Decision model learning device, decision model learning method, decision model learning program, decision device, decision method, decision program, and water purification facility

Abstract

To quickly and accurately determine the forming state of a coagulated flock.SOLUTION: A model learning device (14) comprises a learning unit (41) that performs mechanical learning using an injection rate of a coagulant, water quality data of a liquid on the upstream side of the injection point of the coagulant, and standard deviation of an image feature amount acquired from each of a plurality of flock images photographing a coagulated flock formed by stirring the liquid into which the coagulant has been injected as explanatory variables and the forming state of the coagulated flock at a target point as an objective variable. The target point is positioned in the first half part of a flock forming basin, and the flock images have been captured for each unit time within a prescribed period including the elapsed time of the movement of the liquid from a measurement point of the water quality data to the target point.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a technique for determining the formation state of flocs formed by stirring a liquid into which an agent for flocculating suspended solids has been injected. [Background technology]

[0002] The water treatment system described in Patent Document 1 first measures the water quality of the water to be treated at intervals of a predetermined period or less, and captures images of the treated water at intervals of a predetermined period or less. Next, the water treatment system acquires feature values ​​from processed images obtained by performing predetermined image processing on the captured images. The water treatment system then uses the acquired feature values, the measured water quality values, and learning data to determine whether the treatment of the water to be treated is appropriate and notifies the system of the determination result. The learning data is, for example, a relational expression previously determined based on the feature values, the water quality values, and the coagulation state of the treated water. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2021-047039 Summary of the Invention [Problem to be solved by the invention]

[0004] However, there is still room for improvement in the above-mentioned method for judging the quality of water treatment systems.

[0005] An object of one aspect of the present invention is to quickly and accurately determine the formation state of aggregated flocs formed by stirring a liquid into which an agent has been injected. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, a judgment model learning device according to one aspect of the present invention includes a learning unit that constructs a judgment model by performing machine learning using, as explanatory variables, an injection rate of an agent that is injected into a liquid containing suspended solids to coagulate the suspended solids, water quality data including the turbidity of the liquid upstream of the injection point of the agent, and a degree of variation in image features obtained from multiple floc images taken at a target point of flocs formed by stirring the liquid into which the agent has been injected, and a target variable is the formation state of the flocs at the target point, wherein the target point is located in the front half of a flocculation basin, and the multiple floc images are taken at unit time intervals within a predetermined period that includes the time it takes for the liquid to travel from the water quality data measurement point to the target point.

[0007] A judgment model training method according to another aspect of the present invention includes a learning step of constructing a judgment model by performing machine learning using, as explanatory variables, an injection rate of an agent that is injected into a liquid containing suspended solids to coagulate the suspended solids, water quality data including the turbidity of the liquid upstream of the injection point of the agent, and a degree of variation in image features obtained from each of a plurality of floc images taken at a target point of flocs formed by stirring the liquid into which the agent has been injected, and a target variable being the formation state of the flocs at the target point, wherein the target point is located in the front half of a flocculation basin, and the plurality of floc images are taken at unit time intervals within a predetermined period that includes the time when the liquid travels from the measurement point of the water quality data to the target point.

[0008] According to yet another aspect of the present invention, there is provided a determination device which determines the formation state of flocs using a determination model constructed by machine learning using explanatory variables including an injection rate of an agent that is injected into a liquid containing suspended solids to flocculate the suspended solids, water quality data including the turbidity of the liquid upstream of the injection of the agent, and a degree of variation in image features obtained from a plurality of floc images taken at a target point of flocs formed by stirring the liquid into which the agent has been injected, and a target variable, wherein the target point is located in the front half of a flocculation basin, and the plurality of floc images are taken at unit time intervals within a predetermined period including the time when the liquid has traveled from the water quality data measurement point to the target point.

[0009] A water purification facility according to yet another aspect of the present invention includes the determination device having the above-described configuration.

