JP2021137747A - Operation control input derivation device for water treatment facility - Google Patents

Abstract

To provide a technique capable of accurately deriving control input, even when the number of pieces of data accumulated is small.SOLUTION: A control input derivation device (100) comprises: a learning part (110) for acquiring together, both of learning object data and estimation object data, from trend data and water quality data, and performing learning by data obtained by compressing, both of data with label and data without label, included in the learning object data and the estimation object data, for generating a learning parameter; and a control input derivation part (122) for deriving control input of a control object, using an explanatory variable of the estimation object included in the compressed estimation object data, and the learning parameter.SELECTED DRAWING: Figure 2

Description

The present invention relates to an operation amount derivation device for a water treatment facility.

Conventionally, water quality control in a water treatment facility has been performed by an operator having skill in consideration of various situations in the water treatment facility.
However, since it takes a long time for the aging of the operator and the mastery of the technique, there is a demand to determine the operation amount using artificial intelligence and automate it.
Here, the manipulated variable is mainly determined based on water quality data such as the concentration of ammonia contained in the discharged water.

In Patent Document 1, which is an example of the prior art, by automatically adjusting model parameters, even when an operator without specialized knowledge operates a sewage treatment plant, the operating efficiency, operating cost, etc. are significantly reduced. The challenge is to make it possible to keep the quality of the inflow water within the specified range, and while accumulating the measurement data output from the sewage treatment process and the analysis data obtained by the analysis operation of the water quality analyst, etc. in the database device. , Model parameters set in the model parameter setting unit, plant conditions set in the plant condition setting unit, operating conditions set in the operating condition setting unit, simulation data output from the database device, parameter identification A technique for automatically adjusting model parameters to a process simulator based on data and executing a simulation of a sewage treatment process is disclosed.

JP-A-2007-229550

However, the above-mentioned prior art has a large amount of time-series storage for simulation, including sensor information obtained from sensors installed in the sewage treatment process and water quality analysis result information acquired at the sewage treatment plant. There was a problem that it required the data of.
On the other hand, water quality analysis for acquiring water quality data in a water treatment facility requires cost and time, so it takes a large amount of cost to prepare a large amount of water quality data, and the data acquisition period is long. There was a problem.

The present invention has been made in view of the above, and an object of the present invention is to provide a technique capable of deriving an operation amount with high accuracy even when the amount of accumulated data is small.

The present invention, which solves the above-mentioned problems and achieves the object, has a data storage unit that stores trend data and water quality data, and learning target data and estimation target data from the trend data and water quality data stored in the data storage unit. Both the learning target data and the estimation target data acquisition unit that acquire both of them, and the learning target data and the labeled data and the unlabeled data included in the estimation target data from the learning target data and the estimation target data acquisition unit are both acquired. A data compression unit to be compressed, a learning target data acquisition unit that acquires a portion of the compressed data compressed by the data compression unit that corresponds to the learning target data, and compressed learning from the learning target data acquisition unit. An estimation model generation unit that generates a learning model and learning parameters based on the learning model by performing semi-supervised learning using the target data, and a learning parameter storage unit that stores the learning parameters generated by the estimation model generation unit. A unit, an estimation target data acquisition unit that acquires a portion of the compressed data compressed by the data compression unit that corresponds to the estimation target data, and the estimation of the compressed data from the estimation target data acquisition unit. Water including an explanatory variable of an estimation target included in a portion corresponding to the target data, an operation amount derivation unit for deriving an operation amount of a control target using the learning parameter stored in the learning parameter storage unit, and an operation amount derivation unit. It is an operation amount derivation device for processing facilities.

Alternatively, the present invention has a data storage unit that stores trend data and water quality data, and training target data that acquires training target data used for updating model parameters of an estimation model from the trend data and water quality data stored in the data storage unit. The acquisition unit, the estimation target data acquisition unit that acquires the estimation target data used for deriving the operation amount from the trend data and the water quality data accumulated in the data storage unit, and the learning target data from the learning target data acquisition unit. A model parameter update unit that updates the model parameters of the estimated model by learning using it, a learning parameter storage unit that stores the model parameters updated in the model parameter update unit as learning parameters, and the estimation target data. It is an operation amount derivation device of a water treatment facility including the operation amount derivation unit for deriving the operation amount of a controlled object using the learning parameter.

In the present invention having the above configuration, it is preferable that a data conversion unit for compressing the dimension of data by principal component analysis is provided, and the model parameter update unit performs semi-supervised learning by partial least squares.

In the present invention having the above configuration, the model parameter update unit may perform semi-supervised learning of the transductive SVM.

In the present invention having the above configuration, the model parameter update unit may perform graph-based semi-supervised learning.

According to the present invention, even when the amount of stored data is small, the amount of operation can be derived with high accuracy.

