Learning model generation device and system

The learning model generation device uses machine learning to predict and control membrane fouling in membrane filtration, stabilizing operations by adjusting aeration levels to prevent sudden pressure increases.

JP2024167301AActive Publication Date: 2024-12-03KUBOTA CORP
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
JP2024145214
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-12-03
Estimated Expiration
2040-06-01

AI Technical Summary

Technical Problem

The sudden increase in transmembrane differential pressure (TMP jump) due to membrane fouling in membrane filtration processing can lead to operational issues in membrane separation equipment, which existing control methods fail to predict effectively.

Method used

A learning model generation device that utilizes machine learning to estimate the future state of a separation membrane by acquiring and analyzing operational data, including membrane filtration pressure and aeration amount, to predict and control membrane fouling, thereby stabilizing the filtration process.

Benefits of technology

The system accurately predicts membrane fouling, allowing for stable membrane filtration operations by adjusting aeration levels, reducing the likelihood of sudden pressure increases and maintaining optimal operating conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024167301000001_ABST
    Figure 2024167301000001_ABST
Patent Text Reader

Abstract

To perform a stable membrane filtration operation by estimating a future state of a separation membrane.SOLUTION: A learning model generation device (1) comprises: an input data acquisition unit (21) which acquires input data derived from operation data including a membrane filtration pressure and an amount of diffusion measured during a membrane filtration operation; and a learning unit (13) which generates a learning model (31) to estimate a state of a separation membrane based on a machine learning where the acquired input data is used as input.SELECTED DRAWING: Figure 3
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a learning model generation device and the like that is applied to a membrane filtration process in which treated water is obtained by permeating a separation membrane while diffusing air over the membrane surface of a separation membrane that is immersed in the water to be treated. [Background technology]

[0002] Patent Document 1 discloses a technology for determining the amount of diffused air at the control time for controlling the amount of diffused air by comparing the amount of change, rate of change, or rate of increase in transmembrane pressure from a certain past point in time with a target rate of increase selected from a preset threshold value or organic matter concentration. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 6342101 Summary of the Invention [Problem to be solved by the invention]

[0004] In membrane filtration processing using a membrane separation device, a phenomenon (TMP jump) in which the transmembrane pressure difference suddenly rises may occur due to the progression of fouling on the membrane surface, etc. However, in the technology disclosed in Patent Document 1, if the above-mentioned fouling is progressing at the control time, in other words, if there is an abnormality in the separation membrane, the phenomenon of the transmembrane pressure difference suddenly rising in the future may occur, which may cause a problem in the operation of the membrane separation device.

[0005] One aspect of the present invention aims to realize a learning model generation device for predicting the future state of a separation membrane and performing stable membrane filtration operation. [Means for solving the problem]

[0006] In order to solve the above problems, a learning model generation device according to one aspect of the present invention is provided, which includes a separation membrane immersed in water to be treated and an aeration device that diffuses aeration over the membrane surface of the separation membrane, and includes an input data acquisition unit that acquires input data derived from operational data including a membrane filtration pressure and an aeration amount measured during a membrane filtration operation performed in a membrane separation device that obtains treated water that has permeated the separation membrane while diffusing aeration over the membrane surface using the aeration device; a teacher data generating unit that generates teacher data in which the input data is associated with a label that indicates a state of the separation membrane for the input data, the teacher data generating unit associates a label that indicates a state of the separation membrane after a predetermined time based on a degree of satisfaction of a specific condition at a time when the operating data is acquired and the input data derived from operating data acquired a predetermined time before the time when the operating data is acquired, The specific conditions include a first condition that a fluctuation rate, which is a fluctuation amount per unit time of the transmembrane pressure, is less than a first predetermined value, a second condition that the transmembrane pressure is less than a second predetermined value, a third condition that the fluctuation rate is equal to or greater than the first predetermined value and less than a third predetermined value, a fourth condition that the fluctuation rate is equal to or greater than the third predetermined value, and a fifth condition that the transmembrane pressure is equal to or greater than the second predetermined value, the teacher data generation unit associates a normal label with the input data that satisfies both the first condition and the second condition, associates an intermediate label with the input data that satisfies both the third condition and the second condition, and associates an abnormal label with the input data that satisfies the fourth condition or the fifth condition; The system further comprises a learning unit that generates a learning model for predicting the future state of the separation membrane through machine learning using the acquired input data as input, the learning model outputting the probability that the future state of the separation membrane will be normal or abnormal through supervised learning using the teacher data generated by the teacher data generation unit. Effect of the Invention

[0007] According to one aspect of the present invention, it is possible to predict the future state of the separation membrane and perform stable membrane filtration operation. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram showing an overview of a separation membrane state prediction system according to a first embodiment of the present invention. [Diagram 2] FIG. 13 is a graph showing changes in membrane filtration pressure over time during membrane filtration operation. [Diagram 3] 1 is a block diagram showing an example of a configuration of a main part of a learning model generating device and an estimation device; [Figure 4] FIG. 11 is a diagram showing a specific example of input data. [Diagram 5] FIG. 4 is a diagram illustrating predetermined conditions for determining the state of a separation membrane. [Figure 6] FIG. 13 is a diagram showing a specific example of teacher data. [Figure 7] 13 is a flowchart showing an example of the flow of a learning model generation process. [Figure 8] 13 is a flowchart showing an example of the flow of a learning model update process. [Figure 9] 11 is a flowchart showing an example of the flow of an estimation process and an air diffusion amount control process. [Figure 10] FIG. 13 is a diagram showing a specific example of teacher data. [Figure 11] FIG. 11 is a block diagram showing an example of the configuration of the main parts of a learning model generating device and an inference device according to a second embodiment of the present invention. [Figure 12] 13 is a flowchart showing an example of the flow of a learning model generation process. [Figure 13] 11 is a flowchart showing an example of the flow of an estimation process and an air diffusion amount control process. [Figure 14] FIG. 1 is a schematic diagram showing an outline of a process according to a reference embodiment of the present invention; [Figure 15] FIG. 1 is a block diagram showing an example of a configuration of a main part of a regression model generating device and an estimation device according to a reference embodiment. [Figure 16] 11 is a flowchart showing an example of the flow of an estimation process and an air diffusion amount control process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] [Embodiment 1] <Outline of separation membrane state prediction system> FIG. 1 is a diagram showing an overview of a separation membrane state estimation system 100 according to this embodiment. The state estimation system 100 is a system that estimates the state of a separation membrane 93 used in membrane filtration operation using a learning model generated by machine learning, and then controls the amount of diffused air supplied to the separation membrane 93 according to the estimation result. The "state" of the separation membrane 93 refers to the degree of contamination. A normal state is a state in which the degree of contamination is low and a phenomenon in which a sudden increase in transmembrane pressure is unlikely to occur for the time being.

[0010] The state prediction system 100 includes a learning model generation device 1, an prediction device 2, a memory device 3, an operation data acquisition device 4, an input data calculation device 5, an aeration amount control device 8, and a membrane separation device 90, and may also include a memory device 7.

[0011] The implementation method and location of the learning model generation device 1, the estimation device 2, the memory device 3, the operating data acquisition device 4, the input data calculation device 5, and the memory device 7 are not limited, but as a preferred typical example, the operating data acquisition device 4 and the air diffusion amount control device 8 are implemented as a PLC (Programmable Logic Controller), the estimation device 2, the input data calculation device 5, and the memory device 7 are edge computing, and the learning model generation device 1 and the memory device 3 are cloud computing.

[0012] (Membrane separation device 90) The membrane separation device 90 is a device that performs membrane filtration operation to filter the water to be treated using a separation membrane and obtain treated water that has permeated the separation membrane. The treated water can also be expressed as water to be treated from which impurities have been removed by filtration.

[0013] The membrane separation device 90 includes a membrane separation tank 91, a separation membrane 93, an aeration pipe 94, an aeration device 95, a filtrate water pipe 96, and a filtration pump 97. The membrane separation tank 91 stores the water to be treated 92. The separation membrane 93 is immersed in the water to be treated 92 and filters the water to be treated 92. The filtrate water pipe 96 is connected to the membrane separation tank 91 via the separation membrane 93, and distributes treated water obtained by filtering the water to be treated 92 through the separation membrane 93. The filtration pump 97 is connected to the separation membrane 93 via the filtrate water pipe 96, and discharges the treated water. The aeration device 95 supplies air for peeling off impurities attached to the separation membrane 93. In other words, the aeration device 95 diffuses air on the membrane surface of the separation membrane 93. The aeration pipe 94 is disposed directly below the separation membrane 93, and supplies air bubbles that flow from below to above the separation membrane 93 by air supplied from the aeration device 95.

[0014] The membrane separation tank 91 may be formed of a leak-proof material such as concrete, stainless steel, or resin as long as it is capable of receiving and storing the untreated water 92 flowing into the membrane separation tank 91. Furthermore, the structure of the membrane separation tank 91 may be such that it is leak-proof.

[0015] The separation membrane 93 may be any membrane capable of separating solids and liquids, such as a hollow fiber membrane or a flat membrane. Examples of the separation membrane 93 include, but are not limited to, a reverse osmosis (RO) membrane, a nanofiltration (NF) membrane, an ultrafiltration (UF) membrane, and a microfiltration (MF) membrane.

[0016] The air diffuser 94 may be made of any material as long as it can supply air bubbles, such as glass, stainless steel, sintered metal, or resin. The air diffuser 95 may be any device capable of pumping air, such as a blower.

