JP7020160B2 - Water quality analyzer - Google Patents

Description

The present invention relates to a water quality analyzer and a water quality analysis method for sewage treatment facilities and the like.

In recent years, there has been a demand for more efficient maintenance and management of sewage treatment facilities. Non-Patent Document 1, which is an example of the prior art, discloses a technique for monitoring frock, which is an agglomerate generated in a water purification plant, by image processing instead of visual inspection by an operator. In the technique disclosed in Non-Patent Document 1, the brightness is adjusted by applying a spatial filtering process to a grayscale image, and the flocs are extracted and extracted by binarizing the image with the adjusted brightness. Labeling and particle size distribution calculation are performed for flocs.

Further, in Patent Document 1, which is an example of the prior art, the feature of the image data is extracted by the autoencoder, and the normal value and the outlier are obtained by using a one-class classifier such as One-Class SVM (Support Vector Machine). A technique for identifying and determining an abnormality is disclosed.

Japanese Unexamined Patent Publication No. 2018-5773

Kenji Baba, Mikio Yoda, Masami Tanimoto, "Basic Research on Flock Image Surveillance Technology in Water Purification Plants", IEEJ Journal D, 1987, Vol. 107, No. 7, p. 844-851

However, according to the technique disclosed in Non-Patent Document 1, it is necessary to adjust the threshold value of the binarization process and the filter size and weight of the spatial filtering process, and the model is reconstructed for another object. There was a problem that tuning was required again.

Further, in the technique disclosed in Patent Document 1, since the determination is made by a single one-class classifier, there is a problem that it is difficult to deal with various characteristics of water quality such as color, scum and floc. rice field.

The present invention has been made in view of the above, and an object of the present invention is to perform water quality analysis corresponding to various characteristics of water quality without performing special tuning.

The present invention, which solves the above-mentioned problems and achieves the object, learns the convolution AE (Auto Encoder) of the image data of the settling pond, extracts the feature amount of the hue of the settling pond, and clusters from the feature amount of the hue. The image data is classified into clusters of hues, and the hue analysis unit that draws the feature space and cluster distribution of the hues, and the convolution AE of the image data are learned to extract the features of the suspended matter in the sedimentation pond. It is a water quality analyzer provided with a suspended matter analysis unit that classifies the image data into clusters of suspended matter by clustering from the feature amount of the suspended matter and draws a feature space of the suspended matter and a cluster distribution.

The water quality analyzer having the above configuration learns the convolution AE of the image data, extracts the feature amount of the moving body of the settling pond, and classifies the image data into a cluster of the moving body by clustering from the feature amount of the moving body. It is preferable to provide a moving body analysis unit that draws the feature space and the cluster distribution of the moving body.

Alternatively, the present invention has at least an image processing unit, a convolution AE learning unit, a convolution AE parameter storage unit, a feature amount extraction unit, a clustering unit, a clustering parameter storage unit, and a cluster classification unit. A water quality analysis method of a water quality analyzer including a hue analysis unit and a floating matter analysis unit including a space drawing unit, wherein the image processing unit processes image data of a sedimentation pond, and the convolution AE learning unit When the convolution AE learning is required for the processed image data, the convolution AE learning is performed for the image data, and the convolution AE parameter storage unit stores the convolution AE parameters obtained by the convolution AE learning. The feature amount extraction unit extracts the feature amount from the processed image data to the encoder to which the convolution AE parameter is applied, and when the clustering unit requires clustering for the processed image data, the image data. The clustering is performed, the clustering parameter storage unit stores the clustering parameters obtained by the clustering, and the cluster classification unit classifies the image data from which the feature amount has been extracted into clusters. The space drawing unit is a water quality analysis method including drawing a feature space and a cluster distribution based on the image data classified into clusters.

According to the present invention, there is an effect that water quality analysis corresponding to various characteristics of water quality can be performed without performing special tuning.

