JP7554068B2 - Machine learning device, data processing system, and machine learning method - Google Patents

Description

The present disclosure relates to a machine learning device, a data processing system, and a machine learning method that can be applied to a centrifuge system.

Centrifugal separators that use centrifugal force to separate solids and liquids have been used in water treatment facilities for water supply and sewage, industrial wastewater, and human waste. Various types of centrifuges are known, but a type of centrifuge called a decanter is widely used, particularly in facilities that require continuous treatment for long periods of time.

This decanter generally includes a drive motor, a bowl that is rotated by the drive motor and into which the liquid to be treated is fed, a screw conveyor that is coaxially arranged within the bowl, and a differential speed generating device that generates a differential speed between the rotation speed of the bowl and the rotation speed of the screw conveyor. The solid components in the liquid to be treated fed into the bowl are discharged from one end of the bowl, and the liquid components in the liquid to be treated are discharged from the other end of the bowl, thereby achieving solid-liquid separation. Various suitable methods for controlling the operation of this decanter have been considered. (See, for example, Patent Document 1 below.)

Patent No. 5442099

According to the method described in the above-mentioned Patent Document 1, the differential speed generated by the differential speed generating device and the centrifugal force of the bowl are controlled based on the conveying torque of the screw conveyor, so that automatic operation of the centrifuge device can be partially realized according to the properties of the liquid to be treated. However, there are various information and parameters related to the control of the centrifuge system including such a centrifuge device other than the conveying torque of the screw conveyor, the differential speed generated by the differential speed generating device, and the centrifugal force of the bowl. In particular, the information (state variables) that affect the control of the control object is diverse, and such information may include information that is difficult to quantify. And, at present, a method for effectively and mechanically taking these state variables into consideration has not been established, and as a result, the reality is that such state variables are reflected in the operation control of the centrifuge system based on the judgment of the engineer (operator) based on the experience and tacit knowledge.

In view of the above, the present disclosure aims to provide a machine learning device, a data processing system, and a machine learning method for achieving suitable operational control of a centrifuge system without relying on the judgment of an operator.

In order to achieve the above object, the machine learning device 20 according to the first aspect of the present disclosure is for a centrifuge system 1 including, as shown in FIG. 3, a bowl 2 that applies centrifugal force to the treated liquid PL1 (not shown) and PL2 to separate them into solid matter M (not shown) and separated liquid SL (not shown), a screw conveyor 3 that transports the solid matter M in the bowl 2 toward a discharge outlet 2a, a drive motor 4 that rotates the bowl 2, and a differential speed generating device 5 that rotates the screw conveyor 3 at a differential speed relative to the bowl 2, and includes image data of the treated liquid PL2 before it is supplied to the bowl 2 and after a predetermined additive has been added, taken from a predetermined angle of view. The learning data set storage unit 22 stores a plurality of learning data sets each including input data and output data each including a control parameter associated with the input data, the control parameter including at least one of the additive supply amount, the centrifugal force of the bowl 2, and the differential speed controlled by the differential speed generating device 5; a learning unit 23 that learns a learning model that infers a correlation between the input data and the output data by inputting a plurality of sets of the learning data sets; and a learned model storage unit 24 that stores the learning model learned by the learning unit 23.

By configuring it in this way, it is possible to provide a trained model that can infer the control parameters of a centrifuge system from image data of the floc-containing liquid being treated.

As shown in FIG. 9, the machine learning device 20A according to the second aspect of the present disclosure is the machine learning device according to the first aspect of the present disclosure, in which the learning unit includes a first learning unit 231 that learns a first learning model that infers features of image data by inputting image data in the learning dataset; and a second learning unit 232 that learns a second learning model that infers the control parameters by inputting features of the image data.

In this configuration, when estimating the control parameters of a centrifuge system from image data of the floc-containing treated liquid, the number of learning data sets required to obtain a trained model generated in each learning unit can be relatively reduced by dividing the learning unit into two.

As shown in FIG. 11, for example, a machine learning device 20B according to a third aspect of the present disclosure is a machine learning device according to the first aspect of the present disclosure, in which the input data further includes at least one of the concentration of the separated liquid SL, the moisture content of the liquid-containing solid M discharged from the discharge outlet 2a, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3.

By configuring it in this way, the number of input data increases, making it possible to relatively reduce the number of training datasets required to obtain the trained model generated in the learning unit.

As shown in FIG. 13, the machine learning device 20C according to the fourth aspect of the present disclosure is the machine learning device according to the first aspect of the present disclosure, in which the input data further includes at least one of the concentration of the separated liquid SL, the water content of the liquid-containing solid M discharged from the discharge port 2a, the slurry concentration of the liquid to be treated PL1, and the torque value of the screw conveyor 3, and the learning unit includes a first learning unit 231 that learns a first learning model that infers the feature amount of the image data by inputting the image data in the learning dataset; and a fourth learning unit 234 that learns a fourth learning model that infers the control parameter by inputting the feature amount of the image data, the concentration of the separated liquid SL, the water content of the liquid-containing solid M discharged from the discharge port 2a, the slurry concentration of the liquid to be treated PL1, and the torque value of the screw conveyor 3.

In this configuration, when estimating the control parameters of a centrifuge system from image data of the floc-containing treated liquid, the learning unit is divided into two, and the number of input data of the learning dataset learned in one of the learning units is increased, making it possible to relatively reduce the number of learning datasets required to obtain a trained model generated in each learning unit.

A machine learning device according to a fifth aspect of the present disclosure is a machine learning device according to any one of the first to fourth aspects of the present disclosure, in which the control parameters further include the supply amount of the treated liquid PL1.

By configuring it in this way, the number of control parameters output by the trained model increases, making it possible to achieve more fine-grained parameter adjustment.

A machine learning device according to a sixth aspect of the present disclosure is a machine learning device according to any one of the first to fifth aspects of the present disclosure, in which the control parameters further include a dam set diameter of the bowl 2.

By configuring it in this way, the number of control parameters output by the trained model increases, making it possible to achieve more fine-grained parameter adjustment.

As shown in FIG. 7 and FIG. 10, the data processing system 80, 80A according to the seventh aspect of the present disclosure is used in a centrifuge system 1 including a bowl 2 that applies centrifugal force to the treated liquid PL1, PL2 to separate it into solid matter M and separated liquid SL, a screw conveyor 3 that transports the solid matter M in the bowl 2 toward a discharge port 2a, a drive motor 4 that rotates the bowl 2, and a differential speed generating device 5 that rotates the screw conveyor 3 at a differential speed relative to the bowl 2, and includes an image data acquisition unit 81 for acquiring image data of the treated liquid PL2 taken from a predetermined angle of view before being supplied to the bowl 2 and after a predetermined additive has been added; and an inference unit 87, 871, 872 that infers control parameters of the centrifuge system 1 by inputting the data acquired by the image data acquisition unit 81 into a trained model generated by the machine learning device 20, 20A according to the first or second aspect.

With this configuration, control parameters that can achieve optimal operational control of the centrifuge system can be inferred based on image data of the floc-containing liquid being treated, making it possible to automatically control the operation of the centrifuge system without relying on the judgment of the operator.

Data processing systems 80B and 80C according to an eighth aspect of the present disclosure are used in a centrifuge system 1 including a bowl 2 that applies centrifugal force to the treated liquids PL1 and PL2 to separate them into solids M and a separated liquid SL, a screw conveyor 3 that transports the solids M in the bowl 2 toward a discharge port 2a, a drive motor 4 that rotates the bowl 2, and a differential speed generating device 5 that rotates the screw conveyor 3 at a differential speed relative to the bowl 2, as shown in, for example, FIG. 12 and FIG. 14. The data processing systems 80B and 80C according to an eighth aspect of the present disclosure are used in a centrifuge system 1 including, for example, a bowl 2 that applies centrifugal force to the treated liquids PL1 and PL2 to separate them into solids M and a separated liquid SL, a screw conveyor 3 that transports the solids M in the bowl 2 toward a discharge port 2a, a drive motor 4 that rotates the bowl 2, and a differential speed generating device 5 that rotates the screw conveyor 3 at a differential speed relative to the bowl 2, and an image of the treated liquid PL2 taken from a predetermined angle of view before being supplied to the bowl 2 and after a predetermined additive has been added thereto. The device includes an image data acquisition unit 81 for acquiring data; an additional variable acquisition unit 89 for acquiring at least one of the concentration of the separated liquid SL, the moisture content of the liquid-containing solid M discharged from the discharge port 2a, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3; and inference units 873, 871, and 874 for inferring the control parameters of the centrifuge system 1 by inputting the data acquired by the image data acquisition unit 81 and the additional variable acquisition unit 89 into the trained model generated by the machine learning device 20B or 20C according to the third or fourth aspect.

With this configuration, control parameters that can realize suitable operational control of the centrifuge system can be inferred based on at least one of the concentration of the separated liquid, the moisture content of the liquid-containing solids M, the slurry concentration of the treated liquid, and the torque value of the screw conveyor, as well as image data capturing the floc-containing treated liquid, so that the operational control of the centrifuge system can be performed automatically without relying on the judgment of the operator.