[0017] (Driving data acquisition device 4) The operation data acquisition device 4 acquires operation data measured during membrane filtration operation using various sensors, etc. The operation data according to this embodiment includes at least the membrane filtration pressure and the amount of diffused air, and more preferably includes the membrane filtration flow rate. The membrane filtration pressure is acquired, for example, from a pressure gauge arranged in the filtrate piping 96 between the separation membrane 93 and the filtration pump 97. The amount of diffused air is the amount of air supplied by the air diffuser 95 and is acquired directly from the air diffuser 95. The membrane filtration flow rate is acquired, for example, from a flow meter arranged in the filtrate piping 96. The operation data acquisition device 4 transmits the acquired operation data to the input data calculation device 5.

[0018] (Input data calculation device 5) The input data calculation device 5 derives input data to be input to the learning model generation device 1 and the estimation device 2 from the received operating data. The input data is data representing the feature amount of the operating data, and may be the operating data itself or may be obtained by performing a calculation on the operating data. Then, in the phase of generating the learning model, the input data calculation device 5 transmits the calculated input data directly to the learning model generation device 1, or transmits it to the storage device 7 for storing the input data. Meanwhile, in the phase of estimating the state of the separation membrane 93, the input data calculation device 5 transmits the calculated input data to the estimation device 2. Details of the input data will be described later.

[0019] (Learning model generation device 1) The learning model generation device 1 generates a learning model for predicting the state of the separation membrane 93 by machine learning using the received input data as input, and stores the model in the storage device 3. The details of the learning model generation will be described later.

[0020] (Storage device 3) The storage device 3 stores the learning model generated by the learning model generation device 1. Note that the storage device 3 may store programs and data other than the learning model.

[0021] (Guessing device 2) The estimation device 2 accesses the learning model stored in the storage device 3, and uses the learning model to estimate the state of the separation membrane 93 from the input data received from the input data calculation device 5. Details of estimating the state of the separation membrane 93 will be described later.

[0022] (Air diffusion control device 8) The air diffusion amount control device 8 determines the level of the air diffusion amount of the air diffusion device 95 (hereinafter simply referred to as "air diffusion amount level") according to the estimation result by the estimation device 2, and controls the air diffusion device 95 to diffuse air at the determined air diffusion amount level. As an example, when the estimation result by the estimation device 2 is "normal", the air diffusion amount control device 8 lowers the air diffusion amount level below the current value, while when the estimation result by the estimation device 2 is "abnormal", it raises the air diffusion amount level above the current value.

[0023] The range of increase and decrease of the air diffusion amount level may be a fixed value or a variable value. In the latter case, the air diffusion amount control device 8 may, for example, determine the current variable value according to the most recent increase and decrease of the air diffusion amount level.

[0024] For example, if the predicted result is "normal," the amount of reduction is decreased depending on the number of times the air diffusion level was most recently increased, or the amount of reduction is increased depending on the number of times the air diffusion level was most recently decreased. Conversely, if the predicted result is "abnormal," the amount of increase is increased depending on the number of times the air diffusion level was most recently increased, or the amount of reduction is decreased depending on the number of times the air diffusion level was most recently decreased. Note that hereinafter, the air diffusion amount control device 8 will be described assuming that this example is applied.

[0025] <Membrane filtration operation cycle> Fig. 2 is a diagram showing the change over time in membrane filtration pressure measured during membrane filtration operation performed by the membrane separation device 90. A membrane filtration operation cycle (hereinafter, sometimes referred to as a "unit period") will be described with reference to Fig. 2. A membrane filtration operation cycle consists of an operation period (e.g., about 5 minutes) during which the membrane filtration operation is performed, and a rest period (e.g., about 1 minute) following the operation period during which the membrane filtration operation is not performed. The membrane filtration operation is an intermittent operation in which this cycle is repeated.

[0026] In the condition estimation system 100, as an example, it is preferable that each time an operating period ends and an idle period begins, the input data calculation device 5 derives input data using the operating data acquired by the operating data acquisition device 4 during the operating period. Then, following this, the estimation device 2 estimates the state of the separation membrane 93 and controls the aeration device 95 in accordance with the estimation result. In this way, in the condition estimation system 100, the state of the separation membrane 93 is estimated for each cycle of membrane filtration operation, and the amount of air diffused by the aeration device 95 can be appropriately controlled.

[0027] <Configuration of main parts of learning model generation device 1> FIG. 3 is a block diagram showing an example of the configuration of the main parts of the learning model generating device 1 and the estimation device 2. As shown in FIG.

[0028] The learning model generating device 1 includes a control unit 10. The control unit 10 controls each part of the learning model generating device 1, and is realized by a processor and a memory, for example. In this example, the processor accesses a storage (not shown), loads a program (not shown) stored in the storage into the memory, and executes a series of instructions included in the program. This constitutes each part of the control unit 10.

[0029] As these units, the control unit 10 includes an input data acquisition unit 11, a teacher data generation unit 12, and a learning unit 13.

[0030] (Input data acquisition unit 11) The input data acquisition unit 11 acquires the input data directly from the input data calculation device 5, or acquires the input data from the storage device 7 that stores the input data calculated by the input data calculation device 5. Then, the input data acquisition unit 11 outputs the acquired input data to the teacher data generation unit 12.

[0031] (Example of input data) 4 is a diagram showing a specific example of input data derived from operational data. The input data calculation device 5 calculates, from the operational data, the maximum transmembrane filtration pressure, the minimum transmembrane filtration pressure, the standard deviation of the transmembrane filtration pressure, the average transmembrane filtration pressure, the transmembrane pressure, the fluctuation rate of the transmembrane pressure, the fluctuation amount of the transmembrane pressure, and the fluctuation rate of the transmembrane pressure, for example.

[0032] The maximum value of the membrane filtration pressure (hereinafter referred to as "maximum membrane filtration pressure") is the maximum value of the membrane filtration pressure in a certain cycle of membrane filtration operation (hereinafter referred to as "cycle of interest"). The minimum value of the membrane filtration pressure (hereinafter referred to as "minimum membrane filtration pressure") is the minimum value of the membrane filtration pressure in the cycle of interest. The standard deviation value of the membrane filtration pressure is the standard deviation value of the membrane filtration pressure in the cycle of interest. The average value of the membrane filtration pressure (hereinafter referred to as "average membrane filtration pressure") is the average value of the membrane filtration pressure in the cycle of interest.

[0033] The transmembrane pressure (TMP) is the difference between the pressure applied to the water 92 to be treated and the pressure applied to the treated water in the separation membrane 93. The rate of change of the transmembrane pressure (hereinafter simply referred to as the "rate of change") is calculated as the gradient (ΔTMP / ΔT) of the transmembrane pressure in a predetermined period (hereinafter referred to as "P") from a predetermined time point in a cycle of interest. P is appropriately selected between several hours and several days. As an example, the rate of change may be calculated as the gradient of a regression model (linear regression) in the time-dependent change of the transmembrane pressure in P. In this case, the rate of change may not be a negative value. The amount of change of the transmembrane pressure (hereinafter simply referred to as the "amount of change") refers to the amount of change in P. As an example, the amount of change is calculated as the difference between the value of TMP at a predetermined time point and the value of TMP at the time when P has passed. The rate of change of the transmembrane pressure (hereinafter simply referred to as the "rate of change") refers to the rate of change in P. As an example, the rate of change is calculated by dividing the rate of change by the transmembrane pressure (ΔTMP / (TMP×ΔT)).

[0034] Furthermore, the input data calculation device 5 calculates, for example, an average value of the air diffusion amount and an integrated value of the air diffusion amount from the air diffusion amount, which is the operating data. The average value of the air diffusion amount (hereinafter referred to as the "average air diffusion amount") is the average value of the air diffusion amount in the cycle of interest. The integrated value of the air diffusion amount (hereinafter referred to as the integrated air diffusion amount) is the integrated value of the air diffusion amount at P, and is calculated, for example, as the integral value of the average air diffusion amount at P.

[0035] Furthermore, the input data calculation device 5 calculates, from the membrane filtration flow rate, which is the operational data, an average value of the membrane filtration flow rate and an integrated value of the membrane filtration flow rate, for example. The average value of the membrane filtration flow rate (hereinafter referred to as the "average membrane filtration flow rate") is the average value of the membrane filtration flow rate in the cycle of interest. The integrated value of the membrane filtration flow rate (hereinafter referred to as the "integrated membrane filtration flow rate") is the integrated value of the membrane filtration flow rate at P, and is calculated, for example, as the integral value of the average membrane filtration flow rate at P.

[0036] Although not shown, the input data is associated with time information indicating the time when the driving data from which the input data is derived was acquired.

[0037] (Teacher data generation unit 12) The teacher data generating unit 12 generates teacher data in which input data is associated with a label indicating the state of the separation membrane 93 for the input data. Then, the teacher data generating unit 12 outputs the generated teacher data to the learning unit 13.

[0038] The labels include, for example, a "normal label" that is associated with input data that will result in a normal future state of the separation membrane 93, and an "abnormal label" that is associated with input data that will result in an abnormal future state of the separation membrane 93. The types of labels are not limited to two, and may be three or more types.

[0039] The association between the input data and the labels may be performed manually by a skilled worker or automatically. When performing the association manually, the teacher data generating unit 12 may provide an input / output interface for performing the association. An example of an automatic association method will be described below.

[0040] The teacher data generating unit 12 first identifies the operating condition at the time when the operating data from which the input data of interest is derived was acquired (hereinafter referred to as the "current time"). As an example of the specific condition, for example, the following two conditions are preferable: (First condition) The fluctuation speed is less than a first predetermined value; (Second condition) The transmembrane pressure difference is less than a second predetermined value. These conditions are typical conditions for the state of the separation membrane 93 being normal. If these conditions are satisfied, the state of the separation membrane 93 can be considered to be normal. The first and second predetermined values ​​are, for example, 0.08 kPa / h and 10 kPa, respectively, but are not limited to these.