It is a block diagram which shows the structure of the water quality analyzer which concerns on Embodiment 1. FIG. It is a figure explaining the outline of the convolution AE. It is a figure explaining the outline of the feature amount extraction by the trained encoder. It is a figure explaining the outline of the clustering method. It is a flowchart which shows the operation of the hue analysis part of the water quality analysis apparatus which concerns on Embodiment 1. FIG. It is a block diagram which shows the structure of the mobile body analysis part which is the premise of the water quality analysis apparatus which concerns on Embodiment 2. It is a block diagram which shows the structure of the water quality analyzer which concerns on Embodiment 2.

Hereinafter, a mode for carrying out the water quality analyzer and the water quality analysis method of the present invention will be described with reference to the accompanying drawings. However, the present invention is not limited to the description of the following embodiments.

<Embodiment 1>
FIG. 1 is a block diagram showing a configuration of a water quality analyzer according to the present embodiment. The water quality analysis device 100 shown in FIG. 1 includes a data storage unit 101, a hue analysis unit 110, and a suspended matter analysis unit 120.

The data storage unit 101 stores image data of a settling basin acquired by a camera (not shown) installed in a settling basin such as a sewage treatment facility. The data storage unit 101 can be realized by a recording medium such as a semiconductor memory and a magnetic disk. This image data includes both a still image and a moving image composed of a plurality of continuous still images.

The hue analysis unit 110 analyzes the color of the treated water using the image data of the data storage unit 101. By analyzing the color of the treated water using the hue analysis unit 110, for example, it is possible to determine the change in the color of the settling basin due to the sludge rolled up in the treated water and evaluate the amount of sludge rolled up in the treated water. Is. The suspended matter analysis unit 120 analyzes the suspended matter in the treated water using the image data of the data storage unit 101. By analyzing the treated water using the suspended matter analysis unit 120, for example, it is possible to determine the presence or absence of suspended matter. Further, when a floating substance is present, it is possible to specify the component of the floating substance. Here, as the floating matter, scum and foam can be exemplified. By analyzing the components of the suspended matter, for example, it becomes possible to determine whether or not the suspended matter is a scum. The hue analysis unit 110 and the suspended matter analysis unit 120 perform convolution AE learning, feature quantity extraction, and clustering on the image data stored in the data storage unit 101.

Here, the configuration of the hue analysis unit 110 will be described. The hue analysis unit 110 includes a first image processing unit 111, a first convolution AE learning unit 112, a first convolution AE parameter storage unit 113, a first feature amount extraction unit 114, and a first clustering unit. A unit 115, a first clustering parameter storage unit 116, a first cluster classification unit 117, a first feature space drawing unit 118, and a first learning necessity determination unit 119 are provided to obtain RGB image data. Use to analyze the color of the treated water.

The first image processing unit 111 obtains a processed RGB3 channel image by performing trimming processing or normalization on the image data of the data storage unit 101. The first image processing unit 111 can be realized by a processor such as an MPU (Micro-Processing Unit) and a CPU (Central Processing Unit).

The first convolution AE learning unit 112 learns the convolution AE using the RGB3 channel image processed by the first image processing unit 111. The first convolution AE learning unit 112 can be realized by a processor such as an MPU and a CPU. The first convolution AE parameter storage unit 113 stores the convolution AE parameters learned and obtained by the first convolution AE learning unit 112. The first convolution AE parameter storage unit 113 can be realized by a recording medium such as a semiconductor memory and a magnetic disk.

FIG. 2 is a diagram illustrating an outline of the convolution AE. As shown in FIG. 2, the convolution AE includes an encoder that compresses the dimension and a decoder that restores the dimension. The fully coupled layer that exists between the encoder and the decoder has a lower dimension with respect to the input image, and the trained encoder behaves appropriately as a dimensional compressor. In the first convolution AE learning unit 112, the encoder learns the features of the image by learning so that the error between the image generated by the network and the teacher data is minimized by using the teacher data as an input image. Here, the convolution AE is unsupervised learning, and since the learning is performed using the teacher data as an input image, there is an advantage that the user does not need to label the image of the treated water with respect to colors and suspended matter. In addition, the convolution AE does not require any special tuning in the design, and the learning is performed only by inputting an image.

