JP7537357B2 - Sewage treatment facility operation amount inference basis output device and operation amount inference basis output method - Google Patents

Description

The present invention relates to an operation amount inference basis output device and an operation amount inference basis output method for a wastewater treatment facility.

Conventionally, in wastewater treatment facilities such as sewage treatment plants, it has been common to purify wastewater by activated sludge treatment. In the activated sludge treatment, the wastewater is stored in a reaction tank, where it is exposed to air blown in by aeration or the like, causing aerobic microorganisms to decompose pollutants to produce activated sludge, which is then allowed to grow and settle out, thereby purifying the wastewater.
In such wastewater treatment facilities, daily operations are carried out by workers who determine the amount of sludge to be removed, the amount of dissolved oxygen in the reaction layer, the amount of air blown, etc., and operate various pieces of equipment in the wastewater treatment facility.
When operating the above-mentioned equipment, the decision of the operation amount of the equipment is highly dependent on the intuition, experience, and know-how of a skilled worker, and is a very difficult task. Therefore, it is desired to be able to decide the operation amount without relying on a skilled worker.

In determining the amount of driving operation as described above, it is possible to apply a deep learning method such as a neural network, or a machine learning method including a population learning method such as Random Forest or XGBoost (eXtreme Gradient Boosting), to model the driving operation of a skilled worker.
For example, Patent Document 1 discloses a sewage pump control device for a sewage treatment plant which predicts the amount of sewage that will flow into the sewage storage facility within a certain time period based on the actual inflow volume that has flowed into the sewage storage facility from a certain point in the past to the present and on plant data, and inputs information calculated from a preset target value for pump discharge volume, the predicted inflow volume, the actual inflow volume, the current storage volume, and the current pump discharge volume, and determines the target pump discharge volume through fuzzy inference in accordance with predetermined membership functions and rules.
A neural network is used in the device of Patent Document 1. When time-series data such as the pump well water level, dust removal inflow gate opening, wastewater pump discharge rate, and actual inflow rate are input, the neural network outputs a predicted inflow rate.

Japanese Patent Application Publication No. 9-68170

In general, when applying the above-mentioned deep learning method, its black box nature can be a problem. A neural network that mimics the connections of neurons in the human brain is expressed as a multi-layer structure of linear combinations of many neurons and activation by nonlinear functions, and it is difficult to explain the results of its inference. However, wastewater treatment facilities such as sewage treatment plants are particularly public, so if any problem occurs, it can cause fatal damage to the operating company of the wastewater treatment plant and the users of the facility. Therefore, in order to introduce a deep learning method with black box nature to a wastewater treatment plant, it is necessary to clearly show on what basis the inference model that infers the operation amount made the inference, for example by outputting the basis of the inference, in order to increase the reliability of the results.
When using ensemble learning methods rather than deep learning methods, it is possible to output inference grounds for the entire training data, but this only indicates the overall trend. When making an individual inference, the grounds that primarily contributed to that inference often correspond to and depend on the characteristics of the data that was input to that inference, rather than following the overall trend.
Therefore, for example, when using an ensemble learning method, or when using an inference method other than a deep learning method or an ensemble learning method, it is desirable to increase the reliability of the results by outputting and showing more detailed reasons for the inference.

On the other hand, the treatment in a wastewater treatment facility is very complicated because it involves biological reactions. For example, even if the amount of dissolved oxygen is increased or decreased, the amounts of organic matter, phosphorus, and nitrogen in the water in the reaction tank each increase or decrease in a nonlinear and complicated manner. The amounts of organic matter, phosphorus, and nitrogen also increase or decrease in a complicated manner by changing operation amounts other than the amount of dissolved oxygen, such as return sludge, circulation amount, and excess sludge.
As described above, the processes in wastewater treatment plants are extremely complex, and it is therefore unrealistic to globally represent them with a simple, interpretable model without relying on machine learning techniques.
It is desirable to infer operational values for wastewater treatment facilities using machine learning inference models, while also providing reasons for the inference results to increase reliability of the results.

The problem that the present invention aims to solve is to provide an operation amount inference basis output device and an operation amount inference basis output method for a wastewater treatment facility, which can increase the reliability of the results by showing the basis of the inference results when inferring the operation values of a wastewater treatment facility using a machine learning inference model.

In order to solve the above problems, the present invention adopts the following means. That is, the present invention provides an operation amount inference basis output device for a sewage treatment facility, which infers the operation amount of equipment in the sewage treatment facility and outputs the basis for the inference when changing the operating status of the sewage treatment facility for the purpose of improving water quality, and includes a machine learning device, which inputs learning data corresponding to each of a plurality of facility parameters related to the operating status, trains the machine learning device by machine learning using the operation amount data corresponding to the learning data as teacher data, and generates model parameters related to the inference of the operation amount, an operation amount inference unit inputs inference target data that is the subject of inference of the operation amount to an inference model constructed using the model parameters to infer the operation amount, an approximation model generation unit generates an approximation model that approximates the inference model for the inference target data based on the inference target data and a plurality of associated data having values related to the inference target data, and generates approximation model parameters, and an inference basis output unit outputs the basis for the inference of the operation amount based on the inference target data in the inference model based on the approximation model parameters.

The present invention also provides a method for outputting the basis for inferring the amount of operation of a sewage treatment facility, which infers the amount of operation of equipment in the sewage treatment facility and outputs the basis for the inference when changing the operating status of the sewage treatment facility for the purpose of improving water quality, and which inputs learning data corresponding to each of a plurality of facility parameters related to the operating status, trains a machine learning device using the amount of operation corresponding to the learning data as teacher data, generates model parameters related to the inference of the amount of operation, inputs inference target data that is the subject of inference of the amount of operation of the operation to an inference model constructed using the model parameters to infer the amount of operation of the operation, generates an approximation model that approximates the inference model for the inference target data based on the inference target data and a plurality of associated data having values related to the inference target data, generates approximation model parameters, and outputs the basis for inferring the amount of operation of the operation based on the inference target data in the inference model based on the approximation model parameters.

According to the present invention, it is possible to provide an operation amount inference basis output device and an operation amount inference basis output method for a wastewater treatment facility, which can increase the reliability of the results by showing the basis of the inference results when inferring the operation values of a wastewater treatment facility using a machine learning inference model.

FIG. 1 is a block diagram of a wastewater treatment facility according to an embodiment of the present invention. FIG. 2 is a block diagram of an operation amount inference basis output device provided in the wastewater treatment facility. 4 is an explanatory diagram of a data processing unit of the driving operation amount inference basis output device. FIG. 4 is an explanatory diagram of an inference basis in the driving operation amount inference basis output device. FIG. 10 is a flowchart showing a method for outputting a basis for inference of a vehicle operation amount using the vehicle operation amount inference basis output device, during learning of an inference model. 5 is a flowchart showing the process of inferring a vehicle operation amount in the vehicle operation amount inference basis output method. 5 is a flowchart of an inference basis deriving process in the driving operation amount inference basis output method. FIG. 4 is a block diagram of a driving operation amount inference basis output device according to a first modified example of the embodiment. 5 is a flowchart of a method for outputting basis for inferring vehicle operation amounts using the vehicle operation amount inference basis output device of the first modified example.