FIG. 1 is a diagram showing a configuration of a water treatment facility to which the operation amount derivation device of the water treatment facility according to the first embodiment can be applied. FIG. 2 is a block diagram showing a configuration of an operation amount derivation device of the water treatment facility according to the first embodiment. FIG. 3 is a diagram for explaining an operation image of the operation amount derivation device from the learning target data and the estimation target data to the output of the operation amount. FIG. 4 is a flowchart showing the operation corresponding to FIG. FIG. 5 is a block diagram showing a configuration of an operation amount derivation device of the water treatment facility according to the second embodiment. FIG. 6 is a flowchart showing the operation of the operation amount derivation device from the learning target data and the estimation target data to the output of the operation amount. FIG. 7 is a block diagram showing a configuration of a model parameter update unit. FIG. 8 is a flowchart showing the flow of semi-supervised learning by self-training performed by the model parameter update unit. FIG. 9 is a diagram illustrating an operation image of semi-supervised learning by self-training in the second embodiment. FIG. 10 is a diagram illustrating an operation image of semi-supervised learning of two-class classification by a transductive SVM in the third embodiment. FIG. 11 is a diagram illustrating an operation image of semi-supervised learning of two-class classification based on a graph in the fourth embodiment.

Hereinafter, embodiments for carrying out the present invention will be described with reference to the accompanying drawings.
However, the present invention is not limited to the description of the following embodiments.

(Embodiment 1)
FIG. 1 is a diagram showing a configuration of a water treatment facility 200 to which the operation amount derivation device 100 of the water treatment facility according to the present embodiment can be applied.
The water treatment facility 200 shown in FIG. 1 includes a first settling basin 1, a reaction tank 2, a final settling basin 3, a blower 4, an air volume adjusting valve 5, an air diffuser 6, a first pump 7, and the like. A second pump 8, a measuring instrument 9, a gravity concentration tank 10, a mechanical concentration tank 11, a digestion tank 12, a dehydration tank 13, and pipes 21, 22, 23, 24 are provided.

The first settling basin 1 is a settling basin into which raw water is introduced.
This raw water is wastewater containing organic matter.
In the first settling basin 1, solid-liquid separation of raw water is performed, and the outflow water from the first settling basin 1 is sent to the reaction tank 2 through the pipe 21.
The raw sludge, which is the sludge first settled in the settling basin 1, is sent to the gravity concentrating tank 10 through the pipe 24.

The reaction tank 2 is a tank containing microorganisms and first purifies the effluent from the sedimentation basin 1 by the microorganisms.
In the reaction tank 2, the microorganism first proliferates by assimilating the organic matter contained in the effluent from the sedimentation basin 1, and activated sludge is formed by biological treatment using the microorganism.
The effluent from the reaction tank 2 is sent to the final settling basin 3 through the pipe 22.

The final settling basin 3 is a settling basin that precipitates activated sludge contained in the effluent from the reaction tank 2.
The supernatant of the final settling basin 3 is discharged to the outside of the water treatment facility 200 as treated water.
A part of the sludge settled in the final settling basin 3 is returned to the reaction tank 2 through the pipe 23 by the first pump 7 and reused.
The remaining sludge settled in the final settling basin 3 is sent to the mechanical concentration tank 11 by the second pump 8 as excess sludge.

The blower 4 supplies air to the plurality of air diffusers 6.
The air volume adjusting valve 5 is provided in a pipe that passes through each of the plurality of air diffusers 6, and adjusts the air volume by opening and closing.
The plurality of air diffusers 6 are provided in the lower part of the reaction tank 2, and are connected to a pipe passed through the air volume adjusting valve 5, and the air whose air volume is adjusted by the air volume adjusting valve 5 is introduced into the reaction tank 2. Supply to.
When the amount of air blown to the reaction tank 2 is adjusted in this way, the DO (Dissolved Oxygen) value, which is the amount of dissolved oxygen in the reaction tank 2, is adjusted, and the progress of biological treatment is adjusted.

The first pump 7 is a return sludge pump that sends a part of the sludge settled in the final settling basin 3 to the reaction tank 2 through the pipe 23.
The second pump 8 is a surplus sludge extraction pump that sends the remaining sludge settled in the final settling basin 3 as surplus sludge to the mechanical concentration tank 11.

The measuring instrument 9 is a measuring instrument that measures each parameter indicating the water quality of the reaction tank 2, and the measured value data, which is the measured parameter, is sent to the manipulated variable derivation device 100.
Here, as each parameter indicating the water quality, a DO value which is the amount of dissolved oxygen and an MLSS (Mixed Liquor Suspended Solids) value which is the concentration of suspended solids can be exemplified.
The operator of the water treatment facility 200 determines the amount of operation to be controlled by referring to each parameter indicating the water quality measured by the measuring instrument 9.
Here, the operation amount to be controlled includes the rotation speed of the first pump 7 for adjusting the amount of sludge returned from the final settling pond 3 of the water treatment facility 200, and the withdrawal of excess sludge from the final settling pond 3 of the water treatment facility 200. Examples can be made of the extraction amount per unit time of the second pump 8 for adjusting the amount, the extraction time of excess sludge, and the injection rate of the polymer flocculant into the dehydration tank 13 of the water treatment facility 200.