[0010] A determination method according to yet another aspect of the present invention includes a determination step of determining the formation state of flocs using a determination model constructed by machine learning using explanatory variables including an injection rate of an agent that is injected into a liquid containing suspended solids to flocculate the suspended solids, water quality data including the turbidity of the liquid upstream of the injection point of the agent, and a degree of variation in image feature values ​​obtained from a plurality of floc images taken at a target point of flocs formed by stirring the liquid into which the agent has been injected, and the formation state of the flocs at the target point as a response variable, wherein the target point is located in the front half of a flocculation basin, and the plurality of floc images are taken at unit time intervals within a predetermined period that includes the time it takes for the liquid to travel from the measurement point of the water quality data to the target point. [Effects of the Invention]

[0011] According to one aspect of the present invention, it is possible to quickly and accurately determine the formation state of aggregated flocs that are formed by stirring a liquid into which a chemical agent has been injected. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a diagram showing an overview of a water treatment system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram showing an example of a configuration of a main part of a floc determination system in the water treatment system. [Figure 3] 10 is a flowchart showing an example of the flow of a teacher data creation process in the water treatment system. [Figure 4] 10 is a flowchart showing an example of the flow of a model learning process in the water treatment system. [Figure 5] 10 is a flowchart showing an example of the flow of a determination process in the water treatment system. [Figure 6] 10 is a bar graph showing the accuracy rate of determination of the formation state of flocs for each classification model in an example of the water treatment system and a comparative example for the example. [Figure 7] FIG. 10 is a diagram showing, in tabular form, the accuracy rate of determining the state of formation of flocs in another embodiment of the water treatment system. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, one embodiment of the present invention will be described with reference to FIGS.

[0014] (Water treatment system) Figure 1 is a diagram showing an overview of a water treatment system according to this embodiment. The water treatment system 1 of this embodiment is, for example, a water purification facility, and performs a coagulation-sedimentation process in which suspended solids, such as suspended matter, contained in raw water flowing into the water purification facility from, for example, a river, are coagulated using a coagulant and separated from the raw water by gravitational settling. As shown in Figure 1, the water treatment system 1 includes a receiving well 101, a rapid mixing basin 102, a flocculation basin 103, and a settling basin 104 as facilities for performing the coagulation-sedimentation process.

[0015] The receiving well 101 is a tank for suppressing the agitation of the water surface caused by the raw water sent through the water conveyance pipe and for ensuring the hydraulic stability of the raw water sent to the equipment following the receiving well 101. A water quality meter 111 is provided in the receiving well 101. The water quality meter 111 measures the quality of the raw water that has arrived at the receiving well 101. The water quality is at least turbidity, and other examples include color, water temperature, conductivity, pH (hydrogen ion concentration index), alkalinity (acid consumption), ultraviolet absorbance, etc. Water quality data indicating the water quality measured by the water quality meter 111 is input to the floc determination system 2.

[0016] A flow meter 112 is installed in the water conveyance pipe 105 between the receiving well 101 and the rapid mixing basin 102. The flow meter 112 measures the flow rate of the treated water sent from the receiving well 101 to the rapid mixing basin 102. The flow rate data measured by the flow meter 112 is input to a coagulant injection device 113.

[0017] The rapid mixing basin 102 is a mixing basin for injecting a coagulant into the raw water delivered from the receiving well 101 to rapidly mix it. Here, the coagulant is a chemical agent for flocculating substances contained in the raw water. The coagulant is, for example, polyaluminum chloride (PAC) or aluminum sulfate (aluminum sulfate), and is injected into the rapid mixing basin 102 by a coagulant injector 113. Specifically, the coagulant injector 113 determines the amount of coagulant to be injected based on the flow rate data from the flow meter 112 and a set injection rate, and injects the amount into the rapid mixing basin 102. The injection rate is input to the floc determination system 2.

[0018] The rapid mixing basin 102 is also provided with a rapid mixer 114 for performing the rapid mixing. The rapid mixer 114 is, for example, a flash mixer. The rapid mixer 114 may operate at a constant mixing speed, or may be capable of adjusting the mixing speed by controlling a motor. In the rapid mixing basin 102, fine flocs are formed by the injection of a coagulant and the mixing by the rapid mixer 114. The treated water containing such fine flocs is sent to the subsequent flocculation basin 103, where further agglomeration of the flocs is promoted in the facilities subsequent to the flocculation basin 103.

[0019] The flocculation basin 103 is an agitation tank for forming larger flocs in the treated water. The flocculation basin 103 includes multiple formation basins, each equipped with a slow speed agitator 115. Agitation of the treated water by the slow speed agitators 115 promotes further agglomeration of the flocs. The slow speed agitators 115 are, for example, flocculators. The treated water from the flocculation basin 103 is agitated for a predetermined period of time before being sent to the subsequent settling basin 104.