The machine learning method according to the ninth aspect of the present disclosure is implemented by a computer, for example as shown in FIG. 3 and FIG. 6, and is for a centrifuge system 1 including a bowl 2 that applies centrifugal force to the treated liquid PL1, PL2 to separate the solid matter M and the separated liquid SL, a screw conveyor 3 that transports the solid matter M in the bowl 2 toward a discharge outlet 2a, a drive motor 4 that rotates the bowl 2, and a differential speed generating device 5 that rotates the screw conveyor 3 at a differential speed relative to the bowl 2, and the treated liquid before being supplied to the bowl 2 and after a predetermined additive has been added thereto is The method includes: step S11 of storing multiple sets of learning data sets including input data including image data obtained by capturing an image of PL2 from a predetermined angle of view, and output data including control parameters associated with the input data, the control parameters including at least one of the amount of additives supplied, the centrifugal force of the bowl 2, and the differential speed controlled by the differential speed generating device 5; step S15 of learning a learning model that infers a correlation between the input data and the output data by inputting multiple sets of the learning data sets; and step S17 of storing the learned learning model.

By configuring it in this way, it is possible to provide a trained model that can infer the control parameters of a centrifuge system from image data of the floc-containing liquid being treated.

According to the present disclosure, it becomes possible to obtain control parameters that can mechanically achieve suitable operational control of the centrifuge system without relying on the judgment of the operator.

FIG. 2 is a schematic diagram showing a general structure of a decanter body according to the first embodiment of the present disclosure. FIG. 1 is a schematic diagram showing a centrifugal separation system including a piping structure of a decanter according to a first embodiment of the present disclosure. 1 is a schematic block diagram of a machine learning device according to a first embodiment of the present disclosure. 3 shows an example of image data obtained by the treatment liquid monitoring system according to the first embodiment of the present disclosure. FIG. 2 is a diagram showing an example of a neural network model for supervised learning implemented in the machine learning device according to the first embodiment of the present disclosure. 1 is a flowchart illustrating an example of a machine learning method according to a first embodiment of the present disclosure. 1 is a schematic block diagram illustrating a data processing system according to a first embodiment of the present disclosure. 1 is a flowchart illustrating an example of parameter adjustment by the data processing system according to the first embodiment of the present disclosure. FIG. 11 is a schematic block diagram of a machine learning device according to a second embodiment of the present disclosure. FIG. 11 is a schematic block diagram illustrating a data processing system according to a second embodiment of the present disclosure. FIG. 13 is a schematic block diagram of a machine learning device according to a third embodiment of the present disclosure. FIG. 13 is a schematic block diagram illustrating a data processing system according to a third embodiment of the present disclosure. FIG. 13 is a schematic block diagram of a machine learning device according to a fourth embodiment of the present disclosure. FIG. 13 is a schematic block diagram illustrating a data processing system according to a fourth embodiment of the present disclosure.

Each embodiment for carrying out the present disclosure will be described below with reference to the drawings. Note that the scope necessary for the explanation to achieve the objectives of the present disclosure will be shown in a schematic manner below, and the scope necessary for the explanation of the relevant parts of the present disclosure will be mainly explained, and the parts for which explanation is omitted will be explained as being based on publicly known technology.

First Embodiment
Before describing the machine learning device, data processing system, and machine learning method according to the first embodiment of the present disclosure, a centrifuge system to which the machine learning device, data processing system, and machine learning method are applied will be briefly described. The centrifuge system according to the present embodiment uses a centrifuge device that includes a horizontal decanter 1. Note that the specific aspects of the centrifuge system according to the present disclosure are not limited to those shown below, and can also be applied to centrifuge systems that include a centrifuge device such as a vertical or straight-body decanter.

Figure 1 is a schematic diagram showing the general structure of a decanter body according to a first embodiment of the present disclosure. Also, Figure 2 is a schematic diagram showing a centrifuge system including a piping structure of a decanter according to a first embodiment of the present disclosure. The horizontal decanter 1 shown in this embodiment mainly includes a bowl 2, a screw conveyor 3, a drive motor 4, a differential speed generating device 5, and a casing 6, as shown in Figures 1 and 2.

The bowl 2 is made of a cylindrical member with one end processed into a weight shape, and is supported so as to be rotatable around a horizontal axis. One or more solid discharge outlets 2a are provided at the weight-shaped end of the bowl 2, and a dam 2c with one or more separated liquid discharge outlets 2b is attached to the other end. Solid components contained in the liquid to be treated that is poured into the bowl 2 are mainly discharged from the solid discharge outlet 2a, and liquid components contained in the liquid to be treated that is also poured into the bowl 2 are mainly discharged from the separated liquid discharge outlet 2b. A pulley 4a that transmits power from a drive motor 4, which will be described later, is attached to one end of the bowl 2 (the end processed into a weight shape in FIG. 1). The solid components discharged from the solid discharge outlet 2a are discharged in a state in which most of the liquid components in the liquid to be treated have been dehydrated and removed, and contain a relatively small amount of liquid (separated liquid) components, so hereinafter, these solid components are referred to as liquid-containing solids (or dehydrated solids) M.

The screw conveyor 3 is arranged coaxially within the bowl 2 and is composed of a member with a spiral screw blade 3a formed around it. The screw blade 3a is a member for transporting and/or squeezing the solid matter in the bowl 2. The body 3b of this screw conveyor 3 is provided with a space 3c for receiving the liquid to be treated therein, and a discharge port 3d for pouring the liquid to be treated stored in this space into the bowl 2. In addition, a differential speed generating device 5, which will be described later, is connected to one end of this screw conveyor 3.

The drive motor 4 is a motor for applying a rotational force to the bowl 2, and is connected to a pulley 4a attached to one end of the bowl 2 via a belt 4b. A relatively large motor is preferably used for the drive motor 4 in order to rotate the bowl 2 in the range of, for example, 2000 to 5000 rpm, and the rotation speed is preferably variable as appropriate by inverter control. The differential speed generating device 5 is a device for generating a differential speed between the rotation speed of the bowl 2 and the rotation speed of the screw conveyor 3, and is a device that can rotate the screw conveyor 3 slightly slower (for example, about 1 to 3 rpm) than the bowl 2. The differential speed generating device 5 includes a gear box 5a connected to the screw conveyor 3 and a differential motor (also called a "back drive motor") 5b that applies a braking force to the screw conveyor 3. The specific operating principle of the differential speed generating device 5 is conventionally known, so its description will be omitted.

The casing 6 is provided to cover the bowl 2 and the screw conveyor 3, and guides the liquid-containing solids (dehydrated solids) M, which are solids containing a small amount of liquid and discharged from the solid discharge outlet 2a of the bowl 2, to the solids chute 6a provided below, and similarly guides the separated liquid SL (see FIG. 3) discharged from the separated liquid discharge outlet 2b of the bowl 2 to the separated liquid chute 6b provided below. In addition, the bowl 2 and the screw conveyor 3 covered by the casing 6 are supported together by the frame 6c.

The solid matter chute 6a of the casing 6 is connected to a solid matter discharge conduit 7, and the solid matter discharged from the solid matter chute 6a is transported, for example, via this solid matter discharge conduit 7 to a main transport path 7a to a drying device, incineration equipment, or the like (not shown). In addition, the separated liquid chute 6b of the casing 6 is connected to a separated liquid discharge conduit 8, and the separated liquid discharged from the separated liquid chute 6b is transported, for example, via this separated liquid discharge conduit 8 to a water purification equipment, or the like (not shown).

The decanter 1 further includes a liquid supply pipe 9 for feeding the liquid to be treated into the bowl 2. One end of this liquid supply pipe 9 is connected to the space 3c in the body 3b of the screw conveyor 3, and the other end is connected to a source 10 of the liquid to be treated and a source 11 of additives, and serves as a conduit for feeding the liquid to be treated, etc., that flows in from the other end, into the bowl 2.

The liquid supply pipe 9 and the treated liquid supply source 10 are fluidly connected by a treated liquid pipe 10b equipped with a pump 10a for supplying the treated liquid. By controlling this pump 10a, the amount of treated liquid supplied into the bowl can be controlled. Here, the treated liquid supplied from the treated liquid supply source 10 is often a slurry (suspension) containing sludge as a solid matter.

The liquid supply pipe 9 and the additive supply source 11 are fluidly connected by an additive pipe 11c equipped with a pump 11a for supplying an additive, such as a chemical, to be added to the liquid being treated, and multiple valves 11b for controlling the supply position and timing of the chemical. Here, the chemical supplied from the additive supply source 11 is a flocculant, particularly a polymer flocculant or an inorganic flocculant, which is administered to the slurry to change the sludge contained in the slurry into a floc-like state. In the following, the liquid being treated before the chemical is added will be referred to as the "liquid being treated PL1" and the liquid being treated after the chemical has been added will be referred to as the "floc-containing liquid being treated PL2" to distinguish between the two.