The following describes an embodiment of the present invention in detail with reference to the drawings.

1 is a block diagram of a wastewater treatment facility according to the present embodiment. The wastewater treatment facility 1 purifies wastewater stored in a reaction tank 3 using microorganisms.
The sewage treatment facility 1 is equipped with a primary sedimentation tank 2, a reaction tank 3, a measuring instrument 4, a final sedimentation tank 5, an aeration plate 6, a blower 7, an aeration adjustment valve 8, a return sludge pump 9, an excess sludge extraction pump 10, a gravity thickening tank 11, a mechanical thickening tank 12, a digester tank 13, a dewatering tank 14, an operation amount inference basis output device 20, and first to eighth pipes L1, L2, L3, L4, L5, L6, L7, and L8.

Wastewater containing organic matter is introduced into the primary sedimentation tank 2 as raw water. In the primary sedimentation tank 2, small debris, sand, etc. in the introduced wastewater are removed, and rough solid-liquid separation is performed. The wastewater from which debris, etc. has been removed is sent to the reaction tank 3 via the first pipe L1. The sludge that settles in the primary sedimentation tank 2, so-called raw sludge, is sent to the gravity thickening tank 11 via the second pipe L2.

In the reaction tank 3, the wastewater is biologically treated and purified by microorganisms. During purification, the wastewater is exposed to air by aeration or the like, and aerobic microorganisms decompose organic matter and grow as a result of assimilating the organic matter, forming activated sludge. The water that has been treated as activated sludge is sent to the final sedimentation tank 5 via the third pipe L3.
Air is supplied to the lower part of the reaction tank 3 from a blower 7 via a sixth pipe L6. An aeration control valve 8 is provided in the sixth pipe L6 between the blower 7 and the reaction tank 3. Opening and closing the aeration control valve 8 changes the aeration volume (hereinafter referred to as the air volume), thereby adjusting the amount of dissolved oxygen in the reaction tank 3 and controlling the progress of biological treatment by microorganisms. An aeration plate 6 for increasing the efficiency of dissolving oxygen is provided in the part of the reaction tank 3 to which air is supplied by the sixth pipe L6.
The reaction tank 3 is provided with measuring instruments 4 for measuring various water quality data such as dissolved oxygen (DO), multi-level suspended solids (MLSS), NH4 concentration, and NO3 concentration. In Fig. 1, the measuring instruments are provided in the reaction tank 3, but in reality, the measuring instruments are provided in various locations within the wastewater treatment facility 1, and in some cases, the measurement results from these multiple measuring instruments are summarized as water quality data for the entire treatment system and used for various tasks and decisions such as operation. The same applies to trend data, which will be described later.

In the final settling tank 5, the activated sludge contained in the activated sludge treated water sent from the reaction tank 3 is settled, and the wastewater is purified. The supernatant liquid after the activated sludge is separated in the final settling tank 5 is discharged outside the system as treated water.
A portion of the sludge that has settled in the final settling tank 5 is returned to the reaction tank 3 through the fourth pipe L4 by the return sludge pump 9 and reused in the activated sludge treatment. The remaining sludge is discharged as excess sludge by the sludge withdrawal pump 10 through the fifth pipe L5 and sent to the mechanical thickening tank 12.

The gravity thickening tank 11 and the mechanical thickening tank 12 thicken the raw sludge sent from the primary settling tank 2 and the excess sludge sent from the final settling tank 5 via the excess sludge withdrawal pump 10. The sludge thickened in the gravity thickening tank 11 and the mechanical thickening tank 12 is sent to the digestion tank 13 through the seventh pipe L7.
The digestion tank 13 digests the concentrated sludge, for example, by anaerobic digestion, in which the organic sludge is decomposed by the action of anaerobic microorganisms. The decomposed sludge is sent to the dewatering tank 14 through the eighth pipe L8.
The dewatering tank 14 reduces the moisture content of the decomposed sludge, thereby reducing its volume. The reduced volume sludge is then incinerated.

The operation control amount inference basis output device 20 in this embodiment is targeted at the wastewater treatment facility 1 as described above, and when the operating conditions of the wastewater treatment facility 1 are changed for the purpose of improving water quality, the operation control amount inference basis output device 20 infers the operation control amount of the equipment in the wastewater treatment facility 1 and outputs the basis for the inference.
2 is a block diagram of the driving operation amount inference basis output device 20. The driving operation amount inference basis output device 20 includes a data accumulation unit 24, an inference model learning data acquisition unit 25, an inference target data acquisition unit 26, an approximation model learning data acquisition unit 27, a data processing unit 28, a driving operation amount learning unit 29, an inference model parameter storage unit 30, a driving operation amount inference unit 31, an approximation model generation unit 32, and an inference basis output unit 33.
Among these components of the driving operation amount inference basis output device 20, the inference model learning data acquisition unit 25, the inference target data acquisition unit 26, the approximation model learning data acquisition unit 27, the data processing unit 28, the driving operation amount learning unit 29, the driving operation amount inference unit 31, the approximation model generation unit 32, and the inference basis output unit 33 may be software or programs executed by a CPU in the information processing device, for example. Also, the data accumulation unit 24 and the inference model parameter storage unit 30 may be realized by storage devices such as semiconductor memory and magnetic disks provided inside or outside the information processing device.

The input of the operation amount inference basis output device 20 includes trend data 21 such as sludge concentration, sludge flow rate, water temperature, etc., related to the operating status of the wastewater treatment facility 1. For each of these trend data 21, values acquired in the primary sedimentation tank 2, the reaction tank 3, the final sedimentation tank 5, etc. may be used individually, or a value summarized for the entire treatment system of the wastewater treatment facility 1 may be derived from these values and used.
The input to the operation amount inference basis output device 20 includes water quality data 22 such as dissolved oxygen (DO), MLSS, NH4 concentration, and NO3 concentration, as previously described.
The driving operation amount data 23, which is one of the inputs to the driving operation amount inference basis output device 20, will be described later.

The output of the operation amount inference basis output device 20 is the operation amount 34 of each facility, as described above. In this embodiment, the operation amount 34 includes the inflow amount of wastewater flowing from the primary settling tank 2 to the reaction tank 3, the air supply amount, the returned sludge amount which is the amount of returned sludge returned from the final settling tank 5 to the reaction tank 3 via the return sludge pump 9, and the excess sludge withdrawal amount which is the amount of excess sludge sent from the final settling tank 5 to the mechanical thickening tank 12 via the excess sludge withdrawal pump 10.
For example, with regard to the amount of dissolved oxygen, the airflow rate can be set as a target value in order to maintain the amount of dissolved oxygen in the reaction tank 3 at a certain level or higher. However, a processing system may also be envisaged in which a target value is set for the amount of dissolved oxygen itself in the reaction tank 3, and the airflow rate is adjusted so that the amount of dissolved oxygen is equal to or higher than the target value. In this case, for example, instead of the airflow rate, a set value for the amount of dissolved oxygen may be output as the operation manipulated variable 34.
More specifically, the amount of returned sludge may be, for example, the rotation speed of the return sludge pump 9 .
More specifically, the amount of excess sludge extracted may be, for example, the time it takes for the excess sludge extraction pump 10 to extract the excess sludge.
In this embodiment, the operation input 34 includes the injection rate of the polymer flocculant injected in the dehydration treatment in the dehydration tank 14 .