[0020] The flocculation basin 103 is also provided with a camera 116. Specifically, the camera 116 is located in the front half of the flocculation basin 103, more specifically, in the most upstream flocculation basin of the multiple flocculation basins in the flocculation basin 103.

[0021] The camera 116 photographs the flocs formed in the floc formation basin 103. The photographed image of the flocs (hereinafter referred to as the “floc image”) is input to the image processing device 117.

[0022] The image processing device 117 performs image processing such as binarization on the floc images captured per unit time to extract image feature quantities of the flocs. The image feature quantities are at least one of the number of the flocs, the area of ​​each of the flocs, the total area, and the perimeter of each of the flocs. An operator can check the formation state of the flocs by referring to the captured images and the image feature quantities. The image feature quantity data is input to the floc determination system 2.

[0023] Examples of cases where the formation state of the flocs is good include when the amount, size, etc. of the flocs are appropriate with respect to the water quality data of the raw water, when the shape of the flocs is close to spherical, etc. On the other hand, examples of cases where the formation state of the flocs is poor include when the amount, size, etc. of the flocs are excessive or insufficient with respect to the water quality data of the raw water, when the shape of the flocs is irregular, etc.

[0024] The quality of the formation state of the flocs may be determined using the diameter, area, volume, perimeter, etc. of the flocs, or values ​​calculated by combining these. These values ​​may be obtained by direct measurement from the captured image, or may be calculated by the image processing device 117 using an image analysis method. The formation state of the flocs may be expressed in two stages, good or bad, or in multiple stages such as five or ten stages, or may be expressed by some numerical value corresponding to the formation state.

[0025] The settling basin 104 is a basin for settling and separating aggregated flocs contained in the influent from the flocculation basin 103. When the influent remains in the settling basin 104 for a predetermined time (for example, about 1 to 3 hours), the large-sized flocs formed in the flocculation basin 103 settle by gravity. As a result, the flocs are separated from the treated water, and the supernatant liquid is sent to a filtration basin (not shown) as settled water. Note that a facility may be provided downstream of the settling basin 104 for additional treatment, such as ozone treatment or biological activated carbon treatment, on the treated water sent to the filtration basin. The flocs that have settled in the settling basin 104 are extracted as sludge and sent to a sludge treatment facility (not shown).

[0026] The filter is a pond equipped with a filtration system that filters the settled water flowing in from the settling basin 104. In the filter, minute solids remaining in the settled water are separated by filtration. The filtered settled water is disinfected with a chlorine agent and then supplied to consumers as tap water.

[0027] (Floc detection system) 1, the water treatment system 1 includes a floc determination system 2. The floc determination system 2 determines the formation state of aggregated flocs at a point (target point) where a camera 116 is installed, using the water quality data of the raw water, the injection rate of the coagulant, and the image feature amount of the floc image.

[0028] Here, the injection rate is the ratio of the coagulant injected into the raw water to the raw water. The injection rate is determined using the results of a jar test, past knowledge, a predetermined calculation formula, etc. The floc determination system 2 may be configured by one information processing device equipped with a hardware and software configuration group, or may be configured by multiple information processing devices.

[0029] 2 is a block diagram showing an example of the configuration of the main parts of the floc determination system 2. The floc determination system 2 includes an input device 11, a storage device 12, a teacher data creation device 13, a model learning device 14 (determination model learning device), a determination device 15, and an output device 16.

[0030] The input device 11 accepts input from a user and outputs an input signal based on the input to various devices. The input device 11 also receives information output from external devices (for example, a water quality meter 111, a flocculant injector 113, an image processing device 117, etc.) as input data and outputs the received input data to the storage device 12.

[0031] The output device 16 outputs the information generated by the determination device 15. There are no particular limitations on the output method used by the output device 16. For example, the output device 16 may be a display device that displays the information as an image, an audio output device that outputs the information as audio, or a signal transmission device that transmits an output signal to another input device.

[0032] The storage device 12 stores programs and data used in the floc determination system 2. The storage device 12 includes an input data storage unit 21, a teacher data storage unit 22, and a determination model storage unit 23. The input data storage unit 21 stores input data from the input device 11. Specifically, the input data storage unit 21 stores the water quality data of the raw water, the injection rate of the coagulant, and the image feature quantities of the floc image output from an external device via the input device 11, as well as the formation state of the aggregated flocs output from the user via the input device 11. The data held (stored) in the teacher data storage unit 22 and the determination model storage unit 23 will be described later.

