JP6941083B2 - Water treatment method and water treatment system - Google Patents

Description

The present invention relates to a model using a machine learning algorithm, and a water treatment method and a water treatment system for determining the state of aggregated flocs from a still image or a continuous image of aggregated flocs.

In recent years, it has become the third artificial intelligence boom due to the dramatic improvement of computer computing power, the improvement of communication infrastructure environment, and the evolution of analysis algorithms. The improvement of the communication infrastructure environment has made it possible to easily acquire a large amount of data, and unlike the conventional artificial intelligence boom, expectations for its use in the real world are increasing.

Machine learning / artificial intelligence is expected to be applied to various fields, and the water treatment field is also exploring ways to utilize it. One of the ways to utilize machine learning / artificial intelligence in the water treatment field is to predict an objective variable in an unknown environment, for example, in the future in terms of time. One of the methods for predicting the objective variable in an unknown environment is an analytical approach. This is to theoretically analyze a physical phenomenon and formulate it into an equation. On the other hand, the prediction of the objective variable in an unknown environment by machine learning / artificial intelligence statistics the correlation between the input and the output without analyzing the theoretical relationship between the input and the output according to the natural law. The regression equation is created as an example of the relational equation.

In the field of water treatment, it may be difficult to theoretically analyze what is happening, and machine learning / artificial intelligence that can predict the output from the input without clarifying the theoretical background is effective. It is expected that there will be many situations in which it can be utilized.

As mentioned above, the artificial intelligence boom itself has reached its third boom this time. The leading algorithm of the second artificial intelligence boom that occurred in the late 1980s was neural networks. During the first artificial intelligence boom, a hidden layer was added to the perceptron, which consisted only of an input layer and an output layer, and a learning method called "backpropagation method" was established. It became an artificial intelligence boom. Eventually, the second artificial intelligence boom will be abolished, but it is said that one of the factors was that it was not possible to obtain satisfactory prediction accuracy depending on the issues. In this third artificial intelligence boom, deep learning (deep learning) with multiple layers of hidden layers of neural networks plays a leading role, and due to advances in learning algorithms, etc., it will be possible to demonstrate high accuracy depending on the task to be applied. This is one of the reasons for the boom. In this specification, machine learning using a neural network composed of an input layer, two or more hidden layers, and an output layer is referred to as deep learning. In addition, artificial intelligence is a broader concept than machine learning, and machine learning is one of the means for creating artificial intelligence.

Returning to the field of water treatment, in water treatment, aluminum and iron are used to efficiently remove suspended components such as turbid substances in raw water and soluble components such as biopolymers and humic acid. In many cases, a flocculant treatment containing a polyvalent cation such as is used to agglomerate a suspended component or a soluble component. The mass of dirt formed by agglomeration is called agglomerated floc. This agglutinating floc is settled in a settling tank provided after the agglutination tank, or is directly subjected to sand filtration or membrane filtration without being settled.

At this time, the state of aggregated flocs has a great influence on the processing performance of each processing. For example, in the precipitation treatment, the size and density of the aggregated flocs affect the sedimentation rate of the aggregated flocs. Further, in the filtration treatment, minute components such as picoplankton and soluble chromaticity components that are not sufficiently formed of aggregated flocs and are not incorporated into the aggregated flocs leak into the filtered water. Incomplete cohesive flocs can also clog filter sand and membrane pores, leading to an increase in filtration resistance and severe membrane fouling. Therefore, monitoring the state of aggregated flocs is extremely important in water treatment.

Japanese Unexamined Patent Publication No. 3-175339 Japanese Unexamined Patent Publication No. 2015-192960

It is difficult to judge the state of this aggregated floc in real time from the water quality. That is, was the agglutination operation appropriate until the agglutinating agent was injected, the agglutinating flocs were formed, the treated water was finally obtained through precipitation and filtration, and the quality of the treated water was measured, that is, the agglutinating flocs were formed. It is difficult to determine if it was good.

Occasionally, the raw water condition can fluctuate significantly in a short period of time due to the occurrence of typhoons and guerrilla rainstorms. For example, in recent years, there have been cases where the turbidity suddenly rises from 5 degrees to 1,000 degrees in 10 minutes. In such a situation, even if the water quality of the treated water is confirmed as described above and then the injection rate of the coagulant is reviewed, there are usually dozens of times from the injection of the coagulant to the acquisition of the treated water. It takes minutes to several hours, so it is not in time.

Therefore, when raw water fluctuations actually occur at a water purification plant, a veteran operator determines the coagulant injection rate based on past knowledge based on the raw water quality, and visually confirms the flock state of the coagulation tank. However, it is determined whether or not the determined injection rate is appropriate, and if necessary, the coagulant injection rate is reviewed again. Alternatively, the raw water is actually collected, divided into multiple beakers in equal amounts, and several types of coagulant injection rates are set to perform a coagulation test, a so-called jar test, to obtain an appropriate coagulant injection rate for the current raw water. I have decided.