When the above-mentioned trend data 21 and water quality data 22 are input, the driving operation amount inference basis output device 20 infers a driving operation amount 34 that corresponds to the trend data 21 and the water quality data 22, that is, a driving operation amount 34 that should be executed correspondingly under the circumstances represented by the input trend data 21 and the water quality data 22. In order to perform this inference effectively, the driving operation amount inference basis output device 20 is provided with an inference model 50 generated by machine learning using a machine learning device 40 provided in the driving operation amount learning unit 29. More specifically, the driving operation amount learning unit 29 inputs learning data generated based on the trend data 21 and the water quality data 22 to the machine learning device 40, performs machine learning, and generates model parameters related to the inference of the driving operation amount 34.
That is, the driving operation amount inference basis output device 20 learns the driving operation amounts 34 and infers the driving operation amounts 34. In the present embodiment, the driving operation amount inference basis output device 20 also outputs inference basis 35, which indicates on what basis the inference model 50 infers the driving operation amounts 34 when inferring the driving operation amounts 34. For ease of explanation, the following will first explain each component of the driving operation amount inference basis output device 20 when learning the driving operation amounts 34, then explain the behavior of each component when inferring the driving operation amounts 34, and further explain the behavior of each component when outputting inference basis 35.

When learning the driving operation amount 34, the driving operation amount learning unit 29 causes the machine learning device 40 to perform machine learning based on the learning data. As a result of the machine learning device 40 performing machine learning, an inference model 50 is generated.
The learning data is generated based on and corresponding to trend data (facility parameters) 21 and water quality data (facility parameters) 22 related to the operating conditions, which are input when the operation operation amount inference basis output device 20 infers the operation operation amount 34. More specifically, assuming that the time when the operation operation amount 34 is learned is set as a reference time, the learning data is generated based on trend data 21 and water quality data 22 that are measured and recorded in the actual wastewater treatment facility 1 before the reference time at the time of learning.

When learning data generated based on the trend data 21 and the water quality data 22 is input to the machine learning device 40 and the machine learning device 40 is trained by machine learning, the machine learning device 40 derives an operation operation amount 34 that is considered to be suitable for the learning data in the middle of the learning stage. The operation operation amount 34 output by the machine learning device 40 is compared with the operation operation amount data 23. The operation operation amount data 23 is an actual operation amount that is considered to be performed by a skilled worker on the equipment of the sewage treatment facility 1 when it is assumed that the sewage treatment facility 1 is in a situation indicated by the trend data 21 and the water quality data 22 corresponding to the learning data input to the machine learning device 40. The operation operation amount 34 that is the output of the machine learning device 40 is compared with the operation operation amount data 23 corresponding to the input learning data, and the machine learning device 40 is trained by machine learning so as to reflect the comparison result. In other words, the operation operation amount data 23 is teacher data corresponding to the learning data input to the machine learning device 40.
In this way, the machine learning machine 40 learns to infer the driving operation amount 34 that is close to the driving operation amount data 23.

The data accumulation unit 24 stores and accumulates trend data 21, water quality data 22, and operation amount data 23 used for learning by the machine learning device 40.
The inference model learning data acquisition unit 25 acquires trend data 21 and water quality data 22 used for learning the machine learning device 40 from the data storage unit 24 and transmits them to the data processing unit 28.

データ加工部２８における離散値化の方法はこれに限られず、閾値を設けて閾値との大小関係により２値化するなど、データの分布の態様に応じて加工方法が決定されてよい。
データ加工部２８は、上記のように生成された学習データを、運転操作量学習部２９に送信する。 The data processing unit 28 receives each piece of data from the inference model learning data acquisition unit 25, processes it into a format that is actually used as an input to the machine learning device 40, and generates learning data. FIG. 3 is an explanatory diagram of the data processing unit 28.
The data processing unit 28 in this embodiment discretizes data input as continuous values based on the deviation from the average value. For example, in this embodiment, as shown in Fig. 3, the average value and standard deviation σ of the target data are calculated, and each value in the data is discretized according to the relational expression Ψ shown as the following Equation 1 based on the magnitude relationship between the value x of the random variable corresponding to the value and the standard deviation σ.

The method of discretization in the data processing unit 28 is not limited to this, and the processing method may be determined according to the manner of data distribution, such as setting a threshold and binarizing the data based on its magnitude relationship with the threshold.
The data processing unit 28 transmits the learning data generated as described above to the driving operation amount learning unit 29 .

The driving operation amount learning unit 29 receives the learning data and the teacher data. The driving operation amount learning unit 29 inputs the learning data to the machine learning device 40. In this embodiment, the machine learning device 40 is realized by a neural network.
The machine learning device 40 includes an input layer, a plurality of intermediate layers, and an output layer. The input layer includes input nodes corresponding to each of the learning data. The output layer includes output nodes corresponding to each of the driving operation amounts 34 output by the machine learning device 40.
The driving operation amount learning unit 29 inputs the corresponding learning data value to each input node of the input layer.
The intermediate layer is provided between the input layer and the output layer, and all nodes in adjacent layers are fully connected, with connection weights between the nodes. Based on the connection weights, each node in the intermediate layer and the output layer calculates a weighted sum for each node in the previous layer, and stores the result by applying an output function.

In the machine learning device 40, the input learning data is processed as described above, and the processing result is stored in the output node. The driving operation amount learning unit 29 adjusts the values of each connection weight by an error backpropagation method, a stochastic gradient descent method, or the like, so as to reduce a cost function represented by a squared error or the like between the processing result and the teacher data corresponding to the learning data input to the machine learning device 40, thereby training the machine learning device 40 by machine learning.
As a result, when learning data is input, the machine learning device 40 learns to infer driving operation amounts 34 close to the corresponding teacher data. When learning is completed, the values of each connection weight and the like reflect the extracted features of the learning data.

When the learning is completed, the driving operation amount learning unit 29 stores the adjusted connection weights and other parameters as model parameters in the inference model parameter storage unit 30. The model parameters stored in the inference model parameter storage unit 30 are obtained by a driving operation amount inference unit 31, which will be described later, and an inference model 50 that estimates the driving operation amount 34 is constructed.
That is, the driving operation amount learning unit 29 generates a learned inference model 50 in which appropriate learning parameters have been learned and which is used as a program module that is part of the artificial intelligence software.