The gravity concentration tank 10 is a tank for concentrating the raw sludge that first settled in the settling basin 1.
The mechanical concentration tank 11 is a tank that concentrates the excess sludge drawn from the final settling basin 3 by the second pump 8.
The sludge concentrated in the gravity concentration tank 10 and the mechanical concentration tank 11 is sent to the digestion tank 12.

The digestion tank 12 is a tank for digesting the concentrated sludge.
Here, the digestive treatment may be performed by, for example, an anaerobic digestive treatment method.
In the anaerobic digestion method, organic sludge is decomposed by anaerobic microorganisms.
The sludge decomposed by the digestion process is sent to the dehydration tank 13.

The dehydration tank 13 is a tank that reduces the water content of the sludge and reduces the volume by dehydrating the sludge decomposed by the digestion treatment.

The pipe 21 is initially arranged between the settling basin 1 and the reaction tank 2, and is a pipe that sends the outflow water from the first settling basin 1 to the reaction tank 2.
The pipe 22 is arranged between the reaction tank 2 and the final settling basin 3, and is a pipe that sends the outflow water from the reaction tank 2 to the final settling basin 3.
The pipe 23 is arranged between the first pump 7 and the reaction tank 2, and is a pipe that sends a part of the sludge of the final settling basin 3 to the reaction tank 2.
The pipe 24 is arranged between the first settling basin 1 and the gravity concentrating tank 10, and is a pipe that sends the raw sludge of the first settling basin 1 to the gravity concentrating tank 10.

In the water treatment facility 200 shown in FIG. 1, the main operation items are the inflow amount, which is the amount of water from the first settling pond 1 to the reaction tank 2, the air flow amount, which is the amount of air supplied into the reaction tank 2, and the first. The amount of returned sludge, which is the amount of sludge returned to the reaction tank 2 by the pump 7, and the amount of excess sludge, which is the amount of sludge drawn from the final settling pond 3 by the second pump 8.
The settings of each of these operation items differ depending on the treatment plant.

The reaction tank 2 can be controlled by, for example, constant air flow control, constant ratio control, and constant DO control.
Here, the constant air flow control is a control performed so as to have a constant air flow amount set as a target air flow rate value.
Further, the constant ratio control is a control performed so that the amount of air blown from the settling basin 1 to the reaction tank 2 is initially set to be constant, that is, the ratio of the amount of inflow to the amount of air blown is constant.
Further, the constant DO control is a control performed so that the DO value of the reaction tank 2 becomes the set target DO value.

The main operation items are the rotation speed of the first pump 7 for adjusting the amount of returned sludge, the amount of withdrawal per unit time or the time for withdrawing excess sludge in adjusting the amount of excess sludge withdrawal, and in the dehydration treatment. The injection rate of the polymer flocculant.
The operation amount derivation device 100 according to the present embodiment targets the operation amount when determining these operation items.
Then, the operator operates the controlled object by the operation amount derived by the operation amount derivation device 100 of the water treatment facility.
By determining the manipulated variable to be controlled using the manipulated variable derivation device 100 in this way, it is possible to use a person who does not have intuition, experience, and know-how as an operator.

FIG. 2 is a block diagram showing a configuration of an operation amount derivation device 100 of a water treatment facility according to the present embodiment.
The operation amount derivation device 100 of the water treatment facility shown in FIG. 2 includes a data storage unit 101, a learning target data and estimation target data acquisition unit 102, a data compression unit 103, a learning unit 110, and an operation amount derivation unit 120. , Equipped with.

The data storage unit 101 acquires the measured value data measured by the measuring instrument 9 and stores it as trend data and water quality data.
As trend data, the amount of raw sludge extracted, which is the amount of sludge extracted in the first settling pond 1, the MLSS in the reaction tank 2, and the return sludge which is a part of the sludge settled in the final settling pond 3 sent to the reaction tank 2 are returned. The sludge concentration, the return sludge flow rate, which is the flow rate of the return sludge, the excess sludge withdrawal amount, which is the amount of sludge extracted from the final settling pond 3, the rotation speed of the first pump 7, which is the return sludge pump, and the flow rate of the excess sludge. Excess sludge flow rate and the final sedimentation pond sludge interface, which is the water depth position at the boundary between the accumulated sludge and the supernatant in each layer, can be exemplified.
As the water quality data, NH ₄ concentration, NO ₂ concentration, NO ₃ concentration and PO ₄ concentration can be exemplified.
The data storage unit 101 can be realized by a recording medium such as a semiconductor memory or a magnetic disk.

The learning target data and the estimation target data acquisition unit 102 acquires both the learning target data and the estimation target data from the trend data and the water quality data accumulated in the data storage unit 101.
The learning target data and estimation target data acquisition unit 102 can be realized by combining a processor such as an MPU (Micro-Processing Unit) or a CPU (Central Processing Unit) with a recording medium such as a semiconductor memory or a magnetic disk. can.

The data compression unit 103 reduces the dimension of the data by compressing the learning target data and the learning target data and the estimation target data from the estimation target data acquisition unit 102 together.
The data compression unit 103 can be realized by combining a processor such as an MPU or a CPU and a recording medium such as a semiconductor memory or a magnetic disk.