As shown in FIG. 2, the additive pipe 11c according to this embodiment is branched into two pipes at its midpoint, one of which is connected to the middle of the treated liquid pipe 10b, and the other is connected to the liquid supply pipe 9. When the chemical is supplied via the pipe connected to the middle of the treated liquid pipe 10b, the treated liquid PL1 reacts with the chemical in the treated liquid pipe 10b (so-called "line chemical supply"); when the chemical is supplied via the pipe connected to the liquid supply pipe 9, the treated liquid PL1 reacts with the chemical mainly in the space 3c and the bowl 2 (so-called "in-machine chemical supply"). By adopting such a piping structure, the position and timing at which the chemical is added to the treated liquid PL1 can be adjusted, and the effect of generating sludge flocs (also called "floc aggregates") by the chemical, which is mainly influenced by the state of the treated liquid PL1, can be optimally adjusted. Note that the connection position of the pipes shown in FIG. 2 is only one example, and the piping structure can be appropriately changed in consideration of the state of the treated liquid, etc.

In addition, it is preferable that the various pipes of the decanter 1 according to this embodiment are provided with a monitoring system for monitoring objects passing through the pipes. In detail, this monitoring system includes a separated liquid monitoring system 40 provided in the separated liquid discharge conduit 8, a dehydrated solids monitoring system 50 provided in the solids discharge conduit 7, and a treated liquid monitoring system 60 provided in the treated liquid pipe 10b. In addition to these monitoring systems, a screw conveyor torque monitoring system 70 for monitoring the torque value of the screw conveyor 3 may also be included. Several examples of the specific configuration of each monitoring system will be described in accordance with each embodiment described later. In the first embodiment, no monitoring systems other than the treated liquid monitoring system 60 are used, so monitoring systems other than the treated liquid monitoring system 60 will be described in the second embodiment and onwards.

In the solid-liquid separation by the decanter 1 having the above-mentioned configuration, first, the floc-containing liquid PL2 to be treated is fed through the liquid feed pipe 9 into the bowl 2 rotating at a predetermined rotation speed, and the solids in the floc-containing liquid PL2 (which has become flocculated due to the effect of the chemicals) settle on the inner wall surface of the bowl 2 due to the action of the centrifugal force generated by the drive motor 4. The settled solids are conveyed while being squeezed toward the one end of the bowl 2, which is machined into a weight shape, by the screw blades 3a of the screw conveyor 3, which rotates with the bowl 2 at a rotation speed slightly lower than the rotation speed of the bowl 2 due to the action of the differential speed generating device 5, and are discharged from the solid discharge port 2a to the solid chute 6a together with a predetermined amount of fluid components. In addition, the residue in the bowl 2 after the solids have been settled and removed as described above remains in the bowl 2 as the separated liquid SL for a certain period of time, and then overflows from the separated liquid discharge port 2b and is discharged to the separated liquid chute 6b.

<Machine learning device>
The above-mentioned solid-liquid separation process is performed in a suitable manner by adjusting various control parameters of the decanter 1. The control parameters here mainly consist of the centrifugal force of the bowl 2, the amount of chemicals supplied from the additive supply source 11, and the differential speed between the bowl 2 and the screw conveyor 3. Hereinafter, a machine learning device 20 according to a first embodiment of the present disclosure that learns an inference model (trained model) capable of estimating optimal control parameters in the decanter 1 will be described.

FIG. 3 is a schematic block diagram of a machine learning device 20 according to a first embodiment of the present disclosure. As shown in FIG. 3, the machine learning device 20 according to this embodiment includes a learning dataset acquisition unit 21, a learning dataset storage unit 22, a learning unit 23, and a trained model storage unit 24.

The learning data set acquisition unit 21 is an interface unit that acquires a plurality of data constituting a learning (training) data set, for example, via a wired or wireless communication line. The specific contents of the plurality of data acquired here can be appropriately changed according to the trained model to be generated. In this embodiment, the plurality of data are acquired as image data of the floc-containing treated liquid PL2 and control parameters associated with this image data. The control parameters include the supply amount of the chemical added to the treated liquid PL1, the centrifugal force of the bowl 2, and the differential speed controlled by the differential speed generating device 5. In this embodiment, the above-mentioned three data are acquired as control parameters, but the control parameters constituting the learning data set may include at least one of the above-mentioned three data. In addition, the control parameters are not limited to the above-mentioned three, and may also include, for example, the supply amount of the treated liquid PL1 from the treated liquid supply source 10 and the dam set diameter of the dam 2c of the bowl 2. Therefore, the control parameters constituting the output data may be four or more. In general, if the number of control parameters is increased, more detailed parameter adjustments can be achieved when applied to the operation control of the decanter 1 described below, but as the number of control parameters increases, the number of training data sets required to obtain a trained model capable of inference with sufficient accuracy tends to increase. Therefore, it is preferable to determine the number of control parameters as output data taking into consideration the number of training data sets that can be prepared, etc.

An example of a method for acquiring a plurality of data acquired by the learning data set acquisition unit 21 will be described. As shown in FIG. 3, the learning data set acquisition unit 21 is connected to a computer PC1 and acquires desired data from the computer PC1. The computer PC1 may be, for example, at least a part of the control unit 30 that controls various operations of the decanter 1 described above, or may be a computer communicably connected to the control unit 30. This allows the computer PC1 to acquire various control parameters of the decanter 1. In addition, as shown in FIG. 2, the computer PC1 is directly or indirectly connected to a treated liquid monitoring system 60 provided in the treated liquid piping 10b via the control unit 30, and can acquire image data of the floc-containing treated liquid PL2 from the treated liquid monitoring system 60. The control unit 30 of the decanter 1 referred to here is a device for controlling the entire decanter 1, and may be composed of a well-known computer equipped with a processor, memory, etc., connected to various sensors and driving means provided in the decanter 1.

Here, as shown in FIG. 3, the treated liquid monitoring system 60 may include a sight glass 61 connected to an arbitrary location downstream of the connection position of the treated liquid pipe 10b with at least the additive pipe 11c, and a floc-containing treated liquid imaging camera 63 that is installed at a predetermined angle of view in a window 62 provided in the sight glass 61 and can capture image data of the floc-containing treated liquid PL2 flowing through the sight glass 61. Of these, the floc-containing treated liquid imaging camera 63 may be a well-known camera capable of capturing two-dimensional images. In addition, in this treated liquid monitoring system 60, a light source (not shown) that illuminates the floc-containing treated liquid PL2 in the sight glass 61 and a polarizing filter (not shown) that can be attached to the floc-containing treated liquid imaging camera 63 may be appropriately used so that the state of the floc-containing treated liquid PL2 is easily reflected in the image data captured by the floc-containing treated liquid imaging camera 63. The configuration of the treated liquid monitoring system 60 is not limited to this, and various configurations can be adopted as long as they are capable of acquiring image data of the floc-containing treated liquid PL2. As a specific example, when the chemical is supplied through the additive pipe 11c connected to the liquid supply pipe 9, a configuration may be adopted in which the floc-containing treated liquid imaging camera 63 is disposed at a position capable of imaging the space 3c inside the screw conveyor 3, or a configuration such as the treated liquid monitoring system 60A shown in FIG. 9, which will be described later, may be adopted.

Figure 4 shows an example of image data obtained by the floc-containing treated liquid imaging camera 63 of the treated liquid monitoring system 60 according to the first embodiment of the present disclosure. It is generally known that when performing solid-liquid separation processing using a centrifuge system, if the solid components in the treated liquid fed into the decanter 1 are sufficiently aggregated by the chemical solution, the solid components and the liquid components can be well separated. In addition, in the image data obtained by the treated liquid monitoring system 60, if the floc aggregation of the floc-containing treated liquid PL2 is well performed, one floc aggregate tends to become coarse and the transparency of the area where the floc aggregate is not present tends to increase. Therefore, as shown in Figure 4, it can be inferred that the image data in which the floc aggregation reaction in the floc-containing treated liquid PL2 is well proceeding (for example, the one shown in Figure 4(b)) tends to have a greater amount of color change between pixels than the image data in which the reaction is not well proceeding due to, for example, a shortage of the chemical supply (for example, the one shown in Figure 4(a)).

The learning dataset storage unit 22 is a database for storing a plurality of data constituting the learning dataset acquired by the learning dataset acquisition unit 21, by associating related input data and output data (also called "teacher data") into one learning dataset. In the present embodiment, the learning dataset stored may be one in which the input data is image data captured from a predetermined angle of view of the floc-containing treated liquid PL2 before it is supplied to the bowl 2 and after the addition of the chemical, and the output data is a control parameter associated with the image data as the input data. The specific configuration of the database constituting the learning dataset storage unit 22 may be adjusted as appropriate. For example, in FIG. 3, for convenience of explanation, the learning dataset storage unit 22 and the trained model storage unit 24 described later are shown as separate storage means, but they may also be configured by a single storage medium (database).

As described above, the learning dataset stored in the learning dataset storage unit 22 is composed of one image data and a control parameter corresponding to this one image data. On the other hand, in order to generate one trained model in the learning unit 23, it is necessary to perform learning using many learning datasets (for example, several thousand to tens of thousands of sets). Therefore, in order to prepare a large number of learning datasets in a relatively short time, it is preferable to perform data augmentation in the learning dataset storage unit 22. As a specific method of this data augmentation, for example, a method can be adopted in which a number of square partial image data smaller than the original image data is randomly extracted from one rectangular original image data (for example, 100 pieces), and each of the extracted a pieces of partial image data is associated with a control parameter corresponding to the original image data to obtain a number of learning datasets.