The first feature amount extraction unit 114 applies the convolution AE parameters stored in the first convolution AE parameter storage unit 113, and uses the convolution AE including the trained encoder to use the first image processing unit 111. The feature amount of the processed data is extracted by. The first feature amount extraction unit 114 can be realized by a processor such as an MPU and a CPU.

FIG. 3 is a diagram illustrating an outline of feature amount extraction by the trained encoder. The first feature amount extraction unit 114 inputs the input image to the trained encoder and performs dimensional compression of the input image to extract the feature amount.

The first clustering unit 115 applies a clustering method to the feature amount extracted by the first feature amount extraction unit 114 to perform clustering of image data. The first clustering unit 115 can be realized by a processor such as an MPU and a CPU. Whether or not this clustering is possible is determined by the determination conditions based on the user's knowledge. The first clustering parameter storage unit 116 stores the clustering parameters obtained by the first clustering unit 115. The first clustering parameter storage unit 116 can be realized by a recording medium such as a semiconductor memory and a magnetic disk. The first cluster classification unit 117 determines the cluster to which the image data to which the first feature amount extraction unit 114 has been extracted belongs based on the feature amount, and classifies each image data into clusters. The first cluster classification unit 117 can be realized by a processor such as an MPU and a CPU.

FIG. 4 is a diagram illustrating an outline of the clustering method. The clustering method used here is non-hierarchical clustering, and examples thereof include k-means method, mixed normal distribution, and DBSCAN (Density-Based Spatial Clustering of Applications with Noise). In FIG. 4, the data is classified into cluster 1, cluster 2, or cluster 3, and is drawn in the feature space.

The first feature space drawing unit 118 visualizes information by drawing the feature space and the cluster distribution based on the information of the first clustering unit 115 and the first cluster classification unit 117. The first feature space drawing unit 118 presents the distribution of data and clusters by, for example, compressing the feature space into two or three dimensions by principal component analysis. The user is presented with the characteristics of each cluster, and the user can manually label each cluster. The first feature space drawing unit 118 can be realized by an input / output device such as a touch panel.

The learning by the first convolution AE learning unit 112 and the clustering by the first clustering unit 115 may be performed only when learning is required, that is, at the time of the first operation or when re-learning is required. The first learning necessity determination unit 119 determines the learning necessity of the first convolution AE learning unit 112 and the first clustering unit 115, and determines the learning necessity of the first convolution AE learning unit 112 and the first clustering unit 115. Instruct whether or not to learn. Here, the timing of re-learning is determined by the user and is set at a fixed cycle or when the environment of the sedimentation basin changes. The set fixed cycle is, for example, one week or one month, and the environmental change of the settling basin is caused by, for example, seasonal change or construction of the treatment plant.

Next, the configuration of the suspended matter analysis unit 120 will be described. The floating matter analysis unit 120 includes a second image processing unit 121, a second convolution AE learning unit 122, a second convolution AE parameter storage unit 123, a second feature amount extraction unit 124, and a second. A gray scale image including a clustering unit 125, a second clustering parameter storage unit 126, a second cluster classification unit 127, a second feature space drawing unit 128, and a second learning necessity determination unit 129. Is used to analyze suspended matter on the treated water.

The second image processing unit 121 processes the image data of the data storage unit 101. Specifically, the second image processing unit 121 performs grayscale conversion on the data of the data storage unit 101 for the analysis of suspended matter.

The second convolution AE learning unit 122 learns the convolution AE using the data processed by the second image processing unit 121. For the convolution AE, the explanation of the first convolution AE learning unit 112 is referred to. The second convolution AE parameter storage unit 123 stores the convolution AE parameters learned and obtained by the second convolution AE learning unit 122.