Next, the behavior of each component when the driving operation amount 34 is inferred will be described.
The inference target data acquisition unit 26 acquires trend data 21 and water quality data 22 at a time when the operation amount 34 is to be inferred, for example, the current time, from the data storage unit 24. The inference target data acquisition unit 26 transmits each of the trend data 21 and water quality data 22 to the data processing unit 28.
The data processing unit 28 receives each data from the inference target data acquisition unit 26 and processes it in the manner already described, to generate inference target data that is to be used to infer the driving operation amount 34 .
The data processing unit 28 transmits the inference target data to the driving operation amount inference unit 31 .

The driving operation amount inference unit 31 acquires the inference target data from the data processing unit 28 .
The driving operation amount inference unit 31 acquires model parameters from the inference model parameter storage unit 30 and constructs the inference model 50. The inference model 50 has a structure substantially equivalent to that of the already-described machine learning device 40. The driving operation amount inference unit 31 infers the driving operation amount 34 by executing the inference model 50 as a program on a CPU, for example.
When the driving operation amount inference unit 31 inputs the inference target data to the inference model 50, the inference model 50 executes calculation processing similar to that of the machine learning device 40 in the driving operation amount learning unit 29, tracing in order from the input layer to the output layer. Finally, the driving operation amount 34 corresponding to the inference target data is output from the output node.

Next, the behavior of each component when the inference basis 35 is output will be described.
Conceptually, when explaining on what basis an inference is made in a learning model such as a neural network, it is conceivable to construct an explanatory model that simplifies the entire learning model and use this. However, this is not easy. In particular, since the processing in the wastewater treatment facility 1 as in this embodiment is very complicated, the inference model 50 that infers the operation amount 34 is highly complicated, and therefore, it is not realistic to globally express this with an interpretable and simple model.
In this embodiment, the inference model 50 is not globally represented by a simple model, but instead an approximation model is generated that locally represents the inference model 50. More specifically, the approximation model is generated so as to represent a limited portion around the inference target data whose values are close to the inference target data that was input when the driving operation amount 34 was inferred by the inference model 50. In this way, the approximation model for explaining the inference basis 35 is generated as a relatively simple and interpretable one by remaining a local representation of the inference model 50.
In addition, the approximation model is generated for data related to the inference target data that was input during inference. That is, since the approximation model depends on the inference target data, the approximation model is generated specifically for the inference target data that was input each time the inference model 50 infers the driving operation amount 34.
With the above-mentioned configuration, the approximation model is simple and interpretable at least in the part related to the inference target data. Therefore, there is no need to simplify the inference model 50 and reduce the inference accuracy in order to improve the interpretability for outputting the inference basis 35.

In this embodiment, the generation of the above-mentioned approximation model is performed by applying LIME (Local Interpretable Model-agnostic Explanations), a known machine learning method. LIME is a method for visualizing the basis of inference by approximating the data to be inferred with an analyzer that is locally simple and easy to analyze in the part related to the data.

First, the learning data acquisition unit 27 for the approximation model acquires, from the data storage unit 24, trend data 21 and water quality data 22 related to the trend data 21 and water quality data 22 corresponding to the inference target data generated by the data processing unit 28.
More specifically, in a coordinate space formed by each value of the inference target data a, when a' is a point obtained by converting each component of the inference target data a into a dummy variable by binarizing it, the approximation model learning data acquisition unit 27 samples a certain number of points that include only a part of the non-zero components of a'. This sampling is performed, for example, by randomly replacing the components of a' that are 1 with zero under the condition that a' is not a zero vector in which all components are zero.
The approximation model learning data acquisition unit 27 transmits the trend data 21 and water quality data 22 acquired as described above to the data processing unit 28 .
The data processing unit 28 processes each piece of data received from the approximation model learning data acquisition unit 27 using the procedure already described, and generates multiple pieces of associated data having values related to the inference target data.
The data processing unit 28 transmits the associated data to the approximation model generation unit 32 .

The approximation model generating unit 32 receives the inference target data and a plurality of related data. Based on these data, the approximation model generating unit 32 generates an approximation model that approximates the inference model 50 for the inference target data around the inference target data, and generates approximation model parameters.
As already explained, the model parameters of the approximation model are determined so as to locally approximate the inference model 50 at the portion related to the inference target data. Therefore, the approximation model is generated so as to approximate the portion related to the input inference target data each time the inference model 50 performs inference, rather than approximating the entire inference model 50.

ここで、Ｌ（ｆ、ｇ、π_ａ）は、推論対象データａの周辺で２つのモデルｆ、ｇがどれだけ相違しているかを示す損失関数である。
上記のように、推論対象データａをサンプリングしてできた座標空間をＺ、サンプリングした座標空間Ｚ内の点をｚ´、このダミー変数化されたｚ´をもとの空間に復元したときの点をｚとすると、Ｌ（ｆ、ｇ、π_ａ）は、次の式３として表現される。

π_ａ（ｚ）は、ｚが推論対象データａから離れるほど小さな値を示す関数であり、ｆ（ｚ）とｇ（ｚ´）の誤差に重み付けを行うものである。π_ａ（ｚ）としては、例えば次に式４として示すような指数カーネル関数が採用され得る。

損失関数Ｌが上記のような式３で表されるため、目的関数ξ（ａ）を最小化して得られる近似モデルｇは、推論対象データａに近いほど、推論モデル５０であるモデルｆをよく近似したものとなる。すなわち、このπ_ａ（ｚ）は、ｚが推論対象データａから一定の距離以上離れると、急激にゼロに落ち込む関数となっているため、ｚは推論対象データａの周辺で近似される。 More specifically, when the inference target data is a and the inference model 50 for inferring the driving operation amount 34 is f, the approximation model generation unit 32 generates an approximation model g so as to minimize the objective function ξ(a) described in the following equation 2.

Here, L(f, g, π _a ) is a loss function that indicates how different two models f and g are around the inference target data a.
As described above, if the coordinate space created by sampling the inference target data a is Z, the point in the sampled coordinate space Z is z', and the point when this dummy variable z' is restored to the original space is z, then L(f, g, π _a ) can be expressed as the following equation 3.

π _a (z) is a function that indicates a smaller value as z becomes farther away from the inference target data a, and weights the error between f(z) and g(z'). For π _a (z), for example, an exponential kernel function such as that shown in the following formula 4 can be adopted.

Since the loss function L is expressed by the above formula 3, the closer the approximation model g obtained by minimizing the objective function ξ(a) is to the inference target data a, the better it approximates the model f, which is the inference model 50. In other words, since this π _a (z) is a function that suddenly drops to zero when z is a certain distance or more away from the inference target data a, z is approximated in the vicinity of the inference target data a.

上式において、ｗ＝（ｗ_１、ｗ_２、…、ｗ_ｎ）は回帰係数である。また、δ_ａｉ（ｘ_ｉ）は、次の式６で表される。

すなわち、δ_ａｉ（ｘ_ｉ）は、ｉ番目の変数ｘ_ｉと、推論対象データａのｉ番目の値ａ_ｉが等しいときに１の値となり、異なるときに０の値となる。つまり、ｇはダミー変数化した変数の線形和として表現されることを仮定し、その上で目的関数ξ（ａ）を最小化するように回帰係数ｗを決定する。このとき、ｗの値がそのまま推論モデル５０の推論に寄与している各変数の寄与度と解釈することが可能である。
すなわち、複数の係数ｗ＝（ｗ_１、ｗ_２、…、ｗ_ｎ）が、近似モデルｇを表す近似モデルパラメータである。 For example, when the values of the variables constituting the inference target data a are a ₁ , a ₂ , . . . , a _n , the following equation 5 can be used for the approximation model g.

