JP7172302B2 - Sewage treatment operation status evaluation device and sewage treatment operation status evaluation method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a sewage treatment operation status evaluation device and a sewage treatment operation status evaluation method.

2. Description of the Related Art Conventionally, purification of sewage by activated sludge treatment has been widely performed in sewage treatment facilities such as sewage treatment facilities. In activated sludge treatment, wastewater is exposed to the air by aeration, etc. in a water treatment tank that stores wastewater, and pollutants are decomposed by aerobic microorganisms. be done.

In sewage treatment facilities where activated sludge treatment is carried out, in general, when sewage treatment is being carried out smoothly and when it is being carried out in a deteriorated state rather than smoothly, It is known that the types of microorganisms that appear and dominate are different. For example, when the state of the activated sludge is good, the genus Vorticella, the genus Epistylis, the amoeba genus Arcella, the rotifer genus Philodina and the like are dominant. On the other hand, when the state of activated sludge deteriorates and in the transitional period from good to bad conditions, flagellates such as Bodo, Monas, Amoeba, and Arcella appear.

Therefore, by identifying the microorganisms in the water treatment tank and associating the microorganisms with the state of sewage treatment, it is possible to determine and evaluate whether the sewage treatment is being carried out smoothly.
For example, activated sludge in a water treatment tank is sampled and observed under a microscope to confirm the presence of microorganisms that can be used as a criterion for evaluating whether sewage treatment is proceeding smoothly. However, since these operations are basically performed manually, it takes time. In addition, workers are required to have advanced and specialized knowledge and skills cultivated through long work experience to identify microorganisms.

On the other hand, Patent Document 1 discloses a biota diagnosis support system that identifies and counts microorganisms that appear in activated sludge in an activated sludge treatment process and determines the treatment state of the process using a biota diagnosis device from this data. disclosed. In this biota diagnosis support system, image data of treated water taken with a microscope camera or an underwater camera is captured by the input unit, and the image recognition system automatically identifies microorganisms from this image data by image recognition processing based on the model-based matching method. Identify and count.
Image models of microorganisms required for recognition are registered in advance in a database. Image recognition involves extracting edges from photographed images of microorganisms, identifying the shape of microorganisms using the straight line and arc components of the edge image, registering this shape in a database, and matching it with feature data obtained from photographed images of treated water. Identify the microorganism by

JP-A-8-197084

A biota diagnosis support system such as Patent Document 1 is based on a classical image processing technique such as edge extraction. In other words, it is necessary to effectively extract the characteristics of microorganisms when creating an image model registered in advance in a database. The task of extracting this feature requires advanced knowledge of image processing.
In addition, in order to enable the identification and identification of microorganisms by the biota diagnosis support system, it is necessary to intentionally extract the characteristics of microorganisms so that the characteristics of microorganisms are expressed remarkably. Requires highly specialized knowledge.
In addition, it takes time and effort to extract features unique to each microorganism to the extent that they can be effectively identified and distinguished by the biota diagnosis support system for each image of many microorganisms.
As described above, the biota diagnosis support system as disclosed in Patent Literature 1 requires specialized knowledge and takes time and effort to prepare in advance, so it is not easy to introduce.
Furthermore, since the shape of microorganisms changes due to movement, it is inherently difficult to identify and distinguish microorganisms with high accuracy. The reason why this can be accurately performed by the biota diagnosis support system based on the above-mentioned classical image processing technology is that, for example, the characteristics of microorganisms are not expressed depending on the photographed image used, and it is difficult to identify microorganisms. It is not easy because it is not easy, and in fact, it has not been put into practical use at present.

The problem to be solved by the present invention is to provide a sewage treatment operation status evaluation device and a sewage treatment operation status evaluation method that are easy to introduce and can evaluate the operation status of sewage treatment by accurately identifying microorganisms. It is to be.

In order to solve the above problems, the present invention employs the following means. That is, the present invention is intended for sewage treatment facilities that purify sewage stored in a water treatment tank with microorganisms, and analyzes an input image obtained by imaging the microorganisms living in the water treatment tank, and analyzes the image of the microorganisms. A sewage treatment operation status evaluation device for calculating whether the device belongs to a classification category and evaluating the operation status of sewage treatment, comprising a machine learning device, and inputting learning data including a learning image related to the microorganism to the machine learning device. a microorganism feature learning unit that performs machine learning to generate an identification parameter related to identification of the microorganism; a microorganism identification unit that calculates the identification category to which the microorganism in the input image belongs using the identification parameter; and an operation condition evaluation unit that evaluates the operation condition of the sewage treatment based on the identification result of the microorganism identification unit.

Further, the present invention is directed to sewage treatment facilities that purify sewage stored in a water treatment tank with microorganisms, and analyzes an input image obtained by imaging the microorganisms living in the water treatment tank, and determines which microorganisms are present. A sewage treatment operation status evaluation method for calculating whether the microorganism belongs to the identification category and evaluating the sewage treatment operation status, wherein learning data including a learning image related to the microorganism is input to a machine learning device to perform machine learning, generating an identification parameter for identifying the microorganism, using the identification parameter to calculate the identification category to which the microorganism belongs in the input image, and evaluating the operation status of the wastewater treatment based on the identification result; A treatment operating condition evaluation method is provided.

According to the present invention, it is possible to provide a sewage treatment operation status evaluation device and a sewage treatment operation status evaluation method that are easy to introduce and can evaluate the operation status of sewage treatment by accurately identifying microorganisms. .

1 is a block diagram of a sewage treatment facility according to a first embodiment of the present invention; FIG. It is a block diagram of the control device of the sewage treatment operation status evaluation device provided in the sewage treatment facility. 4 is a block diagram of a machine learning device that implements the learning processing section of the control device; FIG. It is a flowchart at the time of learning of the sewage treatment operation|movement situation evaluation method in the said 1st Embodiment. It is a flow chart at the time of sewage treatment operation state evaluation of the sewage treatment operation state evaluation method. It is a block diagram of the machine learning device which implement|achieves the learning process part of a control apparatus of the sewage treatment operation|movement condition evaluation apparatus in the modification of the said 1st Embodiment. It is a block diagram of the control device of the sewage treatment operating condition evaluation device in the second embodiment of the present invention. It is a block diagram of the machine learning device which implement|achieves the learning process part of the control apparatus of the sewage treatment operation|movement condition evaluation apparatus in the said 2nd Embodiment. It is a flowchart at the time of learning of the sewage treatment operation|movement situation evaluation method in the said 2nd Embodiment. It is a flow chart at the time of sewage treatment operation situation evaluation of the sewage treatment operation situation evaluation method in the said 1st Embodiment.

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
The sewage treatment operating status evaluation apparatus in the present embodiment is intended for a sewage treatment facility that purifies sewage stored in a water treatment tank with microorganisms, and analyzes an input image of microorganisms living in the water treatment tank. It calculates which identification category a microorganism belongs to and evaluates the operation status of sewage treatment, is equipped with a machine learning device, and inputs learning data including learning images related to microorganisms to the machine learning device to perform machine learning. a microorganism feature learning unit that generates identification parameters for identifying microorganisms; a microorganism identification unit that calculates identification categories to which microorganisms in an input image belong using the identification parameters; and an operating condition evaluation unit that evaluates the operating condition of the process.

[First embodiment]
FIG. 1 is a block diagram of a sewage treatment facility in the first embodiment. The sewage treatment facility 1 purifies sewage stored in a water treatment tank using microorganisms.
The sewage treatment facility 1 includes a primary sedimentation tank 2, a biological treatment reaction tank (water treatment tank) 3, a reaction tank measuring instrument 4, a final sedimentation tank 5, an air diffuser plate 6, a blower 7, an aeration control valve 8, and a return sludge pump 9. , a sludge extraction pump 10, a sewage treatment operation status evaluation device 20, and first to fifth pipes L1, L2, L3, L4, and L5.

Wastewater containing organic matter is introduced into the primary sedimentation tank 2 . In the primary sedimentation tank 2, small dust, sand, etc. are removed from the introduced sewage, and rough solid-liquid separation is performed. The sewage from which dust and the like have been removed is sent to the biological treatment reaction tank 3 through the first pipe L1.

In the biological treatment reaction tank 3, sewage is biologically treated and purified by microorganisms. At the time of purification, sewage is exposed to the air by aeration or the like, and aerobic microorganisms decompose organic matter and proliferate along with assimilating the organic matter to form activated sludge. The activated sludge treated water is sent to the final sedimentation tank 5 via the second pipe L2.
Air is supplied from the blower 7 to the lower part of the biological treatment reactor 3 through the third pipe L3. An aeration control valve 8 is provided between the blower 7 and the biological treatment reactor 3 of the third pipe L3. Opening and closing the aeration control valve 8 changes the amount of aeration, thereby adjusting the amount of dissolved oxygen in the biological treatment reaction tank 3 and controlling the progress of biological treatment by microorganisms. A part of the biological treatment reactor 3 to which air is supplied by the third pipe L3 is provided with an air diffusion plate 6 for increasing the efficiency of dissolving oxygen.