The second feature amount extraction unit 124 applies the convolution AE parameters stored in the second convolution AE parameter storage unit 123, and uses the convolution AE including the trained encoder to use the second image processing unit 121. The feature amount of the processed data is extracted by. For the feature amount extraction, the description of the first feature amount extraction unit 114 is referred to.

The second clustering unit 125 applies a clustering method to the feature amount extracted by the second feature amount extraction unit 124 to perform clustering of image data. The second clustering parameter storage unit 126 stores the clustering parameters obtained by the second clustering unit 125. The second cluster classification unit 127 determines the cluster to which the image data to which the second feature amount extraction unit 124 has been extracted belongs based on the feature amount, and classifies each image data into clusters. As for the clustering method, the description of the first clustering unit 115 is referred to.

The second feature space drawing unit 128 visualizes the information by drawing the feature space and the cluster distribution based on the information of the second clustering unit 125 and the second cluster classification unit 127. Regarding the drawing of the feature space and the cluster distribution, the description of the first feature space drawing unit 118 is referred to.

The learning by the second convolution AE learning unit 122 and the clustering by the second clustering unit 125 may be performed only when learning is required, that is, at the time of the first operation or when re-learning is required. The second learning necessity determination unit 129 determines the learning necessity of the second convolution AE learning unit 122 and the second clustering unit 125, and determines the learning necessity of the second convolution AE learning unit 122 and the second clustering unit 125. Instruct whether or not to learn.

As described above, since the hue analysis unit 110 includes the first convolution AE learning unit 112 and the suspended matter analysis unit 120 includes the second convolution AE learning unit 122, it is possible to deal with various characteristics of water quality. Therefore, it is possible to perform feature quantity extraction with higher accuracy than a configuration in which feature quantity extraction is performed by a single convolution AE.

Furthermore, since the first feature space drawing unit 118 and the second feature space drawing unit 128 are provided, the automatically generated cluster distribution can be visually confirmed, and the user can visually confirm the features extracted by the convolution AE. Can be confirmed.

Of the configurations included in the hue analysis unit 110 and the suspended matter analysis unit 120 described above, those realized by the processor may be realized by the same one processor. Further, among the configurations included in the hue analysis unit 110 and the suspended matter analysis unit 120 described above, those realized by the recording medium may be realized by the same one recording medium. Further, among the configurations included in the hue analysis unit 110 and the suspended matter analysis unit 120 described above, those realized by an input / output device such as a touch panel may be realized by the same one input / output device.

Next, the operation of the hue analysis unit 110 will be described. FIG. 5 is a flowchart showing the operation of the hue analysis unit 110. First, the first image processing unit 111 acquires image data from the data storage unit 101 and processes it (S1: image data processing step). The first learning necessity determination unit 119 determines the learning necessity of the first convolution AE learning unit 112 (S2: learning necessity determination step of the convolution AE learning unit). When the first learning necessity determination unit 119 determines that the learning of the first convolution AE learning unit 112 is unnecessary (S2: No), the flow proceeds to S5. When the first learning necessity determination unit 119 determines that the learning of the first convolution AE learning unit 112 is necessary (S2: required), the first AE convolution learning unit 112 is the processed image data. Is used to learn the convolution AE (S3: convolution AE learning step), and the obtained convolution AE parameters are stored in the first convolution AE parameter storage unit 113 (S4: convolution AE parameter storage step). Next, the first feature amount extraction unit 114 acquires the convolution AE parameter from the first convolution AE parameter storage unit 113, inputs the processed image data to the encoder to which the convolution AE parameter is applied, and performs dimensional compression. (S5: Feature amount extraction step). The first learning necessity determination unit 119 determines the learning necessity of the first clustering unit 115 (S6: learning necessity determination step of the clustering unit). When the first learning necessity determination unit 119 determines that the learning of the first clustering unit 115 is unnecessary (S6: No), the flow proceeds to S9. When the first learning necessity determination unit 119 determines that the learning of the first clustering unit 115 is necessary (S6: required), the first clustering unit 115 is the first feature amount extraction unit 114. The extracted features are acquired and clustered (S7: clustering step), and the obtained clustering parameters are stored in the first clustering parameter storage unit 116 (S8: clustering parameter storage step). Next, the first cluster classification unit 117 classifies the image data from which the feature amount has been extracted into clusters (S9: cluster classification step). Next, the first feature space drawing unit 118 draws the feature space and the cluster distribution based on the image data classified into clusters (S10: drawing step), and ends the process. The information drawn in this way is presented to the user, and the user performs labeling processing on the presented information.