In the above equation, w=(w ₁ , w ₂ , . . . , w _n ) is a regression coefficient, and δ _ai (x _i ) is expressed by the following equation 6.

That is, δ _ai (x _i ) is 1 when the i-th variable x _i is equal to the i-th value a _i of the inference target data a, and is 0 when they are different. In other words, it is assumed that g is expressed as a linear sum of dummy variables, and the regression coefficient w is determined so as to minimize the objective function ξ(a). At this time, the value of w can be interpreted as the contribution of each variable that contributes to the inference of the inference model 50.
That is, a plurality of coefficients w=(w ₁ , w ₂ , . . . , w _n ) are approximate model parameters that represent the approximate model g.

In the above formula 2, Ω(g) is a regularization term and is a function indicating the complexity of the approximation model g. As shown in formula 2, Ω(g) is incorporated into the objective function ξ(a) together with the loss function L and is minimized, thereby preventing the approximation model g from having a complex structure.
As explained using Equation 5, the approximation model g is expressed as a linear sum using multiple coefficients w = ( _w1 , _w2 , ..., _wn ) corresponding to each of the values constituting the inference target data a, which are dummy variables. For example, when K is a fixed value and a model having K or less explanatory variables is to be generated as the approximation model g, Ω(g) is set to be infinite when the number of non-zero coefficients among the coefficients _w1 , _w2 , ..., _wn is greater than K, and to be 0 otherwise.

The approximate model generating unit 32 transmits the approximate model parameters w=(w ₁ , w ₂ , . . . , w _n ) generated as described above to the inference basis output unit 33 .

The inference basis output unit 33 receives the approximate model parameters w=(w ₁ , w ₂ , . . . , w _n ).
Regarding the approximate model parameters w = ( _w1 , _w2 , ..., _wn ), when the numerical value of the i-th coefficient w _i is a large positive value, it can be interpreted that the value of the i-th variable x _i being a _i strongly supports that the result of the model f is f(x). Therefore, the approximate model parameters w = ( _w1 , _w2 , ..., _wn ) can be applied as the strength of the basis for the judgment of the machine learning device 40.
Fig. 4 is an explanatory diagram of the inference basis 35. In Fig. 4, the inference target is the amount of change in the dissolved oxygen (DO) that is desirable for operating the wastewater treatment facility 1. The inference model 50 estimates that, among five options for changing the operation control amount 34 of the dissolved oxygen set value, i.e., increasing the DO by 0.2 (mg/L), increasing the DO by 0.1 (mg/L), not changing the DO set value, decreasing the DO set value by 0.1 (mg/L), and decreasing the DO set value by 0.2 (mg/L), that the operation control amount 34 having the highest probability of 0.57 is decreased by 0.1 (mg/L).
In this example, the facility parameters include NO3 concentration, NH4 concentration, total nitrogen (TN), total phosphorus (TP), and PO4 concentration. The values of the corresponding coefficients w _i for each of these facility parameters are displayed as 0.37, 0.21, 0.09, -0.02, and -0.05, respectively. Here, the value of the coefficient w _i corresponding to the NO3 concentration is the largest. This indicates that the NO3 concentration being in the range of 4.4 (mg/L) or more and less than 7.3 (mg/L), which is a range slightly higher than normal, is the most major reason why the inference model 50 infers that the amount of dissolved oxygen is reduced by 0.1 (mg/L). In addition to this NO3 concentration, it is also shown that the NH4 concentration being slightly lower than normal and the total nitrogen being slightly lower than normal are other reasons why the inference model 50 infers that the amount of dissolved oxygen is reduced by 0.1 (mg/L).
In this way, the inference basis output unit 33 outputs the inference basis 35 of the driving operation amount 34 based on the inference target data a in the inference model 50 based on the approximation model parameter w to the outside, for example to a display device not shown.

Next, a method for outputting reasoning basis for operation operations using the operation operation amount inference basis output device 20 for the wastewater treatment facility will be described with reference to Figures 1 to 4 and Figures 5 to 7. Figure 5 is a flowchart of the method for outputting reasoning basis for operation operations when learning an inference model. Figure 6 is a flowchart of the method for outputting reasoning basis for operation operations when inferring operation amounts 34. Figure 7 is a flowchart of the inference basis output process in the method for outputting reasoning basis for operation operations.

When learning the operation control variables 34, first, the trend data 21, the water quality data 22, and the operation control variable data 23 used for learning the machine learning device 40 are stored and accumulated in the data accumulation unit 24 (step S1).
The inference model learning data acquisition unit 25 acquires the trend data 21 and water quality data 22 used for learning the machine learning device 40 from the data storage unit 24, and transmits them to the data processing unit 28 (step S3).
The data processing unit 28 receives each data from the inference model learning data acquisition unit 25, processes it into a format that will actually be used as input to the machine learning device 40, and generates learning data (step S5).
The driving operation amount learning unit 29 uses the learning data and the driving operation amount data 23, i.e., the teacher data, to train the machine learning device 40 and generate model parameters related to the inference of the driving operation amount 34 (step S7).
The driving operation amount learning unit 29 stores the generated model parameters in the inference model parameter storage unit 30 (step S9).

When inferring the operation variables 34, the inference target data acquisition unit 26 acquires the trend data 21 and the water quality data 22 at the time when the operation variables 34 are to be inferred, for example, the current time, from the data storage unit 24. The inference target data acquisition unit 26 transmits each of the trend data 21 and the water quality data 22 to the data processing unit 28 (step S11).
The data processing unit 28 receives each data from the inference target data acquisition unit 26, processes it, and generates inference target data that is to be used as an inference target for the driving operation amount 34 (step S13).
The driving operation amount inference unit 31 acquires model parameters from the inference model parameter storage unit 30 and constructs the inference model 50. The driving operation amount inference unit 31 inputs inference target data to the inference model 50 to infer the driving operation amount 34 (step S15).

Next, the driving operation amount inference basis output device 20 derives the inference basis 35 (step S17). This is performed in detail as follows.
First, the learning data acquisition unit 27 for the approximation model acquires, from the data storage unit 24, trend data 21 and water quality data 22 related to the trend data 21 and water quality data 22 corresponding to the inference target data generated by the data processing unit 28, and transmits them to the data processing unit 28 (step S31).
The data processing unit 28 processes each piece of data received from the approximation model learning data acquisition unit 27 to generate a plurality of associated data having values related to the inference target data (step S33).
The approximation model generating unit 32 receives the inference target data and a plurality of related data. Based on these data, the approximation model generating unit 32 generates an approximation model that approximates the inference model 50 for the inference target data around the inference target data, generates an approximation model parameter w, and transmits it to the inference basis output unit 33 (step S35).
The approximate model parameters w=(w ₁ , w ₂ , . . . , w _n ) serve as inference basis 35 .