However, in the case of the former, it depends largely on the experience and skill of the operator, and it is hard to say that it is a wise method in the water treatment industry, which has a problem of technology inheritance. Further, in the latter case, unevenness due to the operator is reduced, but it takes time and effort, and it takes about 15 minutes for the result of the jar test to be obtained, so that it is not possible to cope with the above-mentioned significant fluctuation of raw water.

Therefore, there is a need for a technique for grasping the state of agglomerated flocs in real time and performing agglomeration treatment more appropriately.
Patent Document 1 describes a flock monitoring device that images a flock under the water surface, performs image processing, quantifies the particle size of each flock, and obtains a particle size distribution. In addition, there is a description about performing operation control using the same device.

In Patent Document 2, the flock particle size is measured at a plurality of locations in the flock forming process by a flock particle size measuring device in the flock forming tank, or in the mixing tank and the flock forming tank, and the difference in the obtained flock particle size is obtained. Alternatively, there is a description of a water treatment system that detects flock dysplasia from the ratio.

In each case, the state of aggregation is determined based on the particle size of the aggregated flocs, but in reality, it is difficult to judge the quality of the aggregated flocs only by the particle size of the flocs. For example, when the flocculant injection rate is 10, 20, 30, 40, 50, 60 mg / L for raw water for which the optimum coagulant injection rate is determined to be 40 mg / L, and the floc particle size of each is measured. When the coagulant injection rate exceeds 40 mg / L, the coagulant becomes excessive, charge neutralization does not occur, floc formation failure occurs, and the floc particle size becomes small. Then, the floc particle size when the flocculant injection rate is 60 mg / L and the floc particle size when the flocculant injection rate is 30 mg / L may be substantially the same. In such a case, if control is performed based on the floc particle size, there is a possibility that the review of the injection rate of the coagulant may be erroneous.

The present invention has been made in view of the above points, and an object of the present invention is to learn the state of aggregated flocs including various elements including particle size by a machine learning algorithm such as deep learning, and to operate a veteran. By constructing a model capable of determining agglomerated flock equal to or higher than that of a member, we will provide a water treatment method that can determine the agglomerated flock state in water treatment in real time during water treatment. be. Another object of the present invention is to provide a water treatment system capable of carrying out such a water treatment method.

In one aspect, a flocculant is injected into raw water to form agglomerated flocs, image data of the agglomerated flocs is generated by an image acquisition device, and at least the image data of the agglomerated flocs is converted into a model constructed by a machine learning algorithm. A water treatment method is provided which inputs and outputs a determination result indicating the state of the aggregated flocs from the model.

As a machine learning algorithm, a deep learning method (deep learning method) is suitable. The deep learning method is a learning method based on a neural network in which hidden layers are multi-layered. In this specification, machine learning using a neural network composed of an input layer, two or more hidden layers, and an output layer is referred to as deep learning. By using the deep learning method, it is possible to determine the state of aggregated flocs, which has been determined based on human eyes and experience, by a computer based on the image data of aggregated flocs.

For example, the floc particle size when the flocculant injection rate is 60 mg / L and the floc particle size when the flocculant injection rate is 30 mg / L may be substantially the same. However, when comparing the actual state of flocs at 30 mg / L and 60 mg / L, there are small differences in the density of flocs, the so-called tightness, the color, the shape, and the like. According to machine learning / deep learning, it is possible to discriminate an image with an accuracy higher than that of a human being by repeatedly learning a large amount of combinations of image data and correct answer data corresponding to the image data to a computer.

In the present specification, water treatment refers to water purification treatment, water treatment, and wastewater treatment, and in particular, polyaluminum chloride (PACl), aluminum sulfate, ferric chloride, polyiron, and polysilica iron (PSI) for water. ), Organic coagulant, inorganic coagulant such as polymer or water treatment with coagulation operation by injecting organic coagulant.

Current agglutinating flocs in a water treatment system are imaged by an image acquisition device during water treatment. The image of the aggregated flocs may be either a still image or a continuous image.

In one aspect, in the step of generating the image data of the aggregated flocs by the image acquisition device, the water containing the aggregated flocs is extracted and introduced into the sampling container, and the image data of the aggregated flocs in the sampling container is imaged. This is a process generated by the acquisition device.

In the construction of a model using the deep learning method among the machine learning algorithms described later, since the image is characterized by a computer rather than a person, information other than aggregate flocs, such as background and image brightness, is also included in the model. May affect construction. Therefore, when an image acquisition device, for example, a camera, is installed in the aquarium, the image of the structure of the aquarium may be reflected or the brightness may differ depending on the weather. Therefore, as described above, water containing agglomerated flocs is extracted from the water treatment area, introduced into a sampling container in which the background and brightness are kept constant, and the agglomerated flocs in the sampling vessel are photographed for deep learning. It is possible to acquire suitable image data.

The image data of the agglutinated flocs obtained by the above means is input to the model constructed by the machine learning algorithm, and the quality of the current state of the agglutinated flocs, that is, whether or not the water treatment is properly performed is aggregated. Judgment is made in real time from the flock state. Machine learning algorithms learn a variety of factors, including the particle size of aggregated flocs. Therefore, the model constructed by the machine learning algorithm (that is, the trained model) can determine the quality of the aggregated floc state with higher accuracy than the conventional aggregation determination technique based on the particle size.