The learning unit 23 inputs a plurality of sets of learning datasets stored in the learning dataset storage unit 22, and learns a learning model that infers the correlation between the input data and the output data in the learning dataset. In this embodiment, as will be described in detail later, supervised learning using a neural network is adopted as a specific method of machine learning. However, the specific method of machine learning is not limited to this, and other learning methods can be adopted as long as they can learn the correlation between input and output from the learning dataset. For example, ensemble learning (random forest, boosting, stacking, etc.) can also be used.

The trained model storage unit 24 is a database for storing trained models generated by the learning unit 23. The trained models stored in the trained model storage unit 24 are applied to the real system upon request via communication lines including the Internet or storage media. Specific aspects of applying the trained models to the real system (data processing system 80) will be described in detail later.

When preparing a learning dataset for the decanter 1 according to this embodiment, it is necessary to identify optimal control parameters that will serve as teacher data. As a method for identifying these optimal control parameters, for example, a method in which the operator EN manually identifies them by taking into consideration the image data acquired by the treated liquid monitoring system 60 and the actual control parameters at that time can be adopted. In this case, it is preferable to assign a skilled engineer or multiple engineers to the operator EN who identifies the optimal control parameters that will serve as teacher data. The optimal control parameters identified in this way are organized and stored in the learning dataset storage unit 22 in the machine learning device 20 as teacher data for the learning dataset in a state associated with the above-mentioned image data.

Here, the optimal control parameters to be identified may be at least one of the centrifugal force of the bowl 2, the amount of drug supply from the additive supply source 11, and the differential speed between the bowl 2 and the screw conveyor 3, as described above. Since these control parameters are considered to be the control parameters that have the greatest effect on the solid-liquid separation process by the decanter 1, it is preferable that the control parameters to be identified here are all of the above-mentioned three control parameters as exemplified in this embodiment. Of course, only one or only two of these three control parameters may be adopted as output data, and it is also possible to further add parameters other than these three control parameters (e.g., drug supply position, etc.) to the output data. Of the above three control parameters, the centrifugal force of the bowl 2 can be adjusted mainly by controlling the rotation speed of the bowl 2 by the drive motor 4, but the control of changing this rotation speed has a characteristic of being less responsive than the other two controls (control of drug supply amount and differential speed). Therefore, if only two of the three control parameters are used as output data, it is advisable to use the amount of drug supplied from the additive supply source 11 and the differential speed between the bowl 2 and the screw conveyor 3.

Next, a learning method in the learning unit 23 using the multiple learning data sets obtained as described above will be briefly described. FIG. 5 is a diagram showing an example of a neural network model for supervised learning implemented in the machine learning device according to the first embodiment of the present disclosure. The neural network in the neural network model shown in FIG. 5 is composed of l neurons (x1 to x1) in the input layer, m neurons (y11 to y1m) in the first hidden layer, n neurons (y21 to y2n) in the second hidden layer, and o neurons (z1 to zo) in the output layer. The first and second hidden layers are also called hidden layers, and the neural network may have multiple hidden layers in addition to the first and second hidden layers, or may have only the first hidden layer as a hidden layer.

In addition, between the input layer and the first hidden layer, between the first hidden layer and the second hidden layer, and between the second hidden layer and the output layer, there are nodes that connect the neurons between the layers, and each node is assigned a weight wi (i is a natural number).

The neural network in the neural network model according to this embodiment uses a learning dataset to learn the correlation between the input data and output data of the learning dataset. Specifically, image data of the floc-containing liquid to be treated PL2 as state variables constituting the input data is associated with the neurons in the input layer, and the values of the neurons in the output layer are calculated using a general method for calculating the output value of a neural network, that is, the value of the output side neuron is calculated as the sum of a sequence of multiplication values of the value of the input side neuron connected to the neuron and the weight wi associated with the node connecting the output side neuron and the input side neuron, for all neurons other than the neurons in the input layer. Note that when the state variables are input to the neurons in the input layer, the format in which the information acquired as the state variables is input can be appropriately set in consideration of the accuracy of the trained model to be generated, etc.

Then, the calculated values of o neurons z1 to zo in the output layer, i.e., the control parameters in this embodiment, are compared with the teacher data t1 to to, which are also composed of control parameters and form part of the learning dataset, to find the error, and the weights wi associated with each node are adjusted (backpropagation) so that the found error becomes small.

Then, when a certain condition is met, such as the above-mentioned series of steps being repeated a certain number of times or the error being smaller than an allowable value, the learning is terminated and the neural network model (all weights wi associated with each of the nodes) is stored as a learned model in the learned model storage unit 24.

<Machine learning method>
In relation to the above, the present disclosure provides a machine learning method. FIG. 6 is a flowchart showing an example of a machine learning method according to a first embodiment of the present disclosure. In the machine learning method described below, the description will be based on the above-mentioned machine learning device 20, but the premise of the configuration is not limited to the above-mentioned machine learning device 20. In addition, the machine learning method is realized by using a computer, but various computers can be applied, such as a computer constituting the control unit 30, a server device arranged on a network, or the computer PC1 shown in FIG. 3. In addition, the specific configuration of the computer can be, for example, one that includes at least a calculation device consisting of a CPU, a GPU, etc., a storage device consisting of a volatile or non-volatile memory such as a RAM or a ROM, a communication device for communicating with a network or other devices, and a bus connecting these devices. Furthermore, the machine learning method may be provided in the form of a program having one or more instructions for causing a computer to execute a predetermined operation, or in the form of a non-transitory computer-readable medium storing the program.

In the supervised learning as a machine learning method according to the present embodiment, a desired number of training datasets are first prepared as a preliminary step before starting machine learning, and the prepared multiple training datasets are stored in the training dataset storage unit 22 (step S11). The number of training datasets to be prepared here should be set taking into consideration the inference accuracy required for the trained model that is ultimately obtained. In addition, an example of a method for preparing training datasets has already been given above, so a description thereof will be omitted here.

After step S11 is completed, a pre-learning neural network model is prepared (S12) to start machine learning in the learning unit 23. The pre-learning neural network model prepared here may have the structure shown in FIG. 4, for example, and the weights of each node may be set to an initial value. Then, for example, one learning dataset is randomly selected from the multiple learning datasets stored in the learning dataset storage unit 22 (step S13), and the input data in the one learning dataset is input to the input layer (see FIG. 4) of the prepared pre-learning neural network model (step S14). Note that various methods can be used to input the input data in the learning dataset to the input layer of the pre-learning neural network model. As a specific example, a method can be used in which the luminance value and/or color value (e.g., RGB value) of each pixel of the image data is input to the neurons of each input layer. In addition, a predetermined pre-processing such as a dimensional reduction process or noise removal process for adjusting the number of data may be performed before inputting the image data as input data to the input layer.

Here, the control parameters of the output layer (see FIG. 4) generated as a result of step S14 are generated by the neural network model before learning, and therefore in most cases are values different from the desired result. Therefore, next, machine learning is performed using the output data as teacher data in one learning dataset acquired in step S13, i.e., the control parameters, and the control parameters of the output layer generated in step S13 (step S15). The machine learning performed here may be, for example, a process (backpropagation) in which the control parameters constituting the teacher data and the control parameters constituting the output layer are compared, an error between the two is detected, and a weight associated with each node in the neural network model before learning is adjusted so that an output layer in which the error is reduced is obtained. In addition, the number and format of the control parameters output to the output layer of the neural network model before learning are the same as those of the teacher data in the learning dataset as the learning target. Therefore, for example, if the control parameters as teacher data in the learning dataset are composed of data corresponding to two control parameters, the control parameters output to the output layer by the neural network model are also data corresponding to the two control parameters.

Once machine learning is performed in step S15, it is determined whether or not it is necessary to continue machine learning, for example, based on the remaining number of unlearned training datasets stored in the training dataset storage unit 22 (step S16). If machine learning is to be continued (No in step S16), the process returns to step S13, and if machine learning is to be terminated (Yes in step S16), the process proceeds to step S17. If the machine learning is to be continued, steps S13 to S15 are performed multiple times on the neural network model being trained using unlearned training datasets. The accuracy of the trained model that is finally generated generally tends to be higher in proportion to the number of times.

When the machine learning is to be terminated (Yes in step S16), the weights associated with each node are adjusted through a series of steps to generate a neural network, which is stored as a trained model in the trained model storage unit 24 (step S17), and the series of learning processes is terminated. The trained model stored here can be applied and used in various data processing systems, and specific examples of such data processing systems will be described later.

In the above-mentioned learning process and machine learning method of the machine learning device, in order to generate one trained model, multiple machine learning processes are repeatedly performed on one (pre-training) neural network model to improve its inference accuracy and obtain a trained model that is suitable for application to a data processing system. However, the present invention is not limited to this acquisition method. For example, a plurality of trained models that have undergone a predetermined number of machine learning processes are stored as candidates in the trained model storage unit 24, and input data from a common dataset for validity judgment is input to the group of multiple trained models to generate an output layer (neuron values), and the accuracy of the control parameters identified in this output layer is compared to select one of the best trained models to be applied to the data processing system. The validity judgment dataset may be composed of a dataset similar to the training dataset used for learning, and may not have been used for learning.

