion in the furnace and the image information are captured. It is generated in the combustion state from the image information obtained by imaging the inside of the furnace at the time of combustion using the estimation model created by machine learning of the training data associated with the state quantity (measured value or estimated value) at the time. Estimate the state quantity (estimated state quantity).

Patent Document 2 is a plant management support device for managing at least one performance management index indicating the performance of the target plant, with the waste incineration plant or the biomass combustion plant as the target plant. This plant management support device collects the state amount and / or the operation amount of the target plant as measurement data, and the measurement data and / or the calculation data obtained by calculation from the collected measurement data at a plurality of first time points. As an explanatory variable, a regression model is constructed using the performance management index at a second time point after any of the first time points as an objective variable, and a plurality of collected data are collected for this regression model. Input the measurement data and / or the calculation data obtained by calculation from the measurement data at the third time point after any of the first time points and at a plurality of fourth time points before the third time point. Thereby, the predicted value of the at least one performance management index at the fifth time point after the third time point is calculated. In this configuration, with respect to the explanatory variables at the past time point t and the time points (t-Δt1), (t-2Δt1), ..., (TnΔt1) retroactive from the time point t, and the past time point t. Prediction of the performance management index at a future time point (t0 + Δt2) with respect to the current time point t0 based on the statistical correlation with the performance management index at the time point (t + Δt2) to which a time width Δt2 larger than 0 is added. The value is calculated without being affected by the time delay of the plant itself or the time delay of the measuring device.

Patent Document 3 is an information processing device that processes information related to an incineration facility. This information processing device acquires a plurality of types of predetermined data regarding the operating state of the incineration facility, and incinerates at least a part of the predetermined data in the past with respect to a predetermined event predicted to occur in the incineration facility. By using it for a factor prediction model constructed based on the correspondence between the above-mentioned predetermined data when the above-mentioned predetermined event occurs in the facility and the cause of the above-mentioned predetermined event, the cause of the above-mentioned predetermined event To identify. Prediction targets include, for example, an event in which the amount of steam decreases, an event in which the temperature inside the incinerator falls below the threshold, and the concentration of harmful substances (carbon monoxide, nitrogen oxides, etc.) in the exhaust gas from the incinerator is the threshold. There are the above events. The probability prediction model is data D regarding the operating state of a predetermined time (for example, several hours) immediately before a certain point in time when the garbage incineration facility is operated in the past, and the occurrence of a decrease in the amount of steam after a predetermined time from that point. It is a model constructed based on the correspondence with the probability. As another probability prediction model, among the data D acquired and accumulated when the garbage incineration facility was operated in the past, the data D for one hour immediately before a certain point in time and the amount of steam in the 30 minutes immediately after that. Identify the correspondence with the probability that

Patent Document 4 is an operation control method for waste treatment plant equipment. In this operation control method, various plant data from the operation of the waste treatment plant equipment are guided to a neural network, the model of the operation control of the waste treatment plant equipment is learned by the neural network, a learning model is created, and the learning model is created. The model performs predictive operation control after a predetermined time. , The neural network derives the correlation of the control target from various plant data, and reconstructs the learning model so that the predicted value approaches the actual value of the control target. The learning model is updated with the reconstructed learning model at a predetermined setting cycle, or the learning model is predicted and evaluated, and when the predicted evaluation value deviates from the predetermined value, the learning model is updated with the reconstructed learning model. Various plant data in the actual operation control of the approaching predetermined period are collected and guided to a neural network, and the neural network learns the correlation with the control target from the various plant data and creates the latest learning model. An operation control simulation is performed by substituting various plant data in the operation control for a predetermined period approaching the created latest learning model, and it is determined whether or not the value by the operation control simulation and the value by the actual operation control are within the predetermined range. If it is within the predetermined range, the current learning model is switched to the latest learning model, and if it is outside the predetermined range, the operation control is continued with the current learning model.

JP-A-2019-74240 JP-A-2018-151771 JP-A-2019-2672 JP-A-2005-249349

Generally, the combustion control operation in the incinerator, various target output value (furnace temperature, the amount of evaporation, such as O ₂ concentration) are set in advance. While performing the operation on the basis of the target output value, the operating personnel (also referred to as "plant data".) Actual measurement value is the target output value (furnace temperature, the amount of evaporation, O ₂ concentration, etc.) the threshold on relative to the And when it approaches the lower threshold, it is judged to be abnormal or it is predicted that it will become abnormal, manual intervention is performed, and various operations (control of air volume, air temperature, waste supply amount, etc.) are performed so as to return to the target output value. ing.
In other words, it is considered that there is a large difference in abnormality prediction depending on the operator's skill level, individual differences (own criteria for judging abnormalities), operator's condition (health condition, emotional change), and the like.
In Patent Documents 1 to 4, it is expected that the variation due to the operator is compensated by machine learning, but there is also a large difference in the prediction accuracy of the machine learning model. In Patent Document 4, in the construction of the re-learning model, (1) the data of a predetermined period is divided into two, one of the data is trained to construct the update learning model, and the other data is used to evaluate the update learning model, and the result is good. If there is, we propose a method to replace the learning model and (2) a method to replace the latest learning model with the latest learning model if the simulation result is better than the simulation and the actual operation control. It is insufficient to eliminate the variation in prediction accuracy caused by the operator.
In addition, there is a large difference between seasonal fluctuations in waste quality and waste quality in the waste collection area, and the operator makes manual interventions, that is, abnormal judgments, taking these fluctuation factors into consideration.