The learning unit 110 includes a learning target data acquisition unit 111, an estimation model generation unit 112, and a learning parameter storage unit 113.
The learning target data acquisition unit 111 acquires a portion corresponding to the learning target data from the compressed data compressed by the data compression unit 103.
The estimation model generation unit 112 generates a learning model and learning parameters based on the learning model using the compressed learning target data from the learning target data acquisition unit 111.
The learning parameter storage unit 113 stores the learning parameters from the estimation model generation unit 112.
The learning target data acquisition unit 111 and the estimation model generation unit 112 can be realized by combining a processor such as an MPU or a CPU and a recording medium such as a semiconductor memory or a magnetic disk.
Further, the learning parameter storage unit 113 can be realized by a recording medium such as a semiconductor memory or a magnetic disk.

The operation amount derivation unit 120 includes an estimation target data acquisition unit 121, an operation amount derivation unit 122, and an operation amount output unit 123, and outputs an operation amount of a control target derived using the estimation target data and learning parameters. do.
The estimation target data acquisition unit 121 acquires a portion corresponding to the estimation target data from the compressed data compressed by the data compression unit 103.
The manipulated variable derivation unit 122 uses the explanatory variables of the estimation target included in the compressed estimation target data from the estimation target data acquisition unit 121 and the learning parameters stored in the learning parameter storage unit 113 to control the target. Derivation of the manipulated variable.
The estimation target data acquisition unit 121 and the operation amount derivation unit 122 can be realized by combining a processor such as an MPU or a CPU and a recording medium such as a semiconductor memory or a magnetic disk.
The operation amount output unit 123 presents the operation amount of the control target to the operator by outputting the operation amount of the control target derived by the operation amount derivation unit 122.
It is also possible that the operation amount output unit 123 outputs the operation amount to be controlled to the operation device, and the operation device automatically operates based on this operation amount.

Here, the operation of the manipulated variable derivation device 100 will be described.
FIG. 3 is a diagram for explaining an operation image of the operation amount derivation device 100 from the learning target data and the estimation target data to the output of the operation amount.
FIG. 4 is a flowchart showing the operation corresponding to FIG.
In FIG. 3, principal component analysis (PCA: Principal Component Analysis) is used for data compression, and partial least squares (PLS: Partial Least Squares) is used for the learning model.

In the learning process, first, trend data and water quality data are stored in the data storage unit 101 (S1).
Then, the learning target data and the estimation target data acquisition unit 102 acquire both the learning target data and the estimation target data from the data accumulated in the data storage unit 101 (S2).
Here, it is important to acquire both the labeled data and the unlabeled data for both the trend data and the water quality data.
The labeled data are water quality data and trend data corresponding to the manipulated variable actually set by the operator, and include both the objective function and the explanatory variables.
The unlabeled data are water quality data and trend data that do not correspond to the manipulated variable, and do not include the objective function but include explanatory variables.
The water quality data and the trend data that do not correspond to the manipulated variable are, for example, the data in the latest estimation time immediately after the introduction of the system.

The data compression unit 103 reduces the dimension by compressing all the data including both the labeled data and the unlabeled data acquired by the learning target data and the estimation target data acquisition unit 102 by the PCA (S3).
As a result, the compressed data obtained after compression reflects not only the past data with the label corresponding to the operation amount but also the latest unlabeled data not corresponding to the operation amount.
By compressing in this way, the latest water quality data and trend data can be reflected in the estimation of the amount of operation after the introduction of the system.
Next, the learning target data acquisition unit 111 acquires a portion corresponding to the labeled data from the compressed data compressed by the data compression unit 103 (S4).
Next, the estimation model generation unit 112 generates an estimation model by PLS using the portion corresponding to the labeled data acquired by the learning target data acquisition unit 111 as an explanatory variable (S5).
Next, the learning parameter storage unit 113 stores the parameters of the estimation model generated by the estimation model generation unit 112 as learning parameters (S6).

The estimation target data acquisition unit 121 acquires a portion to be estimated from the compressed data compressed by the data compression unit 103 (S7).
The operation amount derivation unit 122 derives the operation amount of the control target by inputting the estimation target data into the estimation model using the learning parameter stored in the learning parameter storage unit 113 (S8).
The derived manipulated variable is output from the manipulated variable output unit 123 (S9) and used for the controlled operation.

In general, water quality analysis for acquiring water quality data in a water treatment facility requires cost and time, and the number of water quality data that can be acquired is limited to less than 100 per year.
In addition, since water quality data changes due to seasonal changes and changes in the operation policy of water treatment facilities, when deriving the amount of operation by machine learning of artificial intelligence, appropriate teacher data at that time is both quality and quantity. There is concern that there will be a shortage in.
Estimates in situations where appropriate teacher data are lacking include both online data such as sludge flow and sludge concentration in water treatment facilities and water quality data obtained by water quality analysis, even after the learning model has been constructed and introduced. It is necessary to incorporate as much as possible.
As described in this embodiment, by applying semi-supervised learning, even in a situation where appropriate teacher data is insufficient or when the amount of accumulated data is small, it is possible to achieve high accuracy in real time. It is possible to derive the operation amount.