The inference basis output unit 33 receives the approximate model parameters w=(w ₁ , w ₂ , . . . , w _n ).
The driving operation amount inference unit 31 outputs the inferred driving operation amount 34 to an external device, such as a display device (not shown). The inference basis output unit 33 outputs the approximation model parameters w = ( _w1 , _w2 , ..., _wn ) as inference basis 35 of the driving operation amount 34 by the inference model 50 to an external device, such as a display device (not shown) (step S25).
Based on the thus displayed operation variables 34 and inference grounds 35, the operator finally judges and determines the operation variables for operating the wastewater treatment facility 1, and operates the wastewater treatment facility 1 (step S27).

Next, the effects of the operation amount inference basis output device and operation amount inference basis output method for the above wastewater treatment facility will be explained.

The operation operation amount inference basis output device 20 for a wastewater treatment facility 1 of this embodiment infers operation operation amounts 34 of equipment in the wastewater treatment facility 1 and outputs the basis for the inference when the operating status of the wastewater treatment facility 1 is changed for the purpose of improving water quality. The operation operation amount inference basis output device 20 for a wastewater treatment facility 1 includes a machine learning device 40, and receives learning data corresponding to each of a plurality of facility parameters 21, 22 related to the operating status as input, and performs machine learning on the machine learning device 40 using operation operation amount data 23 corresponding to the learning data as teacher data, thereby generating model parameters related to the inference of the operation operation amounts 34. the driving operation amount inference unit 31 inputting inference target data that is to be inferred for the driving operation amount 34 into an inference model 50 constructed using model parameters to infer the driving operation amount 34; an approximation model generation unit 32 generating an approximation model that approximates the inference model 50 for the inference target data based on the inference target data and a plurality of associated data having values related to the inference target data, and generating approximation model parameters; and an inference basis output unit 33 outputting basis 35 for inference of the driving operation amount 34 based on the inference target data in the inference model 50 based on the approximation model parameters.
The method for outputting an inference basis for an operation amount of a sewage treatment facility 1 according to the present embodiment infers an operation amount 34 of equipment in the sewage treatment facility 1 and outputs the basis for the inference when the operating status of the sewage treatment facility 1 is changed for the purpose of improving water quality, and inputs learning data corresponding to each of a plurality of facility parameters 21, 22 related to the operating status, and performs machine learning on a machine learning device 40 using operation amount data 23 corresponding to the learning data as teacher data to generate model parameters related to the inference of the operation amount 34, inputs inference target data that is the inference target of the operation amount 34 to an inference model 50 constructed using the model parameters to infer the operation amount 34, generates an approximation model that approximates the inference model 50 for the inference target data based on the inference target data and a plurality of associated data having values related to the inference target data, generates approximation model parameters, and outputs basis 35 for inference of the operation amount 34 based on the inference target data in the inference model 50 based on the approximation model parameters.
As already explained, when explaining the basis on which inference is made in a learning model such as a neural network, it is conceivable to construct an explanatory model that simplifies the entire learning model and use this. However, this is not easy. In particular, since the processing in the wastewater treatment facility 1 as in this embodiment is very complicated, the inference model 50 that infers the operation amount 34 is highly complicated, and therefore, it is not realistic to globally express this with a simple model that can be interpreted.
In contrast, according to the above configuration, when inferring the driving operation amount 34 using the inference model 50, an approximation model that approximates the inference model 50 for the inference target data is generated based on the inference target data and a plurality of associated data having values related to the inference target data, and thereby inference grounds 35 for the driving operation amount 34 are output. In other words, the approximation model for explaining the inference grounds 35 is generated as a relatively simple and interpretable model by limiting it to a local expression of the inference model 50.
In this way, even when the operation values 34 of the wastewater treatment facility 1 are inferred by the machine-learned inference model 50, an approximation model can be generated as described above to output the inference basis 35 of the operation values 34 by the inference model 50. This makes it possible to show the basis for the inference result and reduce the black-box nature of the inference result, thereby increasing the reliability of the inference result.

Furthermore, the approximation model is expressed as a linear sum using a plurality of coefficients based on the values constituting the inference target data, and the approximation model parameters are the plurality of coefficients.
According to the above-mentioned configuration, it is possible to appropriately realize the operation amount inference basis output device 20 of the above-mentioned wastewater treatment facility 1.

The facility parameters 21, 22 include trend data 21 including sludge concentration and sludge flow rate, and water quality data 22 of the water treated in the wastewater treatment facility.
According to the above-described configuration, it is possible to analyze and extract the basis for which parts of the trend data 21 including sludge concentration and sludge flow rate, and the water quality data 22 of the water being treated in the sewage treatment facility, in particular, had a significant impact on the inference.

[First Modification of the Embodiment]
Next, a modified example of the operation operation amount inference basis output device and operation operation amount inference basis output method for a wastewater treatment facility shown as the above embodiment will be described with reference to Figures 8 and 9. Figure 8 is a block diagram of the operation operation amount inference basis output device in this modified example. Figure 9 is a flowchart of the operation operation amount inference basis output method using the operation operation amount inference basis output device in this modified example. The operation operation amount inference basis output device 20A in this modified example is different from the operation operation amount inference basis output device 20 in the above embodiment in that it includes an input unit 60 that receives an input of the operation amount from outside, and outputs an inference basis 35 to the outside only when the operation amount input by the input unit 60 differs from the operation amount 34 inferred in the inference model.

The input unit 60 receives, for example, an input of an amount of operation considered by an operator.
When the driving operation amount 34 is inferred by the inference model 50, the inference basis output unit 33A receives the driving operation amount 34 inferred by the inference model 50 from the driving operation amount inference unit 31. The inference basis output unit 33A compares the driving operation amount 34 inferred by the inference model 50 with the driving operation amount input from the input unit 60, and if they differ, outputs the driving operation amount 34 and the inference basis 35 by the inference model 50 to the outside, for example, to a display device (not shown) or the like.

In the driving operation amount inference basis output method of this modification, after step S17 of deriving the inference basis 35 in the driving operation amount inference basis output method of the above embodiment, input of the driving operation amount is accepted by the input unit 60 (step S19).
The inference basis output unit 33A compares the driving operation amount 34 inferred by the inference model 50 with the driving operation amount input from the input unit 60 (step S21).
The inference basis output unit 33A judges the result of the comparison (step S23), and if these values are different (No in step S23), outputs the operation amount 34 inferred by the inference model 50 as well as the inference basis 35 to a display device (not shown) (step S25). After that, an operator judges and determines the operation amount 34 for finally operating the wastewater treatment facility 1, and operates the wastewater treatment facility 1 (step S27).
If they are the same (Yes in step S23), the driving operation amount 34 and the inference basis 35 are not output in step S25, and the process proceeds to step S27.