[0041] Next, the teacher data generating unit 12 identifies input data associated with time information indicating a time a first predetermined time before the current time (hereinafter referred to as "first input data"), and identifies input data associated with time information indicating a time a second predetermined time before the current time (hereinafter referred to as "second input data"). The first and second predetermined times are, for example, 3 hours and 24 hours, respectively, but are not limited thereto.

[0042] Then, the teacher data generating unit 12 associates the satisfaction of the first condition at the current time with the first input data. That is, for example, "whether the fluctuation speed after three hours is less than a first predetermined value" is associated with the first input data.

[0043] Furthermore, the teacher data generating unit 12 associates the satisfaction of the second condition at the current time with the second input data. That is, for example, "whether the transmembrane pressure after 24 hours is less than a second predetermined value" is associated with the second input data.

[0044] Then, the teacher data generating unit 12 repeats the above process for all input data to be processed by replacing the input data of interest. As a result, for each input data, (1) the satisfaction of the first condition after a first predetermined time, and (2) the satisfaction of the second condition after a second predetermined time are identified as the future driving conditions.

[0045] Then, for input data that satisfies both (1) and (2) above, the teacher data generating unit 12 determines that the input data will normalize the state of the separation membrane 93 in the future, and associates a "normal label" with the input data. On the other hand, for input data that does not satisfy either (1) or (2) above, the teacher data generating unit 12 determines that the input data will not normalize the state of the separation membrane 93 in the future, and associates an "abnormal label" with the input data. In this way, the teacher data generating unit 12 automatically associates input data with labels.

[0046] The specific conditions are not limited to the first condition and the second condition. The number of specific conditions is not limited to two, but may be one, or may be three or more. The more specific conditions there are, the more accurate labeling based on multiple scales becomes possible, and it is expected that the estimation accuracy of the estimation device 2 will improve.

[0047] FIG. 5 is a diagram for explaining the specific conditions described above. Graph 81 visualizes the sufficiency of the second condition associated with each input data, that is, whether the transmembrane pressure after 24 hours is less than 10 kPa. Graph 82 visualizes the sufficiency of the first condition associated with each input data, that is, whether the fluctuation rate after 3 hours is less than 0.08 kPa / h. In each graph shown in FIG. 5, the case where the condition is satisfied is indicated as "normal" (gray dots in FIG. 5), and the case where the condition is not satisfied is indicated as "abnormal" (black dots in FIG. 5).

[0048] As an example of a specific condition different from the first and second conditions, a condition that "the gradient does not increase abruptly in the change over time of the transmembrane pressure" may be used. Graph 83 visualizes whether or not the gradient increases abruptly in the transmembrane pressure of each input data based on the condition.

[0049] Fig. 6 is a diagram showing a specific example of teacher data. The teacher data generating unit 12 associates a normal label with input data that satisfies both the first condition and the second condition. That is, the normal label is associated with input data in which the rate of change after 3 hours is less than 0.08 kPa / h and the transmembrane pressure after 24 hours is less than 10 kPa. On the other hand, an abnormal label is associated with input data that does not satisfy at least one of the conditions. In Fig. 6, a label is associated with the right end of each record.

[0050] (Learning Section 13) Returning to FIG. 3, the learning unit 13 will be described. The learning unit 13 generates a learning model 31 for predicting the state of the separation membrane 93 by machine learning using the teacher data generated by the teacher data generation unit 12 as an input. The learning unit 13 stores the generated learning model 31 in the storage device 3. The learning unit 13 uses a plurality of teacher data acquired from the teacher data generation unit 12 to generate the learning model 31 by a known algorithm such as a neural network (NN). In this way, when input data is input, a learning model 31 is generated that outputs the probability that the future state of the separation membrane 93 will be normal (or the probability that it will be abnormal).

[0051] <Main components of the estimation device 2> The inference device 2 includes a control unit 20. The control unit 20 controls each part of the inference device 2, and is realized by a processor and a memory, for example. In this example, the processor accesses a storage (not shown), loads a program (not shown) stored in the storage into the memory, and executes a series of instructions included in the program. In this way, each part of the control unit 20 is configured.

[0052] As these units, the control unit 20 includes an input data acquisition unit 21 and an access unit 22 .

[0053] (Input data acquisition unit 21) The input data acquisition unit 21 acquires input data from the input data calculation device 5, and outputs it to the access unit 22. The input data is preferably derived from operational data measured in the most recent membrane filtration operation.

[0054] (Access part 22) The access unit 22 accesses the learning model 31 stored in the storage device 3. The access unit 22 includes an estimation unit .

[0055] The estimation unit 23 estimates the state of the separation membrane 93 from the input data acquired from the input data acquisition unit 21, using the learning model 31 accessed by the access unit 22. Specifically, the estimation unit 23 acquires the probability that the state of the separation membrane 93 will be normal (or the probability that the state will be abnormal), which is output from the learning model 31 as a result of inputting the input data to the learning model 31. This probability is the probability that the future state of the separation membrane 93 will be normal when membrane filtration operation is continued under the current operating conditions.

[0056] Then, the estimation unit 23 estimates whether the future state of the separation membrane 93 will be normal or abnormal based on the probability value obtained from the learning model 31. Specifically, when the probability value is equal to or greater than a threshold, the estimation unit 23 estimates that the future state of the separation membrane 93 will be normal. On the other hand, when the probability value is less than the threshold, the estimation unit 23 estimates that the future state of the separation membrane 93 will be abnormal. The threshold is, for example, 50(%), but is not limited to this example. Note that the state of the separation membrane 93 to be estimated is not limited to two states, "normal" and "abnormal", but may be three or more states. If there are three states, they are, for example, "normal", "intermediate between normal and abnormal", and "abnormal".

[0057] Then, the estimation unit 23 outputs this estimation result to the air diffusion amount control device 8. Note that the estimation unit 23 may output the probability value itself obtained from the learning model 31 to the air diffusion amount control device 8 as the estimation result.

[0058] In addition, the control unit 20 (input data acquisition unit 21 or estimation unit 23) may store the input data used for estimation in the storage device 7 in order to cause the learning model generation device 1 to update (i.e., re-learn) the learning model 31 using the input data.

[0059] <Learning model generation process flow> Fig. 7 is a flowchart showing an example of the flow of the learning model generation process executed by the learning model generation device 1. Note that the learning model generation process shown in Fig. 7 is a process for generating a new learning model 31, and is not a process for updating the learning model 31. In this example, it is assumed that the state of the separation membrane 93 is normal when both of the above-mentioned (first condition) that the fluctuation rate is less than a first predetermined value and (second condition) that the transmembrane pressure difference is less than a second predetermined value are satisfied. Also, it is assumed that the input data is automatically associated with the label.

[0060] First, the input data acquisition unit 11 acquires a plurality of pieces of input data from the input data calculation device 5 or the storage device 7 (step S1, hereinafter, the description of “step” will be omitted). The input data acquisition unit 11 outputs the acquired plurality of pieces of input data to the teacher data generation unit 12.

[0061] The teacher data generating unit 12 identifies the future driving conditions for each piece of input data by the above-mentioned method (S2). Specifically, the teacher data generating unit 12 identifies, for each piece of input data, (1) whether a first condition is satisfied after a first predetermined time, and (2) whether a second condition is satisfied after a second predetermined time.

[0062] Next, the teacher data generation unit 12 generates teacher data by associating each piece of input data with a label indicating the future state of the separation membrane 93 (S3). Specifically, input data that satisfies both (1) and (2) above is associated with a "normal label," while input data that does not satisfy either (1) or (2) above is associated with an "abnormal label." The teacher data generation unit 12 then outputs the generated teacher data to the learning unit 13.

[0063] Next, the learning unit 13 generates a learning model 31 that outputs the probability that the state of the separation membrane 93 will be normal (or the probability that the state will be abnormal) from the teacher data generated by the teacher data generation unit 12 (S4). Finally, the learning unit 13 stores the generated learning model 31 in the storage device 3 (S5). This ends the learning model generation process.

[0064] <Learning model update process flow> Fig. 8 is a flowchart showing an example of the flow of the learning model update process executed by the learning model generation device 1. The learning model update process shown in Fig. 8 is a process for updating the learning model 31 based on input data newly acquired by performing membrane filtration operation after the learning model 31 is generated by the learning model generation process shown in Fig. 7. Note that, in this example as well, when both of the above-mentioned (first condition) that the fluctuation speed is less than a first predetermined value and (second condition) that the transmembrane pressure difference is less than a second predetermined value are satisfied, the state of the separation membrane 93 is considered to be normal. Also, the input data is automatically associated with the label.

[0065] The input data acquisition unit 11 waits until it acquires the input data used for the estimation by the estimation device 2 from the input data calculation device 5 or the storage device 7 (S11). When the input data is acquired (YES in S11), the input data acquisition unit 11 outputs the input data to the teacher data generation unit 12.

[0066] The teacher data generation unit 12 identifies the driving situation at the time when the driving data from which the acquired input data is derived was acquired (S12). That is, the teacher data generation unit 12 identifies whether the first condition and the second condition are satisfied at that time.

[0067] Next, the teacher data generating unit 12 associates the satisfaction of the first condition with the input data associated with time information indicating a time before the time by a first predetermined time, and associates the satisfaction of the second condition with the input data associated with time information indicating a time before the time by a second predetermined time. Note that these input data were acquired in the past and are stored in the storage device 7.

[0068] The teacher data generating unit 12 labels the input data associated with both (1) the satisfaction of the first condition after a first predetermined time and (2) the satisfaction of the second condition after a second predetermined time according to the satisfaction of (1) and (2) above, and generates the data as teacher data (S13). Then, the teacher data generating unit 12 outputs the generated teacher data to the learning unit 13.

[0069] The learning unit 13 performs re-learning based on the teacher data generated this time by the teacher data generating unit 12, and updates the learning model 31 (S14). Then, the learning model update process returns to S11.