In view of the above circumstances, the present invention provides a combustion abnormality prediction device, a combustion abnormality prediction program, and a combustion including them, which can suitably predict an abnormality by constructing a machine learning model (program) using an operator's abnormality judgment as teacher data. The purpose is to provide a control system.

As a result of examining the actual state of manual intervention by the operator, the present inventors have obtained the following findings.
(1) The operator sets a normal range (referred to as “individual normal range”) based on the target output value. This may be set by the operator at his own discretion, or may be preset by a rule based on the target output value.
Individual normal range, the may be the same as "target output value (furnace temperature, the amount of evaporation, O ₂ concentration, etc.) threshold and a lower threshold on relative to the" greater than the lower limit value lower than the upper threshold value It may be a value.
(2) When the actual measured value (plant data (sometimes referred to as process data)) approaches the upper limit value or the lower limit value of the individual normal range and approaches a predetermined interval (difference), the operator operates the operator. Thinks that if the operation is continued under the same conditions as it is, the actual measured value will be out of the individual normal range (predicted that it will be an abnormal operation). Here, what is important is that, at the time of this judgment, it is not yet in abnormal operation, but in the range of normal operation.
Therefore, in the present invention, the plant data up to the time when the operator determines that it is "abnormal" and the operation data after that are used as the teacher data. In other words, the learning model does not predict (estimate) the future time of abnormal operation, but makes the operator's judgment that it will become abnormal if it continues under the same conditions as it is, and the operator's operation behavior after that judgment. Predict (estimate).
In addition, re-learning using the operator's operation log as teacher data is frequently performed in consideration of regional differences, the operator's situation, the occurrence of new events, and the influence of secular change.

The combustion abnormality prediction device of the present invention
Knowledge using at least training data including state quantities related to the combustion furnace (for example, plant data, operation data, etc.) and at least teacher data including manual intervention data of the operator (for example, abnormality input data when the operator determines that the abnormality is abnormal). Anomaly prediction programs that predict combustion anomalies generated by the information processing technology and / or learning data that includes at least state quantities (eg plant data, operation data, etc.) for the combustion furnace, and at least manual intervention data of the operator (eg Abnormal avoidance means prediction program that predicts combustion abnormalities generated by intelligent information processing technology using teacher data including manually intervened operation logs) and predicts means for avoiding abnormalities (abnormal avoidance means). The first storage unit that memorizes
The combustion abnormality prediction program is used to predict whether or not the state quantity related to the combustion furnace is abnormal, and / or the abnormality avoidance means prediction program is used to predict whether or not the state quantity related to the combustion furnace is abnormal. And the anomaly predictor that predicts the means to avoid the anomaly,
It has a prediction result output unit that outputs the result predicted by the abnormality prediction unit (presence / absence of abnormality and / or abnormality avoidance means).
The abnormality prediction unit or the abnormality avoidance means prediction program may predict a means for avoiding the abnormality and a priority indicating a level at which the means is recommended.

The abnormality prediction unit may be executed in real time or at a preset timing.
The prediction result output unit may output the prediction result in real time, may output at a preset timing, or may output in response to an output command or instruction input.
When the abnormality prediction unit is executing in real time, the prediction result may be output in real time, and the prediction result indicating that it is abnormal continues for a predetermined period (for example, 5 minutes or more, 10 minutes or more). The prediction result may be output to. The output unit can prevent the prediction result from being output if the prediction result indicating that it is abnormal is within a short predetermined period. As a result, the number of manual interventions by the operator can be reduced, and the time lag between the manual interventions and the reflection in the combustion state can be considered.

The above combustion abnormality predictor is
Re-learning to relearn (tune) the combustion abnormality prediction program and / or the abnormality avoidance means prediction program at a predetermined frequency by using the learning data of the past predetermined period and the teacher data of the past predetermined period. Department and
Further having a new combustion abnormality prediction program relearned by the re-learning unit and / or a second storage unit for storing the abnormality avoidance means prediction program.
The abnormality prediction unit predicts using a new combustion abnormality prediction program and / or an abnormality avoidance means prediction program generated each time it is relearned.
The re-learning unit preferably executes re-learning at a predetermined time about once a day (this is referred to as "real-time learning").
The second storage unit may be the same as the first storage unit.

The above combustion abnormality predictor is
It may have a learning data generation unit.
The learning data generation unit preprocesses the data of the state quantity (for example, plant data, operation data, etc.) related to the combustion furnace, performs feature selection such as noise removal, extracts the feature quantity by feature conversion, and extracts the feature quantity. It may be used as training data. The feature conversion includes, for example, principal component analysis, non-negative matrix factorization, factor analysis and the like.