In the construction of the model, the explanatory variable of the training data is the image data of the aggregated floc, and the objective variable of the training data is the numerical value indicating the quality of the aggregated floc state, for example, 1: very bad, 2: bad, 3: slightly good, It can be given as a 5-point scale of 4: good and 5: very good. In one embodiment, a numerical value indicating the quality of the agglomerated floc state can be given as a numerical value indicating an excess or deficiency of the aggregating agent injection rate. For example, a five-grade evaluation of 1: low, 2: slightly low, 3: optimal, 4: slightly high, 5: high can be used as a numerical value indicating excess or deficiency of the coagulant injection rate. However, the present invention is not limited to the 5-grade evaluation, and may be, for example, a 3-grade evaluation or a 10-grade evaluation.

Which of the above five-grade evaluations the actual coagulant injection rate corresponds to is determined by the turbidity and chromaticity of the treated water in the precipitation treatment in the subsequent stage of the coagulation treatment, and the appearance of the coagulation flocs of veteran operators. Further, it can be determined from the quality of the sand filtered water in the subsequent stage, the rate of increase in filtration resistance of sand filtration, and the like. A dataset consisting of a combination of numerical values representing the excess or deficiency of the actual flocculant injection rate (1 to 5 in the above example) and the image data of the corresponding flocculation flocs is used for building and updating the model. Numerical values indicating the excess or deficiency of the actual flocculant injection rate are correct answer data, and are used for machine learning such as deep learning. In one embodiment, the correct answer data may be a numerical value (1 to 5 in the above example) indicating the quality of the actual aggregated floc state.

FIG. 1 is a schematic view showing an embodiment of a model for determining a state of aggregated flocs. The model is a neural network having an input layer 301, a plurality of hidden layers (also referred to as intermediate layers) 302, and an output layer 303. The model shown in FIG. 1 has four hidden layers 302, but the configuration of the model is not limited to the embodiment shown in FIG. Image data of aggregated flocs is input to the input layer 301 of the model. More specifically, the numerical value of each pixel constituting the image data of the aggregated floc is input to the input layer 301. In one example, if the aggregated floc image data is grayscale image data, the gray level of each pixel (usually a number within 0-255) is input to each node (neuron) of the model's input layer 301. When the image data of the aggregated flocs is color image data, numerical values representing red, green, and blue of each pixel are input to the corresponding nodes (neurons) of the input layer 301 of the model.

In the example shown in FIG. 1, the output layer 303 of the model outputs a numerical value indicating the judgment result of the quality of the aggregated floc state. For example, when the quality of the aggregated floc state is expressed by a 5-point scale of 1: very bad, 2: bad, 3: slightly good, 4: good, 5: very good, the output layer 303 of the model. Outputs any numerical value from 1 to 5. However, the configuration of the model shown in FIG. 1 is an example, and the present invention is not limited to the example shown in FIG.

When the image data of the current agglutination flock is input to the model constructed as described above, the computer executes the calculation according to the algorithm of the multi-layer perceptron that constitutes the neural network, and the output layer 303 is in the state of the agglutination flock. A numerical value indicating the judgment result of, or a numerical value indicating the judgment result of excess or deficiency of the coagulant injection rate is output. Based on the judgment result, the operator can aim at optimizing the operation of the water treatment system.

In one aspect, in addition to the image data of the aggregated flocs, information data related to water treatment is input to the model.

FIG. 2 is a schematic diagram showing another embodiment of the model for determining the state of aggregated flocs. In this embodiment, the explanatory variables of the training data include information data related to water treatment in addition to the image data of aggregated flocs. The input layer 301 of the model has a node (neuron) into which image data of aggregated flocs is input and a node (neuron) in which information data is input. According to this embodiment, it is possible to improve the determination accuracy as compared with the model in which only the image data of the aggregated flocs is used as the explanatory variable.

The information data is data representing the operating conditions of the water treatment system. More specifically, the information data includes raw water turbidity, chromaticity, pH, M-alkalinity, total organic carbon (TOC), chemical oxygen demand (COD), ultraviolet absorbance, water temperature, raw water flow rate, and weather. , Includes at least one of rainfall, flocculant injection rate. Considering that the water treatment system is automatically controlled using the model, it is desirable that the items included in the information data are items that can be automatically acquired by the sensor rather than the items obtained by manual analysis.

In particular, in the aggregation operation, pH and M-alkalinity are important for the quality of aggregation. For example, even if the coagulant injection rate is the same, the state of coagulant flocs is different when the pH and / or M-alkalinity is different. Therefore, as described above, a model is constructed according to the machine learning algorithm using not only the image data of the agglomerated flocs but also the information data including the pH of the raw water, the M-alkalinity, the aggregating agent injection rate, and the like. During the actual water treatment, information data such as agglomerated floc image data, raw water pH, M-alkalinity, and flocculant injection rate are input to the trained model.

In one aspect, the determination result indicating the state of the agglutinating floc is the determination result indicating the excess or deficiency of the aggregating agent injection rate.
In one aspect, the determination result is used to control the coagulant injection rate.