Optionally, instead of the pre-trained neural network model prepared in step S12 above in which the node weights are set to initial values, a trained model previously generated for, for example, any image recognition can be used. In this case, the desired trained model is generated by so-called fine tuning or transfer learning. Therefore, a highly accurate trained model can be generated with a smaller number of training datasets than in the normal machine learning process described above.

As described above, by applying the machine learning device 20 and machine learning method according to this embodiment, a trained model capable of deriving optimal control parameters from image data acquired by the treated liquid monitoring system 60 can be obtained.

<Data Processing System>
Next, an application example of the trained model generated by the above-described machine learning device and machine learning method will be described with reference to Fig. 7. Fig. 7 is a schematic block diagram showing a data processing system 80 according to a first embodiment of the present disclosure. As the data processing system 80 according to the present embodiment, a system applied to the control unit 30 of the above-described horizontal decanter 1 is exemplified.

As shown in FIG. 7, the data processing system 80 mainly includes an image data acquisition unit 81, a parameter adjustment unit 82, a calculation unit 83, a database (DB) 84, a user interface 85, and an internal bus 86 for interconnecting these components. Note that FIG. 7 shows only the components related to the data processing system 80, and omits the description of other components provided in the control unit 30 that are not directly related to the data processing system 80 according to this embodiment.

The image data acquisition unit 81 is for acquiring image data, more specifically, image data of the floc-containing treated liquid PL2. Specifically, it is connected to the above-mentioned treated liquid monitoring system 60 and acquires image data captured by the floc-containing treated liquid imaging camera 63 of the treated liquid monitoring system 60.

The parameter adjustment unit 82 adjusts various control units, which are adjusted by the control unit 30 to realize optimal operation control, based on the inference results of the inference unit 87 described later. This parameter adjustment unit 82 is connected to a drive motor control unit 31 that is connected to the drive motor 4 and controls its rotation speed, a pump control unit 32 that is provided in the additive pipe 11c and controls the pump 11a that varies the amount of drug supplied, and a differential motor control unit 33 that is connected to a differential motor 5b and controls its rotation speed to control the rotation speed of the screw conveyor 3 (relative to the bowl 2). Note that the parameter adjustment unit 82 in this embodiment is connected to the three control units described above, but the connection destinations can be appropriately adjusted according to the number and type of control parameters output by the inference unit 87.

The arithmetic unit 83 constitutes a processor for realizing various processes in the data processing system 80, and includes at least an inference unit 87. The arithmetic unit 83 is connected to a learned model storage unit 88 that stores a learned model used in the inference unit 87. The inference unit 87 infers a control parameter to be adjusted by the parameter adjustment unit 82 from image data as a state variable acquired by the image data acquisition unit 81 by taking into account a learned model stored in the learned model storage unit 88. It is preferable that a plurality of learned models are stored in the learned model storage unit 88 according to the purpose and various conditions (for example, environmental conditions such as season, weather, temperature and humidity, or the type of liquid to be treated, etc.). Furthermore, the task of selecting an appropriate learned model from a plurality of learned models may be automatically selected using various sensors, etc., or may be manually selected by an operator EN, etc.

The database 84 is made of a well-known recording medium and is used to temporarily or continuously store various data handled by the data processing system 80, such as image data acquired by the image data acquisition unit 81 and the results of calculations by the calculation unit 83. The user interface 85 is, for example, a GUI (graphical user interface), and is used to display the status of the decanter 1 and receive input operations from the operator EN, etc.

Figure 8 is a flowchart showing an example of parameter adjustment by the data processing system 80 according to the first embodiment of the present disclosure. When the decanter 1 to which the data processing system 80 is applied starts to operate, as shown in Figure 8, the data processing system 80 first determines whether it is time to adjust the parameters (step S21). Here, the timing for adjusting the control parameters can be specified automatically or manually. In addition, the adjustment may be performed periodically at a predetermined time interval while the decanter 1 is operating, or the adjustment may be performed only at a specific timing, such as when the operation of the decanter 1 is started, when the type of the treated liquid to be treated by the decanter 1 is changed, when a predetermined time has elapsed since operation, or when the treated liquid monitoring system 60 or the dehydrated solids monitoring system 50 is capable of detecting a change in the concentration of the treated liquid or the moisture content of the dehydrated solids and the change exceeds a predetermined threshold.

When it is detected that it is time to adjust the parameters (Yes in step S21), the floc-containing treated liquid imaging camera 63 in the treated liquid monitoring system 60 is operated to image the floc-containing treated liquid PL2 flowing in the treated liquid pipe 10b (step S22). The image data captured here is acquired by the image data acquisition unit 81 into the data processing system 80 (step S23).

Then, the image data acquired by the image data acquisition unit 81 is sent to the inference unit 87 via the internal bus 86, and the inference unit 87 infers control parameters by inputting the image data to the input layer of one trained model in the trained model storage unit 88 that has been specified in advance by the inference unit 87 (step S24). Then, the parameter adjustment unit 82 adjusts each control unit using the values of the control parameters inferred here (step S25). After that, the process returns to step S21 and enters a standby state.

As described above, according to the data processing system 80 of this embodiment, the preferred control parameters of the decanter 1 can be estimated based on image data of the floc-containing treated liquid PL2, which is information that is relatively easy to obtain, and therefore this data processing system 80 can be applied to the decanter 1 relatively easily. Furthermore, adjustment of the control parameters can be easily performed without relying on the judgment of the operator EN, etc.

Optionally, in the above embodiment, the trained model taken into consideration in the inference unit 87 is generated through so-called batch learning as shown in FIG. 6, but the data processing system 80 of the present invention is not limited to this. Specifically, for example, the trained model generated through the above-mentioned batch learning may be further subjected to online learning to further improve accuracy. As a specific method, first, the image data as state variables acquired in the image data acquisition unit 81 and the control parameters inferred by the inference unit 87, particularly the control parameters when the solid-liquid separation processing result is improved as a result of reflecting the adjustment in the parameter adjustment unit 82, are temporarily stored in the database 84 as a set of online learning data sets. Then, machine learning (specifically fine tuning) similar to that shown in FIG. 6 can be performed using the online learning data set stored in the database 84 at any time.

Although the above-mentioned data processing system 80 is applied in the control unit 30 of the horizontal decanter 1, it can be applied to, for example, a computer connected to the control unit 30 of the decanter 1 in a communicable manner, or a server device connected to the control unit 30 of the decanter 1 via a network or the like. However, the centrifuge systems used for sludge treatment, which are mainly assumed in this embodiment, usually have significantly different processing contents (type of liquid to be treated, processing amount per unit time, etc.) and surrounding environments (climate, etc.) for each system. Therefore, generating a general-purpose trained model that can be applied to all of these centrifuge systems requires a very large number of training data sets for the generation, and the contents of the data must be obtained under various conditions without bias, which tends to increase the generation cost. Therefore, it is cost-effective and preferable to apply and operate the data processing system 80 and the above-mentioned trained model to each centrifuge system.

Second Embodiment
In the machine learning device, the machine learning method, and the data processing system according to the first embodiment, a case where a control parameter is inferred from image data of the floc-containing liquid to be treated PL2 using one trained model has been described. However, although a correlation is recognized between the image data of the floc-containing liquid to be treated PL2 and the control parameter, the degree of correlation between the two is not extremely high. In general, in a machine learning process of a neural network model, the number of training data sets required to obtain a trained model capable of inference with sufficient accuracy tends to be small in proportion to the degree of correlation between the input data and the output data of the training data set. Therefore, in the first embodiment, in order to learn the correlation between the image data of the floc-containing liquid to be treated PL2 and the control parameter, a relatively large number of training data sets must be prepared. Therefore, as one aspect for generating a trained model capable of inference with sufficient accuracy with a smaller number of training data sets, a case where two trained models are used to infer a control parameter from image data of the floc-containing liquid to be treated PL2 will be described below as a second embodiment of the present disclosure. Among the components of the machine learning device 20A and data processing system 80A according to the second embodiment described below, those common to the components of the machine learning device 20 and data processing system 80 according to the first embodiment are denoted by the same reference numerals and their description is omitted. Similarly to the first embodiment, the machine learning device, machine learning method, and data processing system according to the second embodiment described below are described by taking as an example a case where they are applied to the centrifuge system shown in Figures 1 and 2. Furthermore, all of the modified examples described in the embodiments described above or below are applicable to this embodiment to the extent that no contradiction occurs.

FIG. 9 is a schematic block diagram of a machine learning device 20A according to a second embodiment of the present disclosure. As shown in FIG. 9, the machine learning device 20A according to this embodiment is similar to the machine learning device 20 according to the first embodiment, except that the components of the machine learning device 20A according to this embodiment include a first learning dataset storage unit 221 and a second learning dataset storage unit 222 as learning dataset storage units, and a first learning unit 231 and a second learning unit 232 as learning units. In relation to this, the multiple data acquired in the learning dataset acquisition unit 21 are also different from those in the first embodiment.