[0070] By updating the learning model in this manner, the estimation process by the estimation device 2 can be adapted to the latest driving conditions.

[0071] <Flow of estimation process and air diffusion amount control process> FIG. 9 is a flowchart showing an example of the flow of the estimation process executed by the estimation device 2 and the air diffusion amount control process executed by the air diffusion amount control device 8.

[0072] The input data acquiring unit 21 waits until it acquires input data from the input data calculation device 5 (S21). When the input data is acquired (YES in S21), the input data acquiring unit 21 outputs the input data to the access unit 22.

[0073] When the access unit 22 acquires the input data, it accesses the learning model 31 stored in the storage device 3 (S22). The estimation unit 23 included in the access unit 22 inputs the input data to the learning model 31 (S23) and acquires, for example, the probability that the state of the separation membrane 93 is normal from the learning model 31 (S24).

[0074] Then, as one example, the estimation unit 23 determines whether the acquired probability value is equal to or greater than a threshold value (S25). If it is determined that the probability value is equal to or greater than the threshold value (YES in S25), the estimation unit 23 estimates that the state of the separation membrane 93 will be normal (S26). On the other hand, if it is determined that the acquired probability value is less than the threshold value (NO in S25), the estimation unit 23 estimates that the state of the separation membrane 93 will be abnormal (S27). The estimation unit 23 outputs the estimation result to the air diffusion amount control device 8.

[0075] The air diffusion amount control device 8 determines the next air diffusion amount level based on the acquired estimation result and the most recent air diffusion amount level (S28). Then, the air diffusion amount control device 8 controls the air diffusion device 95 to diffuse air at the determined air diffusion amount level (S29). Then, the process returns to S21.

[0076] <Effects> As described above, the learning model generation device 1 according to this embodiment includes an input data acquisition unit 11 that acquires input data derived from operational data measured during membrane filtration operation. The learning model generation device 1 also includes a learning unit 13 that generates a learning model 31 for predicting the future state of the separation membrane 93 by machine learning using the input data as input.

[0077] The estimation device 2 according to this embodiment also includes an input data acquisition unit 21 that acquires input data derived from operational data measured during membrane filtration operation. The estimation device 2 also includes an access unit 22 that accesses a learning model 31. The estimation device 2 also includes an estimation unit 23 that uses the learning model 31 to estimate a future state of the separation membrane 93 from the input data.

[0078] As a result, the learning model generation device 1 generates a learning model 31 that infers the state of the separation membrane 93, and the estimation device 2 can infer the state of the separation membrane 93 using the learning model 31. If the state of the separation membrane 93 can be inferred, the amount of aeration can be controlled based on that, so that stable membrane filtration operation can be performed while maintaining the separation membrane 93 in a normal state (for example, a state in which the possibility of a sudden increase in transmembrane pressure is reduced).

[0079] The learning model generation device 1 also includes a teacher data generation unit 12 that generates teacher data in which input data is associated with a label indicating a future state of the separation membrane 93. The learning unit 13 generates the learning model 31 by supervised learning using the teacher data. As a result, the learning model 31 is generated by supervised learning, making it possible to make highly accurate predictions.

[0080] The labels include a normal label and an abnormal label. This makes it possible to accurately predict whether the separation membrane 93 will be normal or abnormal in predicting the state of the separation membrane 93. Therefore, based on the prediction result, the amount of diffused air can be appropriately controlled so that the separation membrane 93 will be normal.

[0081] Furthermore, the learning unit 13 updates the learning model by machine learning using the input data used by the estimation device 2. As a result, the learning model is updated by the input data used to estimate the state of the separation membrane 93, so that the estimation accuracy of the learning model 31 can be improved every time estimation is performed.

[0082] Furthermore, the input data acquisition unit 21 acquires input data for each cycle of membrane filtration operation, which is composed of an operating period and an idle period. Furthermore, the estimation unit 23 estimates the future state of the separation membrane 93 for each cycle. As a result, the future state of the separation membrane 93 is estimated for each cycle of membrane filtration operation, so that a sudden change in the state of the separation membrane 93 can be quickly confirmed.

[0083] Furthermore, the air diffusion amount control device 8 determines the air diffusion amount level according to the state of the separation membrane 93 estimated by the estimation device 2, and controls the air diffusion device 95 to diffuse air at the determined air diffusion amount level. As a result, the air diffusion amount level is automatically determined according to the estimated state of the separation membrane 93, so that the operator performing the membrane filtration operation does not need to adjust the air diffusion amount level. As a result, the workload of the operator can be reduced.

[0084] The air diffusion amount control device 8 further determines the current air diffusion amount level in accordance with the air diffusion amount level determined most recently by the device itself. This allows a more appropriate air diffusion amount level to be determined compared to a configuration that does not consider the most recent air diffusion amount level.

[0085] The input data derived from the membrane filtration pressure includes the maximum membrane filtration pressure, the minimum membrane filtration pressure, the standard deviation value of the membrane filtration pressure, the average membrane filtration pressure, and the transmembrane pressure in a certain cycle, and at least one of the fluctuation speed, the fluctuation amount, and the fluctuation rate at P before the cycle. The input data derived from the air diffusion amount includes at least one of the average air diffusion amount and the accumulated air diffusion amount in the cycle. The operating data further includes the membrane filtration flow rate, and the input data derived from the membrane filtration flow rate includes at least one of the average membrane filtration flow rate and the accumulated membrane filtration flow rate in the cycle. These can improve the estimation accuracy using the learning model 31.

[0086] <Modification of the first embodiment> As described above, the labels are not limited to two types, normal labels and abnormal labels, and may include labels indicating "intermediate states between normal and abnormal" (hereinafter, referred to as "intermediate labels"). An "intermediate state" refers to a state that is neither "normal" nor "abnormal."

[0087] FIG. 10 is a diagram showing a specific example of teacher data. Each piece of data shown in FIG. 10 is the same as each piece of data shown in FIG. 6. In the example shown in FIG. 10, the teacher data generating unit 12 associates a normal label with input data in which the fluctuation rate after 3 hours is less than 0.04 kPa / h and the transmembrane pressure after 24 hours is less than 10 kPa. The teacher data generating unit 12 also associates an intermediate label with input data in which the fluctuation rate after 3 hours is 0.04 kPa / h or more and less than 0.08 kPa and the transmembrane pressure after 24 hours is less than 10 kPa. The teacher data generating unit 12 also associates an abnormal label with input data in which the fluctuation rate after 3 hours is 0.08 kPa / h or more and the transmembrane pressure after 24 hours is less than 10 kPa. The teacher data generating unit 12 also associates an abnormal label with input data in which the transmembrane pressure after 24 hours is 10 kPa or more, regardless of the value of the fluctuation rate after 3 hours.

[0088] The estimation unit 23 may be configured to estimate the state of the separation membrane 93 as either "normal", "abnormal", or "intermediate". For example, when the probability value is equal to or greater than a first threshold, the estimation unit 23 estimates that the state of the separation membrane 93 will be normal. When the probability value is less than the first threshold and equal to or greater than a second threshold, the estimation unit 23 estimates that the state of the separation membrane 93 will be intermediate. When the probability value is less than the second threshold, the estimation unit 23 estimates that the state of the separation membrane 93 will be abnormal. The first threshold is, for example, 70%, and the second threshold is, for example, 40%, but is not limited thereto. When the estimation result is "intermediate", the air diffusion amount control device 8 may maintain the current air diffusion amount level.

[0089] Furthermore, the estimation unit 23 may be configured to output the probability value that the separation membrane 93 will be normal, output by the learning model 31, itself to the air diffusion amount control device 8. The air diffusion amount control device 8 may determine whether to increase or decrease the air diffusion amount level and the amount by which the air diffusion amount level will be increased or decreased, depending on the probability value. For example, the air diffusion amount control device 8 may be configured to decrease the air diffusion amount level when the probability is 50% or more, and to increase the amount by which the air diffusion amount level is decreased as the probability value increases. On the other hand, the air diffusion amount control device 8 may be configured to increase the air diffusion amount level when the probability is less than 50%, and to increase the amount by which the air diffusion amount level is increased as the probability value decreases.

[0090] [Embodiment 2] Other embodiments of the present invention will be described below. For ease of explanation, the same reference numerals will be used to designate members having the same functions as those described in the first embodiment, and the description thereof will not be repeated.

[0091] FIG. 11 is a block diagram showing an example of the configuration of the main parts of a learning model generating device 1A and an estimation device 2A according to this embodiment.

[0092] In the first embodiment, the learning model generation device 1 performs so-called supervised learning to generate the learning model 31. On the other hand, the learning model generation device 1A according to the present embodiment performs unsupervised learning to generate the learning model 31A.

[0093] <Configuration of Main Parts of Learning Model Generation Device 1A> Hereinafter, only differences will be described for components with the same names as those included in the learning model generation device 1. The learning model generation device 1A includes a control unit 10A. The control unit 10A includes an input data acquisition unit 11 and a learning unit 13A.

[0094] The learning unit 13A generates a learning model 31A by unsupervised learning using the input data acquired from the input data acquisition unit 11. Specifically, the learning unit 13A determines a boundary for dividing the distribution of the input data, and creates, as a learning result, a cluster (hereinafter referred to as a "normal cluster") in which input data in which the state of the separation membrane 93 is normal is classified, and a cluster (hereinafter referred to as an "abnormal cluster") in which input data in which the state of the separation membrane 93 is abnormal is classified. Then, the created cluster is stored in the storage device 3A as the learning model 31A.

[0095] In unsupervised learning, a large amount of unlabeled input data is used to learn how the input data is distributed. As the cluster analysis algorithm, a known algorithm such as K-means can be used.