[0033] (Teacher data creation device) The teacher data creation device 13 creates a group of teacher data used when the model learning device 14 performs machine learning. The teacher data creation device 13 may be configured as a single information processing device having a group of hardware and software components, or may be configured as multiple information processing devices. In this embodiment, for simplicity of explanation, the teacher data creation device 13 will be described as being configured as a single information processing device.

[0034] The teacher data creation device 13 includes a control unit 30. The control unit 30 controls each unit of the teacher data creation device 13. The control unit 30 is realized by, for example, a processor and a memory. 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 configures each unit included in the control unit 30.

[0035] The control unit 30 includes an explanatory variable creation unit 31 and a response variable creation unit 32. The explanatory variable creation unit 31 creates explanatory variables using input data stored in the input data storage unit 21. The response variable creation unit 32 creates response variables using the formation state of flocs stored in the input data storage unit 21. A combination of the explanatory variables created by the explanatory variable creation unit 31 and the response variable created by the response variable creation unit 32 is stored in the training data storage unit 22 as training data.

[0036] Specifically, the explanatory variables of the training data include water quality data and injection rates at a certain measurement time point. Furthermore, the explanatory variables include the standard deviation (hereinafter referred to as the "standard deviation at the elapsed time point") of multiple image features calculated from multiple floc images taken at each unit time (e.g., 10 seconds) within a predetermined period (e.g., 1 minute) including the time point (hereinafter referred to as the "elapsed time point") when raw water from the measurement point (the installation point of the water quality meter 111) for the water quality data at the measurement time point has been treated and moved to the target point (the installation point of the camera 116). The objective variable of the training data is the formation state of aggregated flocs at the elapsed time point. In addition to the standard deviation, various degrees of variation of the multiple image features, such as variance, maximum and minimum values, kurtosis, and skewness, can also be used.

[0037] (Model learning device) The model learning device 14 performs machine learning to generate a determination model for determining the formation state of flocs at the above-mentioned time point, using a training data group, which is a collection of training data, stored in the training data storage unit 22. The model learning device 14 may be configured as a single information processing device having a hardware and software configuration group, or may be configured as a plurality of information processing devices. In this embodiment, for the sake of simplicity, the model learning device 14 will be described as being configured as a single information processing device.

[0038] Model learning device 14 includes a control unit 40. Control unit 40 controls all the components of model learning device 14. Note that control unit 40 has the same hardware configuration as control unit 30, and therefore a description thereof will be omitted.

[0039] The control unit 40 includes a learning unit 41. The learning unit 41 performs machine learning using a group of teacher data read from the teacher data storage unit 22, and constructs a determination model as a learning model. For each teacher data in the group of teacher data, the water quality data and injection rate at the time of measurement and the standard deviation at the time point are explanatory variables, and the formation state of flocs at the time point is the objective variable.

[0040] The learning unit 41 generates a classification model obtained by, for example, a random forest (RF), a gradient boosted tree regression (GBR), or a neural network (NN) as the determination model. The determination model storage unit 23 of the storage device 12 stores the determination model.

[0041] As described above, the model learning device 14 of this embodiment includes a learning unit 41 that constructs a judgment model by performing machine learning using the following explanatory variables: the injection rate of the flocculant injected by the flocculant injector 113; the water quality data measured by the water quality meter 111 installed upstream of the flocculant injection point; and the standard deviation of image features calculated by the image processing device 117 from multiple floc images taken by the camera 116 at a target point of flocs formed by stirring the liquid into which the flocculant has been injected by the rapid mixer 114 in the rapid mixing basin 102; and the formation state of the flocs at the target point as a response variable. The target point is located in the front half of the flocculation basin 103, and the multiple floc images are taken at unit time intervals within a predetermined period that includes the time it takes for the liquid to travel from the water quality data measurement point to the target point.

[0042] According to the above configuration, since the target point for photographing flocs is located in the first half of the flocculation basin 103, it is possible to construct a judgment model that can quickly judge the formation state of flocs compared to when the target point is located in the second half of the flocculation basin 103. As a result, even when the formation state is poor, it is possible to respond quickly and improve the formation state. Furthermore, by performing machine learning using explanatory variables that include not only water quality data including the turbidity of the liquid (raw water) before the injection of the flocculant, but also the injection rate of the flocculant and the standard deviation of the image features obtained from each floc image, it is possible to construct a judgment model with high accuracy.