(Embodiment 2)
In the first embodiment, a mode using PCA and PLS as semi-supervised learning will be described, and the present invention is not limited thereto. In the present embodiment, self-training is performed as semi-supervised learning. The form using the above will be described.

In this embodiment, discretized values are assumed for the objective variable, which is the manipulated variable to be estimated.
For example, for the DO set value, it is common to adjust the operation amount in units of 0.1 mg / L.
Therefore, here, there are three values, that is, (-0.1) indicating that the operation amount is decreased, (0) indicating that the status quo is maintained, and (+0.1) indicating that the operation amount is increased. , (-0.2) indicating that the manipulated variable is further reduced, and (+0.2) indicating that the manipulated variable is further increased. It can be treated as.
In addition, the amount of excess sludge withdrawn, the amount of raw sludge withdrawn, etc. are generally set as an integer value within a certain range, and can be treated as a multi-class classification problem in the same way as the DO set value. be.

FIG. 5 is a block diagram showing the configuration of the operation amount derivation device 150 of the water treatment facility according to the present embodiment.
The operation amount derivation device 150 of the water treatment facility shown in FIG. 5 is estimated to include a data storage unit 151, a data conversion unit 152, a learning target data acquisition unit 161, a model parameter update unit 162, and a learning parameter storage unit 163. It includes a target data acquisition unit 171, an operation amount derivation unit 172, and an operation amount output unit 173.

The data storage unit 151 acquires the measured value data measured by the measuring instrument 9 and stores it as trend data and water quality data.
As trend data, the amount of raw sludge extracted, which is the amount of sludge extracted in the first settling pond 1, the MLSS in the reaction tank 2, and the return sludge which is a part of the sludge settled in the final settling pond 3 sent to the reaction tank 2 are returned. The sludge concentration, the return sludge flow rate, which is the flow rate of the return sludge, the excess sludge withdrawal amount, which is the amount of sludge extracted from the final settling pond 3, the rotation speed of the first pump 7, which is the return sludge pump, and the flow rate of the excess sludge. Excess sludge flow rate and the final sedimentation pond sludge interface, which is the water depth position at the boundary between the accumulated sludge and the supernatant in each layer, can be exemplified.
The water quality data, can be exemplified NH ₄ concentration.
The data storage unit 151 can be realized by a recording medium such as a semiconductor memory or a magnetic disk.

The learning target data acquisition unit 161 acquires the training data necessary for updating the model parameters of the estimation model from the water quality data and the trend data accumulated in the data storage unit 151.
The data conversion unit 152 processes and converts the learning target data from the learning target data acquisition unit 161 into learnable data, and processes and converts the estimation target data from the estimation target data acquisition unit 171 into estimable data.
The model parameter update unit 162 updates the model parameters of the estimation model using the converted learning target data from the data conversion unit 152, and generates model parameters as learning parameters.
The learning parameter storage unit 163 stores the learning parameters from the model parameter update unit 162.
The learning target data acquisition unit 161 and the model parameter update unit 162 can be realized by combining a processor such as an MPU or a CPU and a recording medium such as a semiconductor memory or a magnetic disk.
Further, the learning parameter storage unit 163 can be realized by a recording medium such as a semiconductor memory or a magnetic disk.

The estimation target data acquisition unit 171 acquires the estimation target data necessary for deriving the manipulated variable from the water quality data and the trend data accumulated in the data storage unit 151.
The operation amount derivation unit 172 derives the operation amount of the control target by using the converted estimated target data from the data conversion unit 152 and the learning parameters stored in the learning parameter storage unit 163.
The estimation target data acquisition unit 171 and the operation amount derivation unit 172 can be realized by combining a processor such as an MPU or a CPU and a recording medium such as a semiconductor memory or a magnetic disk.
The operation amount output unit 173 presents the operation amount of the control target to the operator by outputting the operation amount of the control target derived by the operation amount derivation unit 172.
It is also possible that the operation amount output unit 173 outputs the operation amount to be controlled to the operation device, and the operation device automatically operates based on this operation amount.

Here, the operation of the manipulated variable derivation device 150 will be described.
FIG. 6 is a flowchart showing the operation of the operation amount derivation device 150 from the learning target data and the estimation target data to the output of the operation amount.

In the learning process, first, trend data and water quality data are accumulated in the data storage unit 151 (S11).
Then, the learning target data acquisition unit 161 acquires the learning target data from the data accumulated in the data storage unit 151 (S12).
Here, it is important to acquire both the labeled data and the unlabeled data for both the trend data and the water quality data.
The labeled data are water quality data and trend data corresponding to the manipulated variable actually set by the operator, and include both the objective function and the explanatory variables.
The unlabeled data are water quality data and trend data that do not correspond to the manipulated variable, and do not include the objective function but include explanatory variables.
The water quality data and the trend data that do not correspond to the manipulated variable are, for example, the data in the latest estimation time immediately after the introduction of the system.