It goes without saying that this modified example provides the same effects as the embodiment already described.
In particular, this modified example is provided with an input unit 60 that accepts external input of driving operation amounts, and outputs inference grounds 35 to the outside only when the driving operation amount input by the input unit 60 differs from the driving operation amount 34 inferred in the inference model 50.
According to the above-mentioned configuration, when the driving operation amount input by the input unit 60 matches the driving operation amount 34 inferred by the inference model 50, these values are not output to the outside. This reduces the effort required for the operator to check the output results.

[Second Modification of the Embodiment]
Next, a modified example of the operation amount inference basis output device and operation amount inference basis output method for a wastewater treatment facility shown in the above embodiment will be described. The operation amount inference basis output device in this modified example is a further modified example of the above first modified example, and differs in that when the operation amount input by the input unit 60 differs from the operation amount 34 inferred in the inference model 50, the amount of learning data of the approximation model is changed according to the magnitude of the difference between them so that the subsequent output of the inference basis 35 is more appropriately executed.

For example, when the amount of dissolved oxygen is inferred as the inference basis 35, the learning data is replicated α times based on the value of α defined by the following formula, and then registered as the learning data.

It goes without saying that this modified example provides the same effects as the embodiment already described.
In particular, in this modified example, when the driving operation amount input by the input unit differs from the driving operation amount 34 inferred in the inference model 50, the larger the difference between these values is, the more the multiple related data used when generating the approximation model are duplicated and increased.
According to the above-mentioned configuration, the more the data deviates from the amount of driving operation assumed by the operator, the stronger the learning process is. Therefore, in the subsequent inference, the inference basis 35 closer to the operator's assumption is derived.

[Third Modification of the Embodiment]
Next, a modified example of the operation operation amount inference basis output device and operation operation amount inference basis output method for a wastewater treatment facility shown in the above embodiment will be described. The operation operation amount inference basis output device in this modified example is mainly different in the method of generating an approximation model in the approximation model generating unit 32. The block diagram of the operation operation amount inference basis output device in this modified example is the same as that of the above embodiment, so the description will be mainly made with reference to FIG. 2.
Here, a description will be given of the behavior of each component when the inference basis 35 is output. The behavior of each component when the driving operation amount 34 is learned and when the driving operation amount 34 is inferred is the same as in the above embodiment.
In this modified example, as in the above embodiment, an approximation model is generated that locally represents the inference model 50. The approximation model is generated for data related to the inference target data that was input during inference.

As in the above embodiment, the approximation model learning data acquisition unit 27 acquires, from the data storage unit 24, trend data 21 and water quality data 22 related to the trend data 21 and water quality data 22 corresponding to the inference target data.
The approximation model learning data acquisition unit 27 transmits the acquired trend data 21 and water quality data 22 to the data processing unit 28 .
The data processing unit 28 processes each piece of data received from the approximation model learning data acquisition unit 27 to generate a plurality of associated data having values related to the inference target data.
The data processing unit 28 transmits the associated data to the approximation model generation unit 32 .

The approximation model generating unit 32 receives the inference target data and a plurality of related data. Based on these data, the approximation model generating unit 32 generates an approximation model that approximates the inference model 50 for the inference target data, and generates approximation model parameters.
The model parameters of the approximation model are determined so as to locally approximate the inference model 50 at a portion related to the inference target data. Therefore, the approximation model does not approximate the entire inference model 50, but is generated so as to approximate a portion related to the input inference target data each time the inference model 50 performs inference.
In this modification, the above-mentioned approximation model is generated by applying SHAP (Shapley Additive explanations). SHAP is an approach that quantifies the importance, i.e., the contribution, of each feature of the data to be inferred based on the degree of change in the inference result when each feature is changed for the data to be inferred.

上式における複数の係数ｗ＝（ｗ_０、ｗ_１、…、ｗ_ｎ）が、上記実施形態と同様に、近似モデルｇを表す近似モデルパラメータである。
ここで、ｚ（ａ）は、次式９で表される。

上式におけるδ_ａｉは、上記実施形態における式６と同様、所謂クロネッカーのデルタである。すなわち、上式においては、説明変数ｘ＝（ｘ_１、ｘ_２、…、ｘ_ｎ）のｉ番目の要素ｘ_ｉが、推論対象データのｉ番目の要素の値ａ_ｉと等しい場合に１と、等しくない場合に０と、それぞれなるように設定されている。 More specifically, if the data to be inferred is a and the inference model 50 for inferring the driving operation amount 34 is f, the approximation model generation unit 32 generates a simplified model g, as expressed by the following equation 8, for the explanatory variable x = a = ( _a1 , _a2 , ..., _an ) to be inferred, as an approximation model g that approximates f.

The multiple coefficients w=(w ₀ , w ₁ , . . . , w _n ) in the above equation are approximation model parameters representing the approximation model g, similarly to the above embodiment.
Here, z(a) is expressed by the following equation 9.

In the above formula, δ _ai is the so-called Kronecker delta, as in formula 6 in the above embodiment. That is, in the above formula, the i-th element x _i of the explanatory variable x = (x ₁ , x ₂ , ..., x _n ) is set to 1 when it is equal to the value a _i of the i-th element of the inference target data, and is set to 0 when it is not equal.

In SHAP, the calculation is restricted by three properties: local accuracy, missingness, and consistency.
Local accuracy is a constraint that the values of the inference model f to be explained and the approximation model g(f) are equal when the explanatory variable x is equal to the inference target data a and x = a. That is, the following formula 10 holds for the inference target x = a = (a ₁ , a ₂ , ..., a _n ).

Missingness is a constraint that the contribution w _i of a non-existent explanatory factor z _i is zero. That is, in the above formula 8, for i such that z _i (a)=0, w _i =0 holds.

ここで、ｚ^（ｉ）（ａ）は、上式９で表されるｚ（ａ）のうち、ｉ番目の説明因子ｚ_ｉ（ａ）をゼロとしたベクトルである。すなわち、本制約は、２つの推論モデルｆ_１、ｆ_２を比較したときに、ｉ番目の説明因子が存在することによる推論モデルｆ_１の変化が推論モデルｆ_２の変化よりも大きければ、推論モデルｆ_１の、ｉ番目の説明因子による貢献度合いｗ_ｉ（ｆ_１）は、推論モデルｆ_２の、ｉ番目の説明因子による寄与度ｗ_ｉ（ｆ_２）よりも大きいことを示すものである。 Consistency is a constraint as described below. First, between an inference model f and an approximation model g(f), it is assumed that f(a) and g(f)(z(a)) are equal.
In this case, the following equation 11 holds for the two inference models f ₁ and f ₂ .

Here, z ⁽ⁱ⁾ (a) is a vector in which the i-th explanatory factor z _i (a) in z(a) expressed by the above formula 9 is set to zero. In other words, this constraint indicates that when comparing two inference models f ₁ and f ₂ , if the change in inference model f ₁ due to the presence of the i-th explanatory factor is greater than the change in inference model f ₂ , then the contribution degree w _i (f ₁ ) of the i-th explanatory factor in inference model f ₁ is greater than the contribution degree w _i (f ₂ ) of the i-th explanatory factor in inference model f ₂ .