In the final sedimentation tank 5, the activated sludge contained in the activated sludge-treated water sent from the biological treatment reaction tank 3 is precipitated to purify sewage. The supernatant from which the activated sludge has been separated in the final sedimentation tank 5 is discharged out of the system as treated water.
A part of the sludge settled in the final sedimentation tank 5 is returned to the biological treatment reaction tank 3 through the fourth pipe L4 by the return sludge pump 9 and reused for activated sludge treatment. The remaining sludge is discharged as surplus sludge by the sludge extraction pump 10 through the fifth pipe L5, sent to a sludge treatment facility or the like, and incinerated after water is removed.

The sewage treatment operation status evaluation device 20 in the first embodiment is intended for the sewage treatment facility 1 that purifies the sewage stored in the biological treatment reaction tank 3 as described above using microorganisms.
In the sewage treatment facility 1 as described above, when sewage treatment is performed smoothly and when it is performed in a deteriorated state rather than smoothly, it appears in the biological treatment reaction tank 3 and becomes dominant. different types of microorganisms may be used. For example, when the state of the activated sludge is good, the genus Vorticella, the genus Epistylis, the amoeba genus Arcella, the rotifer genus Philodina and the like are dominant. On the other hand, when the state of activated sludge deteriorates and in the transitional period from good to bad conditions, flagellates such as Bodo, Monas, Amoeba, and Arcella appear.

The sewage treatment operation status evaluation device 20 in the first embodiment identifies microorganisms in the biological treatment reaction tank 3 and associates the microorganisms with the state of sewage treatment to determine whether sewage treatment is being performed smoothly. Judge and evaluate. More specifically, the sewage treatment operation status evaluation device 20 analyzes an input image of microorganisms inhabiting the biological treatment reaction tank 3, calculates which identification category the microorganisms belong to, and evaluates the sewage treatment operation status. Evaluate.
That is, the sewage treatment operation status evaluation device 20 analyzes an input image of microorganisms to identify the microorganisms, and thus identifies the microorganisms depending on their appearance characteristics.

Here, as an identification classification for the sewage treatment operation status evaluation device 20 to identify and classify microorganisms, each microorganism belonging to one class, for example, one "genus" in the biological classification has the same characteristics in the shape of these microorganisms. In addition, when it has a common effect on activated sludge, that is, it has either good or bad effects, the "genus" may correspond to one identification category.
In addition, for example, in each of the microorganisms belonging to one "genus", the characteristics of the shape of these microorganisms are roughly divided into a plurality of types, and in each of the plurality of types, each microorganism belongs to the activated sludge situation. In the case of common, ie, good or bad effects, each of the broadly classified types may correspond to one identification division.
Alternatively, if there are a plurality of microorganisms that have common shape features and actions on activated sludge across the classification by "genus", they may be collectively classified as one identification category.

That is, the sewage treatment operation status evaluation device 20 in the first embodiment is configured so that a plurality of microorganisms having the same characteristics in shape and having a common action on activated sludge belong to one identification category. is divided and classified into a plurality of identification categories, and when evaluating the operation status of sewage treatment, it is calculated to which identification category the microorganisms captured in the input image belong.

Details of the sewage treatment operation status evaluation device 20 will be described below. The sewage treatment operating condition evaluation device 20 includes an imaging device 21 and a control device 22 .

In the first embodiment, the imaging device 21 is, for example, an underwater camera provided inside the biological treatment reaction tank 3, and takes enlarged images of the activated sludge inside the biological treatment reaction tank 3. The imaging device 21 is not limited to this, and may be a system that automatically collects the activated sludge in the imaging device 21, enlarges it with a microscope, and takes an image, for example.

The control device 22 is, for example, an information processing device such as a personal computer. FIG. 2 is a block diagram of the controller 22. As shown in FIG. The control device 22 includes a microorganism imaging unit 23 , an image storage unit 24 , a microorganism feature learning unit 25 , a judgment image acquisition unit 26 , a microorganism identification unit 27 and an operating condition evaluation unit 28 . The microbial feature learning unit 25 includes a learning image acquisition unit 30 , a labeling unit 31 , a learning processing unit 32 and a learning parameter storage unit 33 .
Among the components of the control device 22, the microorganism imaging unit 23, the determination image acquisition unit 26, the microorganism identification unit 27, the driving situation evaluation unit 28, the learning image acquisition unit 30, the labeling unit 31, and the learning processing unit 32 are, for example, It may be software or a program executed by the CPU in the information processing apparatus. Also, the image storage unit 24 and the learning parameter storage unit 33 may be realized by a storage device such as a semiconductor memory or a magnetic disk provided inside or outside the information processing apparatus.

As will be described later, the microorganism identification unit 27 calculates identification categories to which the microorganisms in the input image input to the control device 22 belong, and the operation condition evaluation unit 28 evaluates the operation condition of sewage treatment based on this. . In order to effectively identify the microorganisms and evaluate the operation status of the sewage treatment, the microorganism feature learning unit 25 is equipped with a machine learning device as will be described later. is input to a machine learning machine to perform machine learning to generate identification parameters for identifying microorganisms.
That is, the control device 22 can be roughly divided into two types of operations: learning of identification categories of microorganisms and evaluation of operating conditions. To simplify the explanation, first, each component of the control device 22 during learning of the identification category will be described, and then the behavior of each component during evaluation of the driving situation will be described.

Microorganism imaging unit 23 controls imaging device 21 to image activated sludge in biological treatment reaction tank 3 or automatically collected from biological treatment reaction tank 3 . Microorganism photographing unit 23 further cuts out, for example, a portion in which some object is photographed from the photographed image, enlarges it to an appropriate size, and transmits it to image storage unit 24 as a learning image.
Microorganism imaging unit 23 performs imaging and transmission to image storage unit 24 until the number of these learning images reaches the number required for proper learning of the machine learning device.

The image accumulation unit 24 accumulates and stores the plurality of learning images transmitted from the microorganism imaging unit 23 .

The learning image acquisition unit 30 of the microorganism feature learning unit 25 acquires the learning images from the image storage unit 24 and transmits them to the labeling unit 31 .

The labeling unit 31 receives the learning images transmitted from the learning image acquisition unit 30 .
The labeling unit 31 sequentially displays the learning images on a display device (not shown). The operator views the learning image displayed on the display device, and if the image is a microorganism, associates the label corresponding to the identification section corresponding to the microorganism with the learning image.
The learning image associated with the identification segment label becomes learning data for a machine learning device to be described later. That is, the machine learning device in the first embodiment performs supervised learning by inputting, as teacher data, learning images related to microorganisms and information on identification categories of microorganisms corresponding to the learning images.
The labeling unit 31 transmits learning data to the learning processing unit 32 .

The learning processing unit 32 receives learning data from the labeling unit 31 .
The learning processing unit 32 comprises a machine learner 40, shown as a block diagram in FIG. The machine learning device 40 in the first embodiment performs convolutional deep learning, and is realized particularly by a convolutional neural network. The machine learner 40 generates a trained model 40A trained with appropriate learning parameters, which is used as a program module that is part of artificial intelligence software.
The machine learning device 40 includes a convolution processing section 41 and a full coupling section 42 . The convolution processing unit 41 includes a first convolution layer 44 , a second convolution layer 45 and a third convolution layer 46 . The first convolutional layer 44 comprises an input layer 47 and an output layer 48 .

The learning processing unit 32 inputs the learning image 43 to the machine learning device 40 . In FIG. 3, the training image 43 is shown as an image with three R, G, B channels 43R, 43G, 43B.
The first convolutional layer 44 comprises a predetermined number of first filters, not shown. For each first filter, the machine learning device 40 positions it on the learning image 43, and for each pixel value of each pixel of the learning image 43 of the first filter, finds a pixel position in the first filter corresponding to the pixel position. Convolution filtering is performed by calculating the sum with the set weights. Thereby, the pixel value of one pixel in the first convolutional layer 44 is calculated. The machine learning device 40 calculates a plurality of pixel values by performing such convolution filtering while moving the first filter on the learning image 43 in predetermined resolution increments. generates a single image corresponding to This image is referred to as the first feature map 44m because the feature corresponding to the first filter is extracted by the applied first filter.
The machine learning device 40 executes this process for all first filters to generate first feature maps 44m corresponding to the number of first filters.

The first filter actually expresses the enhancement, smoothing, etc. of the pixel values of the learning image 43 as weights. In the first feature map 44m generated by executing the convolution filtering process using such a first filter, the density pattern of the image such as the edge feature is effectively extracted. In addition, since the features are extracted from the local area of the learning image 43 through the first filter, it is robust against displacement of the object existing in the learning image 43 .

In the first convolutional layer 44, pooling processing is performed after convolution filtering. More specifically, each first feature map 44m is divided into, for example, a plurality of small regions having a size of 2×2, and the maximum value of the pixel values in each small region is calculated to obtain one By using the pixel values of pixels, each of the 2×2 small regions of each first feature map 44m is converted into 1×1 pixels, thereby reducing the information. That is, in the first embodiment, the pooling process is the maximum pooling process. Since the pooling process selects the maximum pixel value from the local area of each first feature map 44m as described above, it is possible to leave only appropriate features specialized for the image, and subsequent processes can be performed. can be done efficiently.