Since the operation of the suspended matter analysis unit 120 is the same as that of the hue analysis unit 110, the description of the hue analysis unit 110 is incorporated.

According to the present embodiment described above, it is not necessary to adjust the threshold value of the binarization process and the filter size and weight of the spatial filtering process, which are necessary in the prior art, and the model can be re-modeled even if the target is different. No need to build and retune. Furthermore, since clustering is performed by providing a plurality of convolutional AE learning units, flexible feature extraction and cluster classification are possible. Therefore, it is possible to perform water quality analysis corresponding to various characteristics of water quality without performing special tuning. Further, according to the present embodiment, by linking the outputs of the hue analysis unit 110 and the suspended matter analysis unit 120, highly accurate analysis becomes possible.

<Embodiment 2>
In the first embodiment, the water quality analyzer including the hue analysis unit and the suspended matter analysis unit has been described, but the water quality analyzer according to the present invention is not limited to this, and further includes a mobile body analysis unit. It is also possible to do.

<Prerequisite technology>
FIG. 6 is a block diagram showing a configuration of a mobile body analysis unit which is a premise of the water quality analysis device according to the present embodiment. The mobile body analysis unit 210 shown in FIG. 6 includes an image processing unit 211, a flow estimation unit 212, and a statistic calculation unit 213. The mobile analysis unit 210 is configured to be able to supply the data of the data storage unit 201. Similar to the data storage unit 101, the data storage unit 201 stores moving image data of the settling basin acquired by a camera installed in the settling basin. The data storage unit 201 can be realized by a recording medium such as a semiconductor memory and a magnetic disk. The image processing unit 211 converts the moving image data of the data storage unit 201 into grayscale and processes it into a plurality of grayscale image data. The flow estimation unit 212 estimates the flow of a moving object by an optical flow that displays the movement of an object as a vector. The mobiles are scum and flock. Here, as the calculation algorithm of the optical flow, a dense optical flow (Dense Optical Flow) that analyzes the moving object of the entire pixel is used. Examples of the dense optical flow (Dense Optical Flow) include the Gunnar Farneback method, the TV-L1 method, and the Brox method. The statistic calculation unit 213 calculates the statistic from the speed and movement direction information obtained by the flow estimation unit 212. In general, scum and frock move in a messy direction and speed as compared with the treated water itself, so that it is possible to evaluate the presence or absence of scum and frock from the calculated statistics. The image processing unit 211, the flow estimation unit 212, and the statistic calculation unit 213 can be realized by a processor such as an MPU (Micro-Processing Unit) and a CPU (Central Processing Unit).

As described above, by providing the moving body analysis unit using moving images, it is possible to quantitatively evaluate the positions, moving directions and moving speeds of the flocs and scums.

FIG. 7 is a block diagram showing the configuration of the water quality analyzer according to the present embodiment. The water quality analysis device 100A shown in FIG. 7 includes a data storage unit 101, a hue analysis unit 110, a suspended matter analysis unit 120, and a mobile body analysis unit 130. Since the hue analysis unit 110 and the suspended matter analysis unit 120 are the same as those shown in FIG. 1, the description of the first embodiment is incorporated.