[0096] <Main components of the estimation device 2A> Hereinafter, only the differences will be described for components with the same names as those included in the estimation device 2. The estimation device 2A is equipped with a control unit 20A. The control unit 20A includes an input data acquisition unit 21 and an access unit 22A. The access unit 22A accesses a learning model 31A stored in the storage device 3A. The access unit 22A includes a estimation unit 23A.

[0097] The estimation unit 23A inputs the input data acquired by the input data acquisition unit 21 to the learning model 31A accessed by the access unit 22A. Then, the estimation unit 23A acquires, as an output value output from the learning model 31A, either "normal" indicating that the input data belongs to a normal cluster, or "abnormal" indicating that the input data belongs to an abnormal cluster. The estimation unit 23A outputs the acquired output value to the air diffusion amount control device 8.

[0098] The air diffusion amount control device 8 lowers the air diffusion amount level when the acquired output value is "normal". On the other hand, the air diffusion amount control device 8 raises the air diffusion amount level when the acquired output value is "abnormal".

[0099] <Learning model generation process flow> Fig. 12 is a flowchart showing an example of the flow of a learning model generation process executed by the learning model generation device 1 A. Note that the learning model generation process shown in Fig. 12 is a process for generating a new learning model 31A.

[0100] The input data acquisition unit 11 acquires a plurality of pieces of input data from the input data calculation device 5 or the storage device 7 (S31). The input data acquisition unit 11 outputs the acquired plurality of pieces of input data to the learning unit 13A.

[0101] The learning unit 13A creates normal and abnormal clusters by setting boundaries on the distribution of the acquired input data (S32). The learning unit 13A then generates a learning model 31A including these clusters (S33) and stores it in the storage device 3A (S34). The learning unit 13 also stores the acquired teacher data in the storage device 3A. This completes the learning model generation process.

[0102] Although not shown, the learning model update process executed by the learning model generating device 1A will be described below.

[0103] When the input data acquisition unit 11 acquires input data from the input data calculation device 5, it outputs the input data to the learning unit 13A. The learning unit 13A updates the normal cluster and the abnormal cluster by redefining a boundary for separating a distribution consisting of the acquired input data and the input data stored in the storage device 3A. Then, the learning unit 13A generates a learning model 31A including the updated clusters, and overwrites the learning model 31A stored in the storage device 3A with the generated learning model 31A.

[0104] <Flow of estimation process and air diffusion amount control process> FIG. 13 is a flowchart showing an example of the flow of the estimation process executed by the estimation device 2A and the air diffusion amount control process executed by the air diffusion amount control device 8.

[0105] The input data acquiring unit 21 waits until it acquires input data from the input data calculation device 5 (S41). When the input data is acquired (YES in S41), the input data acquiring unit 21 outputs the input data to the access unit 22A.

[0106] When the access unit 22A acquires the input data, it accesses the learning model 31A stored in the storage device 3A (S42). The estimation unit 23A inputs the input data acquired by the input data acquisition unit 21 to the learning model 31A accessed by the access unit 22A (S43), and acquires the output value "normal" or "abnormal" output from the learning model 31A (S44). The estimation unit 23A outputs the acquired output value to the air diffusion amount control device 8.

[0107] The air diffusion amount control device 8 determines the next air diffusion amount level based on the acquired output value and the most recent air diffusion amount level (S45). Then, the air diffusion amount control device 8 controls the air diffusion device 95 to diffuse air at the determined air diffusion amount level (S46). Then, the process returns to S41.

[0108] <Effects> According to the learning model generation device 1A of this embodiment, the learning unit 13A generates a learning model 31A including normal clusters and abnormal clusters as a learning result by unsupervised learning. This makes it possible to easily generate a learning model even in a situation where sufficient teacher data cannot be prepared. Therefore, the learning by the learning model generation device 1A and the inference by the inference device 2A can be performed as a substitute for the learning by the learning model generation device 1 and the inference by the inference device 2, or as a pre-stage thereof.

[0109] <Example of outlier detection> The learning model 31A may create only normal clusters as learning results and perform so-called outlier detection. In this example, the learning model 31A identifies the distance between the input data and a representative point (e.g., center of gravity) in the normal cluster. The learning model 31A outputs that the input data is an outlier if the distance is equal to or greater than a predetermined value, and outputs that the input data is a normal value (not an outlier) if the distance is less than the predetermined value. When an outlier is output, the estimation unit 23A estimates that the state of the separation membrane 93 will be abnormal, and when a normal value is output, the estimation unit 23A estimates that the state of the separation membrane 93 will be normal. Note that the local outlier factor method (LOF), one-class support vector machine (OC-SVM), or the like may be used as an algorithm for outlier detection.

[0110] [Reference form] The present invention will be described below with reference to the embodiment. For the sake of convenience, the same reference numerals are given to members having the same functions as those described in the first and second embodiments, and the description thereof will not be repeated.

[0111] In this reference embodiment, a simulation is performed by repeatedly changing parameters based on long-term predictions made by regression analysis, with the aim of optimizing the operating costs involved in membrane filtration operation. The operating costs include the cost of energy required for aeration by the aeration device 95 (hereinafter referred to as "energy cost") and the cost required for chemical washing of the contaminated separation membrane 93 (hereinafter referred to as "chemical washing cost"). The chemical washing cost includes the cost of purchasing the chemicals used in the chemical washing and the labor costs of the workers who perform the chemical washing.

[0112] [Process 1] First, long-term prediction by regression analysis of this embodiment will be outlined with reference to FIG. 14. In the regression analysis of this embodiment, the input data obtained from the input data calculation device 5 or the storage device 7 is used as an explanatory variable, and data on the transmembrane pressure of the separation membrane 93 a predetermined n hours (n is a positive integer) after the time associated with the input data (hereinafter, referred to as "transmembrane pressure related data") is used as a response variable. The transmembrane pressure related data is, for example, at least one of the transmembrane pressure itself, the rate of change of the transmembrane pressure, the amount of change of the transmembrane pressure, and the rate of change of the transmembrane pressure, as described in the first embodiment. This regression analysis is used to perform a process of predicting the transmembrane pressure n hours later when the current value of the amount of diffused air is maintained. Then, data (hereinafter, referred to as "updated data") is generated by updating the transmembrane pressure related data of the input data using the predicted transmembrane pressure related data n hours later, and regression analysis is performed again on the updated data. This process is repeated N times (N is an integer of 2 or more). In other words, the process of "estimating transmembrane pressure related data n hours later through regression analysis, and updating the input data by changing the transmembrane pressure related data included in the input data to the transmembrane pressure related data n hours later" is executed N times.

[0113] Specifically, in the first repetition, the transmembrane pressure-related data n hours after the time associated with the input data is estimated by regression analysis using the input data as explanatory variables, and the transmembrane pressure-related data in the input data is updated using the estimated transmembrane pressure-related data to obtain updated data U (1,1) Generate.

[0114] In the Xth iteration (X is an integer between 2 and N), update data U (1,X-1) By using a regression analysis with the explanatory variables, the transmembrane pressure-related data X×n hours after the time associated with the input data is estimated, and the transmembrane pressure-related data is used to generate the updated data U (1,X-1) Updated data U that has updated the transmembrane pressure related data (1,X) Generate.

[0115] In the Nth iteration, the updated data U (1,N-1)By using a regression analysis with the explanatory variables, the transmembrane pressure-related data N×n hours after the time associated with the input data is estimated. The transmembrane pressure-related data is used to generate the updated data U (1,N-1) Updated data U that has updated the transmembrane pressure related data (1,N) may be generated.

[0116] Using the above, a total of N transmembrane pressures are estimated, n hours, 2n hours, ....N x n hours after the time associated with the input data. Using the above, the change in transmembrane pressure over time up to N x n hours when the current value of the air diffusion rate is maintained is estimated.

[0117] [Process 2] Next, in this reference embodiment, the above process 1 is performed using data (hereinafter referred to as "simulation data") obtained by changing data related to the amount of diffused air (part of the data, hereinafter referred to as "aeration amount related data") from the input data. The aeration amount related data is, as an example, at least one of the average amount of diffused air and the integrated amount of diffused air, as described in the first and second embodiments. This process is performed M times while changing the amount of diffused air related data (M is an integer of 2 or more). In other words, the above process 1 is performed for M pieces of data (input data and M-1 pieces of simulation data) that are different from each other. In other words, in process 2, the above process 1 is performed M times while changing part of the data included in the input data, thereby obtaining M estimated results regarding the change over time in the transmembrane pressure up to N×n hours.

[0118] The first process among M processes is the above-mentioned process 1 on the input data, and is specifically as described above.

[0119] For the Yth processing of M times (Y is an integer between 2 and M), the simulation data S in which the data related to the amount of diffused air among the input data has been changed is Y Specifically, in the first iteration of the process 1, the simulation data S Y By regression analysis with explanatory variables, the simulation data S YThe transmembrane pressure-related data n hours after the time associated with the time is estimated, and the transmembrane pressure-related data is used to generate simulation data S Y Updated data U that has updated the transmembrane pressure related data (Y,1) In the Xth iteration of the above process 1, the updated data U (Y,X-1) By regression analysis with explanatory variables, the simulation data S Y , and then, using the intermembrane pressure related data, update data U (Y,X-1) Updated data U that has updated the transmembrane pressure related data (Y,X) In the Nth iteration of the above process 1, the updated data U (Y,N-1) By regression analysis with explanatory variables, the simulation data S Y The transmembrane pressure related data N×n hours after the time associated with the time is estimated.

[0120] By carrying out the above processes 1 and 2, it is possible to estimate the change in transmembrane pressure over time for N×n hours when the current value of the aeration amount is maintained for each of the M aeration amounts. From these M estimation results, the timing of chemical washing and the aeration amount are determined to keep the operating cost reasonable.