The pooled first feature map 44m generated in the first convolutional layer 44 becomes the input image of the second convolutional layer 45 .

In the second convolution layer 45, similarly to the first convolution layer 44, convolution filter processing and pooling processing are sequentially performed.
Like the first convolution layer 44, the second convolution layer 45 includes a predetermined number of second filters. A predetermined number of pooled second feature maps 45m are generated according to the number of filters.

Also in the third convolution layer 46, the convolution filtering process and the pooling process are sequentially executed.
The third convolution layer 46 also has a predetermined number of third filters, and performs convolution filtering using these filters, and further performs pooling processing to obtain a predetermined number of filters according to the number of the third filters. , to generate a pooled third feature map 46m.
The weights of each first, second and third filter are adjusted by machine learning.
The pooled pixel value information of the third feature map 46 m is input to the input layer 47 .

The input layer 47 has a predetermined number of input nodes 47n. Each input node 47n is connected to all pixel values of the pooled third feature map 46m and has a connection weight between layers. In the input layer 47, based on this connection weight, a weighted sum is calculated for each pixel value information of the pooled third feature map 46m, and an output function such as ReLU (Rectified Liner Unit) is applied to the result. The applied value is stored in each input node 47n.

The machine learning device 40 in the first embodiment learns to which of a plurality of classification categories the microorganisms in the input image belong, and the trained model 40A in which appropriate learning parameters have been learned by machine learning is It performs multi-class classification. Therefore, if the number of identification divisions of microorganisms is N, the output layer 48 is configured to have N output nodes 48n. This value N can be set, for example, for all types of microorganisms that are likely to appear within one year in the sewage treatment plant 1 in which the sewage treatment operational status evaluation device 20 is installed.
Each output node 48n is coupled to all input nodes 47n of the input layer 47 with a connection weight between the layers. In the output layer 48, a weighted sum is calculated for the input node 47n based on this connection weight, and a value obtained by applying the output function to the result is stored in the output node 48n.

In the machine learning device 40, the learning image 43 input to the first convolution layer 44 is processed by the convolution processing unit 41 and the full connection unit 42 as described above, and the processing result is stored in the output node 48n.
In the first embodiment, the value of the output node 48n corresponding to the identification section to which the microorganisms in the input image belong is a predetermined first judgment value, for example close to 1, and the value of the other output node 48n is a predetermined value. It is designed to be a second judgment value, for example, a value close to zero.

In this case, the label corresponding to the learning image 43 input to the machine learning device 40, that is, the information of the identification division to which the microorganism belongs, is used as the correct value when the machine learning device 40 learns using the learning image 43. used. More specifically, the output node 48n corresponding to the identification section to which the microorganism in the learning image 43 belongs should originally output the first determination value, and the other output nodes 48n should output the second determination value. A squared error between these values and the values actually output from the machine learning device 40 is used as a cost function.
Then, by adjusting the weight value of each filter of the convolution processing unit 41, the value of each connection weight of the all-connection unit 42, etc. by the error back propagation method or the like so as to reduce this cost function, the machine The learner 40 is machine-learned.
The learning processing unit 32 stores the weight value of each filter of the convolution processing unit 41, the value of each connection weight of the all-connection unit 42, and the like at the time when the machine learning is completed as learning parameters in the learning parameter storage unit 33. Send to

The learning parameter storage unit 33 receives learning parameters from the learning processing unit 32 and accumulates and stores them.

As described above, the machine learning device 40 performs machine learning by reflecting the characteristics of microorganisms in the learning parameters and generating the trained model 40A. The learning parameters stored in the learning processing unit 32 are used as identification parameters for identification of microorganisms in the evaluation of driving conditions, which will be described below. More specifically, in the evaluation of the driving situation, the learned model 40A having the learning parameters generated by the learning processing unit 32 is executed as a program on the CPU, for example, so that the microorganisms in the image to be evaluated Identify the identification category to which the belongs.
In this way, the microorganism feature learning unit 25 inputs learning data to the machine learning device 40 to perform machine learning and generate identification parameters for identifying microorganisms.

Next, the behavior of each component at the time of evaluation of the driving situation in the control device 22 will be described.

The microorganism imaging unit 23 controls the imaging device 21 to image the activated sludge in the biological treatment reaction tank 3 or automatically collected from the biological treatment reaction tank 3 to generate an image of microorganisms, and input It is transmitted to the image storage unit 24 as an image.
Microorganism imaging unit 23 performs imaging and transmission to image storage unit 24 until the number of input images reaches the number required for evaluating the driving situation.

The image accumulation unit 24 accumulates and stores a plurality of input images transmitted from the microorganism imaging unit 23 .

The determination image acquisition unit 26 acquires the input image from the image storage unit 24 and transmits it to the microorganism identification unit 27 .

The microorganism identification section 27 receives the input image transmitted from the determination image acquisition section 26 .
The microorganism identifying unit 27 further acquires learning parameters from the learning parameter storage unit 33, that is, identification parameters relating to identification of microorganisms.
The microorganism identification unit 27 uses the identification parameters to calculate the identification category to which the microorganisms in the input image belong. More specifically, the microorganism identification unit 27 uses the machine learning unit 40 in which the values of the weights of the filters of the convolution processing unit 41, the values of the connection weights of the all-connection unit 42, and the like are learned values as described above. That is, the learned model 40A is executed as a program on a CPU, for example, and an input image is input to calculate the identification segment.

That is, as shown in FIG. 3, when the input image 49 is input to the first convolution layer 44 of the trained model 40A, the first convolution layer 44 sequentially performs convolution filtering and pooling. Subsequently, the second and third convolution layers 45, 46 perform convolution filtering and pooling, respectively, to generate a third pooled feature map 46m.
Furthermore, the pooled third feature map 46m is used as an input, and the weighted sum calculation is executed in the all-connector 42, and finally the processing result is stored in each output node 48n.

For each output node 48n, as already explained, the value of the output node 48n corresponding to the identification section to which the microorganism in the input image belongs is close to the predetermined first judgment value, for example 1, and the other output node 48n is designed to be a predetermined second judgment value, for example, a value close to zero. Therefore, for example, the identification section corresponding to the output node 48n closest to the first determination value is identified and specified as belonging to the microorganism in the input image 49 that has been input.
In this way, the microorganism identification unit 27 inputs the input image 49 to the trained model 40A, and calculates identification categories to which microorganisms in the input image 49 belong.
The microorganism identification unit 27 calculates identification categories corresponding to microorganisms for all input images 49 and transmits all calculated identification categories to the driving situation evaluation unit 28 .

The operating condition evaluation unit 28 receives all calculated identification categories from the microorganism identification unit 27 .
The operation status evaluation unit 28 evaluates the operation status of sewage treatment based on the identification result of the microorganism identification unit 27 . More specifically, all the received identification divisions are statistically analyzed, for example, microorganisms that dominate when the state of activated sludge is good, such as the genus Vorticella, the genus Epistylis, the amoeba genus Arcella, and the rotifer Philodina genus is equal to or greater than a predetermined first threshold value, the driving condition is evaluated as good. In addition, for example, when the state of activated sludge such as flagellate genus Bodo, Monas and Amoeba is deteriorated, and when the state of activated sludge is deteriorated, and microorganisms that appear in the transitional period from good to bad conditions, identification categories corresponding to microorganisms such as the genus Arcella is equal to or greater than a predetermined second threshold value, it is evaluated that the driving condition is deteriorating.
In this way, the operation status evaluation unit 28 evaluates the operation status of sewage treatment based on the ratio of microorganisms for each identification category calculated by the microorganism identification unit 27 for a plurality of input images 49, that is, the appearance rate of microorganisms. do.

Next, referring to FIGS. 1 to 3, 4, and 5, a sewage treatment operation status evaluation method for evaluating the operation status using the sewage treatment operation status evaluation device 20 will be described. FIG. 4 is a flow chart of the sewage treatment operation status evaluation method during learning. FIG. 5 is a flowchart of the sewage treatment operation status evaluation method of the sewage treatment operation status evaluation method.
This sewage treatment operation status evaluation method is intended for a sewage treatment facility 1 that purifies sewage stored in a water treatment tank with microorganisms. It calculates to which identification category it belongs and evaluates the operation status of sewage treatment. Learning data including learning images related to microorganisms is input to a machine learning device 40 to perform machine learning, and identification related to microorganism identification. A parameter is generated, the identification parameter is used to calculate the identification section to which the microorganisms in the input image belong, and the operation status of the sewage treatment is evaluated based on the identification result.
First, with reference to FIG. 4, the operation during learning of the identification segment will be described.