The data conversion unit 152 processes and converts the learning target data into a learnable form among all the data including both the labeled data and the unlabeled data acquired by the learning target data acquisition unit 161 (S13).
Here, the processing and conversion of the data may be performed as needed, and as the processing and conversion method, compression of the data dimension by principal component analysis, factor analysis, independent component analysis, or the like can be exemplified.
Next, the model parameter update unit 162 updates the parameters of the estimation model by performing semi-supervised learning with the learning target data in a learnable form (S14).
Here, the form of the parameters of the estimation model differs depending on the learning method used.
Next, the learning parameter storage unit 163 stores the parameters of the estimated model updated by the model parameter updating unit 162 as learning parameters (S15).

Next, the estimation target data acquisition unit 171 acquires the estimation target data from the data accumulated in the data storage unit 151 (S16).
The data conversion unit 152 processes and converts the estimation target data into an estimable form among all the data including both the labeled data and the unlabeled data acquired by the estimation target data acquisition unit 171 (S17).
The method of processing and converting into a presumable shape is the same as the method of processing and converting into a learnable shape.
The operation amount derivation unit 172 derives the operation amount of the control target by inputting the estimation target data into the estimation model using the learning parameter stored in the learning parameter storage unit 163 (S18).
The derived manipulated variable is output from the manipulated variable output unit 173 (S19) and used for the controlled operation.

Here, in the present embodiment, self-training is applied as semi-supervised learning performed by the model parameter updating unit 162.
In self-training, the estimation accuracy P (y | x, θ) is set on the premise that the manipulated variable is derived by a classifier that outputs the estimation result and the estimation accuracy indicating the reliability of the estimation result. The discrimination is repeated recursively while sequentially incorporating the data having a value or more into the learning data.

FIG. 7 is a block diagram showing the configuration of the model parameter update unit 162.
The model parameter update unit 162 shown in FIG. 7 includes a data extraction unit 1621, a data presence / absence determination unit 1622, an operation amount estimation unit 1623, an estimation accuracy determination unit 1624, a labeling unit 1625, and an update parameter generation unit 1626. , Equipped with.
FIG. 8 is a flowchart showing the flow of semi-supervised learning by self-training performed by the model parameter updating unit 162.
First, the data extraction unit 1621 extracts unlabeled data from the training data (S21), and the data presence / absence determination unit 1622 determines whether or not there is unlabeled data (S22).
When there is unlabeled data (S22: Y), the manipulated variable estimation unit 1623 estimates all the manipulated variables of the unlabeled data (S23), and the estimated accuracy determination unit 1624 estimates the estimated accuracy is equal to or higher than the threshold value. It is determined whether or not there is the data of (S24).
The model parameter update unit 162 ends learning when there is no unlabeled data (S22: N).
When there is data whose estimated accuracy is equal to or higher than the threshold value (S24: Y), the labeling unit 1625 assigns a label to the data whose estimated probability is equal to or higher than the threshold value and adds the data to the training data (S25). When the update parameter generation unit 1626 generates the update parameter, the model parameter is updated (S26) and returns to S1.
The model parameter update unit 162 ends learning when there is no data whose estimated accuracy is equal to or greater than the threshold value (S24: N).

FIG. 9 is a diagram illustrating an operation image of semi-supervised learning by self-training in the present embodiment.
In FIG. 9, three DO setting values of −0.1 mg / L, ± 0.0 mg / L, and +0.1 mg / L are assumed.
Here, for example, a logistic regression model is classified as a model according to the classification of (-1) indicating that the operation amount is decreased, (0) indicating that the status quo is maintained, and (+1) indicating that the operation amount is increased. When selected, the estimated accuracy P (y | x, θ) is expressed by the following equation (1).

When the estimated accuracy P (y = 1 | x, θ) is equal to or greater than the set threshold value, it is added to the training data as class 1 data to increase the operation amount, and the estimated accuracy P (y = -1 | x, θ) is added. ) Is above the set threshold, the operation amount is reduced. It is added to the training data as class-1 data, and when the estimated accuracy P (y = 0 | x, θ) is above the set threshold. Is added to the training data as class 0 data that does not change the manipulated variable.
In this way, by adding data with high estimation accuracy to the labeled training data and repeating re-learning, the reliability of the estimation model generated from the few labeled training data in the initial state shown on the left side of FIG. 9 can be improved. As the number increases, a highly reliable model generated from many labeled training data is constructed when the training shown on the right side of FIG. 9 is completed.

As described above, by using self-training as semi-supervised learning, it is possible to derive the amount of operation with high accuracy even when the amount of accumulated data is small.

(Embodiment 3)
In the first embodiment, a mode using PCA and PLS as semi-supervised learning has been described, and in the second embodiment, a mode using self-training as semi-supervised learning has been described, but the present invention is limited thereto. is not it.
In this embodiment, a mode in which a transductive SVM (Transductive SVM (Support Vector Machine)) is used as semi-supervised learning will be described.
The manipulated variable derivation device according to the present embodiment is the manipulated variable derivation device 150 shown in FIG. 5, and only the semi-supervised learning method is different.