It is possible to automatically optimize the operation of the water treatment system from the judgment result of excess or deficiency of the coagulant injection rate. For example, the determination result is output at a predetermined cycle, for example, every minute. At a certain time T, if the judgment result is 1: "less", the current flocculant injection rate is increased by + 50% in 0.5 minutes, and if 2: "slightly less", the current flocculant injection rate is increased. Increase the rate by + 25% in 0.5 minutes, 3: if "optimal", the current coagulant injection rate is unchanged, 4: if "slightly high", the current coagulant injection rate is 0. .Reduce by -25% in 5 minutes, 5: If "high", reduce current flocculant injection rate by -50% in 0.5 minutes. Subsequently, the determination result is output again at time T + 1, and the coagulant injection rate is adjusted in the same manner. This step is repeated until the optimum judgment result is output.

In the above example, the numerical value representing the determination result of excess or deficiency of the coagulant injection rate is any one of 1 to 5, but the present invention is not limited to the above example. In one embodiment, the numerical values representing the judgment result of excess or deficiency of the coagulant injection rate are + 50% (corresponding to "less"), + 25% (corresponding to "slightly less"), and 0% (corresponding to "optimal"). , -25% (corresponding to "slightly more") or -50% (corresponding to "more").

Here, at time 0, that is, when switching from the existing control in the water treatment system to the control based on the above model, or when the operation of the water treatment system is newly started, the coagulant injection rate at the time when the control is started is set. The value may be determined from the past knowledge of the water treatment system, the judgment of an expert, the batch test such as the jar test, and the like. The control of the water treatment system is started by the coagulant injection rate determined here. After that, after considering the water flow time from the injection of the coagulant to the point to be photographed by the image acquisition device, the control for optimizing the coagulant injection rate using the above model is started. good.

In one aspect, in the water treatment method, a reference value of the coagulant injection rate, which is irrelevant to the state of the coagulation flocs, is determined by a reference value calculation formula, and a correction value of the reference value is used as the coagulant injection rate. The step of correcting the coagulant injection rate from the reference value and the correction value, which is determined based on the determination result indicating the excess or deficiency of the above, is further included.

If the quality of the raw water flowing into the water treatment system changes suddenly in a short time, such as when a guerrilla rainstorm occurs, control may be delayed and the quality of the treated water may deteriorate. Therefore, the computer has a reference value calculation formula for determining the reference value of the coagulant injection rate regardless of the state of the agglomerate floc, and the reference value of the coagulant injection rate is used by using this reference value calculation formula. To determine. Further, the computer determines the correction value of the reference value from the determination result indicating the excess or deficiency of the coagulant injection rate.

_{The reference value calculation formula is, for example, the reference value (D 0} ) of the coagulant injection rate from the value of the raw water quality (turbidity, chromaticity, pH, M-alkalinity, etc.) that is constantly monitored by the raw water quality meter. Is an expression for determining. This standard value calculation formula is, for example, from the data set of the past operation results in the water treatment system, that is, the value of the raw water quality and the corresponding coagulant injection rate, and from the value of the raw water quality by the statistical analysis method. It is possible to use a regression equation to calculate the corresponding flocculant injection rate. The flocculant is injected based on the reference value of the coagulant injection rate thus obtained, and agglomerated flocs are formed.

By imaging the aggregated flocs by the method described above and inputting the image data of the aggregated flocs into the model described above, the excess or deficiency of the flocculant injection rate is determined. Here _{, let D 0 be} a reference value (mg / L) of the coagulant injection rate, and d be a correction value represented by a positive or negative percentage (%) indicating the judgment result of excess or deficiency of the coagulant injection rate. The corrected flocculant injection rate (D, mg / L) is represented by the following correction formula (1).
D = D ₀ x (1 + d / 100) (1)

The correction value d represents the judgment result of excess or deficiency of the coagulant injection rate, + 50% (when "small"), + 25% (when "slightly low"), 0% (when "optimal"), -25%. Either (when "somewhat more") or -50% (when "more"). For example, when D ₀ = 20 mg / L and d = -25%, the corrected flocculant injection rate D is 15 mg / L. According to the above-described embodiment, it is possible to control the flocculant injection rate based on the image data of the flocculant flocs more quickly.

In one aspect, the water treatment method further comprises updating the model with a dataset consisting of a combination of the image data and numerical values representing the quality of the actual aggregated floc state.

The above model using the machine learning algorithm is continuously updated based on newly obtained data in the water treatment system performing the operation using the model. As a result, even if the properties of the raw water in the water treatment system change significantly from the time the model is first constructed, or even if the operation method is changed, the model that is updated according to the actual situation is appropriate. It is possible to continuously output various judgment results.

In one aspect, a computer having a stirring tank that stirs raw water infused with a flocculant to form agglomerated flocs, an image acquisition device that generates image data of the agglomerated flocs, and a model constructed by a machine learning algorithm. The computer includes a storage device in which the model is stored, and a processing device that inputs at least the image data to the model and executes an operation for outputting a determination result indicating the state of the aggregated flocs from the model. A water treatment system is provided.