Microorganism imaging unit 23 controls imaging device 21 to image activated sludge (step S1). Microorganism photographing unit 23 further cuts out, for example, a portion in which some object is photographed from the photographed image, enlarges it to an appropriate size, and transmits it to image storage unit 24 as a learning image.
The image accumulation unit 24 accumulates and stores the plurality of learning images transmitted from the microorganism imaging unit 23 .

The learning image acquisition unit 30 of the microorganism feature learning unit 25 acquires the learning images from the image storage unit 24 and transmits them to the labeling unit 31 .
The labeling unit 31 receives the learning images transmitted from the learning image acquisition unit 30 .
The labeling unit 31 sequentially displays the learning images on a display device (not shown). The operator browses the learning image displayed on the display device, and if it is a microorganism, associates the label corresponding to the identification section corresponding to the microorganism with the learning image (step S3).
The labeling unit 31 transmits learning data to the learning processing unit 32 .

The learning processing unit 32 receives learning data from the labeling unit 31 .
The learning processing unit 32 inputs the learning image 43 to the machine learning device 40 . In the machine learning device 40, the input learning image 43 is processed in the convolution processing section 41 and the full connection section 42, and the processing result is stored in the output node 48n.
By adjusting the weight value of each filter of the convolution processing unit 41, the value of each connection weight of the total coupling unit 42, etc. by the error back propagation method or the like so as to reduce the cost function for this processing result, Machine learning is performed by the machine learning device 40 (step S5).
The learning processing unit 32 stores the weight value of each filter of the convolution processing unit 41, the value of each connection weight of the all-connection unit 42, and the like at the time when the machine learning is completed as learning parameters in the learning parameter storage unit 33. Send to

The learning parameter storage unit 33 receives the learning parameters from the learning processing unit 32, and accumulates and stores them (step S7).

Next, with reference to FIG. 5, the operation during evaluation of the driving situation will be described.

Microorganism imaging unit 23 controls imaging device 21 to image activated sludge to generate an image of microorganisms, and transmits the generated image to image storage unit 24 as an input image.
The image accumulation unit 24 accumulates and stores a plurality of input images transmitted from the microorganism imaging unit 23 .
The determination image acquisition unit 26 acquires the input image 49 from the image storage unit 24 and transmits it to the microorganism identification unit 27 (step S11).

The microorganism identification section 27 receives the input image 49 transmitted from the judgment image acquisition section 26 .
The microorganism identifying unit 27 further acquires learning parameters from the learning parameter storage unit 33, that is, identification parameters relating to identification of microorganisms.
The microorganism identification unit 27 uses the identification parameters to calculate identification categories to which the microorganisms in the input image 49 belong (step S13). More specifically, the microorganism identification unit 27 uses the machine learning unit 40 in which the values of the weights of the filters of the convolution processing unit 41, the values of the connection weights of the all-connection unit 42, and the like are learned values as described above. That is, the learned model 40A is executed as a program on a CPU, for example, and the input image 49 is input to calculate the identification segment.
The microorganism identification unit 27 calculates identification categories corresponding to microorganisms for all input images 49 and transmits all calculated identification categories to the driving situation evaluation unit 28 .

The operating condition evaluation unit 28 receives all calculated identification categories from the microorganism identification unit 27 .
The operation status evaluation unit 28 evaluates the operation status of sewage treatment based on the identification result of the microorganism identification unit 27 (step S15).

Next, the effects of the sewage treatment operating condition evaluation device and the sewage treatment operation condition evaluation method will be described.

The sewage treatment operation status evaluation device 20 in the first embodiment is targeted for the sewage treatment facility 1 that purifies the sewage stored in the biological treatment reaction tank (water treatment tank) 3 with microorganisms. A sewage treatment operation status evaluation device 20 that analyzes an input image 49 that captures microorganisms inhabiting a sewage treatment system, calculates which identification category the microorganisms belong to, and evaluates the operation status of sewage treatment, and is a machine learning device 40 A microorganism feature learning unit 25 that inputs learning data including a learning image 43 related to microorganisms to a machine learning device 40 to perform machine learning and generates identification parameters for identifying microorganisms, and an input image using the identification parameters A microorganism identification unit 27 that calculates identification categories to which microorganisms in 49 belong, and an operation status evaluation unit 28 that evaluates the operation status of sewage treatment based on the identification results of the microorganism identification unit 27 are provided.
According to the above configuration, the microorganism feature learning unit 25 performs machine learning by inputting learning data including the learning image 43 related to microorganisms to the machine learning device 40, so that the feature of microorganisms in the image can be effectively recognized. It is possible to extract and automatically generate identification parameters that reflect this. For this reason, when introducing the sewage treatment operating condition evaluation device 20, advanced knowledge of image processing is not particularly required.
In addition, since it is not necessary for the operator to intentionally extract the features so that the unique features of the microorganisms can be conspicuously expressed, highly specialized knowledge of the microorganisms is not particularly required. In the first embodiment, the operator is required to associate labels with microorganisms, but does not require particularly advanced knowledge required for extracting characteristics of microorganisms.
Therefore, introduction of the sewage treatment operating condition evaluation device 20 is easy.
Also, the accuracy of the identification parameter can be easily improved by preparing a large amount of learning data to be input to the machine learning device 40 .
Furthermore, the microorganism identification unit 27 uses the identification parameter generated as described above when calculating the identification category to which the microorganism belongs. This identification parameter reflects the features of the appearance of microorganisms abstracted by the machine learning device 40 . Therefore, when calculating the identification category to which the microorganism belongs, it is possible to suppress the influence of the imaged angle, size, color, and brightness depending on the weather and time of the microorganism.
Therefore, it is possible to improve the calculation accuracy of the identification category to which microorganisms belong, that is, the accuracy of evaluation of the operation status of sewage treatment.

Further, the machine learning device 40 performs convolutional deep learning. Particularly, in the first embodiment, the machine learning device 40 is implemented by a convolutional neural network.
According to the configuration as described above, the sewage treatment operation status evaluation device 20 can be properly implemented.

In addition to the learning image 43, the machine learning device 40 inputs information on the classification of microorganisms corresponding to the learning image 43 as teacher data, reflects the characteristics of the microorganisms in the learning parameters, and generates a trained model 40A. and the discriminating parameter is the learning parameter.
Furthermore, the microorganism identification unit 27 inputs the input image 49 to the learned model 40A and calculates identification categories to which microorganisms in the input image 49 belong.
For example, in a conventional system based on a classical image processing technique, the brightness, image quality, etc. of a captured image that is input to the system greatly depends on the environment of the sewage treatment facility 1 . Therefore, when evaluating the driving situation, it may be necessary to appropriately adjust various parameters in image processing according to the imaging environment.
On the other hand, according to the configuration as described above, the machine learning device 40 executes machine learning to generate the learned model 40A, and the microorganism identification unit 27 has the learned model 40A in which the learning parameter, that is, the identification parameter is reflected. input image 49 to calculate the identification section to which the microorganism in the input image 49 belongs. That is, the input image 49 used in the evaluation of the driving situation is processed by the learned model 40A that reflects the characteristics of the microorganisms, so that the characteristics are efficiently extracted, so that the influence of the imaging environment can be reduced. can be done. Therefore, operation of the sewage treatment operating condition evaluation device 20 is facilitated.

In addition, the machine learning device 40 learns to which of a plurality of identification categories the microorganisms in the learning image 43 belong.
According to the configuration as described above, it is possible to calculate the identification category to which microorganisms belong using only one machine learning device 40, so the number of machine learning devices 40 provided in the sewage treatment operation status evaluation device 20 can be reduced. . That is, since the number of machine learning devices 40 to be machine-learned is reduced, introduction of the sewage treatment operating condition evaluation device 20 is easy.

The operation status evaluation unit 28 also evaluates the operation status of sewage treatment based on the appearance rate of microorganisms for each identification category calculated by the microorganism identification unit 27 for the plurality of input images 49 .
According to the configuration as described above, it is possible to evaluate the operation status of sewage treatment with high accuracy.

In particular, the sewage treatment operation status evaluation device 20 as described above is effective when there are many types of microorganisms. As the number of types of microorganisms increases, even similar microorganisms may have different effects on operating conditions. This may result in an erroneous assessment of the driving situation. In addition, in the conventional system, workers must extract features according to the type of microorganism, so the labor required for introduction increases in proportion to the type of microorganism.
On the other hand, in the sewage treatment operation status evaluation device 20, even when the number of types of microorganisms increases, the machine learning device 40 learns to accurately identify the characteristics of microorganisms and to identify and calculate for each identification category. It is possible to suppress erroneous evaluation of the driving situation, and since the feature extraction itself is automatically performed once the learning data is created, the increase in introduction cost is minimal. In this way, each of the effects already described with respect to the sewage treatment operational status evaluation device 20 can be more pronounced when there are many types of microorganisms compared to conventional systems.

[Modification of First Embodiment]
Next, with reference to FIG. 6, a modification of the sewage treatment operating condition evaluation device and the sewage treatment operation condition evaluation method shown as the first embodiment will be described. FIG. 6 is a block diagram of the machine learning device 50 included in the learning processing unit of the sewage treatment operation status evaluation device in this modified example. The sewage treatment operating condition evaluation device in this modification differs from the sewage treatment operation condition evaluation device 20 of the first embodiment in the configuration of the machine learning device 50 .