The multiple data acquired by the learning data set acquisition unit 21 in this embodiment are acquired from the computer PC1, and in addition to the image data of the floc-containing liquid PL2 to be treated and the control parameters, the multiple data further include feature quantities of the image data of the floc-containing liquid PL2 to be treated (hereinafter also referred to as "feature quantities of the floc-containing liquid PL2 to be treated"). As the feature quantities of the floc-containing liquid PL2 to be treated, various pieces of information identified based on the image of the floc-containing liquid PL2 to be treated can be adopted. Specifically, for example, one or more of the size, color value, density, etc. of the floc aggregates can be included. The feature quantities can be identified, for example, by actually sampling the floc-containing liquid PL2 to be treated and measuring and analyzing it using various measuring instruments, or by the operator EN visually judging the feature quantities.

In order to measure the above-mentioned characteristic quantities, it is preferable that the treated liquid monitoring system 60A according to this embodiment adopts a configuration in which the floc-containing treated liquid PL2 is sampled and image data is obtained. Specifically, as shown in FIG. 9, the treated liquid monitoring system 60A can adopt a system including a branch pipe 64 provided at an appropriate position in the treated liquid piping 10b, a valve 65 provided in the branch pipe 64, a transparent container 66 disposed at the end of the branch pipe 64, and a floc-containing treated liquid imaging camera 63 installed on the side of the transparent container 66. Note that the arrangement of the floc-containing treated liquid imaging camera 63 is not limited thereto, and may be installed at a position where the container 66 can be imaged from above, for example. In the treated liquid monitoring system 60A having such a configuration, the floc-containing treated liquid PL2 flowing in the treated liquid piping 10b is sampled and extracted into the transparent container 66 by opening the valve 65 at any timing. The floc-containing treated liquid PL2 in the container 66 is then imaged by the floc-containing treated liquid imaging camera 63 to generate image data of the floc-containing treated liquid PL2, and the corresponding desired control parameters are identified by the operator EN. Furthermore, by measuring and analyzing the floc-containing treated liquid PL2 sampled in the container 66, it is possible to identify the characteristic quantities of the floc-containing treated liquid PL2. Of course, the treated liquid monitoring system 60 described in the first embodiment above can also be used in this embodiment as an alternative.

The multiple data acquired in the learning data set acquisition unit 21 are stored separately as two learning data sets in the first learning data set storage unit 221 and the second learning data set storage unit 222, taking into consideration their respective correspondences. The first learning data set stored in the first learning data set storage unit 221 may include, as first input data, image data of the floc-containing treated liquid PL2 before it is supplied to the bowl 2 and after the chemical is added, taken from a predetermined angle of view, and may include, as first output data, a feature quantity of the floc-containing treated liquid PL2 associated with the first input data. The second learning data set stored in the second learning data set storage unit 222 may include, as second input data, a feature quantity of the floc-containing treated liquid PL2, and may include, as second output data, a control parameter associated with the second input data. When splitting and generating two learning data sets from a plurality of data in this manner, for example, the image data of the floc-containing liquid to be treated PL2, the feature quantities of the floc-containing liquid to be treated PL2, and the control parameters associated with the same floc-containing liquid to be treated PL2 may be split into a set of the image data of the floc-containing liquid to be treated PL2 and the feature quantities of the floc-containing liquid to be treated PL2, and a set of the feature quantities of the floc-containing liquid to be treated PL2 and the control parameters, and each set may be used as a first and second learning data set. Note that the first and second learning data sets obtained after splitting in this case do not need to be stored in a format that maintains their association, since each set is referenced in a different learning unit, which will be described later.

The first and second learning data sets stored in the first learning data set storage unit 221 and the second learning data set storage unit 222, respectively, are referred to only by different learning units. The first learning unit 231 learns a learning model that infers the correlation between the first input data and the first output data by inputting a plurality of sets of the first learning data set. In other words, the first learning unit 231 learns a first learning model that infers the feature amount of the floc-containing treated liquid PL2 in the image data, i.e., the feature amount of the image data, by inputting the image data of the floc-containing treated liquid PL2 in the first learning data set. The second learning unit 232 learns a learning model that infers the correlation between the second input data and the second output data by inputting a plurality of sets of the second learning data set. In other words, the second learning unit 232 learns a learning model that infers the control parameter by inputting the feature amount of the image data in the second learning data set.

The specific machine learning methods in the first and second learning units 231, 232 use different learning datasets, but the process is similar to the supervised learning process shown in FIG. 6. The first and second trained models obtained through a series of machine learning processes are then stored in the trained model storage unit 24, respectively.

Figure 10 is a schematic block diagram showing a data processing system 80A according to a second embodiment of the present disclosure. As shown in Figure 10, the data processing system 80A according to this embodiment has the same components as the data processing system 80 according to the first embodiment described above, except that the data processing system 80A according to this embodiment has a first inference unit 871 and a second inference unit 872 as inference units in the arithmetic unit 83.

The first inference unit 871 performs inference using the first trained model generated by the above-mentioned machine learning device 20A and stored in the trained model storage unit 88. Therefore, when image data of the floc-containing treated liquid PL2 acquired by the image data acquisition unit 81 is input, the first inference unit 871 outputs the feature amount of the image data to the output layer. Moreover, the second inference unit 872 performs inference using the second trained model generated by the above-mentioned machine learning device 20A and stored in the trained model storage unit 88. Therefore, when the feature amount of the image data inferred by the first inference unit 871 is input, the second inference unit 872 outputs a control parameter to the output layer.

When adjusting the control parameters in the decanter 1 to which the data processing system 80A including the above-mentioned first and second inference units 871, 872 is applied, the same process as that shown in FIG. 8 described above can be performed. However, when inferring the control parameters in step S24 of the process shown in FIG. 8, first, image data of the floc-containing treated liquid PL2 is input to the first inference unit 871, and the feature quantities of the output image data are input to the second inference unit 872.

As described above, according to the machine learning device, machine learning method, and data processing system of this embodiment, the feature quantities of image data that have a higher degree of correlation with the image data of the floc-containing treated liquid PL2 than the control parameters are inferred, and then the control parameters are inferred using the feature quantities of image data that have a higher degree of correlation with the control parameters than the image data of the floc-containing treated liquid PL2. This is expected to relatively reduce the number of learning datasets required to generate each trained model that enables inference with sufficient accuracy.

Third Embodiment
In the above-described first and second embodiments, the machine learning device and the machine learning method have been described in terms of the case where only image data of the floc-containing liquid to be treated PL2 is used as input data of the learning data set. However, when only image data of the floc-containing liquid to be treated PL2 is used as input data, the number of learning data sets required to generate a trained model capable of inference with sufficient accuracy tends to increase. Therefore, as one aspect for generating a trained model capable of inference with sufficient accuracy with a smaller number of learning data sets, a case where the number of input data of the learning data set shown in the above-described first embodiment is increased will be described below as a third embodiment of the present disclosure. Note that, among the components of the machine learning device 20B and the data processing system 80B according to the third embodiment described below, those components common to the components of the machine learning device 20 and the data processing system 80 according to the first embodiment are denoted by the same reference numerals and their description will be omitted. In addition, the machine learning device, the machine learning method, and the data processing system according to the third embodiment described below are described by taking as an example a case where they are applied to the centrifuge system shown in FIG. 1 and FIG. 2, as in the first and second embodiments. Furthermore, all of the modifications described in the embodiments described above or below are applicable to this embodiment as well to the extent that no contradiction occurs.

FIG. 11 is a schematic block diagram of a machine learning device 20B according to a third embodiment of the present disclosure. As shown in FIG. 11, the machine learning device 20B according to this embodiment is similar to the machine learning device 20 according to the first embodiment described above, except that the machine learning device 20B according to this embodiment includes a third learning data set storage unit 223 as a learning data set storage unit and a third learning unit 233 as a learning unit. In addition, the multiple data acquired in the learning data set acquisition unit 21 are different from those according to the first embodiment. As shown in FIG. 11, the learning data set acquisition unit 21 according to this embodiment is connected to a computer PC1 and acquires desired data from this computer PC1. Like the computer PC1 according to the first embodiment, this computer PC1 is connected to the control unit 30 and the treated liquid monitoring system 60, but in addition to these, it is also connected to the separated liquid monitoring system 40, the dewatered solids monitoring system 50, the treated liquid concentration sensor 67, and the screw conveyor torque monitoring system 70.

As shown in Figures 2 and 11, the separated liquid monitoring system 40 is preferably provided, for example, in the separated liquid discharge conduit 8, and monitors the solid matter concentration of the separated liquid SL passing through this separated liquid discharge conduit 8. A specific method for monitoring the separated liquid concentration may be to directly detect the concentration value of the separated liquid SL using, for example, a well-known laser, optical, or ultrasonic concentration sensor.

As shown in Figs. 2 and 11, the dehydrated solids monitoring system 50 is preferably installed in the solids discharge conduit 7, for example, and is capable of monitoring the moisture content of the dehydrated solids (also called "dehydrated cake") M passing through the solids discharge conduit 7. The dehydrated solids monitoring system 50 measures or estimates the moisture content of the dehydrated solids M passing through the solids discharge conduit 7. Various methods can be used to measure or estimate the moisture content of the dehydrated solids M, but in this embodiment, a method is used to directly measure the moisture content of the dehydrated solids M passing through the solids discharge conduit 7 using a well-known moisture content meter that uses microwaves or the like. In Fig. 2, the solids discharge conduit 7 is shown as an example of the installation location of the dehydrated solids monitoring system 50, but instead, for example, the main conveying path 7a or the solids chute 6a may be used.