The above combustion abnormality predictor is
It may have a teacher data generation unit.
The teacher data generation unit compares the prediction result of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program with the abnormality determination (or the operator's intervention operation) determined by the operator, and selects one of them. It may be used as teacher data.
"Compare and select one of them" may be selected by the operator's instruction input, or may be selected by the teacher data generation unit based on a predetermined selection criterion. As the predetermined selection criterion, for example, a decision tree may be used. In addition, the operation data after the selected time is compared with the operation data including the selected time and before this, and only when (the state quantity related to the combustion furnace) is improved, it is used as the teacher data (selected person). ) May be adopted.
The above combustion abnormality predictor is
It has a user interface that accepts operator input.
The user interface may have an abnormality determination input unit that receives an input for an abnormality determination determined by the operator.
The above combustion abnormality predictor is
It has a user interface that accepts operator input.
The user interface may have an intervention operation input unit that receives an input of an intervention operation by an operator.
The combustion abnormality prediction device may include a determination reason input unit for inputting an operator's determination reason (reason for selecting the manually intervened operation item) in response to the display of the operation log.
The operator's determination reason may include an operator's operation error (erroneous operation). Manual intervention operations related to operational errors (erroneous operations) may be configured to be excluded from the teacher data used for real-time learning or re-learning.

The above combustion abnormality predictor is
It may have an accuracy calculation unit that compares the prediction result of the abnormality avoidance means prediction program with the abnormality determination determined by the operator or the intervention operation of the operator, and calculates the accuracy of the prediction result.
The prediction results of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program are associated with the operator's judgment (abnormality judgment, intervention operation) as follows. "Normal" is a state in which "abnormality determination or intervention operation" is not performed.
a: The prediction result is "abnormal" and the operator's judgment is "abnormal".
b: The prediction result is "abnormal" and the operator's judgment is "normal".
c: Prediction result is "normal", operator's judgment is "abnormal"
d: The prediction result is "normal" and the operator's judgment is "normal".
As an evaluation index of accuracy, the correct answer rate (the ratio of the data in which the prediction result and the operator's judgment match in all the data) and the recall rate (the data in which the operator's judgment was abnormal, both the prediction result and the operator's judgment) Percentage of data that is abnormal) is used.
Correct answer rate = (a + d) / (a + b + c + d) (1)
Recall rate = a / (a + c) (2)
The accuracy calculation unit may calculate the correct answer rate and / or the recall rate.
The accuracy calculation unit may compare the timing, operation item, and operation amount (increase / decrease) of the operator's operation when both the prediction result and the operator's intervention operation are the result of performing the abnormality avoidance operation.
(I) Determine whether the operator's operation timing is within the specified period, and
(Ii) If it is within the predetermined period in (i) above, the matching rate of each operation item is calculated.
(Iii) The correct answer rate of the operation amount of the operation item that matches in (ii) above may be calculated.
Here, the correct answer may be, for example, a correct answer as long as the predicted result of the program is an operation amount within a predetermined% based on the intervention operation of the operator.
The above combustion abnormality predictor is
It may have an accuracy result output unit that outputs the result of the accuracy calculated by the accuracy calculation unit (for example, correct answer rate, recall rate).

The combustion control system of another invention functions as the main control device of the combustion furnace.
Combustion control system
Equipped with the above combustion abnormality prediction device
It has a combustion control unit that controls combustion in a combustion furnace based on the abnormal result and / or means for avoiding the abnormality.
Combustion control systems of other inventions
A receiving unit that receives various data sent from the above-mentioned combustion abnormality prediction device, a storage unit that stores various data, and a storage unit.
It has a combustion control unit that controls combustion in a combustion furnace based on the abnormal result and / or means for avoiding the abnormality.
The above combustion control system
It has a user interface that accepts operator input.
The user interface may have an abnormality determination input unit that receives an input for an abnormality determination determined by the operator.
The above combustion control system
It has a user interface that accepts operator input.
The user interface may have an intervention operation input unit that receives an input of an intervention operation by an operator.
The user interface may have a determination reason input unit for inputting an operator's determination reason (reason for selecting a manually intervened operation item) corresponding to the display of the operation log.
The operator's determination reason may include an operator's operation error. Manual intervention operations related to operational errors may be configured to be excluded from the teacher data used for real-time learning or re-learning.
The combustion control system may include the accuracy calculation unit and the accuracy result output unit.

The combustion control system presents the operator with "abnormality prediction" and / or "means for avoiding the abnormality", and assists the operator in determining whether or not to execute the means, and "avoids the abnormality". Of the fully automatic operation mode that automatically executes "means for avoiding abnormalities" and the part that executes some means or sets a weight of 0% to 100% among "means for avoiding abnormalities". It has an operation mode, and one of the operation modes can be selected.
The combustion control unit may control the combustion of the combustion furnace according to the selected operation mode.