[0043] Furthermore, the multiple floc images are taken at unit time intervals within a predetermined period, including the time elapsed during which the liquid travels from the water quality data measurement point to the target point. Therefore, machine learning is performed using the injection rate, the water quality data, and the standard deviation of the image features for the same liquid as explanatory variables, allowing for the construction of a more accurate judgment model. Therefore, by using the constructed judgment model, the formation state of aggregated flocs can be quickly and accurately judged from the water quality data and injection rate at the time of measurement and the standard deviation of the image features at the time of elapse.

[0044] (judgment device) The determination device 15 determines the most recent formation state of agglomerated flocs at a target location using a determination model stored in the determination model storage unit 23 of the storage device 12. The determination device 15 may be configured by one information processing device including a hardware and software configuration group, or may be configured by multiple information processing devices. In this embodiment, for the sake of simplicity, the determination device 15 will be described as being configured by one information processing device.

[0045] The determination device 15 includes a control unit 50. The control unit 50 controls all the components of the determination device 15. The control unit 50 has the same hardware configuration as the control unit 30, and therefore a description thereof will be omitted.

[0046] The control unit 50 includes a determination unit 51. The determination unit 51 creates the most recent explanatory variables using the input data stored in the input data storage unit 21. The most recent explanatory variables are the standard deviation at the most recent time point, and the water quality data and injection rate at the measurement time point corresponding to the most recent time point. Next, the determination unit 51 determines the formation state of flocs at the target location at the most recent time point from the most recent explanatory variables using a determination model read out from the determination model storage unit 23. Then, the determination unit 51 outputs the determination result to the output device 16.

[0047] As described above, the determination device 15 of this embodiment is equipped with a determination unit 51 that determines the formation state of flocs using a determination model constructed by machine learning by the model learning device 14, using as explanatory variables the injection rate of the flocculant injected by the flocculant injector 113, water quality data measured by the water quality meter 111 installed upstream of the injection point of the flocculant, and the standard deviation of image features calculated by the image processing device 117 from multiple floc images taken by the camera 116 at a target point of flocs formed by stirring the liquid into which the flocculant has been injected in the rapid mixing basin 102 with the rapid mixer 114, and the formation state of the flocs at the target point as a response variable. The target point is located in the front half of the flocculation basin 103, and the multiple floc images are taken at unit time intervals within a predetermined period that includes the time it takes for the liquid to travel from the measurement point of the water quality data to the target point.

[0048] According to the above configuration, the target point for photographing flocs is located in the front half of the flocculation basin 103, so the formation state of flocs can be determined more quickly than when the target point is located in the rear half of the flocculation basin 103. As a result, even if the formation state is poor, a quick response can be taken to improve the formation state.

[0049] Furthermore, by using a judgment model constructed by machine learning that includes as explanatory variables not only water quality data including the turbidity of the liquid (raw water) before the injection of the coagulant, but also the injection rate of the coagulant and the standard deviation of the image features obtained from each floc image, the formation state can be accurately determined.

[0050] Furthermore, the plurality of floc images are taken at unit time intervals within a predetermined period, including the time when the liquid travels from the measurement point of the water quality data to the target point. Therefore, by using a determination model constructed by machine learning using the injection rate, the water quality data, and the standard deviation of the image features for the same liquid as explanatory variables, the formation state can be determined with even greater accuracy.

[0051] (Teacher data creation process) 3 is a flowchart showing an example of the flow of the training data creation process in the water treatment system 1 configured as described above. As shown in FIG. 3, first, the explanatory variable creation unit 31 acquires water quality data and an injection rate at a certain measurement time point from the input data storage unit 21 (step S11; hereinafter, the word "step" will be omitted). The explanatory variable creation unit 31 also acquires from the input data storage unit 21 a plurality of image feature amounts within a predetermined period of time that includes a time point that corresponds to the measurement time point, and calculates the standard deviation of the plurality of image feature amounts (S12).

[0052] Meanwhile, the response variable creating unit 32 acquires the formation state of the aggregated flocs at the above-mentioned time point from the input data storage unit 21 (S13). Note that the order of S11 to S13 is arbitrary.