In the first embodiment, the learning processing unit 32 has only one machine learning device 40 that learns N class classification when the number of classification categories of microorganisms is N. The machine learning device 50 included in the learning processing unit in this modification includes N machine learning devices 50-1 to 50-N.
Each of the machine learning units 50-1 to 50-N corresponds to each identification division of microorganisms, and learns whether or not the microorganism in the input image belongs to the corresponding identification division. That is, each trained model in which appropriate learning parameters have been learned by machine learning in this modified example performs two-class classification.

More specifically, the output layer 58 of each of the machine learners 50-1 to 50-N has two output nodes 58a, 58b. Each of machine learning devices 50-1 to 50-N determines that the value of one output node 58a is close to a predetermined first judgment value, for example, 1, when there is no microorganism in the corresponding identification section in the input image, so that the value of the other output node 58b becomes a predetermined second judgment value, for example, a value close to 0, and when the microorganisms are present, the value of the output node 58a is close to the second judgment value and the value of the output node 58b is designed to be close to the first judgment value.
For each of the machine learning devices 50-1 to 50-N, training images related to microorganisms and information as to whether or not the microorganisms in the learning images belong to the identification section corresponding to the machine learning device are provided as teacher data. , and supervised learning is performed.
The learning parameter storage unit stores the weight values of the filters of the convolution processing unit 51 and the connection weights of the total coupling unit 52 of each of the plurality of machine learners 50-1 to 50-N, which are machine-learned in this way. Values and the like are accumulated and stored as learning parameters.

The microorganism identification unit inputs each of the input images received from the determination image acquisition unit to all the learned models corresponding to each of the machine learning devices 50-1 to 50-N, and the corresponding By aggregating the determination results of the presence or absence of microorganisms belonging to the identification section, the identification section to which the microorganism in the input image belongs is calculated.

It goes without saying that this modified example has the same effect as the already described first embodiment.
In particular, in this modification, machine learning devices 50-1 to 50-N are provided corresponding to each classification of microorganisms, and each of the machine learning devices 50-1 to 50-N determines whether the microorganism in the learning image is It learns whether it belongs to the corresponding identification category.
According to the configuration as described above, each of the machine learners 50-1 to 50-N is machine-learned specifically for the corresponding identification classification, so that the identification classification of microorganisms can be calculated more accurately. .
Further, for example, when an image in which a plurality of microorganisms are captured is input as an input image, in the sewage treatment operation status evaluation device 20 of the first embodiment, the learned model 40A determines the classification classification to which the microorganisms belong. As a result of hesitation, a situation may arise in which a plurality of output nodes 48n corresponding to each of the identification divisions to which these plurality of microorganisms belong output intermediate values, for example, between the first judgment value and the second judgment value. On the other hand, in the configuration of this modified example, each trained model operates to detect only microorganisms belonging to the corresponding classification division. The output nodes 58b of the trained models 50-1 to 50-N corresponding to the segment output values close to the first judgment value. Therefore, even when a plurality of microorganisms are imaged in the input image, it is possible to accurately calculate and identify the identification section corresponding to each microorganism.

[Second embodiment]
Next, the sewage treatment operating condition evaluation device 60 in the second embodiment will be described. In the second embodiment, the same reference numerals as in the first embodiment are used in the drawing, and the description thereof is omitted.
The sewage treatment operating condition evaluation device 60 is used in the sewage treatment plant 1 described with reference to FIG. 1, like the sewage treatment operation condition evaluation device 20 in the first embodiment.
The sewage treatment operating condition evaluation device 60 includes an imaging device 21 and a control device 62 .
The imaging device 21 is similar to that of the first embodiment.

The control device 62 is, for example, an information processing device such as a personal computer. FIG. 7 is a block diagram of the controller 62. As shown in FIG. The control device 62 includes a microorganism imaging unit 23, an image storage unit 24, a microorganism feature learning unit 65, a determination image acquisition unit 26, a feature extraction unit 68, a feature storage unit 69, a determination feature acquisition unit 70, and a microorganism identification unit 71. , and a driving situation evaluation unit 28 . The microorganism feature learning section 65 includes a machine learning section 66 and a clustering section 67 . The machine learning unit 66 includes a learning image acquisition unit 30 , a learning processing unit 81 and a learning parameter storage unit 82 . The clustering unit 67 includes a learned feature quantity acquisition unit 83 , a clustering execution unit 84 , a clustering parameter storage unit 85 , a result presentation unit 86 and an association storage unit 88 .
Among the components of the control device 22, the microorganism imaging unit 23, the determination image acquisition unit 26, the feature amount extraction unit 68, the determination feature amount acquisition unit 70, the microorganism identification unit 71, the driving situation evaluation unit 28, and the learning image acquisition unit 30 , the learning processing unit 81, the learning feature amount acquisition unit 83, the clustering execution unit 84, and the result presentation unit 86 may be software or programs executed by the CPU in the information processing apparatus, for example. Further, the image storage unit 24, the feature amount storage unit 69, the learning parameter storage unit 82, the clustering parameter storage unit 85, and the association storage unit 88 are stored in semiconductor memories, magnetic disks, etc. provided inside or outside the information processing apparatus. It may be realized by an apparatus.

As in the first embodiment, the control device 62 can be broadly classified into two types of operations: learning of identification categories of microorganisms and evaluation of driving conditions. To simplify the explanation, first, each component of the control device 62 during learning of the identification category will be described, and then the behavior of each component during evaluation of the driving situation will be described.

Microorganism imaging unit 23 controls imaging device 21 to image activated sludge. Microorganism photographing unit 23 further cuts out, for example, a portion in which some object is photographed from the photographed image, enlarges it to an appropriate size, and transmits it to image storage unit 24 as a learning image.
Microorganism imaging unit 23 performs imaging and transmission to image storage unit 24 until the number of these learning images reaches the number required for proper learning of the machine learning device.

The image accumulation unit 24 accumulates and stores the plurality of learning images transmitted from the microorganism imaging unit 23 .

The learning image acquisition section 30 of the machine learning section 66 of the microorganism feature learning section 65 acquires the learning image from the image storage section 24 and transmits it to the learning processing section 81 as learning data.
The learning image acquisition unit 30 also transmits the learning image to the feature amount extraction unit 68, which will be described later.

The learning processing unit 81 receives learning data from the learning image acquisition unit 30 .
The learning processing unit 81 includes a machine learning device 90, which is shown as a block diagram in FIG. The machine learning device 90 in the second embodiment performs convolutional deep learning, and is particularly realized by a convolutional autoencoder. The machine learner 90 is used as a program module that is part of the artificial intelligence software to generate a trained model trained with appropriate learning parameters.
The machine learning device 90 includes a convolution processing section 91 , a fully connected layer 92 and a deconvolution processing section 93 . The convolution processing unit 91 includes first to third convolution layers 91a, 91b, and 91c. The all-bonded layer 92 includes first to third fully-bonded layers 92a, 92b, and 92c. The deconvolution processing unit 93 includes first to third deconvolution layers 93a, 93b, and 93c.

The learning processing unit 81 inputs the learning image 43 to the machine learning device 90 . The first to third convolution layers 91a, 91b, and 91c of the convolution processing unit 91 generate a feature map 91m to be input to the fully connected layer 92 by the same processing as the convolution processing unit 41 in the first embodiment.
The third convolutional layer 91c is connected to the first fully connected layer 92a of the fully connected layer 92 with a predetermined number of nodes 92l. The values of the feature map 91m are stored in each node 92l of the first fully connected layer 92a, as in the first embodiment.

The second fully-connected layer 92b also has a predetermined number of nodes 92m, like the first fully-connected layer 92a. The number of nodes 92m in the second fully connected layer 92b is less than the number of nodes 92l in the first fully connected layer 92a. Each node 92m of the second fully connected layer 92b is connected to all nodes 92l of the first fully connected layer 92a, providing a connection weight between the layers. In the second fully connected layer 92b, based on this connection weight, a weighted sum is calculated for the node 92l of the first fully connected layer 92a, and the value obtained by applying the output function to the result is the second fully connected layer stored in each node 92m of 92b.

The third fully connected layer 92c has the same number of nodes 92n as the first fully connected layer 92a. Each node 92n of the third fully-connected layer 92c is connected to all nodes 92m of the second fully-connected layer 92b, providing a connection weight between the layers. In the third fully connected layer 92c, based on this connection weight, a weighted sum is calculated for the node 92m of the second fully connected layer 92b. stored in each node 92n of 92c.