Further, the combustion abnormality prediction method of another invention is
Knowledge using at least learning data including state quantities related to the combustion furnace (for example, plant data, operation data, etc.) and at least teacher data including manual intervention data of the operator (for example, abnormality input data when the operator determines that the abnormality is abnormal). Predict whether or not the state quantity related to the combustion furnace is abnormal by using an abnormality prediction program that predicts combustion abnormality generated by the information processing technology, and / or
Generated by intelligent information processing technology using at least training data including state quantities related to the combustion furnace (eg plant data, operation data, etc.) and at least teacher data including manual intervention data of the operator (eg, manual intervention operation log). A means for predicting whether or not the state quantity of the combustion furnace is abnormal and avoiding the abnormality by using the abnormality avoidance means prediction program for predicting the combustion abnormality to be performed and predicting the means for avoiding the abnormality. Anomaly prediction steps and
Includes an output step that outputs the result predicted in the abnormality prediction step (the prediction result of the abnormality and / or the abnormality avoiding means).
The combustion abnormality prediction method uses the learning data of the past predetermined period and the teacher data of the past predetermined period to re-execute the combustion abnormality prediction program and / or the abnormality avoidance means prediction program at a predetermined frequency. Including re-learning steps to learn (tune)
The abnormality prediction step is predicted by using a new combustion abnormality prediction program and / or an abnormality avoidance means prediction program generated each time it is relearned.
The above combustion abnormality prediction method is
It may include a teacher data generation step of comparing the prediction result of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program with the judgment result of the operator and selecting one of them to generate the teacher data.
In the teacher data generation step, when there is no abnormality determination (or operator intervention operation) determined by the operator, the prediction results of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program may be used as teacher data as they are. ..
If the automatic operation based on the prediction result of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program is appropriate, the operator does not need to perform manual intervention, and therefore, the combustion abnormality prediction program and / or the abnormality avoidance If the prediction result of the means prediction program is appropriate (correct), only this prediction result is used as the training data as data. In other words, if there is no operator intervention operation, there is an advantage that the prediction can be used as it is as teacher data as it is correct.
As a method of determining that the prediction result is correct, the prediction result is obtained without manual intervention of the operator during a predetermined period before or after the prediction (for example, it may be set in the range of several tens of seconds to several tens of minutes). Is correct and used as teacher data. If the prediction is incorrect, the operator manually intervenes for a predetermined period before or after the prediction (for example, it may be set in the range of tens of seconds to tens of minutes) to compensate (cover) it. carry out. In this way, it may be determined that the prediction result is correct.
The above combustion abnormality prediction method is
An accuracy calculation step of comparing the prediction result of the abnormality avoidance means prediction program with the abnormality determination determined by the operator or the intervention operation of the operator and calculating the accuracy of the prediction result may be included.
The above combustion abnormality prediction method is
An accuracy result output step that outputs the result of the accuracy calculated in the accuracy calculation step may be included.

Further, the information processing device of another invention is
With at least one processor
Includes memory for storing instructions that can be executed by the processor.
The processor is an information processing device that realizes the combustion abnormality prediction method by executing an executable instruction.

Further, the combustion abnormality prediction program of another invention is a program that realizes the above-mentioned combustion abnormality prediction method by using at least one processor.
Further, it is a computer-readable recording medium in which computer instructions of another invention are stored, and when the computer instructions are executed by a processor, the computer-readable recording medium realizes the steps of the combustion abnormality prediction method described above. Recording medium.

Further, the method for generating an abnormality prediction program of another invention is
A training data generation step that generates training data including at least the state quantity related to the combustion furnace,
A teacher data generation step that generates teacher data, including at least manual intervention data for the operator,
It includes a model generation step of generating an abnormality prediction program for predicting a combustion abnormality generated by an intelligent information processing technique using the learning data and the teacher data.
Further, the method of generating the abnormality avoidance means prediction program of another invention is
A training data generation step that generates training data including at least the state quantity related to the combustion furnace,
A teacher data generation step that generates teacher data, including at least manual intervention data for the operator,
A model generation step of predicting a combustion abnormality generated by an intelligent information processing technique using the learning data and the teacher data and generating an abnormality avoidance means prediction program for predicting means for avoiding the abnormality. Including.
In the teacher data generation step, one of the prediction results of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program is compared with the abnormality determination (or the operator's intervention operation) determined by the operator, and one of them is selected. It may be used as teacher data.
In the teacher data generation step, when there is no abnormality determination (or operator intervention operation) determined by the operator, the prediction results of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program may be used as teacher data as they are. ..
If the automatic operation based on the prediction result of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program is appropriate, the operator does not need to perform manual intervention, and therefore, the combustion abnormality prediction program and / or the abnormality avoidance If the prediction result of the means prediction program is appropriate (correct), only this prediction result is used as the training data as data. In other words, if there is no operator intervention operation, there is an advantage that the prediction can be used as it is as teacher data as it is correct.
As a method of determining that the prediction result is correct, the prediction result is obtained without manual intervention of the operator during a predetermined period before or after the prediction (for example, it may be set in the range of several tens of seconds to several tens of minutes). Is correct and used as teacher data. If the prediction is incorrect, the operator manually intervenes for a predetermined period before or after the prediction (for example, it may be set in the range of tens of seconds to tens of minutes) to compensate (cover) it. carry out. In this way, it may be determined that the prediction result is correct.

The forms output by the "prediction result output unit" and the "accuracy result output unit" are, for example, display on a display device, print output to a printer, sound output to a speaker, data transmission to an external device, and storage. Including data storage and so on.