In this embodiment, for the sake of simplification of the description, it is shown that the operation amount is decreased (DO set value −0.1 mg / L) and that the operation amount is increased (DO set value + 0.1 mg / L). There are two classes, L) and.

FIG. 10 is a diagram illustrating an operation image of semi-supervised learning of two-class classification by a transductive SVM in the present embodiment.
Here, the transducible SVM is applied to the labeled data including the objective function and the explanatory variables and the unlabeled data not including the objective function and including the explanatory variables.
If necessary, the data conversion unit 152 may process and convert the data as a preprocessing for applying the transductive SVM.
Here, as a processing and conversion method, compression of data dimensions by principal component analysis, factor analysis, independent component analysis, or the like can be exemplified.
Further, FIG. 10A shows an initial state including a separation boundary surface using only labeled data, and FIG. 10B shows a learning completion time including a separation boundary surface using both labeled data and unlabeled data. It is shown.

In learning all target data is normal is labeled data SVM, when learning data _{(x i, y i) (} y i ∈ {-1,1}, 1 ≦ i ≦ n) is given, the following The parameters w and b indicating the position of the separation boundary surface are determined by the minimization problem represented by the equation (2).

Where || w || represents the Euclidean norm of w, φ (x _{i ) represents the mapping of x i} to the higher dimensional space determined by the SVM kernel, and C represents the soft margin parameter.
In the transductive SVM, the function to minimize includes unlabeled data as shown in the following equation (3).

Here, the training data (x _i , y _i ) (y _i ∈ {-1,1}, 1 ≦ i ≦ l) represents the labeled data associated with the operation amount, and x _i (l + 1 ≦ i). ≤n) represents unlabeled data that is not associated with the manipulated data, and C ₁ and C ₂ represent hyperparameters.
The third term in the above equation (3), for the unlabeled data parameter as output _{w · φ (x i) +} b of the discriminant function approaches 1 or -1 is determined.
That is, for data for which the label is unknown, the parameters w and b of the separation boundary surface are determined so that the determination results are clearly separated.
As a result, it is possible to correctly discriminate the label for the labeled data and to reflect the distribution of the unlabeled data for the discrimination.
Hyperparameters C ₁ and C ₂ adjust the weights of labeled data or unlabeled data, respectively. When hyperparameter C ₁ is increased, the operation amount with emphasis on labeled data is derived, and hyperparameter C _{When 2} is increased, the operation amount with emphasis on unlabeled data is derived.

As explained above, by using Transductive SVM (Transductive SVM (Support Vector Machine)) for semi-supervised learning, even when the amount of accumulated data is small, the amount of operation can be derived with high accuracy. can.

(Embodiment 4)
The first embodiment describes a form using PCA and PLS as semi-supervised learning, the second embodiment describes a form using self-training as semi-supervised learning, and the third embodiment explains as semi-supervised learning. Although the form using the transductive SVM has been described, the present invention is not limited thereto.
In this embodiment, graph-based semi-supervised learning is used as semi-supervised learning.
The manipulated variable derivation device according to the present embodiment is the manipulated variable derivation device 150 shown in FIG. 5, and only the semi-supervised learning method is different.

Here, in the present embodiment, the graph-based semi-supervised learning is applied as the semi-supervised learning performed by the model parameter updating unit 162.
In this embodiment, for the sake of simplification of the description, as in the case of the third embodiment, it is shown that the operation amount is reduced (DO set value −0.1 mg / L) and that the operation amount is increased (DO setting value −0.1 mg / L). There are two classes, DO set value + 0.1 mg / L).

FIG. 11 is a diagram illustrating an operation image of semi-supervised learning of two-class classification based on a graph in the present embodiment.
Here we apply graph-based semi-supervised learning to all data, including labeled and unlabeled data.
If necessary, the data conversion unit 152 may process and convert the data as a preprocessing to apply the graph-based semi-supervised learning.
Further, FIG. 11A shows a state before estimation, and FIG. 11B shows a state after estimation.

First, the model parameter update unit 162 calculates the similarity between the data for all the data including the labeled data and the unlabeled data, and the similarity between the data becomes the model parameter.

Here, an example of calculating the degree of similarity will be described.
The model parameter update unit 162 has two vectors x _i = (x _i ¹ , x _i ² , ..., x _i ^D ) and x _j = (x _j ¹ , x _j ² , ...) From the learning target data acquisition unit 161. , X _j ^D ), the similarity _wij is defined by the following equation (4).

Here, σ _d (1 ≦ d ≦ D) represents a hyperparameter.
In this way, the similarity obtained by the model parameter update unit 162 by learning is stored in the learning parameter storage unit 163 as a learning parameter.
The model parameter update unit 162 propagates the label information of the labeled data to the unlabeled data by the parameters stored in the learning parameter storage unit 163, and estimates the label of the unlabeled data.
Specifically, suppose that the labeled data is (x _i , y _i ) (y _i ∈ {-1,1}, 1 ≦ i ≦ l) and the unlabeled data is x _i (l + 1 ≦ i ≦ l + u). , The matrix W representing the similarity and the diagonal matrix D are represented by the following equation (5).