In one aspect, a step of inputting at least aggregated floc image data into the input layer of a model consisting of a neural network having an input layer, a plurality of hidden layers, and an output layer, and a multilayer perceptron constituting the neural network. By executing the calculation according to the algorithm of, the output layer provides a computer-readable recording medium recording a program for causing the computer to execute a step of outputting a numerical value indicating the quality of the aggregated floc state.

According to the water treatment method according to the present invention, aggregation that has been conventionally determined by the human eye from a model constructed by a machine learning algorithm and image data of aggregation flocs obtained by an image acquisition device such as a small camera. The flock state can be automatically determined in real time with high accuracy, enabling more optimal operation of the water treatment system.

It is a schematic diagram which shows one Embodiment of the model which determines the state of aggregated flocs. It is a schematic diagram which shows the other embodiment of the model which determines the state of aggregated flocs. It is a figure which shows one Embodiment of a water treatment system. It is a figure which shows the other embodiment of a water treatment system. It is a front view which shows typically the sampling container and the image acquisition apparatus. It is a side view which shows typically the sampling container and the image acquisition apparatus. 3 is a schematic diagram showing an embodiment of a computer constituting at least a part of the arithmetic system shown in FIGS. 3 and 4.

FIG. 3 is a diagram showing an embodiment of a water treatment system. In this water treatment system, a trained model is used in a water treatment system using a rapid filtration method to determine the state of coagulation flocs (excess or deficiency of coagulant injection rate) and control the coagulant injection rate. All the descriptions including the construction of the model, the configuration of the model, etc. described in " Means for Solving the Problem " are applied to each embodiment of the water treatment system described below.

The water treatment system includes a landing well 1 for receiving raw water, a rapid stirring tank 2 and a slow stirring tank 3 as stirring tanks connected to the landing well 1, and a coagulant injection pump 5 for injecting a coagulant into the raw water. ing. In the present embodiment, the coagulant injection pump 5 is connected to the rapid stirring tank 2 and is arranged so as to inject the coagulant into the raw water in the rapid stirring tank 2. In one embodiment, the coagulant injection pump 5 may be connected to a pipe connecting the landing well 1 and the rapid stirring tank 2.

The rapid stirring tank 2 is connected to the landing well 1, and the slow speed stirring tank 3 is connected to the rapid stirring tank 2. The rapid stirring tank 2 is a stirring tank that stirs the raw water into which the coagulant is injected and agitates the suspended solids in the raw water to form aggregated flocs. The slow-speed stirring tank 3 is a stirring tank that agitates water containing agglomerated flocs to coarsen the agglomerated flocs.

A precipitation tank 7 is connected to the slow stirring tank 3. The water containing the agglutinating flocs grown in the slow stirring tank 3 is sent to the settling tank 7, and the agglutinating flocs are settled in the settling tank 7. The structure of the settling tank 7 is not particularly limited, but a case where an inclined plate or an inclined pipe is provided inside the sedimentation tank 7 and the aggregated flocs are settled on the inclined plate or the inclined pipe to remove the aggregated flocs. There is also. A filtration tank 8 is connected to the settling tank 7, and the supernatant water of the settling tank 7 is sent to the filtration tank 8. In the filtration tank 8, impurities such as coagulant residue and fine flocs are removed from the supernatant water to purify the water. The configuration of the filtration tank 8 is not particularly limited, but is generally configured to remove impurities by sand filtration or membrane filtration. The purified water is sent from the settling tank 7 to the water purification tank 9. In one embodiment, the precipitation tank 7 may be omitted, and the slow stirring tank 3 may also be omitted. In this case, water containing agglomerated flocs is sent to the filtration tank 8, and impurities such as agglomerated flocs are removed in the filtration tank 8.

The water treatment system inputs the image acquisition device 12 that generates the image data of the aggregated flocs formed in the rapid stirring tank 2 and at least the image data into the model constructed by the machine learning algorithm, and is in the state of the aggregated flocs. It further includes a calculation system 100 that outputs a pass / fail judgment result from the model. The arithmetic system 100 includes a program for constructing and updating a model, a storage device 110 in which the model is stored, and a processing apparatus 120 (such as a GPU or a CPU) that executes arithmetic according to the program. The processing device 120 inputs at least the image data to the model, and executes an operation for outputting the determination result of the quality of the aggregated floc state from the model.

The arithmetic system 100 is composed of at least one computer. The at least one computer may be one server or a plurality of servers. The model for determining the excess or deficiency of the aggregating agent injection rate (that is, the quality of the agglomerating floc state) based on at least the image data of the agglomerating flocs is stored in the storage device 110 of the arithmetic system 100. Further, the above-mentioned reference value calculation formula is also stored in the storage device 110 of the calculation system 100.

The computing system 100 may be an edge server connected to the image acquisition device 12 by a communication line, a cloud server connected to the image acquisition device 12 by a network such as the Internet, or an image acquisition device. It may be a fog computing device (gateway, fog server, router, etc.) installed in the network connected to 12. The arithmetic system 100 may be a plurality of servers connected by a network such as the Internet. For example, the arithmetic system 100 may be a combination of an edge server and a cloud server.