The "past" is required for the past one year from the current day, and may be longer than that (for example, two years, five years, ten years, etc.).
The "predetermined period" is set automatically or manually as a fixed period such as 1 to 60 days.
"State quantity related to the combustion furnace" is, for example, furnace gas temperature, the main steam flow rate, the furnace exit gas concentration _{(O 2, CO, NOx)} , etc. combustion image data in the furnace and the like.
Examples of the predicted "combustion abnormality" include an increase in the furnace temperature, a decrease in the furnace temperature, an increase in the evaporation amount, a decrease in the evaporation amount, an increase in CO, and an increase in NOx.
The "manual intervention data" includes an operation record (manual intervention operation log) of the operator when the operator determines that the abnormality is abnormal. All data are linked to time.
The "means for avoiding an abnormality (abnormality avoiding means)" includes an operator's operation item and an operation amount (increase / decrease) of the operation item.
Examples of the operation items include air amount, air temperature, combustion raw material supply amount, combustion raw material supply speed, dust supply device speed, combustion stoker speed, furnace water spray amount, exhaust gas treatment chemical supply amount and the like.
Examples of the "intelligent information processing technology" include machine learning, deep learning, reinforcement learning, and deep reinforcement learning.
The algorithms for machine learning, deep learning, reinforcement learning, and deep reinforcement learning are not particularly limited, and conventional algorithms may be used. As supervised learning, for example, various algorithms such as linear regression, generalized linear model, support vector regression, Gaussian process regression, ensemble method, decision tree, neural network, support vector machine, discriminant analysis, naive bayes, nearest neighbor method, etc. It may be adopted.
The "display device" is not particularly limited, and examples thereof include a liquid crystal monitor, an organic EL monitor, a CRT monitor, a smartphone, a tablet, and a monitor of a general-purpose personal computer.

The above-mentioned components may be composed of an information processing device (for example, a computer, a server) having a memory, a processor, and a software program, a dedicated circuit, firmware, and the like. The information processing device may be either on-premises or cloud, or a combination of both.

It is a figure which shows an example of the combustion abnormality prediction apparatus and combustion control system of Embodiment 1. FIG. It is an example of the functional block diagram of the component of the combustion abnormality prediction device and the combustion control system of Embodiment 1. This is an example of a flow for explaining the real-time learning of the first embodiment. This is an example of the output (screen, data structure) of the prediction result of the first embodiment. This is an example of the input screen for the abnormality prediction of the first embodiment. It is an example of the figure explaining the calculation method of the prediction accuracy of Embodiment 1. It is an example of the figure explaining the prediction accuracy of Embodiment 1.

(Embodiment 1)
FIG. 1 shows a configuration example of the combustion abnormality prediction device 30 and the combustion control system 40, and FIG. 2 shows a functional block diagram of the components of the combustion abnormality prediction device 30 and the combustion control system 40.
In the first embodiment, the combustion control system 40 controls the combustion of the stoker type incinerator 1. The stoker combustion furnace 1, various sensors and instruments (laser type O, etc. ₂ analyzer) and the like are installed in place, and the data measured (plant data), the combustion state in the furnace Image data captured by the camera is sent to the data server 20 via the communication unit 11, and the data server 20 stores the data as time-series data.
The combustion control system 40 is set with a predetermined target output value related to combustion, and is automatically operated so as to reach this target output value, so that the operator can manually intervene when the target output value is about to deviate significantly. It has become.
In the combustion control system 40, the operation data (target output value) and the operation log (manual intervention operation data) are stored in a memory (not shown) and sent to the data server 20 via the communication unit 41. Accumulate as time series data. The data server 20 stores plant data and image data in association with operation data and operation logs as time-series data.

The plant data (including image data) stored in the data server 20 is sent to the combustion abnormality prediction device 30.
The combustion abnormality prediction device 30 has a learning data generation unit 31, a teacher data generation unit 32, and a model generation unit 33.
The learning data generation unit 31 generates learning data including plant data and operation data. The learning data generation unit 31 uses various data such as plant data and operation data as training data by performing preprocessing such as noise removal, selecting features, and extracting feature quantities by feature conversion.
The teacher data generation unit 32 generates teacher data including the operator's manual intervention data (for example, abnormality input data when the operator determines that it is abnormal).
The teacher data generation unit 32 generates teacher data including the operator's manual intervention data (for example, the manual intervention operation log). Teacher data is basically an operator's manual intervention operation.
As another embodiment, the teacher data generation unit 32 compares the prediction result of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program with the abnormality determination (or the operator's intervention operation) determined by the operator. Select one to use as teacher data. The selection method may be a selection input by an operator's instruction, or may be selected by the teacher data generation unit 32 based on a predetermined selection criterion.

The model generation unit 33 receives learning data including state quantities (for example, plant data, operation data, etc.) related to the stoker type incinerator 1 and at least manual intervention data of the operator (for example, abnormality input data when the operator determines that it is abnormal). An abnormality prediction program that predicts combustion abnormalities is generated by intelligent information processing technology using the included teacher data. The learning data is generated by the learning data generation unit 31, and the teacher data is generated by the teacher data generation unit 32.
Further, the model generation unit 33 includes learning data including state quantities (for example, plant data, operation data, etc.) related to the stoker type incinerator 1, and teacher data including at least manual intervention data of the operator (for example, operation log of manual intervention). A combustion abnormality is predicted by intelligent information processing technology using the above, and an abnormality avoidance means prediction program for predicting means for avoiding the abnormality (abnormality avoidance means) is generated. The learning data is generated by the learning data generation unit 31, and the teacher data is generated by the teacher data generation unit 32.
The first storage unit 34 stores the abnormality prediction program and the abnormality avoidance means prediction program.