At this time, the label of the unlabeled data is estimated by the following equation (6).

However, f _l = (y _i ) (y _i ∈ {-1,1}, 1 ≦ i ≦ l) is an l-dimensional vector consisting of labeled data, and in particular, when a binary manipulated variable is derived. the, it is possible to determine the sign of f _u.
Estimation result f _u for unlabeled data in the above equation (6) is derived by minimizing the L of the formula (7).

That is, the amount of operation that is the label of the unlabeled data is determined so that the data having a high degree of similarity as a whole of the data including the labeled data and the unlabeled data has the same label.

As explained above, by using Graph-based semi-supervised learning as semi-supervised learning, it is possible to derive the amount of operation with high accuracy even when the amount of accumulated data is small. ..

In the first to fourth embodiments, when processing all the data including both the labeled data and the unlabeled data, it is not always necessary to process all the data, and both the labeled data and the unlabeled data are included. If it is possible to process it, a part of all the data may be missing.

Further, the present invention is not limited to the above-described embodiment, and includes various modifications in which components are added, deleted, or converted from the above-described configuration.

1 First settling pond 2 Reaction tank 3 Final settling pond 4 Blower 5 Adjusting valve 6 Air diffuser 7 First pump 8 Second pump 9 Measuring instrument 10 Gravity concentration tank 11 Mechanical concentration tank 12 Digestion tank 13 Dehydration tank 21 and 22 , 23, 24 Piping 100 Operation amount derivation device 101 Data storage unit 102 Learning target data and estimation target data acquisition unit 103 Data compression unit 110 Learning unit 111 Learning target data acquisition unit 112 Estimated model generation unit 113 Learning parameter storage unit 120 Operation amount Derivation unit 121 Estimated target data acquisition unit 122 Operation amount derivation unit 123 Operation amount output unit 150 Operation amount derivation device 151 Data storage unit 152 Data conversion unit 161 Learning target data acquisition unit 162 Model parameter update unit 163 Learning parameter storage unit 171 Estimation target Data acquisition unit 172 Operation amount derivation unit 173 Operation amount output unit 200 Water treatment facility 1621 Data extraction unit 1622 Data presence / absence judgment unit 1623 Operation amount estimation unit 1624 Estimation accuracy judgment unit 1625 Labeling unit 1626 Update parameter generation unit

Claims (5)

A data storage unit that stores trend data and water quality data,
The learning target data and the estimation target data acquisition unit that acquire both the learning target data and the estimation target data from the trend data and the water quality data accumulated in the data storage unit, and the estimation target data acquisition unit.
A data compression unit that compresses both the learning target data and the labeled data and the unlabeled data included in the learning target data and the estimation target data from the learning target data and the estimation target data acquisition unit, and a data compression unit.
A learning target data acquisition unit that acquires a portion of the compressed data compressed by the data compression unit that corresponds to the learning target data, and a learning target data acquisition unit.
A learning model and an estimation model generation unit that generates learning parameters based on the learning model by performing semi-supervised learning using the compressed learning target data from the learning target data acquisition unit.
A learning parameter storage unit that stores the learning parameters generated by the estimation model generation unit, and a learning parameter storage unit.
An estimation target data acquisition unit that acquires a portion of the compressed data compressed by the data compression unit that corresponds to the estimation target data, and an estimation target data acquisition unit.
Control using the explanatory variables of the estimation target included in the portion of the compressed data from the estimation target data acquisition unit corresponding to the estimation target data, and the learning parameters stored in the learning parameter storage unit. An operation amount derivation device for a water treatment facility including an operation amount derivation unit for deriving a target operation amount. A data storage unit that stores trend data and water quality data,
A learning target data acquisition unit that acquires learning target data used for updating model parameters of an estimation model from trend data and water quality data accumulated in the data storage unit, and a learning target data acquisition unit.
An estimation target data acquisition unit that acquires estimation target data used for deriving the manipulated variable from the trend data and water quality data accumulated in the data storage unit, and an estimation target data acquisition unit.
A model parameter update unit that updates the model parameters of the estimation model by learning using the learning target data from the learning target data acquisition unit, and
A learning parameter storage unit that stores the updated model parameters as learning parameters in the model parameter update unit,
An operation amount derivation device for a water treatment facility including an operation amount derivation unit for deriving an operation amount of a control target using the estimation target data and the learning parameter. Equipped with a data conversion unit that compresses the dimensions of data by principal component analysis
The operation amount derivation device for a water treatment facility according to claim 2, wherein the model parameter update unit performs semi-supervised learning by partial least squares. The operation amount deriving device for a water treatment facility according to claim 2, wherein the model parameter updating unit performs semi-supervised learning of a transductive SVM. The operation amount derivation device for a water treatment facility according to claim 2, wherein the model parameter update unit performs graph-based semi-supervised learning.

Images