In the present embodiment, the agglomerated flocs in the rapid stirring tank 2 after the coagulant injection is photographed by the image acquisition device 12. The image acquisition device 12 generates image data of aggregated flocs and sends the image data to the calculation system 100. The arithmetic system 100 acquires image data from the image acquisition device 12 and stores it in the storage device 110. The image acquisition device 12 is installed above the rapid stirring tank 2 (above the water surface). An illuminator such as an LED light illuminates the aggregated flocs in the rapid stirring tank 2 so that the image acquisition device 12 can clearly image the aggregated flocs. The installation position of the image acquisition device 12 may be inside the rapid stirring tank 2 (below the water surface) or inside the pipe connecting the rapid stirring tank 2 and the slow speed stirring tank 3 as long as the aggregated flocs can be clearly imaged. No.

The image acquisition device 12 is a digital camera provided with an image sensor (for example, a CCD image sensor or a CMOS image sensor) capable of generating a still image or a continuous image. As the specifications of the digital camera, in order to keep the focal length constant, it is desirable to have a manual focus function, and the number of pixels is preferably 5 million pixels or more, preferably 10 million pixels or more.

In the present embodiment, the agglutinating floc to be imaged is the agglutinating floc in rapid agitation. This is because the control delay can be shortened, which is preferable from the viewpoint of controlling the coagulant injection rate. If the water purification plant has little fluctuation in raw water, the coarsened aggregated flocs in the slow stirring tank 3 may be the target of imaging.

The calculation system 100 issues a command to the coagulant injection pump 5 to adjust the coagulant injection rate based on the determination result of excess or deficiency of the coagulant injection rate output from the model. Specifically, when the determination result indicates that the current coagulant injection rate is low, the calculation system 100 issues a command to the coagulant injection pump 5 to increase the coagulant injection rate. When the determination result indicates that the current coagulant injection rate is high, the calculation system 100 issues a command to the coagulant injection pump 5 to reduce the coagulant injection rate. In this way, the calculation system 100 can optimize the injection rate of the coagulant by controlling the operation of the coagulant injection pump 5 based on the determination result output from the model. At the beginning of the optimization of the coagulant injection rate using the model, the coagulant injection rate is determined from the knowledge based on the past operation of the water treatment system and the jar test result.

As shown in FIG. 3, the landing well 1 is provided with a turbidity sensor 21, a pH sensor 22, and an M-alkalinity measuring device 23 for acquiring information data related to water treatment. The turbidity sensor 21, pH sensor 22, and M-alkali measuring device 23 are connected to the calculation system 100, and the turbidity measurement value, the pH measurement value, and the M-alkali measurement value are sent to the calculation system 100. It is supposed to be. Further, as information data, at least one of total organic carbon (TOC), chemical oxygen demand (COD), ultraviolet absorbance, water temperature, raw water flow rate, weather, rainfall amount, and coagulant injection rate is input to the calculation system 100. May be done. The numerical value input to the calculation system 100 as information data may be a measured value transmitted from the sensor or a value input to the calculation system 100 by the worker.

In one embodiment, the above-mentioned information data is input to the input layer of the above-mentioned trained model as an explanatory variable together with the image data of the aggregated flocs. The model outputs the judgment result of excess or deficiency of the coagulant injection rate, which is the objective variable. Since the quality of the raw water flowing to the image acquisition device 12 at the time of agglomeration flock photography and the quality of the raw water at the landing well 1 at that time may be different, the raw water quality corresponding to the acquired image data of the agglomerate flock. It is desirable to adopt a value in consideration of the arrival time from the landing well 1 to the image acquisition device 12.

The model is constructed by acquiring various data with the water treatment system shown in FIG. 3 before the operation using the model by using a deep learning method or the like. Here, in a model using a machine learning algorithm, the prediction accuracy for input data that has never been experienced is basically low, so the data in the model construction includes changes in raw water quality due to seasonal fluctuations. desirable. Therefore, it is desirable to build a model using the data acquired over one year or more. However, if the raw water quality is stable, it is possible to shorten the data acquisition period. In addition, if a model has already been constructed in a neighboring water treatment system that makes the raw water the same, or a water treatment system that treats similar raw water, the model can be diverted.

The agglomerated flock information output from the model relates to the quality of the agglomerated flock state, that is, the excess or deficiency of the aggregating agent injection rate. The operator of the water treatment system may review the coagulant injection rate with reference to this information, but as shown in Fig. 3, if this information is incorporated into the coagulant control, the coagulant injection rate can be continuously adjusted. It is possible to optimize. In this way, the main water treatment system can be continuously operated stably.

However, if the operation of the water treatment system is prolonged, the quality of the raw water may change or a part of the operation flow may be changed, and the model may not be able to output an appropriate judgment result. Therefore, it is necessary to update the model as appropriate. Whether or not the determination result of excess or deficiency of the coagulant injection rate was correct can be determined based on the value of the turbidity (precipitated water turbidity) of the supernatant water at the outlet of the settling tank 7. If the agglomeration treatment is not performed properly, the sedimentation property of the agglomerated flocs deteriorates, and the amount of suspended solids remaining in the supernatant water without agglomeration increases, resulting in a high degree of sedimentation turbidity.