[0121] In view of the fact that in the later stages of membrane filtration operation when fouling has progressed, a situation may arise in which the total operating costs cannot necessarily be reduced simply by continuing to suppress an increase in the transmembrane pressure difference, this reference embodiment aims to optimize the total operating costs by avoiding the long-term extension of the separation membrane 93 by controlling the amount of diffused air.

[0122] FIG. 15 is a block diagram showing an example of a configuration of a main part of a regression model generating device 6 and an estimation device 2B according to this embodiment.

[0123] <Configuration of Main Parts of Regression Model Generator 6> The regression model generating device 6 includes a control unit 60. The control unit 60 includes an input data acquiring unit 61, an association unit 62, and a regression model generating unit 63.

[0124] The input data acquisition unit 61 acquires input data from the input data calculation device 5 or the storage device 7 , and outputs the acquired input data to the association unit 62 .

[0125] The association unit 62 associates each piece of input data with transmembrane pressure-related data n hours after the time associated with the input data, where the value of n is, for example, 12 or 24, but is not limited to this example.

[0126] The associating unit 62 outputs the input data associated with the transmembrane pressure related data after n hours to the regression model generating unit 63. Note that the input data to which the transmembrane pressure related data has not been associated because the transmembrane pressure related data after n hours does not yet exist may be held in the associating unit 62 until the transmembrane pressure related data can be acquired.

[0127] The regression model generating unit 63 generates a regression model 32 in which the input data is used as an explanatory variable and the transmembrane pressure related data after n hours is used as a response variable, and stores the model in the storage device 3B.

[0128] <Main components of the estimation device 2B> Hereinafter, only the differences will be described for components having the same names as those included in the estimation device 2. The estimation device 2B includes a control unit 20B. The control unit 20B includes an input data acquisition unit 21, an access unit 22B, and a cost calculation unit .

[0129] The access unit 22B accesses the regression model 32 stored in the storage device 3B. The access unit 22B includes an estimation unit 23B. The estimation unit 23B first performs the above-mentioned process 1. Specifically, the estimation unit 23B inputs the input data acquired by the input data acquisition unit 21 to the regression model 32 accessed by the access unit 22B, thereby acquiring data related to transmembrane pressure after n hours from the regression model 32. Next, the estimation unit 23B inputs updated data obtained by updating the input data using the acquired data related to transmembrane pressure after n hours to the regression model 32, thereby acquiring data related to transmembrane pressure after 2n hours from the regression model 32. This process is repeated N times to acquire data related to transmembrane pressure up to N×n hours.

[0130] Then, the estimation unit 23B performs the above-mentioned process 2. Specifically, the estimation unit 23B performs estimation using the regression model by the above-mentioned process 1 for the input data and the simulation data, and acquires data related to the transmembrane pressure up to N×n hours after each. As a result, the estimation unit 23B can obtain M estimation results regarding the change over time of the transmembrane pressure up to N×n hours after.

[0131] Then, the estimation unit 23B estimates (A) the input data, (B) the time course of the transmembrane pressure change up to N×n hours after the input data, and (C) the simulation data S2 to S M , (D) Simulation data S2 to S M The time course of the transmembrane pressure change up to N×n hours estimated based on each of the above is output to the cost calculation unit 24. Note that there are M−1 of each of the above (C) and (D).

[0132] The cost calculation unit 24 calculates the expected cost for membrane filtration operation based on the change over time estimated by the estimation unit 23B. Specifically, the energy cost is calculated from the input data acquired from the estimation unit 23B and the data related to the amount of diffused air in the simulation data. The cost calculation unit 24 also identifies the chemical washing time from the change over time in the transmembrane pressure acquired from the estimation unit 23B. The chemical washing time may be, for example, the time when a sudden rise in the transmembrane pressure (TMP jump) occurs. Then, the cost calculation unit 24 calculates the chemical washing cost based on the number of chemical washing times identified.

[0133] Furthermore, the cost calculation unit 24 identifies an appropriate combination based on the calculated combination of energy cost and chemical washing cost. For example, the cost calculation unit 24 refers to predetermined energy cost conditions and chemical washing cost conditions, and identifies a combination that is closest to these conditions. The cost calculation unit 24 outputs the diffused air amount related data corresponding to the identified combination to the diffused air amount control device 8.

[0134] The air diffusion amount control device 8 of this embodiment determines the air diffusion amount level based on the air diffusion amount related data acquired from the cost calculation unit 24, and controls the air diffusion device 95 so as to diffuse air at the determined air diffusion amount level.

[0135] <Flow of estimation process and air diffusion amount control process> FIG. 16 is a flowchart showing an example of the flow of the estimation process executed by the estimation device 2B and the air diffusion amount control process executed by the air diffusion amount control device 8.

[0136] The input data acquiring unit 21 waits until it acquires input data from the input data calculation device 5 (S51). When the input data is acquired (YES in S51), the input data acquiring unit 21 outputs the input data to the access unit 22B.

[0137] When the access unit 22B acquires the input data, it accesses the regression model 32 stored in the storage device 3B. Then, the estimation unit 23B executes a regression analysis process (S52). Specifically, the estimation unit 23B inputs the input data acquired by the input data acquisition unit 21 to the regression model 32 accessed by the access unit 22B, and acquires the transmembrane pressure related data after n hours from the regression model 32.

[0138] Next, the estimation unit 23B judges whether the number of regression analyses, which is the number of times the regression analysis process has been performed, has reached N times (S53). If it has not reached N times (NO in S53), the estimation unit 23B executes a data update process (S54). Specifically, the estimation unit 23B uses the acquired transmembrane pressure related data after n hours to generate update data by updating the transmembrane pressure related data among the input data. Then, the estimation unit 23B executes the process of S52 again using the update data. Note that the target of the process of S52 thereafter is the update data generated in the most recent process of S54. That is, the estimation unit 23B inputs the generated update data to the regression model 32. The estimation unit 23B repeats the execution of the process of S54 and the subsequent process of S52 until it is determined in the process of S53 that the number of regression analyses has reached N times. When the number of regression analyses reaches N, the estimation unit 23B acquires, for the input data, data relating to the transmembrane pressure up to N×n hours after the current value of the amount of diffused air is maintained.

[0139] When the number of regression analyses reaches N (YES in S53), the estimation unit 23B judges whether the number of times the simulation data has been generated reaches M-1 times (S55). Note that when S55 is reached for the first time, the number of times the simulation data has been generated is 0, so the estimation unit 23B judges that the number of times the simulation data has been generated has not reached M-1 times.

[0140] If the number of times the simulation data has been generated has not reached M-1 times (NO in S55), the estimation unit 23B executes a simulation data generation process (S56). Specifically, the estimation unit 23B changes the diffused air amount related data among the input data to generate simulation data. Then, the estimation unit 23B executes the processes from S52 to S54 on the generated simulation data. As a result, the estimation unit 23B acquires transmembrane pressure related data for the generated simulation data up to N×n hours after the current value of the diffused air amount is maintained. Furthermore, the estimation unit 23B repeats the execution of the process of S56 and the subsequent processes from S52 to S54 until it is determined in the process of S55 that the number of times the simulation data has been generated has reached M-1 times. When the number of times the generation has reached M-1 times, the estimation unit 23B acquires M pieces of transmembrane pressure related data for the N×n hours after the current value of the diffused air amount is maintained.

[0141] When the number of times the simulation data has been generated reaches M-1 times (YES in S55), the estimation unit 23B generates a time-dependent change in the transmembrane pressure from each of the transmembrane pressure related data up to N×n hours. Then, the estimation unit 23B generates a time-dependent change in the transmembrane pressure up to N×n hours, which is estimated based on (A) the input data, (B) the time-dependent change in the transmembrane pressure up to N×n hours, which is estimated based on the input data, and (C) the simulation data S2 to S M , (D) Simulation data S2 to S M The time-dependent change in transmembrane pressure up to N×n hours estimated based on each of the above is output to the cost calculation unit 24.

[0142] Next, the cost calculation unit 24 executes an expected cost specification process (S57). Specifically, the cost calculation unit 24 calculates the energy cost from the input data acquired from the estimation unit 23B and the data related to the amount of diffused air in the simulation data. The cost calculation unit 24 also specifies the chemical washing time from the change over time in the transmembrane pressure acquired from the estimation unit 23B. The cost calculation unit 24 then calculates the chemical washing cost based on the number of specified chemical washing times. The cost calculation unit 24 also specifies an appropriate combination based on the calculated combination of energy cost and chemical washing cost. For example, the cost calculation unit 24 refers to predetermined energy cost conditions and chemical washing cost conditions, and specifies a combination that is closest to these conditions.

[0143] Next, the cost calculation unit 24 executes an air diffusion amount related data output process (S58). Specifically, the cost calculation unit 24 identifies input data or simulation data corresponding to the optimal combination identified in the expected cost identification process. Then, the cost calculation unit 24 outputs the air diffusion amount related data of the identified input data or simulation data to the air diffusion amount control device 8.

[0144] Next, the air diffusion amount control device 8 executes the air diffusion amount control process (S59). Specifically, the air diffusion amount control device 8 determines the air diffusion amount level based on the air diffusion amount related data acquired from the cost calculation unit 24, and controls the air diffusion device 95 to diffuse air at the determined air diffusion amount level.

[0145] <Effects> Operating the membrane filtration while maintaining a state in which the possibility of a sudden rise in transmembrane pressure is reduced may not lead to cost reduction or energy saving. In response to this problem, the estimation device 2B according to this embodiment can identify the operating conditions of the aeration device 95 when the energy cost for aeration and the cost for chemical washing are appropriate. For example, by performing chemical washing at an appropriate timing, the energy cost for aeration can be reduced, and as a result, the operating cost can be reduced. In other words, the estimation device 2B can optimize the total cost for membrane filtration operation.