The abnormality prediction unit 35 predicts whether or not the state quantity of the stoker type incinerator 1 is abnormal by using the combustion abnormality prediction program.
Further, the abnormality prediction unit 35 predicts whether or not the state quantity of the stoker type incinerator 1 is abnormal by using the abnormality avoidance means prediction program, and predicts the abnormality avoidance means.
Further, the abnormality prediction unit 35 uses the abnormality avoidance means prediction program to predict the abnormality avoidance means and the priority indicating the level at which the abnormality avoidance means is recommended.

The prediction result output unit 36 outputs the result (presence / absence of abnormality, abnormality avoidance means, priority) predicted by the abnormality prediction unit 35. In the present embodiment, the predicted result is transmitted to the combustion control system 40, stored in a storage unit (not shown) in the combustion control system 40, and displayed on the display device 45 in a predetermined format.
Further, in the present embodiment, the prediction result output unit 36 outputs the prediction result in real time. When the abnormality prediction unit 35 executes the prediction in real time, the prediction result (for example, the data of FIG. 4) is output in real time.
As another embodiment, even when the abnormality prediction unit 35 executes the prediction in real time, the prediction result output unit 36 continues the prediction result indicating that it is abnormal for a predetermined period (for example, 5 minutes or more, 10 minutes or more, etc.). ), The prediction result may be output. The prediction result output unit 36 can reduce the number of manual interventions by the operator by not outputting the prediction result if the prediction result indicating that it is abnormal is within a short predetermined period, and reflects it in the combustion state after the manual intervention. You can consider the time lag until it is done.

(Real-time learning; re-learning)
FIG. 3 is a flow for explaining real-time learning.
As the time axis progresses from top to bottom, the prediction result (S1) of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program is compared with the abnormality determination (S2) determined by the operator and determined (S3). To. The determination method may be selected according to the instruction of the operator. For example, by using a machine learning means that employs classification or regression processing such as random forest, the better one is selected and used as teacher data (teacher data generation unit). 32 functions). As a result, the better operation is selected as the teacher data, so the teacher data immediately before re-learning is defined as correct. It is up to you whether the better operation is performed by manual intervention or not (S4). For example, re-learning is executed at midnight (S5). By constantly performing re-learning at a fixed time, a combustion abnormality prediction program and an abnormality avoidance means prediction program with improved accuracy from yesterday are generated.
Further, as another embodiment of the above determination method, either one of the prediction result of the program and the abnormality determination by the operator is selected, and the operation data after the selected time and the operation data including the selected time and before this are included. It may be adopted as the teacher data (the selected one) only when (the amount of the state related to the combustion furnace) is improved in comparison with.

As another embodiment, teacher data may be generated as follows.
When there is no abnormality determination (or operator intervention operation) determined by the operator, the teacher data generation unit 32 may use the prediction results of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program as teacher data as they are.
If automatic operation based on the prediction results of the combustion anomaly prediction program and / or the anomaly avoidance means prediction program is appropriate, the operator does not need to perform manual intervention, and therefore the combustion anomaly prediction program and / or the anomaly avoidance means prediction program. If the prediction result is appropriate (correct), only this prediction result is used as the training data as data. In other words, if there is no operator intervention operation, there is an advantage that the prediction can be used as it is as teacher data as it is correct.
The teacher data generation unit 32 determines that the prediction result is correct if there is no manual intervention of the operator in a predetermined period before or after the prediction (for example, it may be set in the range of several tens of seconds to several tens of minutes). , Used as teacher data. The teacher data generation unit 32 makes a mistake in the prediction result if the operator performs manual intervention during a predetermined period before or after the prediction (for example, it may be set in the range of several tens of seconds to several tens of minutes). Use manual intervention for teacher data.

The re-learning unit 331 determines a combustion abnormality prediction program and / or an abnormality avoidance means prediction program using the learning data of the past predetermined period (for example, 30 days before the current day) and the teacher data of the past predetermined period. Relearn (tune) at the frequency of (for example, once a day). In this embodiment, performing daily tuning using the latest data is referred to as real-time learning. Further, as the teacher data, the teacher data selected by the method of FIG. 3 is used. The shorter the tuning time is, the more preferable it is, and examples thereof include within 1 hour, within 30 minutes, and within 10 minutes.
The second storage unit 341 stores a new combustion abnormality prediction program and / or an abnormality avoidance means prediction program relearned by the relearning unit 331. The past program stored in the second storage unit 341 may be stored as it is, or may be deleted at a predetermined timing. The same applies to the program of the first storage unit 34.
The abnormality prediction unit 35 makes a prediction using a new combustion abnormality prediction program and / or an abnormality avoidance means prediction program generated each time it is relearned.

(Display of forecast results)
The combustion control system 40 has a user interface, and the user interface operates with an abnormality determination input unit 42 that receives an input of an abnormality determination determined by the operator, and an intervention operation input unit 43 that receives an input of an intervention operation by the operator. Corresponding to the display of the log, it has a determination reason input unit 44 for inputting an operator's determination reason (reason for selecting a manually intervened operation item), and a display device 45 for displaying prediction result data. Specific input screens of the abnormality determination input unit 42 and the intervention operation input unit 43 are omitted.