The treated liquid concentration sensor 67 may be disposed, for example, at a predetermined position in the treated liquid pipe 10b, particularly before it merges with the additive pipe 11c, and may directly measure the slurry concentration of the treated liquid PL1 supplied from the treated liquid supply source 10. As with the above-mentioned separation liquid monitoring system 40, the treated liquid concentration sensor 67 may be, for example, a well-known laser type, optical type, or ultrasonic type concentration sensor.

As shown in Figures 2 and 11, the screw conveyor torque monitoring system 70 is attached to, for example, the rotating shaft of the screw conveyor 3, and monitors the torque value of the screw conveyor 3 by detecting the reaction force acting on the screw conveyor 3.

Various data acquired by these monitoring systems, specifically, image data of the floc-containing treated liquid PL2, the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 are sent directly or via the control unit 30 to the computer PC1, which then sends them to the learning dataset acquisition unit 21 together with corresponding desired control parameters. The multiple data constituting the learning dataset acquired by the learning dataset acquisition unit 21 are stored in the third learning dataset storage unit 223 in the form of a third learning dataset in which the image data of the floc-containing treated liquid PL2, the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 are third input data, and the control parameters associated with these data are third output data. Then, in the third learning unit 233, machine learning is performed using the third learning data set in a manner similar to the machine learning method of the first embodiment described above, and the obtained third learned model is stored in the learned model storage unit 24. In this embodiment, among the third input data, the concentration of the separated liquid SL, the water content of the dewatered solid matter M, the slurry concentration of the liquid to be treated PL1, and the torque value of the screw conveyor 3 are data of one value, whereas the image data of the floc-containing liquid to be treated PL2 are the brightness and/or color values of all pixels constituting the image data, and therefore the total number of data is significantly different between them. Therefore, if these input data are directly associated with the input layer, the obtained inference result will be extremely influenced by the image data of the floc-containing liquid to be treated PL2. Therefore, in this embodiment, it is particularly preferable to perform preprocessing of the data to adjust the number of data before associating the third input data with the input layer, so that the degree of influence of input data other than the image data of the floc-containing liquid to be treated PL2 is not too small. Furthermore, this preprocessing can also be performed during inference by storing the preprocessing in the trained model storage unit 24 together with the neural network model obtained as part of the trained model.

Figure 12 is a schematic block diagram showing a data processing system 80B according to a third embodiment of the present disclosure. As shown in Figure 12, the data processing system 80B according to this embodiment has the same components as the data processing system 80 according to the first embodiment described above, except that it has a third inference unit 873 as an inference unit and an additional variable acquisition unit 89.

As shown in FIG. 12, the additional variable acquisition unit 89 is connected to the separated liquid monitoring system 40, the dehydrated solids monitoring system 50, the treated liquid concentration sensor 67, and the screw conveyor torque monitoring system 70, and acquires the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 from these monitoring systems and sensors. It is preferable that the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 acquired by this additional variable acquisition unit 89 are acquired at the same time as the image data acquired by the image data acquisition unit 81 from the floc-containing treated liquid PL2 monitoring system 60.

When adjusting the control parameters in the decanter 1 to which the data processing system 80B having the additional variable acquisition unit 89 described above is applied, the same process as that shown in FIG. 8 described above may be performed. However, in the process shown in FIG. 8, the additional variable acquisition unit 89 acquires the concentration of the separated liquid SL, the water content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 at the same time as step S23 or at a predetermined timing before or after it. These four values are then associated with the input layer of the trained model in the third inference unit 873, together with the image data of the floc-containing treated liquid PL2.

In the above-mentioned third embodiment, the computer PC1 is connected to the control unit 30, the treated liquid monitoring system 60, the separated liquid monitoring system 40, the dehydrated solids monitoring system 50, the treated liquid concentration sensor 67, and the screw conveyor torque monitoring system 70, and the four values acquired from these are used as input data for the learning data set together with image data, but the present invention is not limited to this. In more detail, it is sufficient that at least one of the values acquired by the separated liquid monitoring system 40, the dehydrated solids monitoring system 50, the treated liquid concentration sensor 67, and the screw conveyor torque monitoring system 70 is used as input data for the learning data set. The same applies to the number of pieces of information acquired by the additional variable acquisition unit 89.

As described above, according to the machine learning device, machine learning method, and data processing system of this embodiment, in addition to image data of the floc-containing treated liquid PL2, at least one of the concentration of the separated liquid SL, the moisture content of the dewatered solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 is adopted as input data, making it easier to identify the correlation between the input data and the output data in the learning stage of the neural network model. This is expected to reduce the number of learning data sets required to generate each trained model capable of inference with sufficient accuracy.

<Fourth embodiment>
As described above, the second and third embodiments are both examples showing a mode in which the number of learning data sets required to generate a trained model capable of inference with sufficient accuracy can be reduced. The two embodiments can be combined. Therefore, a case in which the technical ideas shown in the second and third embodiments described above are combined will be described below as a fourth embodiment of the present disclosure. Note that, among the components of the machine learning device 20C and the data processing system 80C according to the fourth embodiment described below, those components common to the machine learning devices 20, 20A, and 20B and the data processing systems 80, 80A, and 80B according to the first to third embodiments are given the same reference numerals and their description will be omitted. In addition, the machine learning device, the machine learning method, and the data processing system according to the third embodiment described below are described by taking as an example a case in which they are applied to the centrifuge system shown in FIG. 1 and FIG. 2, as in the first to third embodiments. Furthermore, all the modified examples described in the above or later-described embodiments can be applied to this embodiment to the extent that no contradiction occurs.

FIG. 13 is a schematic block diagram of a machine learning device 20C according to a fourth embodiment of the present disclosure. As shown in FIG. 13, the machine learning device 20C according to this embodiment is similar to the machine learning device 20 according to the first embodiment, except that the components of the machine learning device 20C according to this embodiment include a first learning dataset storage unit 221 and a fourth learning dataset storage unit 224 as learning dataset storage units, and a first learning unit 231 and a fourth learning unit 234 as learning units. In relation to this, the multiple data acquired in the learning dataset acquisition unit 21 also differ from those of the first embodiment.

As shown in FIG. 13, the learning data set acquisition unit 21 according to this embodiment is connected to a computer PC1 and acquires desired data from the computer PC1. Like the computer PC1 according to the second embodiment, this computer PC1 is connected to the control unit 30 and the treated liquid monitoring system 60A, but in addition to these, it is also connected to at least one of the separated liquid monitoring system 40, the dehydrated solids monitoring system 50, the treated liquid concentration sensor 67, and the screw conveyor torque monitoring system 70. The specific configurations of the control unit 30, treated liquid monitoring system 60A, separated liquid monitoring system 40, dehydrated solids monitoring system 50, treated liquid concentration sensor 67, and screw conveyor torque monitoring system 70 connected to the computer PC1 are the same as those exemplified in the other embodiments already described above. In addition to the image data of the floc-containing treated liquid PL2 and the control parameters, the multiple data acquired by the learning data set acquisition unit 21 in this embodiment further includes at least one of the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3, as well as the feature value of the floc-containing treated liquid PL2. In this embodiment, as in the third embodiment, the computer PC1 is connected to all of the separated liquid monitoring system 40, the dehydrated solids monitoring system 50, the treated liquid concentration sensor 67, and the screw conveyor torque monitoring system 70, and the multiple data acquired by the learning data set acquisition unit 21 includes all of the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3.

The multiple data acquired in the learning data set acquisition unit 21 are stored separately as two learning data sets in the first learning data set storage unit 221 and the fourth learning data set storage unit 224 while taking into consideration their respective correspondences. The first learning data set stored in the first learning data set storage unit 221 may include image data of the floc-containing treated liquid PL2 before it is supplied to the bowl 2 and after the chemical agent is added, taken from a predetermined angle of view, as the first input data, and may include a feature amount of the floc-containing treated liquid PL2 associated with the first input data, as the first output data. The fourth learning data set stored in the fourth learning data set storage unit 224 may include a feature amount of the floc-containing treated liquid PL2, the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3, as the fourth input data, and may include a control parameter associated with the fourth input data, as the fourth output data. In this way, when dividing and generating two learning data sets from a plurality of data, for example, the image data of the floc-containing treated liquid PL2, the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, the torque value of the screw conveyor 3, the feature values and the control parameters of the floc-containing treated liquid PL2 associated with the same floc-containing treated liquid PL2 may be divided into a set of the image data of the floc-containing treated liquid PL2 and the feature values of the floc-containing treated liquid PL2, and a set of the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, the torque value of the screw conveyor 3, and the feature values and the control parameters of the floc-containing treated liquid PL2, and each may be set as a first and fourth learning data set.