[0146] <Modification> In the above, the input data is changed to generate simulation data S2 to S M In the above description, the data to be changed is data related to the amount of diffused air, but the data is not limited to this and may be any controllable input data. For example, if the membrane filtration flow rate in the membrane filtration operation is controllable, the average membrane filtration flow rate and the integrated membrane filtration flow rate may be changed.

[0147] The cost calculation unit 24 may select the operating cost based on an operation by an operator of the membrane filtration operation. In other words, the operator may be able to specify an appropriate operating cost index.

[0148] In the flowchart shown in Fig. 16, the nested structure of the repetitive processes is not limited to this example. That is, in Fig. 16, the regression analysis process is executed in an inner loop, and the simulation data generation process is executed in an outer loop, but instead, the simulation data generation process may be executed in an inner loop, and the regression analysis process may be executed in an outer loop.

[0149] [Modifications common to the embodiment and the reference embodiment] The input data may include at least one of the maximum transmembrane pressure, minimum transmembrane pressure, standard deviation of transmembrane pressure, average transmembrane pressure, transmembrane pressure fluctuation rate, transmembrane pressure fluctuation amount, and transmembrane pressure fluctuation rate, which are calculated from the transmembrane pressure. The input data may also include at least one of the average diffused air amount and the integrated diffused air amount, which are calculated from the diffused air amount.

[0150] The input data may include data other than the data shown in Fig. 4. For example, the input data may include data on the water to be treated 92. Such data may include, but are not limited to, average water temperature, viscosity, organic matter concentration (e.g., wastewater organic matter concentration, total organic carbon concentration, UV260), pH, sludge concentration (MLSS), suspended solids (SS), dissolved oxygen concentration (DO), oxidation-reduction potential (ORP), ammonium ion concentration, and nitrate ion concentration. The input data may also include a membrane filtration resistance value obtained by dividing the membrane filtration pressure by the membrane filtration flow rate, and a fluctuation value thereof.

[0151] The estimation devices 2, 2A, and 2B may also have the function of the air diffusion amount control device 8. In this case, the air diffusion amount control device 8 is not necessary.

[0152] The state estimation system 100 may have a function of notifying the operation manager of the estimation results of the estimation devices 2, 2A, and 2B, thereby supporting stable operation of the membrane filtration operation.

[0153] The learning models 31, 31A and the regression model 32 may be stored in a memory unit (not shown) of the estimation devices 2, 2A, 2B.

[0154] The learning model generation devices 1, 1A and the regression model generation device 6 may obtain input data from the estimation devices 2, 2A, 2B in the process of updating the learning model or the regression model.

[0155] The learning model generating device 1 and the estimation device 2, the learning model generating device 1A and the estimation device 2A, and the regression model generating device 6 and the estimation device 2B may each be configured as separate entities or separate systems and communicatively connected, or may be configured as an integrated entity or a single system.

[0156] [Software implementation example] The control blocks (control units 10, 10A, 20, 20A, 20B, 60) of the learning model generating devices 1, 1A, the estimation devices 2, 2A, 2B and the regression model generating device 6 may be realized by logic circuits (hardware) formed in an integrated circuit (IC chip) or the like, or by software.

[0157] In the latter case, the learning model generating device 1, 1A, the estimation device 2, 2A, 2B, and the regression model generating device 6 are provided with a computer that executes instructions of a program, which is software that realizes each function. This computer is provided with, for example, one or more processors, and a computer-readable recording medium that stores the above program. In the computer, the processor reads the above program from the recording medium and executes it, thereby achieving the object of the present invention. For example, a CPU (Central Processing Unit) can be used as the processor. For the recording medium, a "non-transient tangible medium", for example, a ROM (Read Only Memory) or the like, as well as a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like can be used. In addition, the computer may further include a RAM (Random Access Memory) that expands the above program. In addition, the above program may be supplied to the computer via any transmission medium (such as a communication network or a broadcast wave) that can transmit the program. Note that one aspect of the present invention may also be realized in the form of a data signal embedded in a carrier wave, in which the above program is embodied by electronic transmission.

[0158] <Notes> [Means for solving the problem] In order to solve the above problems, a learning model generation device according to one embodiment of the present invention comprises a separation membrane immersed in the water to be treated, and an aeration device that diffuses aeration over the membrane surface of the separation membrane, and is equipped with an input data acquisition unit that acquires input data derived from operating data including membrane filtration pressure and aeration amount measured during membrane filtration operation performed in a membrane separation device that obtains treated water that has permeated the separation membrane while diffusing the membrane surface with the aeration device, and a learning unit that generates a learning model for predicting the state of the separation membrane by machine learning using the acquired input data as input.

[0159] According to the above configuration, a learning model can be generated that predicts the state of the separation membrane based on operation data including the membrane filtration pressure and the amount of diffused air. If the state of the separation membrane can be predicted, the amount of diffused air can be controlled based on the prediction, so that stable membrane filtration operation can be performed while maintaining the separation membrane in a normal state (for example, a state in which the possibility of a sudden increase in transmembrane pressure is reduced).

[0160] In addition, in a learning model generating device according to one aspect of the present invention, the membrane filtration operation is an intermittent operation, and the input data derived from the membrane filtration pressure includes at least one of the maximum membrane filtration pressure, the minimum membrane filtration pressure, the standard deviation value of the membrane filtration pressure, the average membrane filtration pressure, and the transmembrane pressure in a unit period consisting of an operation period and a pause period following the operation period, as well as the rate of change of the transmembrane pressure, the amount of change of the transmembrane pressure, and the rate of change of the transmembrane pressure in a specified period prior to the unit period, and the input data derived from the air diffusion amount may include at least one of the average value of the air diffusion amount in the unit period and the integrated value of the air diffusion amount in the specified period.

[0161] According to the above configuration, a learning model is generated using at least one of input data having various feature quantities derived from the membrane filtration pressure and the aeration amount, so that the estimation accuracy of the learning model can be improved.

[0162] In addition, in the learning model generating device according to one aspect of the present invention, the operating data may further include a membrane filtration flow rate measured during the membrane filtration operation.

[0163] According to the above configuration, the learning model is further generated using input data derived from the membrane filtration flow rate, thereby improving the estimation accuracy of the learning model.

[0164] In addition, in a learning model generating device according to one embodiment of the present invention, the membrane filtration operation is intermittent operation, and the input data derived from the membrane filtration flow rate may include at least one of an average value of the membrane filtration flow rate in a unit period consisting of an operation period and a pause period following the operation period, and an integrated value of the membrane filtration flow rate in a predetermined period prior to the unit period.

[0165] According to the above configuration, the learning model is further generated using either the average value of the membrane filtration flow rate in a unit period derived from the membrane filtration flow rate or the integrated value of the membrane filtration flow rate in a specified period, thereby improving the estimation accuracy of the learning model.

[0166] In addition, a learning model generation device according to one embodiment of the present invention may further include a teacher data generation unit that generates teacher data in which the input data is associated with a label indicating the state of the separation membrane for the input data, and the learning unit may generate the learning model by supervised learning using the generated teacher data.

[0167] According to the above configuration, a learning model is generated by supervised learning, making it possible to make highly accurate predictions.

[0168] In addition, in a learning model generation device according to one aspect of the present invention, the labels may include a normal label that is associated with the input data when the state of the separation membrane is normal, and an abnormal label that is associated with the input data when the state of the separation membrane is abnormal.

[0169] According to the above-mentioned configuration, in estimating the state of the separation membrane, it is possible to accurately estimate whether the separation membrane will be normal or abnormal. Therefore, based on the estimation result, the amount of diffused air can be appropriately controlled (to an amount close to the minimum required) so that the separation membrane will be normal.

[0170] In addition, in the learning model generating device according to one aspect of the present invention, the labels may further include an intermediate label that is associated with the input data when the state of the separation membrane is an intermediate state between normal and abnormal.

[0171] According to the above-described configuration, it is possible to predict not only whether the separation membrane will be normal or abnormal, but also whether it will be in an intermediate state between normal and abnormal, thereby enabling the state of the separation membrane to be predicted with even greater accuracy.

[0172] In addition, in the learning model generation device according to an aspect of the present invention, the learning unit may generate the learning model including clusters as a learning result by unsupervised learning.

[0173] According to the above configuration, even in a situation where sufficient training data cannot be prepared, a learning model can be easily generated.

[0174] In addition, in a learning model generation device according to one embodiment of the present invention, the learning unit may acquire data input into the learning model when the state of the separation membrane is predicted using the learning model, and update the learning model by machine learning using the acquired data as input.

[0175] According to the above configuration, the learning model is updated by machine learning using the input data input in the estimation process, so that the estimation accuracy of the learning model can be improved every time estimation is performed.

[0176] In order to solve the above problems, an estimation device according to one embodiment of the present invention includes a separation membrane immersed in the water to be treated, and an aeration device that diffuses aeration over the membrane surface of the separation membrane, and is equipped with an access unit that accesses a learning model for estimating the state of the separation membrane, which is generated by machine learning based on operating data including membrane filtration pressure and diffused aeration amount measured during membrane filtration operation performed in a membrane separation device that obtains treated water that has permeated the separation membrane while diffusing the membrane surface with the aeration device, an input data acquisition unit that acquires input data derived from the operating data measured during the membrane filtration operation, and an estimation unit that uses the accessed learning model to estimate the state of the separation membrane from the acquired input data.

[0177] According to the above-mentioned configuration, the state of the separation membrane can be estimated based on the operation data including the membrane filtration pressure and the amount of diffused air. If the state of the separation membrane can be estimated, the amount of diffused air can be controlled based on the estimated state, so that stable membrane filtration operation can be performed while maintaining the separation membrane in a normal state (e.g., a state in which the possibility of a sudden rise in transmembrane pressure difference is reduced).