FIG. 4 shows an example of the prediction result predicted by the combustion abnormality prediction program and / or the abnormality avoidance means prediction program.
The prediction result includes the date and time, the type of abnormality, the priority, the operation item, the change item, the difference in increase / decrease, the amount of change, the unit, and the like.
The priority indicates the level at which the abnormality avoidance means is recommended.
The combustion control unit 49 controls the combustion of the stoker type incinerator 1 based on the abnormality result and / or the abnormality avoidance means.
Here, the operator may select some of the abnormality avoidance means to perform the intervention operation, or each means may be weighted from 0% to 100% to perform the intervention operation, and the abnormality avoidance means. None of the means need to be used for the intervention operation. The operator's intervention operation is recorded as an operation log and used as teacher data.

FIG. 5 shows an example of the display of the operation log and the input screen of the determination reason selected by the operator.
In No. 2, the operator performed the intervention operation "as expected", so the operator performed the check "■ as expected".
In No. 1, there was an abnormality of "lowering the furnace temperature", and the operator performed an intervention operation to "increase" the "combustion air temperature", so the operator checked "■ lowered the furnace temperature".
The operator's judgment reason includes the item of "operation error" of the operator. Manual intervention operations related to "operation errors" are configured to be excluded from the teacher data used for real-time learning.

(Prediction accuracy)
The accuracy calculation unit 46 compares the prediction result of the abnormality avoidance means prediction program with the abnormality determination determined by the operator or the intervention operation of the operator, and calculates the accuracy of the prediction result.
FIG. 6A is an example for explaining a method of calculating the prediction accuracy. The prediction results of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program are associated with the operator's judgment (abnormality judgment, intervention operation) as follows. "Normal" is a state in which "abnormality determination or intervention operation" is not performed.
a: The prediction result is "abnormal" and the operator's judgment is "abnormal".
b: The prediction result is "abnormal" and the operator's judgment is "normal".
c: Prediction result is "normal", operator's judgment is "abnormal"
d: The prediction result is "normal" and the operator's judgment is "normal".
As an evaluation index of accuracy, the correct answer rate (the ratio of the data in which the prediction result and the operator's judgment match in all the data) and the recall rate (the data in which the operator's judgment was abnormal, both the prediction result and the operator's judgment) Percentage of data that is abnormal) is used.
Correct answer rate = (a + d) / (a + b + c + d) (1)
Recall rate = a / (a + c) (2)

In the present embodiment, the accuracy calculation unit 46 calculates the correct answer rate and the recall rate. FIG. 6B shows the correct answer rate and the recall rate as a result of the prediction accuracy in the first embodiment. The correct answer rate was 97% or more, and the recall rate was 66% or more.
The accuracy result output unit 47 outputs the accuracy result (correct answer rate, recall rate) calculated by the accuracy calculation unit 46 (for example, displayed on the display device 45).

(Separate embodiment)
The data server 20 may be omitted, and the function of the data server may be realized by the storage unit of the combustion abnormality prediction device 30 or the storage unit of the combustion control system 40.
Although the model generation unit 33 has generated the abnormality prediction program and the abnormality avoidance means prediction program, either one may be generated depending on the usage pattern.
The combustion abnormality prediction device 30 may be incorporated in the combustion control system 40 instead of a separate configuration.
The combustion control system 40 may store an abnormality prediction program and an abnormality avoidance means prediction program in a memory and have an abnormality prediction unit.
The combustion abnormality prediction device 30 may have the functions of an abnormality determination input unit, an intervention operation input unit, a determination reason input unit, an accuracy calculation unit, an accuracy result output unit, and a display device of the combustion control system 40.

(Embodiment 2)
The combustion control system 40 indicates to the operator "abnormality prediction" and / or "means for avoiding the abnormality", and assists the operator in determining whether or not to execute the means, and "avoids the abnormality". Of the fully automatic operation mode that automatically executes "means for avoiding abnormalities" and the part that executes some means or sets a weight of 0% to 100% among "means for avoiding abnormalities". It has an operation mode, and any operation mode can be selected.
The combustion control unit 49 controls the combustion of the stoker type incinerator according to the selected operation mode.

1 Stoker type incinerator 20 Data server 30 Combustion abnormality prediction device 31 Learning data generation unit 32 Teacher data generation unit 33 Model generation unit 34 First storage unit 331 Re-learning unit 341 Second storage unit 35 Abnormality prediction unit 36 Prediction result output unit 40 Combustion control system 42 Abnormality judgment input unit 43 Intervention operation input unit 44 Display device 45 Accuracy calculation unit 46 Accuracy result output unit

Claims (17)