The value of each node 92n of the third fully connected layer 92c is input to the deconvolution processing section 93. FIG. In the deconvolution processing unit 93, processing symmetrical to that of the convolution processing unit 91 is executed. That is, the learning image 43 was compressed to a low dimension by the convolution processing unit 91 until reaching the fully connected layer 92, but in the deconvolution processing unit 93, it is restored from the low-dimensional compressed state. . More specifically, a first deconvolution layer 93a that performs processing substantially symmetrical with the third convolution layer 91c, a second deconvolution layer 93b that performs processing substantially symmetrical with the second convolution layer 91b, and the first convolution layer An image 94 having the same resolution as the learning image 43 is generated by sequentially passing through a third deconvolution layer 93c that performs processing substantially symmetrical to that of 91a.
The machine learning device 90 performs machine learning such that the generated image 94 becomes a reproduced image 94 that reproduces the learning image 43 . That is, the machine learning device 90 in the second embodiment uses only the learning images 43 related to microorganisms as learning data, and performs learning without particularly requiring information generated by the worker's label association work.

The learning of the machine learning device 90 is performed in order to generate a trained model used for calculation of the identification division to which microorganisms belong, as in the first embodiment. Differently, not all parameters set in the learning of the machine learner 90 are used as the trained model.
As already explained, the machine learning device 90 is a convolutional autoencoder, and the learning image 43 is compressed to a low dimension in the convolution processing unit 91 and input to the first fully connected layer 92a of the fully connected layer 92, and further to the second fully connected layer 92a. After being maximally compressed in fully connected layer 92b, it is decompressed into reproduced image 94. FIG. That is, the second fully connected layer 92b is a portion in which the features of the input image are most compressed, and the value of the node 92m is a feature amount that abstractly condenses the features of microorganisms in the image. can handle.

Conceptually, the second embodiment utilizes this characteristic of the convolutional autoencoder, inputs each of the learning images 43 to the machine learning device 90 that has completed learning, and performs the 2 Store the value of each node 92m of the fully connected layer 92b as a feature amount. In this state, the learned model is executed for the input image when evaluating the driving situation to derive the value of each node 92m of the second fully connected layer 92b for the image, that is, the feature quantity, which is stored in advance. By comparing with the feature amount of the learning image 43 obtained, the identification section of the microorganism in the input image is calculated.
Therefore, in the second embodiment, the portion from the first convolutional layer 91a to the second fully connected layer 92b surrounded by the chain double-dashed line in FIG. 8 corresponds to the trained model 95. Therefore, the value of each parameter in this portion, that is, the value of the weight of each filter in the convolution processing unit 91, the value of each connection weight up to the second fully connected layer 92b of the fully connected layer 92, etc. becomes a learning parameter in
The learning processing unit 81 transmits the learning parameters to the learning parameter storage unit 82 when the machine learning is completed.

In this way, the machine learning unit 66 of the microorganism feature learning unit 65 inputs the learning images 43 related to microorganisms to the machine learning device 90 as learning data to perform machine learning. In addition, the machine learning device 90 executes machine learning by inputting learning data and reflecting the characteristics of microorganisms in learning parameters to generate a trained model 95 .

The learning parameter storage unit 82 receives learning parameters from the learning processing unit 81, and accumulates and stores them.

The feature quantity extraction unit 68 receives each learning image 43 from the learning image acquisition unit 30 .
Also, the feature amount extraction unit 68 acquires learning parameters from the learning parameter storage unit 82 .
The feature quantity extraction unit 68 extracts the learned parameters, that is, the values of the weights of the filters of the convolution processing unit 91, the values of the connection weights up to the second fully connected layer 92b of the fully connected layer 92, and the like as described above. The learned model 95, which has been converted to a value, is executed as a program, for example, on a CPU to calculate the value of each node 92m of the second fully connected layer 92b, that is, the feature amount corresponding to each learning image 43. FIG.
In this way, the feature amount extraction unit 68 inputs the learning image 43 again to the trained model 95 and extracts the feature of the microorganism in the learning image 43 as a feature amount.
The feature amount extraction unit 68 transmits the combination of the learning image 43 and the calculated feature amount for each learning image 43 to the feature amount accumulation unit 69 .

The feature amount accumulation unit 69 receives the combination of the learning image 43 and the feature amount from the feature amount extraction unit 68, and accumulates and stores them.
Combinations of the learning images 43 and feature amounts stored in the feature amount accumulation unit 69 are used in the clustering unit 67 described below. The clustering unit 67 divides a plurality of feature quantities corresponding to each of the learning images 43 into a plurality of sets corresponding to each identification section by unsupervised clustering.

The learning feature amount acquisition unit 83 of the clustering unit 67 of the microorganism feature learning unit 65 acquires all combinations of learning images 43 and feature amounts from the feature amount storage unit 69 and transmits them to the clustering execution unit 84 .

The clustering execution unit 84 receives all combinations of learning images 43 and feature quantities from the learning feature quantity acquisition unit 83 .
The clustering execution unit 84 executes clustering on all feature quantities. As a clustering method, a method assuming a statistical model such as mixed normal distribution, a machine learning method such as SVM (Support Vector Machine), or the like can be applied.
As a result of clustering, each feature amount is classified into a plurality of sets, and each feature amount is assigned a cluster ID associated with the corresponding set.
The clustering execution unit 84 transmits to the clustering parameter storage unit 85 the clustering parameters set as the result of the classification by the clustering unit, and all combinations of the learning images 43, feature amounts, and cluster IDs.

The clustering parameter storage unit 85 receives clustering parameters from the clustering execution unit 84 and all combinations of learning images 43, feature amounts, and cluster IDs, and accumulates and stores them.

The result presenting unit 86 acquires the combination of the learning image 43 , the feature amount, and the cluster ID from the clustering parameter storage unit 85 .
Based on the acquired information, the result presenting unit 86 presents the learning image 43 representing each set and the corresponding cluster ID to the operator using a display device or the like (not shown).
In the above, as the learning image 43 representing each set, for example, in the case of clustering assuming a mixture normal distribution, the image having the parameter closest to (mean, variance) of each mixture component is the image having the parameter closest to (mean, variance) of each mixture component in the case of classification by SVM , the image located furthest from the class boundary plane in the parameter space is considered.
Based on this, the operator uses an input device or the like (not shown) to give each cluster ID a label indicating the identification classification of the corresponding microorganism, for example, the type.
The result presenting unit 86 transmits all combinations of cluster IDs and assigned labels to the correspondence storage unit 88 .

The association storage unit 88 receives all combinations of cluster IDs and labels from the result presentation unit 86, and accumulates and stores them.

Next, the behavior of each component during evaluation of the driving situation in the control device 62 will be described.

Microorganism imaging unit 23 controls imaging device 21 to generate an image of microorganisms, and transmits the generated image to image storage unit 24 as an input image.
Microorganism imaging unit 23 performs imaging and transmission to image storage unit 24 until the number of input images reaches the number required for evaluating the driving situation.

The image accumulation unit 24 accumulates and stores a plurality of input images transmitted from the microorganism imaging unit 23 .

The determination image acquisition section 26 acquires the input image from the image storage section 24 and transmits it to the feature quantity extraction section 68 .

The feature quantity extraction unit 68 receives the input image transmitted from the judgment image acquisition unit 26 .
Also, the feature amount extraction unit 68 acquires learning parameters from the learning parameter storage unit 82 .
The feature amount extraction unit 68 executes the learned model 95 in which the learning parameters are reflected as a program on, for example, a CPU, and extracts the feature amount corresponding to the input image 49, which is the value of each node 92m of the second fully connected layer 92b. to calculate That is, the feature quantity extraction unit 68 inputs the input image 49 to the trained model 95 and extracts the features of the microorganisms in the input image 49 as feature quantities.
The feature quantity extraction unit 68 transmits the feature quantity for each input image 49 to the feature quantity accumulation unit 69 .

The feature amount accumulation unit 69 receives the feature amount corresponding to each input image 49 from the feature amount extraction unit 68, and accumulates and stores the feature amount.

The determination feature quantity acquisition unit 70 acquires the feature quantity corresponding to each input image 49 from the feature quantity accumulation unit 69 and transmits the feature quantity to the microorganism identification unit 71 .

The microorganism identification unit 71 receives all the feature values transmitted from the determination feature value acquisition unit 70 .
Microorganism identification unit 71 acquires clustering parameters set as a result of classification in the clustering unit from clustering parameter storage unit 85 .
The microorganism identification unit 71 further acquires all combinations of cluster IDs and labels from the correspondence storage unit 88 .
The microorganism identification unit 71 uses clustering parameters to calculate a cluster ID for each feature quantity generated from the input image 49 . The microorganism identification unit 71 calculates the label corresponding to this cluster ID, that is, the identification category to which the microorganism belongs, for each feature amount.
In this way, the microorganism identification unit 71 calculates the identification category to which the microorganism belongs by calculating to which of a plurality of groups the feature amount of the microorganism in the input image 49 corresponds based on the clustering parameter. .

As described above, in the second embodiment, the microorganism identifying unit 71 uses the clustering parameters generated by the microorganism feature learning unit 65 and stored in the clustering parameter storage unit 85 as identification parameters for identifying microorganisms. , to compute the discriminative section to which the microorganisms in the input image 49 belong.
As already explained, the identification parameter, that is, the clustering parameter in the second embodiment is generated by clustering the feature amount of each learning image generated based on the learning parameter in the machine learning device 90 that has already been trained. ing.