[0178] In addition, in the estimation device according to one aspect of the present invention, the membrane filtration operation is intermittent operation, the input data acquisition unit acquires the input data derived from the operating data for each unit period consisting of an operating period and a pause period following the operating period, and the estimation unit may estimate the state from the acquired input data for each unit period.

[0179] According to the above-mentioned configuration, the state of the separation membrane is estimated for each unit period during membrane filtration operation, so that a sudden change in the state of the separation membrane can be quickly confirmed.

[0180] In addition, in the estimation device according to one aspect of the present invention, the membrane filtration operation is intermittent operation, and the input data derived from the membrane filtration pressure includes at least one of the maximum membrane filtration pressure, the minimum membrane filtration pressure, the standard deviation value of the membrane filtration pressure, the average membrane filtration pressure, and the transmembrane pressure in a unit period consisting of an operation period and a pause period following the operation period, as well as the rate of fluctuation of the transmembrane pressure, the amount of fluctuation of the transmembrane pressure, and the rate of fluctuation of the transmembrane pressure in a specified period prior to the unit period, and the input data derived from the air diffusion amount may include at least one of the average value of the air diffusion amount in the unit period and an integrated value of the air diffusion amount in the specified period.

[0181] According to the above configuration, estimation is performed using a learning model generated using at least one of input data having various feature quantities derived from the membrane filtration pressure and the amount of diffused air, thereby improving the accuracy of estimation.

[0182] In addition, in the estimation device according to one aspect of the present invention, the operational data may further include a membranous filtration flow rate measured during the membranous filtration operation.

[0183] According to the above configuration, estimation is performed using a learning model generated using input data derived from the membrane filtration flow rate, thereby improving the estimation accuracy.

[0184] In addition, in the estimation device of one embodiment of the present invention, the membrane filtration operation is intermittent operation, and the input data derived from the membrane filtration flow rate may include at least one of an average value of the membrane filtration flow rate in a unit period consisting of an operation period and a pause period following the operation period, and an integrated value of the membrane filtration flow rate in a predetermined period prior to the unit period.

[0185] According to the above configuration, further, estimation is performed using a learning model generated using either the average membrane filtration flow rate over a unit period derived from the membrane filtration flow rate or the integrated value of the membrane filtration flow rate over a specified period, thereby improving the estimation accuracy.

[0186] In addition, in the estimation device according to one aspect of the present invention, the learning model may be generated by supervised learning using training data including a label indicating a state of the separation membrane.

[0187] According to the above configuration, a learning model generated by supervised learning is used, so that highly accurate estimation is possible.

[0188] In addition, in the estimation device according to one aspect of the present invention, the labels may include a label indicating that the state of the separation membrane is normal, and a label indicating that the state of the separation membrane is abnormal.

[0189] According to the above-mentioned configuration, in estimating the state of the separation membrane, it is possible to accurately estimate whether the separation membrane will be normal or abnormal. Therefore, based on the estimation result, the amount of diffused air can be appropriately controlled (to an amount close to the minimum required) so that the separation membrane will be normal.

[0190] In addition, in the estimation device according to one aspect of the present invention, the label may further include a label indicating that the state of the separation membrane is an intermediate state between normal and abnormal.

[0191] According to the above-described configuration, it is possible to predict not only whether the separation membrane will be normal or abnormal, but also whether it will be in an intermediate state between normal and abnormal, thereby enabling the state of the separation membrane to be predicted with even greater accuracy.

[0192] In addition, in the estimation device according to one aspect of the present invention, the learning model may be a learning model that is generated by unsupervised learning and includes clusters as a learning result.

[0193] According to the above-described configuration, even in a situation where sufficient training data cannot be prepared, the state of the separation membrane can be easily estimated.

[0194] In addition, in an estimation device according to one aspect of the present invention, an output value of the learning model may be an outlier or not, and the estimation unit may infer that the state of the separation membrane will be abnormal when the output value is an outlier, and may infer that the state of the separation membrane will be normal when the output value is not an outlier.

[0195] According to the above-described configuration, the state of the separation membrane can be easily estimated using a learning model capable of detecting outliers.

[0196] In addition, an air diffusion amount control device according to one embodiment of the present invention may determine the level of the air diffusion amount in accordance with the state estimated by the estimation device, and control the air diffusion device to diffuse air at the determined level.

[0197] According to the above-mentioned configuration, the level of the amount of diffused air is automatically determined according to the estimated state of the separation membrane, so that the operator performing the membrane filtration operation does not need to adjust the level of the amount of diffused air, and as a result, the workload of the operator can be reduced.

[0198] In addition, the air diffusion amount control device according to one aspect of the present invention may determine the current level of the air diffusion amount based on the state estimated by the estimation device and the level most recently determined by the device itself.

[0199] According to the above configuration, the level of the air diffusion amount is determined taking into consideration the most recent air diffusion amount level, so that a more appropriate air diffusion amount level can be determined compared to a configuration that does not take into consideration the most recent air diffusion amount level.

[0200] The learning model generation device, the estimation device and the air diffusion amount control device according to each aspect of the present invention may be realized by a computer. In this case, the control programs for the learning model generation device, the estimation device and the air diffusion amount control device that realize the learning model generation device, the estimation device and the air diffusion amount control device by operating a computer as each part (software element) of the learning model generation device, the estimation device and the air diffusion amount control device, and the computer-readable recording medium on which the same is recorded, also fall within the scope of the present invention.

[0201] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of symbols]

[0202] 1. 1A Learning model generation device 2, 2A guessing device 8. Air diffusion control device 11 Input data acquisition section 12 Teacher data generation unit 13, 13A Learning Section 21 Input data acquisition section 22, 22A Access section 23, 23A Estimation part 31, 31A Learning Model 90 Membrane separation equipment 92 Untreated water 93 Separation membrane 95 Air diffuser

Claims

1. an input data acquisition unit that acquires input data derived from operational data including a membrane filtration pressure and an aeration amount measured during a membrane filtration operation performed in a membrane separation device that is provided with a separation membrane immersed in the water to be treated and an aeration device that diffuses aeration over the membrane surface of the separation membrane and obtains treated water that has permeated the separation membrane while diffusing aeration over the membrane surface using the aeration device; a teacher data generating unit that generates teacher data in which the input data is associated with a label that indicates a state of the separation membrane for the input data, the teacher data generating unit associates a label that indicates a state of the separation membrane after a predetermined time based on a degree of satisfaction of a specific condition at a time when the operating data is acquired and the input data derived from operating data acquired a predetermined time before the time when the operating data is acquired, The specific conditions include a first condition that a fluctuation rate, which is a fluctuation amount per unit time of the transmembrane pressure, is less than a first predetermined value, a second condition that the transmembrane pressure is less than a second predetermined value, a third condition that the fluctuation rate is equal to or greater than the first predetermined value and less than a third predetermined value, a fourth condition that the fluctuation rate is equal to or greater than the third predetermined value, and a fifth condition that the transmembrane pressure is equal to or greater than the second predetermined value, the teacher data generation unit associates a normal label with the input data that satisfies both the first condition and the second condition, associates an intermediate label with the input data that satisfies both the third condition and the second condition, and associates an abnormal label with the input data that satisfies the fourth condition or the fifth condition; A learning model generation device comprising: a learning unit that generates a learning model for predicting the future state of the separation membrane through machine learning using the acquired input data as input, the learning model outputting the probability that the future state of the separation membrane will be normal or abnormal through supervised learning using the teacher data generated by the teacher data generation unit.

2. The learning model generating device according to claim 1 , wherein the input data derived from the membrane filtration pressure includes a membrane filtration resistance value obtained by dividing the membrane filtration pressure by a membrane filtration flow rate and a fluctuation value thereof.

3. The membrane filtration operation is an intermittent operation, The input data derived from the transmembrane pressure includes at least one of the maximum transmembrane pressure, the minimum transmembrane pressure, the standard deviation of the transmembrane pressure, the average transmembrane pressure, and the transmembrane pressure in a unit period consisting of an operation period and a pause period following the operation period, and the rate of change of the transmembrane pressure, the amount of change of the transmembrane pressure, and the rate of change of the transmembrane pressure in a predetermined period prior to the unit period; The learning model generating device according to claim 1 , wherein the input data derived from the amount of air diffusion includes at least one of an average value of the amount of air diffusion in the unit period and an integrated value of the amount of air diffusion in the specified period.

4. The learning model generating device according to claim 1 , wherein the operational data further includes a membrane filtration flow rate measured during the membrane filtration operation.

5. The membrane filtration operation is an intermittent operation, The learning model generating device of claim 4, wherein the input data derived from the membrane filtration flow rate includes at least one of an average value of the membrane filtration flow rate in a unit period consisting of an operating period and an idle period following the operating period, and an integrated value of the membrane filtration flow rate in a predetermined period prior to the unit period.

6. The learning unit acquires data input into the learning model when the state of the separation membrane is predicted using the learning model, and updates the learning model through machine learning using the acquired data as input. A learning model generation device according to any one of claims 1 to 5.

7. A learning model generating device according to any one of claims 1 to 6, A storage device that stores the learning model generated by the learning model generation device; an estimation device that estimates a state of the separation membrane from the acquired input data using the learning model; and an air diffusion amount control device that controls the air diffusion device in accordance with the state estimated by the estimation device.

Citation Information

Patent Citations

  • Control method and control system for eliminating membrane bioreactor pollution in real time, and storage medium

    CN110668562A

  • Operation control device of water making plant

    JP1996039065A

  • Method for operation of membrane filter

    JP2000300968A

  • Determination method of operation condition of membrane filtration apparatus, operation method of membrane filtration apparatus using it and membrane filtration apparatus

    JP2008142675A

  • Sensor interface apparatus, measurement information communication system, measurement information communication method, and measurement information communication program

    JP2018077779A