[0053] Next, the explanatory variable creation unit 31 and the response variable creation unit 32 use the water quality data and injection rate at the measurement time point and the standard deviation at the elapsed time point as explanatory variables, and the formation state of the flocs at the elapsed time point as a response variable, and store the combination of the explanatory variables and the response variable as training data in the training data storage unit 22 (S14). Thereafter, the process returns to S11 and the training data creation process is repeated. As a result, a group of training data is stored in the training data storage unit 22.

[0054] (Model learning process) 4 is a flowchart showing an example of the flow of the model learning process (determination model learning method) in the water treatment system 1 configured as described above. As shown in FIG. 4, first, the learning unit 41 acquires a group of teacher data from the teacher data storage unit 22 (S21). Next, the learning unit 41 uses the acquired group of teacher data to machine-learn a determination model (S22, learning step). Then, the learning unit 41 stores the machine-learned determination model in the determination model storage unit 23 (S23). Thereafter, the model learning process ends.

[0055] (Determination process) 5 is a flowchart showing an example of the flow of the determination process (determination method) in the water treatment system 1 configured as described above. As shown in FIG. 5, first, the determination unit 51 acquires, from the input data storage unit 21, a plurality of image feature amounts within a predetermined period including the most recent elapsed time point, and calculates the standard deviation of the plurality of image feature amounts (S31). The determination unit 51 also acquires water quality data and an injection rate at a measurement time point corresponding to the most recent elapsed time point (S32). Note that the order of S31 and S32 is arbitrary.

[0056] Next, the determination unit 51 acquires a determination model from the determination model storage unit 23, and uses the determination model to determine the formation state of flocs at the target point at the most recent time point based on the water quality data, the injection rate, and the standard deviation (S33, determination step).The determination unit 51 then outputs the determination result to the output device 16 (S34).Then, the determination process ends.

[0057] Example 1 An example of the floc determination system 2 having the above configuration and a comparative example will be described with reference to FIG.

[0058] In this example, the standard deviation of the image feature values ​​over a predetermined period is the standard deviation of the area of ​​the flocs calculated from six floc images taken over one minute. Various classification models were machine-learned as determination models using explanatory variables and objective variables stored in the training data storage unit 22 shown in Figure 2. The machine-learned determination models were then used to determine the formation state of flocs, and the accuracy of the predictions compared with the actual formation state was examined.

[0059] The classification models used were LDA (Linear Discriminant Analysis), LSVM (Linear Support Vector Machine), NLSVM (Non-Linear Support Vector Machine), QDA (Quadratic Discriminant Analysis), kNNC (k-Nearest Neighbor Classification), NB (Naive Bayes classifier), DT (Decision Tree), RF (Random Forests), and GPC (Gaussian Process Classification).

[0060] On the other hand, the comparative example differs from the above-mentioned example in that the above-mentioned standard deviation is omitted from the explanatory variables in the training data, but the rest is the same.

[0061] Fig. 6 is a bar graph showing the accuracy rate (precision) of determining the formation state of flocs for each classification model in the Examples and Comparative Examples. In Fig. 6, the graphs for the Examples are shown in white, and the graphs for the Comparative Examples are shown in hatched.

[0062] As shown in Figure 6, for most of the classification models, the accuracy rate of the above judgment was higher in the Example than in the Comparative Example. Furthermore, for the remaining classification models, the accuracy rate of the above judgment was the same in the Example and the Comparative Example. In other words, the accuracy rate of the above judgment did not decrease in the Example compared to the Comparative Example.

[0063] Example 2 Another embodiment of the floc determination system 2 having the above configuration will be described with reference to FIG.

[0064] The explanatory variables stored in the training data storage unit 22 shown in FIG. 2 include the "standard deviation of image features over a predetermined period of time including the elapsed time point." In this example, the number of the aggregated flocs, the area of ​​each of the aggregated flocs, the sum of the areas, and the perimeter of each of the aggregated flocs were used as the image features. In addition, instead of the standard deviation, the 10th, 25th, 50th, 75th, and 90th percentile values, maximum and minimum values, and the degree of variation including the average value and standard deviation were added to the explanatory variables. The degree of variation of image features over a predetermined period of time is the degree of variation of image features calculated from six floc images taken over one minute.