Anomalous prediction programs that predict combustion anomalies generated by intelligent information processing technology using at least training data including state quantities for the combustion furnace and at least teacher data including manual intervention data of the operator, and / or at least combustion. Abnormalities that predict combustion anomalies generated by intelligent information processing technology using learning data including state quantities related to the furnace and at least teacher data including manual intervention data of operators, and predict means for avoiding the anomalies. The first storage unit that stores the workaround prediction program,
The combustion abnormality prediction program is used to predict whether or not the state quantity related to the combustion furnace is abnormal, and / or the abnormality avoidance means prediction program is used to predict whether or not the state quantity related to the combustion furnace is abnormal. And the anomaly predictor that predicts the means to avoid the anomaly,
A combustion abnormality prediction device including a prediction result output unit that outputs a result predicted by the abnormality prediction unit. A re-learning unit that relearns a combustion abnormality prediction program and / or an abnormality avoidance means prediction program at a predetermined frequency using the learning data in the past predetermined period and the teacher data in the past predetermined period.
It further has a new combustion abnormality prediction program and / or a second storage unit that stores the abnormality avoidance means prediction program relearned by the re-learning unit.
The combustion abnormality prediction device according to claim 1, wherein the abnormality prediction unit predicts using a new combustion abnormality prediction program and / or an abnormality avoidance means prediction program generated each time it is relearned. The combustion abnormality prediction device according to claim 1 or 2,
A combustion control system comprising a combustion control unit that controls combustion in a combustion furnace based on anomalous results and / or means for avoiding anomalies. A receiving unit that receives various data sent from the combustion abnormality prediction device according to claim 1 or 2, a storage unit that stores various data, and a storage unit.
A combustion control system comprising a combustion control unit that controls combustion in a combustion furnace based on the abnormal result and / or means for avoiding the abnormality. It has a user interface that accepts operator input.
The combustion control system according to claim 3 or 4, wherein the user interface has an abnormality determination input unit that receives an input for an abnormality determination determined by an operator. The combustion control system according to claim 5, wherein the user interface has an intervention operation input unit that receives an input of an intervention operation by an operator. The combustion control system according to claim 5 or 6, wherein the user interface has a determination reason input unit for inputting an operator's determination reason in response to display of an operation log. An assistive operation mode that indicates to the operator the means for predicting anomalies and / or avoiding the anomaly and determines whether or not the operator executes the means.
Fully automatic operation mode that automatically executes measures to avoid abnormalities,
It has one or more operation modes, including a partial operation mode in which some means are executed or a weighting of 0% to 100% is set and executed among the means for avoiding an abnormality.
Either operation mode is configured to be selectable,
The combustion control system according to any one of claims 3 to 7, wherein the combustion control unit controls combustion in a combustion furnace according to a selected operation mode. The combustion furnace is used with an anomaly prediction program that predicts combustion anomalies generated by intelligent information processing technology using at least learning data including state quantities related to the combustion reactor and teacher data including at least manual intervention data of the operator. Predict whether the state quantity with respect to is abnormal and / or
Predict combustion anomalies generated by intelligent information processing technology using at least training data including state quantities related to the combustion furnace and at least teacher data including manual intervention data of the operator, and predict the means for avoiding the anomalies. Anomaly prediction step and anomaly prediction step that predicts whether or not the state quantity related to the combustion furnace is abnormal and predicts the means for avoiding the abnormality by using the abnormality avoidance means prediction program.
A combustion abnormality prediction method including an output step for outputting the result predicted in the abnormality prediction step. Includes a re-learning step in which a combustion abnormality prediction program and / or an abnormality avoidance means prediction program is relearned at a predetermined frequency using the learning data in the past predetermined period and the teacher data in the past predetermined period. ,
The combustion abnormality prediction method according to claim 9, wherein the abnormality prediction step is predicted by using a new combustion abnormality prediction program and / or an abnormality avoiding means prediction program generated each time the abnormality is relearned. The prediction result of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program is compared with the judgment result of the operator, and one of them is selected to generate the teacher data, or there is no abnormality judgment judged by the operator. In this case, the prediction results of the combustion abnormality prediction program and / or the abnormality avoidance means prediction program are used as they are as training data.
The combustion abnormality prediction method according to claim 9 or 10, which includes a teacher data generation step. Any of claims 9 to 12, including an accuracy calculation step of comparing the prediction result of the abnormality avoidance means prediction program with the abnormality determination determined by the operator or the intervention operation of the operator to calculate the accuracy of the prediction result. The combustion abnormality prediction method according to item 1. With at least one processor
Includes memory for storing instructions that can be executed by the processor.
The processor is an information processing device that realizes the combustion abnormality prediction method according to any one of claims 9 to 12 by executing an executable instruction. It is a combustion abnormality prediction program
A program that realizes the combustion abnormality prediction method according to any one of claims 9 to 12 by using at least one processor. A computer-readable recording medium that stores computer instructions.
A computer-readable recording medium that realizes the step of the combustion abnormality prediction method according to any one of claims 9 to 12, when the computer instruction is executed by the processor. A training data generation step that generates training data including at least the state quantity related to the combustion furnace,
A teacher data generation step that generates teacher data, including at least manual intervention data for the operator,
An abnormality prediction program generation method including a model generation step of generating an abnormality prediction program for predicting a combustion abnormality generated by an intelligent information processing technique using the learning data and the teacher data. A training data generation step that generates training data including at least the state quantity related to the combustion furnace,
A teacher data generation step that generates teacher data, including at least manual intervention data for the operator,
A model generation step of predicting a combustion abnormality generated by an intelligent information processing technique using the learning data and the teacher data and generating an abnormality avoidance means prediction program for predicting means for avoiding the abnormality. Abnormal avoidance means prediction program generation method including.

Images