The first and fourth learning data sets stored in the first learning data set storage unit 221 and the fourth learning data set storage unit 224, respectively, are referenced only by different learning units. The first learning unit 231 learns a learning model that infers the correlation between the first input data and the first output data by inputting multiple sets of the first learning data set. In other words, the first learning unit learns a first learning model that infers the feature amount of the image data by inputting image data of the floc-containing treated liquid PL2 in the first learning data set. And the fourth learning unit 234 learns a learning model that infers the correlation between the fourth input data and the fourth output data by inputting multiple sets of the fourth learning data set. In other words, by inputting the features of the image data in the fourth learning data set, the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3, a fourth learning model is trained to infer the control parameters.

The specific machine learning methods in the first and fourth learning units 231, 234 use different learning datasets, but the process is similar to the supervised learning process shown in FIG. 6. The first and fourth trained models obtained through a series of machine learning processes are then stored in the trained model storage unit 24.

Figure 14 is a schematic block diagram showing a data processing system 80C according to a fourth embodiment of the present disclosure. As shown in Figure 14, the data processing system 80C according to this embodiment has the same components as the data processing system 80 according to the first embodiment described above, except that it has a first inference unit 871 and a fourth inference unit 874 as inference units in the arithmetic unit 83, and also has an additional variable acquisition unit 89.

As shown in FIG. 14, the additional variable acquisition unit 89 is connected to the separated liquid monitoring system 40, the dewatered solids monitoring system 50, the treated liquid concentration sensor 67, and the screw conveyor torque monitoring system 70, and acquires the concentration of the separated liquid SL, the water content of the dewatered solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 from these monitoring systems and sensors. The first inference unit 871 performs inference using the first trained model generated by the above-mentioned machine learning device 20C and stored in the trained model storage unit 88. Therefore, when the image data of the floc-containing treated liquid PL2 acquired by the image data acquisition unit 81 is input, the first inference unit 871 outputs the feature amount of this image data to the output layer. The fourth inference unit 874 performs inference using the fourth trained model generated by the above-mentioned machine learning device 20C and stored in the trained model storage unit 88. Therefore, when the feature quantities of the image data inferred by the first inference unit 871, the concentration of the separated liquid SL, the moisture content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 are input, the fourth inference unit 874 outputs control parameters to the output layer.

When adjusting the control parameters in the decanter 1 to which the data processing system 80C including the above-mentioned additional variable acquisition unit 89 and the first and fourth inference units 871 and 874 is applied, the same process as that shown in FIG. 8 may be performed. However, the process shown in FIG. 8 further includes a step in which the additional variable acquisition unit 89 acquires the concentration of the separated liquid SL, the water content of the dehydrated solids M, the slurry concentration of the treated liquid PL1, and the torque value of the screw conveyor 3 at the same time as step S23 or at a predetermined timing before or after it. In addition, when inferring the control parameters in step S24, first, the image data of the floc-containing treated liquid PL2 is input to the first inference unit 871, and the feature amount of the output image data is input to the fourth inference unit 874 together with the four values acquired by the additional variable acquisition unit 89.

As described above, according to the machine learning device, machine learning method, and data processing system of this embodiment, it is expected that the number of training data sets required to generate each trained model capable of inference with sufficient accuracy can be further reduced compared to the second and third embodiments described above.

The above-described embodiment is merely an example, and the present disclosure is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present disclosure. All of these are included in the technical ideas of the present disclosure. In addition, in the present disclosure, each component may exist in one or more than one form, as long as no contradiction arises.

1 Decanter (Centrifugal Separator System)
2 Bowl 2a Solids discharge outlet 2b Separated liquid discharge outlet 2c Dam 3 Screw conveyor 3c Space 4 Drive motor 5 Differential speed generating device 6 Casing 6a Solids chute 6b Separated liquid chute 7 Solids discharge conduit 7a Main conveying path 8 Separated liquid discharge conduit 9 Supply pipe 10 Treated liquid supply source 10b Treated liquid piping 11 Additive supply source 11c Additive piping 20, 20A to C Machine learning device 21 Learning data set acquisition unit 22, 221 to 224 Learning data set storage unit 23, 231 to 234 Learning unit 24 Learned model storage unit 30 Control unit 40 Separated liquid monitoring system 50 Dewatered solids monitoring system 60, 60A Treated liquid monitoring system 67 Treated liquid concentration sensor 70 Screw conveyor torque monitoring system 80, 80A to C Data processing system 81 Image data acquisition unit 82 Parameter adjustment unit 83 Arithmetic units 87, 871 to 874 Inference unit 88 Learned model storage unit 89 Additional variable acquisition unit SL Separation liquid M Liquid-containing solid (dehydrated solid)
PL1: Liquid to be treated PL2: Floc-containing liquid to be treated

Claims (9)

A machine learning device for a centrifugation system including: a bowl that applies centrifugal force to a liquid to be treated to separate the liquid into solids and a separated liquid; a screw conveyor that transports the solids in the bowl toward a discharge outlet; a drive motor that rotates the bowl; and a differential speed generating device that rotates the screw conveyor at a differential speed relative to the bowl,
a learning dataset storage unit that stores a plurality of learning datasets each including input data including image data obtained by capturing an image of the liquid to be treated from a predetermined angle of view before the liquid is supplied to the bowl and after a predetermined additive has been added thereto, and output data including control parameters associated with the input data, the control parameters including at least one of an amount of the additive supplied, a centrifugal force of the bowl, and a differential speed controlled by the differential speed generating device;
a learning unit that learns a learning model that infers a correlation between the input data and the output data by inputting a plurality of sets of the learning data sets;
A learned model storage unit that stores the learning model learned by the learning unit;
Machine learning device. The learning unit includes:
a first learning unit that learns a first learning model that infers features of the image data by inputting the image data in the learning dataset;
a second learning unit that learns a second learning model that infers the control parameter by inputting a feature amount of the image data;
The machine learning device according to claim 1 . The input data further includes at least one of the concentration of the separated liquid, the moisture content of the liquid-containing solid matter discharged from the discharge port, the slurry concentration of the liquid to be treated, and the torque value of the screw conveyor.
The machine learning device according to claim 1 . The input data further includes at least one of the concentration of the separated liquid, the moisture content of the liquid-containing solid matter discharged from the discharge port, the slurry concentration of the liquid to be treated, and the torque value of the screw conveyor;
The learning unit includes:
a first learning unit that learns a first learning model that infers features of the image data by inputting the image data in the learning dataset;
a fourth learning unit that learns a fourth learning model that infers the control parameters by inputting the feature amount of the image data, the concentration of the separated liquid, the moisture content of the liquid-containing solid matter discharged from the discharge port, the slurry concentration of the treated liquid, and a torque value of the screw conveyor;
The machine learning device according to claim 1 . The control parameters further include a supply amount of the liquid to be treated.
The machine learning device according to claim 1 . The control parameters further comprise a dam set diameter of the bowl.
The machine learning device according to claim 1 . A data processing system for use in a centrifuge system including a bowl for applying centrifugal force to a liquid to be treated to separate the liquid into solids and a separated liquid, a screw conveyor for transporting the solids in the bowl toward a discharge outlet, a drive motor for rotating the bowl, and a differential speed generating device for rotating the screw conveyor at a differential speed relative to the bowl,
an image data acquisition unit for acquiring image data obtained by capturing an image of the liquid to be treated from a predetermined angle of view before the liquid is supplied to the bowl and after a predetermined additive is added;
an inference unit that infers a control parameter of the centrifuge system by inputting the data acquired by the image data acquisition unit into a trained model generated by the machine learning device according to claim 1 or 2;
Data processing system. A data processing system for use in a centrifuge system including a bowl for applying centrifugal force to a liquid to be treated to separate the liquid into solids and a separated liquid, a screw conveyor for transporting the solids in the bowl toward a discharge outlet, a drive motor for rotating the bowl, and a differential speed generating device for rotating the screw conveyor at a differential speed relative to the bowl,
an image data acquisition unit for acquiring image data obtained by capturing an image of the liquid to be treated from a predetermined angle of view before the liquid is supplied to the bowl and after a predetermined additive is added;
an additional variable acquisition unit for acquiring at least one of the concentration of the separated liquid, the moisture content of the liquid-containing solid matter discharged from the discharge port, the slurry concentration of the treated liquid, and the torque value of the screw conveyor;
and an inference unit that infers a control parameter of the centrifuge system by inputting the data acquired by the image data acquisition unit and the additional variable acquisition unit into a trained model generated by the machine learning device according to claim 3 or 4.
Data processing system. A machine learning method using a computer for a centrifuge system including a bowl that applies centrifugal force to a liquid to be treated to separate the liquid into solids and a separated liquid, a screw conveyor that transports the solids in the bowl toward a discharge outlet, a drive motor that rotates the bowl, and a differential speed generating device that rotates the screw conveyor at a differential speed relative to the bowl, the method comprising:
a step of storing a plurality of learning data sets including input data including image data obtained by capturing an image of the liquid to be treated from a predetermined angle of view before being supplied to the bowl and after a predetermined additive has been added thereto, and output data including control parameters associated with the input data, the control parameters including at least one of the supply amount of the additive, the centrifugal force of the bowl, and a differential speed controlled by the differential speed generating device;
learning a learning model that infers a correlation between the input data and the output data by inputting a plurality of sets of the learning data sets;
and storing the trained model.
Machine learning methods.