Therefore, in the present embodiment, during the operation of the model, the value of the sedimented water turbidity is continuously acquired by the turbidity sensor 30, and the data set of the above-mentioned explanatory variables and objective variables is taken in consideration of the inflow time. Add as supplementary information to. This supplementary information is compared with the output result of the model to verify the accuracy of the model on a regular basis. When the accuracy of the model is lower than that at the beginning of operation, the correct answer data is given to the objective variable of the data set again based on the supplementary information, and the model is updated by the deep learning method or the like.

When the arithmetic system 100 includes an edge server at the edge (field) of the water treatment system and a cloud server connected to the edge server by a network (for example, the Internet), the model update work is performed by the cloud server. It is desirable to do this and then send the updated model (ie the trained model) to the edge server and store it in the edge server.

FIG. 4 is a diagram showing another embodiment of the water treatment system. Since the configuration of the present embodiment not particularly described is the same as that of the embodiment shown in FIG. 3, the duplicated description will be omitted. In this embodiment, similarly to the above-described embodiment, the excess or deficiency of the coagulant injection rate is determined by using the model learned in the processing system using the rapid filtration method, and the coagulant injection rate is controlled. This is an example of the method.

In the present embodiment, the imaging of the aggregated flocs is not performed in the water treatment area of the water treatment system, but in the sampling container 41 arranged in the light-shielding housing 40. That is, the water containing the aggregated flocs is extracted from the water treatment region and introduced into the sampling container 41 through the extraction pipe 43 branching from the water treatment region. Then, the image acquisition device 12 images the aggregated flocs in the sampling container 41. In the present embodiment, one end of the extraction pipe 43 is connected to a pipe connecting the rapid stirring tank 2 and the slow speed stirring tank 3. In one embodiment, one end of the extraction tube 43 may be connected to a rapid stirring tank 2, a slow speed stirring tank 3, or a pipe connecting the slow speed stirring tank 3 and the precipitation tank 7.

FIG. 5 is a front view schematically showing the sampling container 41 and the image acquisition device 12, and FIG. 6 is a side view schematically showing the sampling container 41 and the image acquisition device 12. The sampling container 41 and the image acquisition device 12 are arranged in the light-shielding housing 40. A dark room is formed inside the light-shielding housing 40. The sampling container 41 has a transparent side wall 41a. As the material of the transparent side wall 41a, quartz, a transparent acrylic resin, transparent vinyl chloride, or the like can be used. A flow cell can be used as such a sampling container 41.

An illuminator 46 such as an LED light is arranged in the light-shielding housing 40. The illuminator 46 illuminates the aggregated flocs in the sampling container 41 so that the image acquisition device 12 can clearly photograph the aggregated flocs. The illuminator 46 is arranged above the sampling container 41 and illuminates the sampling container 41 from above. The image acquisition device 12 is arranged so as to face the transparent side wall 41a of the sampling container 41, and images the aggregated flocs in the sampling container 41.

As the power for introducing water containing agglomerated flocs from the water treatment area into the sampling container 41, the method using the water level difference is the most gentle and does not damage the flocs, so it is desirable. , Mono pumps, positive displacement pumps, peristaltic pumps and the like may be used.

Water containing agglomerated flocs flows in from the lower part of the sampling container 41, passes over the overflow weir 41b, and flows out from the sampling container 41. Further, the flow rate of the water sent to the sampling container 41 may be such that the aggregated flocs do not settle in the sampling container 41. The water containing agglomerated flocs flowing out of the sampling container 41 may be returned to the rapid stirring tank 2 or the slow stirring tank 3 or may be discarded.

According to this embodiment, it is possible to eliminate the influence of sunlight and the reflection of structures such as a stirring tank. Further, by keeping the brightness in the dark room constant by the illuminator 46, it is possible to always image the aggregated flocs under the same conditions. Further, maintenance such as lens cleaning of the image acquisition device 12 can be easily performed.

Although it depends on the state of the raw water, dirt adheres to the transparent side wall 41a of the sampling container 41 within one to several days, which hinders the acquisition of aggregated flock images. To prevent this, the sampling container 41 is manually cleaned or replaced once a day. Alternatively, a chemical for removing stains, such as a sulfuric acid solution or a sodium hypochlorite solution, is automatically poured into the sampling container 41 at a frequency (for example, once every 3 hours) according to the stains on the transparent side wall 41a. Dirt may be removed. Further, a mechanism for automatically cleaning the transparent side wall 41a by a physical means such as a sponge or a wiper may be provided.

FIG. 7 is a schematic diagram showing an embodiment of a computer constituting at least a part of the arithmetic system 100 shown in FIGS. 3 and 4. The computer includes a storage device 110 that stores programs and data, a processing device 120 such as a CPU (central processing unit) or GPU (graphic processing unit) that performs calculations according to the program stored in the storage device 110, and data. , A program, and an input device 130 for inputting various information to the storage device 110, an output device 140 for outputting processing results and processed data, and a communication device 150 for connecting to a network such as the Internet. I have.