[0065] Various classification models were machine-learned as judgment models using the explanatory variables of this example and the objective variables stored in the training data storage unit 22 shown in Figure 2. The above-mentioned LDA, LSVM, NLSVM, kNNC, NB, DT, and RF were used as the classification models. The machine-learned judgment models were then used to judge the formation state of flocs, and the accuracy of the judged results compared with the actual formation state was examined. Furthermore, in this example, the machine-learned judgment models were used to judge the formation state of flocs by excluding some of the standard deviations of the image features, and the accuracy of the judged results compared with the actual formation state was examined.

[0066] Fig. 7 is a diagram showing, in tabular form, the accuracy rate of the determination of the formation state of flocs in this example. In Fig. 7, the item "test0" shows the case where all of the standard deviations of the image features (the number of flocs, the area of ​​each of the flocs, the sum of the areas, and the standard deviation of the perimeter of each of the flocs) were used. Furthermore, the item "test1" shows the case where the standard deviation of the number of flocs was excluded from the standard deviations of the image features.

[0067] Furthermore, item "test2" indicates a case where the standard deviation of each area of ​​the agglomerated floc is excluded from the standard deviations of the image features. Furthermore, item "test3" indicates a case where the standard deviation of the total value of the area of ​​each of the agglomerated floc is excluded from the standard deviations of the image features. Furthermore, item "test4" indicates a case where the standard deviation of each perimeter of the agglomerated floc is excluded from the standard deviations of the image features. Furthermore, item "test5" indicates a case where all of the standard deviations of the image features are excluded.

[0068] As shown in Fig. 7, the accuracy rate of the above judgment when all of the standard deviations of the image features were used (test0) was the same as or higher than the other cases (test1 to test5). In contrast, when the standard deviation of the number of flocs was excluded from the standard deviations of the image features (test1) and when the standard deviation of the total area of ​​each of the flocs was excluded (test3), the accuracy rate of the above judgment using the LSVM and DT classification models decreased compared to test0.

[0069] Furthermore, when the standard deviation of each area of ​​the agglomerated flocs was excluded from the standard deviations of the image features (test 2) and when the standard deviation of each perimeter of the agglomerated flocs was excluded (test 4), the accuracy rates of the above judgments using the LSVM, kNNC, and DT classification models decreased compared to test 0. Furthermore, when all standard deviations of the image features were excluded (test 5), the accuracy rates of the above judgments using the LSVM, kNNC, and DT classification models decreased compared to test 0.

[0070] Incidentally, ensemble learning, which has recently attracted attention, is a method in which multiple estimation results estimated using multiple machine-learned estimation models are ultimately determined by majority vote. Considering the use of such ensemble learning, it is believed that the more estimation models there are in which the accuracy rate of the estimation result decreases when a certain explanatory variable is excluded, the higher the contribution of the explanatory variable to the accuracy rate. Therefore, with reference to Figure 7, it can be seen that the standard deviation of the image features has a high contribution to the accuracy rate of the determination, and that, among these, the standard deviation of the area of ​​each of the agglomerated flocs and the standard deviation of the perimeter of each of the agglomerated flocs have an even higher contribution to the accuracy rate of the determination.

[0071] (Additional notes) When the output device 16 outputs a determination result that the state of formation of flocs is poor, the user may reset the injection rate based on the determination result and instruct the flocculant injector 113 to do so via the floc determination system 2. In this case, the state of formation of flocs can be improved.

[0072] [Software implementation example] The functions of the floc determination system 2 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (particularly each unit included in the control units 30, 40, 50).

[0073] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0074] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0075] Furthermore, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of the control blocks can also be realized by, for example, a quantum computer.

[0076] 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]

[0077] 1. Water treatment system 2. Floc Judgment System 11 Input Devices 12 Storage device 13 Teacher data creation device 14 Model learning device (decision model learning device) 15 Judgment device 16 Output Devices 21 Input data storage unit 22 Teacher data storage unit 23 Decision model memory unit 30, 40, 50 Control section 31 Explanatory variable creation section 32 Objective variable creation section 41 Learning Department 51 Judgment section 101 Water landing well 102 Rapid mixing basin 103 Flocculation pond 104 Sedimentation pond 105 Water Pipe 111 Water quality meter 112 Flow meter 113 Coagulant injection device 114 Rapid Mixer 115 Slow Stirrer 116 Camera 117 Image Processing Device