The microorganism identification unit 71 transmits all the calculated identification categories to the driving condition evaluation unit 28 .

The driving condition evaluation unit 28 receives all calculated identification categories from the microorganism identification unit 71 .
The operation status evaluation unit 28 evaluates the operation status of sewage treatment based on the identification result of the microorganism identification unit 71 . More specifically, all the received identification divisions are statistically analyzed, for example, microorganisms that dominate when the state of activated sludge is good, such as the genus Vorticella, the genus Epistylis, the amoeba genus Arcella, and the rotifer Philodina genus is equal to or greater than a predetermined first threshold value, the driving condition is evaluated as good. In addition, for example, when the state of activated sludge such as flagellate genus Bodo, Monas and Amoeba is deteriorated, and when the state of activated sludge is deteriorated, and microorganisms that appear in the transitional period from good to bad conditions, identification categories corresponding to microorganisms such as the genus Arcella is equal to or greater than a predetermined second threshold value, it is evaluated that the driving condition is deteriorating.
In this way, the operation status evaluation unit 28 evaluates the operation status of sewage treatment based on the ratio of microorganisms for each identification category calculated by the microorganism identification unit 71 for a plurality of input images 49, that is, the appearance rate of microorganisms. do.

Next, referring to FIGS. 1, 7, 8, 9, and 10, a sewage treatment operation status evaluation method for evaluating the operation status by the sewage treatment operation status evaluation device 60 will be described. FIG. 9 is a flow chart of the sewage treatment operation status evaluation method during learning. FIG. 10 is a flow chart of the sewage treatment operation status evaluation method at the time of sewage treatment operation status evaluation.
First, with reference to FIG. 9, the operation during learning of the identification segment will be described.

The microorganism imaging unit 23 controls the imaging device 21 to image the activated sludge (step S21). Microorganism photographing unit 23 further cuts out, for example, a portion in which some object is photographed from the photographed image, enlarges it to an appropriate size, and transmits it to image storage unit 24 as a learning image.
The image accumulation unit 24 accumulates and stores the plurality of learning images transmitted from the microorganism imaging unit 23 .

The learning image acquisition section 30 of the machine learning section 66 of the microorganism feature learning section 65 acquires the learning image from the image storage section 24 and transmits it to the learning processing section 81 as learning data.
The learning image acquisition unit 30 also transmits the learning images to the feature amount extraction unit 68 .

The learning processing unit 81 receives learning data from the learning image acquisition unit 30 .
The learning processing unit 81 inputs the learning image 43 to the machine learning device 90 and causes the machine learning device 90 to perform machine learning.
The learning processing unit 81 transmits the learning parameters to the learning parameter storage unit 82 when the machine learning is completed.

The learning parameter storage unit 82 receives learning parameters from the learning processing unit 81, and accumulates and stores them.

The feature quantity extraction unit 68 receives each learning image 43 from the learning image acquisition unit 30 .
Also, the feature amount extraction unit 68 acquires learning parameters from the learning parameter storage unit 82 .
The feature quantity extraction unit 68 extracts the learned parameters, that is, the values of the weights of the filters of the convolution processing unit 91, the values of the connection weights up to the second fully connected layer 92b of the fully connected layer 92, and the like as described above. The learned model 95, which has been converted to a value, is executed as a program, for example, on a CPU to calculate the value of each node 92m of the second fully connected layer 92b, that is, the feature amount corresponding to each learning image 43. FIG.
In this way, the feature amount extraction unit 68 inputs the learning image 43 to the trained model 95 and extracts the feature of the microorganism in the learning image 43 as a feature amount (step S23).
The feature amount extraction unit 68 transmits the combination of the learning image 43 and the calculated feature amount for each learning image 43 to the feature amount accumulation unit 69 .

The feature amount accumulation unit 69 receives the combination of the learning image 43 and the feature amount from the feature amount extraction unit 68, and accumulates and stores them.

The learning feature amount acquisition unit 83 of the clustering unit 67 of the microorganism feature learning unit 65 acquires all combinations of learning images 43 and feature amounts from the feature amount storage unit 69 and transmits them to the clustering execution unit 84 .

The clustering execution unit 84 receives all combinations of learning images 43 and feature quantities from the learning feature quantity acquisition unit 83 .
The clustering execution unit 84 executes clustering for all feature amounts (step S25).
As a result of clustering, each feature amount is classified into a plurality of sets, and each feature amount is assigned a cluster ID associated with the corresponding set.
The clustering execution unit 84 transmits to the clustering parameter storage unit 85 the clustering parameters set as the result of the classification by the clustering unit, and all combinations of the learning images 43, feature amounts, and cluster IDs.

The clustering parameter storage unit 85 receives clustering parameters from the clustering execution unit 84 and all combinations of learning images 43, feature amounts, and cluster IDs, and accumulates and stores them (step S27).

The result presenting unit 86 acquires the combination of the learning image 43 , the feature amount, and the cluster ID from the clustering parameter storage unit 85 .
Based on the acquired information, the result presenting unit 86 presents the learning image 43 representing each set and the corresponding cluster ID to the operator using a display device or the like (not shown).
The operator uses an input device or the like (not shown) to give each cluster ID a label indicating the identification classification of the corresponding microorganism, for example, the type (step S29).
The result presenting unit 86 transmits all combinations of cluster IDs and assigned labels to the correspondence storage unit 88 .

The association storage unit 88 receives all combinations of cluster IDs and labels from the result presentation unit 86, and accumulates and stores them.

Next, with reference to FIG. 10, the operation during evaluation of the driving situation will be described.

Microorganism imaging unit 23 controls imaging device 21 to generate an image of microorganisms, and transmits the generated image to image storage unit 24 as an input image.
The image accumulation unit 24 accumulates and stores a plurality of input images transmitted from the microorganism imaging unit 23 .
The determination image acquisition section 26 acquires the input image from the image storage section 24 and transmits it to the feature quantity extraction section 68 .

The feature quantity extraction unit 68 receives the input image transmitted from the judgment image acquisition unit 26 .
Also, the feature amount extraction unit 68 acquires learning parameters from the learning parameter storage unit 82 .
The feature amount extraction unit 68 executes the learned model 95 in which the learning parameters are reflected as a program on, for example, a CPU, and extracts the feature amount corresponding to the input image 49, which is the value of each node 92m of the second fully connected layer 92b. to calculate That is, the feature quantity extraction unit 68 inputs the input image 49 to the trained model 95 and extracts the features of microorganisms in the input image 49 as feature quantities (step S31).
The feature quantity extraction unit 68 transmits the feature quantity for each input image 49 to the feature quantity accumulation unit 69 .

The feature amount accumulation unit 69 receives the feature amount corresponding to each input image 49 from the feature amount extraction unit 68, and accumulates and stores the feature amount.
The determination feature quantity acquisition unit 70 acquires the feature quantity corresponding to each input image 49 from the feature quantity accumulation unit 69 and transmits the feature quantity to the microorganism identification unit 71 .

The microorganism identification unit 71 receives all the feature values transmitted from the determination feature value acquisition unit 70 .
Microorganism identification unit 71 acquires clustering parameters set as a result of classification in the clustering unit from clustering parameter storage unit 85 .
The microorganism identification unit 71 further acquires all combinations of cluster IDs and labels from the correspondence storage unit 88 .
The microorganism identification unit 71 uses clustering parameters to calculate a cluster ID for each feature quantity generated from the input image 49 . The microorganism identification unit 71 calculates the label corresponding to this cluster ID, that is, the identification category to which the microorganism belongs for each feature amount (step S33).
The microorganism identification unit 71 transmits all the calculated identification categories to the driving condition evaluation unit 28 .

The driving condition evaluation unit 28 receives all calculated identification categories from the microorganism identification unit 71 .
The operation status evaluation unit 28 evaluates the operation status of sewage treatment based on the identification result of the microorganism identification unit 71 (step S35).

Next, the effects of the sewage treatment operating condition evaluation device and the sewage treatment operation condition evaluation method will be described.