The storage device 110 includes a main storage device 111 accessible to the processing device 120 and an auxiliary storage device 112 for storing data and programs. The main storage device 111 is, for example, a random access memory (RAM), and the auxiliary storage device 112 is a storage device such as a hard disk drive (HDD) or a solid state drive (SSD).

The input device 130 includes a keyboard and a mouse, and further includes a recording medium reading device 132 for reading data from the recording medium and a recording medium port 134 to which the recording medium is connected. The recording medium is a computer-readable recording medium that is a non-temporary tangible object, and is, for example, an optical disk (for example, a CD-ROM or a DVD-ROM) or a semiconductor memory (for example, a USB flash drive or a memory card). be. Examples of the recording medium reading device 132 include an optical drive such as a CD-ROM drive and a DVD-ROM drive, and a card reader. An example of the recording medium port 134 is a USB port. The program and / or data stored in the recording medium is introduced into the computer via the input device 130 and stored in the auxiliary storage device 112 of the storage device 110. The output device 140 includes a display device 141 and a printing device 142.

The trained model constructed by the machine learning algorithm is stored in the storage device 110. As shown in FIG. 1 or 2, this trained model is a neural network having an input layer 301, a plurality of hidden layers (also referred to as intermediate layers) 302, and an output layer 303. The computer operates according to a program electrically stored in the storage device 110. That is, the computer has a step of acquiring image data of aggregated flocs from the image acquisition device 12, a step of inputting at least image data into a trained model constructed by a machine learning algorithm, and a multi-layer perceptron constituting a neural network. By executing the calculation according to the algorithm, the step of outputting the numerical value indicating the judgment result of the quality of the coagulation floc state or the judgment result of the excess or deficiency of the coagulant from the output layer 303 is executed. In addition, the computer performs a step of building the model and a step of updating the model. The program for causing the computer to perform these steps is recorded on a computer-readable recording medium, which is a non-temporary tangible object, and is provided to the computer via the recording medium. Alternatively, the program and the model may be input to the computer from the communication device 150 via a communication network such as the Internet.

The above-described embodiment is described for the purpose of enabling a person having ordinary knowledge in the technical field to which the present invention belongs to carry out the present invention. Various modifications of the above embodiment can be naturally made by those skilled in the art, and the technical idea of the present invention can be applied to other embodiments. Therefore, the present invention is not limited to the described embodiments, but is construed in the broadest range according to the technical idea defined by the claims.

1 Water landing well 2 Rapid stirring tank 3 Slow speed stirring tank 5 Coagulant injection pump 7 Settling tank 8 Filtration tank 9 Water purification tank 12 Image acquisition device 21 Turbidity sensor 22 pH sensor 23 M-alkalinity measuring device 30 Turbidity sensor 40 Shading Housing 41 Sampling container 43 Extractor tube 46 Illuminator 100 Computational system 110 Storage device 111 Main storage device 112 Auxiliary storage device 120 Processing device 130 Input device 132 Recording medium reading device 134 Recording medium port 140 Output device 141 Display device 142 Printing device 150 Communication Device 301 Input layer 302 Hidden layer 303 Output layer

Claims (5)

A coagulant is injected into raw water to form coagulation flocs,
The image data of the aggregated flocs is generated by an image acquisition device, and the image data is generated.
The image data before Symbol flocs, information data related to water treatment using the learning data including the explanatory variables to construct a learned model according the machine learning algorithm, the information data, pH of the raw water, M- alkalinity , Including coagulant injection rate,
From the trained model, a judgment result indicating excess or deficiency of the coagulant injection rate, which is an objective variable indicating the state of coagulation flocs, is output in real time, and the injection rate of the coagulant is adjusted based on the judgment result. A characteristic water treatment method. The water treatment method according to claim 1.
In the step of generating the image data of the aggregated flocs by the image acquisition device, the water containing the aggregated flocs is extracted and introduced into the sampling container, and the image data of the aggregated flocs in the sampling container is generated by the image acquisition device. Water treatment method, which is a process to be performed. The water treatment method according to claim 1.
The reference value of the aggregating agent injection rate, which is irrelevant to the state of the agglutinating floc, is determined by the reference value calculation formula.
The correction value of the reference value is determined based on the determination result indicating the excess or deficiency of the coagulant injection rate.
A water treatment method further comprising a step of correcting the flocculant injection rate from the reference value and the correction value. The water treatment method according to any one of claims 1 to 3.
A water treatment method further comprising a step of updating the trained model using a data set consisting of a combination of the image data and a numerical value indicating the quality of the actual aggregated floc state. A stirring tank that stirs the raw water infused with the coagulant to form coagulating flocs,
An image acquisition device that generates image data of the aggregated flocs, and
A computer having a trained model constructed according to a machine learning algorithm using the image data of the aggregated flocs and training data including information data related to water treatment as explanatory variables is provided.
The information data includes the pH of raw water, M-alkalinity, and flocculant injection rate.
The computer outputs in real time a storage device in which the trained model is stored and a determination result indicating excess or deficiency of the coagulant injection rate, which is an objective variable indicating the state of coagulation flocs, from the trained model. A water treatment system including a treatment device that executes a calculation for adjusting the injection rate of the coagulant based on a determination result.

Images