上式において、｜Ｓ｜は、集合Ｓの元の数である。すなわち、ｉ番目の要因を加えたときと加えなかったときの差を、全ての組み合わせの分だけ総和して平均したものとして、ｉ番目の説明変数の寄与度は計算される。
この寄与度、すなわち複数の係数ｗ＝（ｗ_０、ｗ_１、…、ｗ_ｎ）が、近似モデルｇを表す近似モデルパラメータである。 The contribution degree w _i (f) that satisfies each of the above constraints can be uniquely determined as shown in the following equation 12.

In the above formula, |S| is the number of elements in set S. In other words, the contribution of the i-th explanatory variable is calculated by averaging the sum of the differences between when the i-th factor is added and when it is not added for all combinations.
The contributions, that is, the multiple coefficients w=(w ₀ , w ₁ , . . . , w _n ), are approximate model parameters that represent the approximate model g.

Thus, in this modified example, as in the above embodiment, the approximation model generation unit 32 generates an approximation model that approximates the inference model 50 for the inference target data based on the inference target data and a plurality of related data having values related to the inference target data, and generates approximation model parameters.
The approximate model generating unit 32 transmits the approximate model parameters w=(w ₀ , w ₁ , . . . , w _n ) generated as described above to the inference basis output unit 33 .

The inference basis output unit 33 receives the approximate model parameters w=(w ₀ , w ₁ , . . . , w _n ).
In this modified example, as in the above embodiment, with respect to the approximate model parameters w = ( _w0 , _w1 , ..., _wn ), when the numerical value of the i-th coefficient _wi is a large positive value, it can be interpreted that the value of the i-th variable _xi being _ai strongly supports that the result of the model f is f(x). For this reason, the approximate model parameters w = ( _w0 , _w1 , ..., _wn ) can be applied as the strength of the basis for the judgment of the machine learning device 40.
As explained using Figure 4 in the above embodiment, the inference basis output unit 33 outputs the inference basis 35 of the driving operation amount 34 based on the inference target data a in the inference model 50 based on the approximation model parameter w to the outside, for example to a display device not shown.

It goes without saying that this modification provides the same effects as the embodiment already described.

The wastewater treatment facility operation amount inference basis output device and operation amount inference basis output method of the present invention are not limited to the above-mentioned embodiment and each modified example described with reference to the drawings, and various other modified examples are conceivable within the technical scope.

For example, in the above first modified example, the determination as to whether the driving operation amount input by the input unit 60 differs from the driving operation amount 34 inferred in the inference model is performed by the inference basis output unit 33A, but it may be performed by another part. For example, the approximation model learning data acquisition unit may make the determination, and only if they differ, a series of processes for outputting the inference basis may be performed, including acquisition of data by the approximation model learning data acquisition unit, processing of the data by the data processing unit, generation of the approximation model by the approximation model generation unit, and output of the results by the inference basis output unit.
In addition, in the above embodiment and each modified example, the machine learning machine is realized by a neural network, but may be realized by other machine learning methods, including ensemble learning methods such as Random Forest and XGBoost (eXtreme Gradient Boosting).

In addition, without departing from the spirit of the present invention, it is possible to select and discard the configurations given in the above embodiment and each modified example, or to change them to other configurations as appropriate.
For example, since the second modified example is a further modified example of the first modified example, similarly to the first modified example, the inference basis 35 is configured to be output to the outside only when the driving operation amount input by the input unit 60 differs from the driving operation amount 34 inferred in the inference model. Alternatively, in the second modified example, similarly to the above embodiment, the inference basis 35 may be configured to be output to the outside even when the driving operation amount input by the input unit 60 and the driving operation amount 34 inferred in the inference model are the same.

1 Sewage treatment facility 20, 20A Operation operation amount inference basis output device 21 Trend data (facility parameters)
22 Water quality data (facility parameters)
23 Driving operation amount data 29 Driving operation amount learning unit 31 Driving operation amount inference unit 32 Approximation model generation unit 33, 33A Inference basis output unit 34 Driving operation amount 35 Inference basis 40 Machine learning device 50 Inference model 60 Input unit

Claims (6)

A sewage treatment facility operation amount inference basis output device that infers an operation amount of equipment in the sewage treatment facility and outputs a basis for the inference when changing an operation state of the sewage treatment facility for the purpose of improving water quality,
a driving operation amount learning unit including a machine learning device, the machine learning device learning data corresponding to each of a plurality of facility parameters related to the driving situation as input, and driving operation amount data corresponding to the learning data as teacher data to train the machine learning device by machine learning, thereby generating model parameters related to the inference of the driving operation amount;
a driving operation amount inference unit that inputs inference target data that is to be inferred for the driving operation amount into an inference model constructed using the model parameters, and infers the driving operation amount;
an approximation model generating unit that generates an approximation model that approximates the inference model for the inference target data based on the inference target data and a plurality of related data having values related to the inference target data, and generates approximation model parameters;
an inference basis output unit that outputs the basis for inferring the driving operation amount based on the inference target data in the inference model based on the approximation model parameters;
An operation quantity inference basis output device for a sewage treatment facility, comprising: The sewage treatment facility operation quantity inference basis output device according to claim 1, wherein the approximation model is expressed as a linear sum using multiple coefficients based on values constituting the inference target data, and the approximation model parameters are the multiple coefficients. An input unit that receives an input of the driving operation amount from an outside,
An operation operation quantity inference basis output device for a sewage treatment facility as described in claim 1 or 2, which outputs the basis of the inference to the outside only when the operation operation quantity input by the input unit differs from the operation operation quantity inferred in the inference model. The operation amount inference basis output device for a wastewater treatment facility according to claim 3, which, when the operation amount input by the input unit differs from the operation amount inferred in the inference model, increases the number of copies of the multiple related data used in generating the approximation model as the difference between them increases. The operation quantity inference basis output device for a wastewater treatment facility according to any one of claims 1 to 4, wherein the plurality of facility parameters include trend data including sludge concentration and sludge flow rate, and water quality data of the water treated in the wastewater treatment facility. A method for outputting an inference basis for an operation amount of a sewage treatment facility, which infers an operation amount of equipment in the sewage treatment facility and outputs a basis for the inference when changing an operation state of the sewage treatment facility for the purpose of improving water quality, comprising:
learning data corresponding to each of a plurality of facility parameters related to the driving situation is input, and a machine learning device is trained by using the driving operation amount data corresponding to the learning data as teacher data to generate model parameters related to the inference of the driving operation amount;
inputting inference target data that is to be inferred as the driving operation amount into an inference model constructed using the model parameters, and inferring the driving operation amount;
generating an approximation model that approximates the inference model for the inference target data based on the inference target data and a plurality of associated data having values related to the inference target data, and generating approximate model parameters;
An operation quantity inference basis output method for a sewage treatment facility, which outputs the basis for inferring the operation quantity based on the inference target data in the inference model based on the approximation model parameters.

Images