The sewage treatment operation status evaluation device 60 in the second embodiment is intended for a sewage treatment facility 1 that purifies sewage stored in a biological treatment reaction tank (water treatment tank) 3 with microorganisms. A sewage treatment operating status evaluation device 60 for analyzing an input image 49 of captured microorganisms, calculating which identification category the microorganisms belong to, and evaluating the operating status of sewage treatment, comprising a machine learning device 90 , a microbial feature learning unit 65 that inputs learning data including a learning image 43 related to microorganisms to a machine learning device 90 to perform machine learning and generates identification parameters for identifying microorganisms; and an operation condition evaluation unit 28 for evaluating the operation condition of sewage treatment based on the identification result of the microorganism identification unit 71 .
According to the above configuration, the microorganism feature learning unit 65 performs machine learning by inputting learning data including the learning image 43 related to microorganisms to the machine learning device 90, so that the features of microorganisms in the image can be effectively recognized. It is possible to extract and automatically generate identification parameters that reflect this. For this reason, when introducing the sewage treatment operating condition evaluation device 60, advanced knowledge of image processing is not particularly required.
In addition, since it is not necessary for the operator to intentionally extract the features so that the unique features of the microorganisms can be conspicuously expressed, highly specialized knowledge of the microorganisms is not particularly required. In the second embodiment, it is necessary for the operator to associate a label with each set classified by clustering, but particularly advanced knowledge required for extracting the characteristics of microorganisms is required. do not have.
Therefore, introduction of the sewage treatment operating condition evaluation device 60 is easy.
Also, the accuracy of the identification parameters can be easily improved by preparing a large amount of learning data to be input to the machine learning device 90 . For example, in the second embodiment, by preparing a large amount of learning data, the accuracy of the learning parameters of the machine learning device 90 is improved, and this increases the accuracy of the feature amount of the learning image 43 output by the trained model 95, It affects the accuracy of the clustering parameters or identification parameters. Therefore, it is possible to improve the calculation accuracy of the identification category to which microorganisms belong, that is, the accuracy of evaluation of the operation status of sewage treatment.

Further, the machine learning device 90 performs convolutional deep learning.
According to the configuration as described above, the sewage treatment operation status evaluation device 60 can be properly realized.

In addition, the machine learning device 90 executes machine learning by inputting learning data and reflecting the characteristics of microorganisms in learning parameters to generate a trained model 95 .
It also has a feature quantity extraction unit 68 that inputs the learning image 43 and the input image 49 to the trained model 95 and extracts the features of the microorganisms in the learning image 43 and the input image 49 as feature quantities.
Further, the microorganism feature learning unit 65 includes a clustering unit 67 that classifies a plurality of feature amounts corresponding to each of the plurality of learning images 43 into a plurality of sets by clustering, and each of the plurality of sets corresponds to the identification division. , the identification parameter is a clustering parameter set as a result of classification in the clustering unit 67 .
Based on the clustering parameters, the microorganism identification unit 71 also calculates the identification category to which the microorganism belongs by calculating to which of a plurality of sets the feature amount of the microorganism in the input image 43 belongs.
According to the configuration as described above, the input image 49 used in the evaluation of the driving situation is processed by the learned model 95 in which the characteristics of microorganisms are reflected, so that the characteristics are efficiently extracted. Environmental impact can be reduced. Therefore, operation of the sewage treatment operating condition evaluation device 20 is facilitated.
In addition, the feature amount of the learning image 43 is accumulated and classified into a plurality of sets by clustering, and then the identification section is calculated by calculating which set the feature amount of the input image 49 corresponds to, thereby increasing the accuracy. High identification becomes possible.

Also, the machine learning device 90 is realized by a convolutional autoencoder.
According to the configuration as described above, when an image is input to the trained model 95 of the machine learning device 90, it is output as an appropriately compressed feature amount, so the feature amount is accumulated and clustering processing is performed. The amount of data to be handled is reduced. This facilitates clustering processing.

The operation status evaluation unit 28 also evaluates the operation status of sewage treatment based on the appearance rate of microorganisms for each identification category calculated by the microorganism identification unit 71 for the plurality of input images 49 .
According to the configuration as described above, it is possible to evaluate the operation status of sewage treatment with high accuracy.

In the sewage treatment operation status evaluation device 60, as in the first embodiment, even if the number of types of microorganisms increases, the machine learning device 90 learns to accurately capture the characteristics of microorganisms, so that erroneous evaluation of the operation status is prevented. can be suppressed, and since the feature extraction itself is automatically performed by creating learning data, the increase in introduction cost is minimal. Thus, each of the effects already described with respect to the sewage treatment operational status evaluation device 60 can be more pronounced when there are many types of microorganisms compared to conventional systems.

It should be noted that the sewage treatment operation status evaluation apparatus and the sewage treatment operation status evaluation method of the present invention are not limited to the above-described embodiments and modifications described with reference to the drawings, and other Various modifications are conceivable.

For example, it goes without saying that the configurations of the machine learning devices 40, 50, and 90 are not limited to those described in the above embodiments and modifications. For example, in the machine learning device 40, many convolutional layers are provided, pooling is performed only in limited convolutional layers, one or more intermediate layers are provided between the input layer 47 and the output layer 48, and so on. However, any configuration may be employed as long as it can be set to extract features of microorganisms in an input image. The same applies to the machine learning devices 50 and 90 as well.

In addition, although the controllers 22 and 62 have been described as being realized as one information processing device, each component is divided into a plurality of information processing devices and installed, and these operate in cooperation with each other. It goes without saying that For example, in the control devices 22 and 62, the part for learning and the part for evaluating driving conditions may be divided into two corresponding information processing devices.

In addition, the driving condition evaluation unit 28 was explained to evaluate whether the driving condition is good or bad, but based on the appearance rate of microorganisms, the evaluation can be subdivided into more stages, for example, five stages. may be implemented to perform
Alternatively, the evaluation by the driving condition evaluation unit 28 may be classified into a plurality of conditions that are not superior or inferior to each other, instead of being based on the evaluation axis of good or bad driving conditions.

In addition to this, it is possible to select the configurations mentioned in each of the above-described embodiments and modified examples, or to change them to other configurations as appropriate without departing from the gist of the present invention.

1 Sewage treatment facility 3 Biological treatment reaction tank (water treatment tank)
20, 60 Sewage treatment operation status evaluation device 21 Imaging device 22, 62 Control device 23 Microorganism imaging unit 24 Image storage unit 25, 65 Microorganism feature learning unit 26 Determination image acquisition unit 27, 71 Microorganism identification unit 28 Operation condition evaluation unit 30 Learning Image acquisition unit 31 Labeling units 32, 81 Learning processing units 33, 82 Learning parameter storage units 40, 50, 50-1 to 50-N, 90 Machine learning devices 40A, 95 Trained model 43 Learning image 49 Input image 66 Machine learning Unit 67 Clustering unit 68 Feature amount extraction unit 69 Feature amount accumulation unit 70 Determined feature amount acquisition unit 83 Learning feature amount acquisition unit 84 Clustering execution unit 85 Clustering parameter storage unit 86 Result presentation unit 88 Correspondence storage unit 94 Reproduced image

Claims (6)

Targeting a sewage treatment facility that purifies sewage stored in a water treatment tank with microorganisms, an input image obtained by imaging the microorganisms living in the water treatment tank is analyzed to determine which identification category the microorganisms belong to. A sewage treatment performance evaluation device for calculating and evaluating sewage treatment performance, comprising:
A plurality of machine learning devices corresponding to each of the identification divisions is provided, learning data including learning images related to the microorganisms is input to each of the plurality of machine learning devices, and the microorganisms in the learning images correspond to the a microorganism feature learning unit that performs machine learning on whether or not the microorganism belongs to an identification division, generates learning parameters related to identification of the microorganism, and generates a plurality of trained models for each identification division;
The input image is input to each of the plurality of trained models, the presence or absence of the microorganism belonging to the corresponding identification section is determined, and the determination results are aggregated to determine the microorganism belonging to the input image. a microorganism identification unit that calculates the identification category;
an operation status evaluation unit that evaluates the operation status of the sewage treatment based on the identification result of the microorganism identification unit;
A sewage treatment operation status evaluation device. 2. The sewage treatment operation status evaluation device according to claim 1, wherein each of said plurality of machine learners performs convolutional deep learning. Each of the plurality of machine learning devices inputs, in addition to the learning image, information as to whether or not the microorganism in the learning image belongs to the identification section corresponding to the machine learning device as teacher data. 3. The sewage treatment operation status evaluation device according to claim 1 or 2, which performs machine learning. The sewage treatment according to any one of claims 1 to 3, wherein each of the plurality of trained models performs two-class classification of whether or not microorganisms in the input image belong to the corresponding identification division. Driving condition evaluation device. 2. The operation status evaluation unit evaluates the operation status of the sewage treatment based on the appearance rate of the microorganisms for each of the identification categories calculated by the microorganism identification unit for the plurality of input images. 5. The sewage treatment operation status evaluation device according to any one of 4 from to 4. Targeting a sewage treatment facility that purifies sewage stored in a water treatment tank with microorganisms, an input image obtained by imaging the microorganisms living in the water treatment tank is analyzed to determine which identification category the microorganisms belong to. A wastewater treatment operating condition evaluation method for calculating and evaluating a wastewater treatment operating condition, comprising:
Learning data including learning images about the microorganisms are input to each of a plurality of machine learning devices provided corresponding to the identification divisions, and the microorganisms in the learning images belong to the corresponding identification divisions. machine learning whether or not there is, generating learning parameters related to the identification of the microorganisms, generating a plurality of trained models for each of the identification categories,
The input image is input to each of the plurality of trained models, the presence or absence of the microorganism belonging to the corresponding identification section is determined, and the determination results are aggregated to determine the microorganism belonging to the input image. calculating the identification class;
A sewage treatment operation status evaluation method for evaluating the sewage treatment operation status based on the identification result.

Images