Claims

1. a learning unit that constructs a judgment model by performing machine learning using, as explanatory variables, an injection rate of an agent that is injected into a liquid containing suspended solids to coagulate the suspended solids, water quality data including turbidity of the liquid upstream of the injection point of the agent, and a degree of variation in image feature values ​​obtained from each of a plurality of floc images of coagulated flocs formed by stirring the liquid into which the agent has been injected, taken at a target point downstream of the injection point of the agent, and a target variable being the formation state of the coagulated flocs at the target point; The target point is located in the front half of the flocculation basin; the plurality of floc images are taken at unit time intervals within a predetermined period including a time point at which the liquid has traveled from the measurement point of the water quality data to the target point, The image feature is included in one of the flock images. the number of the flocs; the area of ​​each of the flocs; The sum of the areas, and the perimeter of each of said flocs; The decision model learning device is at least one of the above.

2. The flocculation basin is composed of a plurality of flocculation basins, The judgment model learning device according to claim 1 , wherein the target point is included in the upstream formation pond.

3. The determination model learning device according to claim 1 , wherein the degree of variation is a standard deviation.

4. a learning step of constructing a judgment model by performing machine learning using as explanatory variables the injection rate of an agent that is injected into a liquid containing suspended solids to coagulate the suspended solids, water quality data including the turbidity of the liquid upstream of the injection point of the agent, and the degree of variation in image feature values ​​obtained from each of a plurality of floc images of coagulated flocs formed by stirring the liquid into which the agent has been injected, taken at a target point downstream of the injection point of the agent, and the formation state of the coagulated flocs at the target point as a response variable; The target point is located in the front half of the flocculation basin; the plurality of floc images are taken at unit time intervals within a predetermined period including a time point at which the liquid has traveled from the measurement point of the water quality data to the target point, The image feature is included in one of the flock images. the number of the flocs; the area of ​​each of the flocs; The sum of the areas, and the perimeter of each of said flocs; A decision model learning method, which is at least one of the above.

5. 2. A judgment model learning program for causing a computer to function as the judgment model learning device according to claim 1, the judgment model learning program causing a computer to function as the learning unit.

6. a determination unit that determines the formation state using a determination model constructed by machine learning using explanatory variables including an injection rate of an agent that is injected into a liquid containing suspended solids to coagulate the suspended solids, water quality data including turbidity of the liquid upstream of the injection of the agent, and a degree of variation in image feature values ​​obtained from each of a plurality of floc images of coagulated flocs formed by stirring the liquid into which the agent has been injected, taken at a target point downstream of the injection point of the agent, and the formation state of the coagulated flocs at the target point as a response variable; The target point is located in the front half of the flocculation basin; the plurality of floc images are taken at unit time intervals within a predetermined period including a time point at which the liquid has traveled from the measurement point of the water quality data to the target point, The image feature is included in one of the flock images. the number of the flocs; the area of ​​each of the flocs; The sum of the areas, and the perimeter of each of said flocs; The determination device is at least one of the above.

7. The flocculation basin is composed of a plurality of flocculation basins, The determination device according to claim 6 , wherein the target point is included in the upstreammost formation pond.

8. The determination device according to claim 6 , wherein the degree of variation is a standard deviation.

9. A water purification facility comprising the determination device according to any one of claims 6 to 8.

10. the method includes a determination step of determining the formation state using a determination model constructed by machine learning using explanatory variables including an injection rate of an agent that is injected into a liquid containing suspended solids to coagulate the suspended solids, water quality data including the turbidity of the liquid upstream of the injection point of the agent, and a degree of variation in image feature values ​​obtained from each of a plurality of floc images of coagulated flocs formed by stirring the liquid into which the agent has been injected, taken at a target point downstream of the injection point of the agent, and the formation state of the coagulated flocs at the target point as a response variable; The target point is located in the front half of the flocculation basin; the plurality of floc images are taken at unit time intervals within a predetermined period including a time point at which the liquid has traveled from the measurement point of the water quality data to the target point, The image feature is included in one of the flock images. the number of the flocs; the area of ​​each of the flocs; The sum of the areas, and the perimeter of each of said flocs; The determination method is at least one of the above.

11. 7. A determination program for causing a computer to function as the determination device according to claim 6, the determination program causing a computer to function as the determination unit.

Images