The moving body analysis unit 130 includes a third image processing unit 131, a third convolution AE learning unit 132, a third convolution AE parameter storage unit 133, a third feature amount extraction unit 134, and a third. The clustering unit 135, the third clustering parameter storage unit 136, the third cluster classification unit 137, the third feature space drawing unit 138, the third learning necessity determination unit 139, and the flow estimation unit 140. It is possible to evaluate the presence or absence of scum and floc.

Next, the configuration of the mobile body analysis unit 130 will be described. The third image processing unit 131 grayscale-converts the moving image data of the data storage unit 101 and processes it into a plurality of grayscale image data, similarly to the image processing unit 211 shown in FIG. 6 in the prerequisite technology. The flow estimation unit 140 uses the data processed by the third image processing unit 131 to generate the flow of the moving body by the optical flow, as in the flow estimation unit 212 shown in FIG. 6 in the prerequisite technology. presume. As a result, information on the moving direction and moving speed of the moving body is extracted.

The third convolution AE learning unit 132 learns the convolution AE by using the information of the moving direction and the moving speed of the moving body obtained by the flow estimation unit 140. As for the convolution AE, the description of the first convolution AE learning unit 112 in the first embodiment is referred to. The third convolution AE parameter storage unit 133 stores the convolution AE parameters learned and obtained by the third convolution AE learning unit 132.

The third feature amount extraction unit 134 is processed by the third image processing unit 131 using a convolution AE including a trained encoder to which the convolution AE parameter stored in the third convolution AE parameter storage unit 133 is applied. Extract the features of the completed data. For the feature amount extraction, the description of the first feature amount extraction unit 114 in the first embodiment is referred to.

The third clustering unit 135 applies a clustering method to the feature amount extracted by the third feature amount extraction unit 134 to perform clustering of image data. The third clustering parameter storage unit 136 stores the clustering parameters obtained by the third clustering unit 135. The third cluster classification unit 137 determines the cluster to which the image data to which the third feature amount extraction unit 134 has been extracted belongs based on the feature amount, and classifies each image data into clusters. As for the clustering method, the description of the third clustering unit 135 is referred to.

The third feature space drawing unit 138 visualizes the information by drawing the feature space and the cluster distribution based on the information of the third clustering unit 135 and the third cluster classification unit 137. Regarding the drawing of the feature space and the cluster distribution, the description of the first feature space drawing unit 118 is referred to.

The learning by the third convolution AE learning unit 132 and the clustering by the third clustering unit 135 may be performed only when learning is required, that is, at the time of the first operation or when re-learning is required. The third learning necessity determination unit 139 determines the learning necessity of the third convolution AE learning unit 132 and the third clustering unit 135, and determines the learning necessity of the third convolution AE learning unit 132 and the third clustering unit 135. Instruct whether or not to learn.

Of the configurations included in the mobile body analysis unit 130 described above, those realized by the processor may be realized by the same one processor, or may be realized by the same processor, or the processor included in the hue analysis unit 110 and the suspended matter analysis unit 120. It may be combined. Further, among the configurations included in the mobile body analysis unit 130 described above, those realized by the recording medium may be realized by the same one recording medium, and the hue analysis unit 110 and the suspended matter analysis unit 120 include. It may also be used as a recording medium. Of the configurations included in the mobile body analysis unit 130 described above, those realized by an input / output device such as a touch panel may be realized by the same one touch panel, and the hue analysis unit 110 and the suspended matter analysis unit may be used. It may also be used as a touch panel included in the 120.

Since the operation of the mobile body analysis unit 130 is the same as that of the hue analysis unit 110, the description of the hue analysis unit 110 is incorporated.

According to the present embodiment, by performing convolution AE and clustering on the estimation result of the optical flow, it is possible to extract the features of the moving body and classify the clusters. Further, according to the present embodiment, unlike the precondition technique for analyzing a moving body by statistics, since feature quantity extraction is directly performed for the moving direction and moving speed information of the moving body, more accurate analysis is possible. It becomes. Further, according to the present embodiment, by linking the outputs of the hue analysis unit 110, the suspended matter analysis unit 120, and the mobile body analysis unit 130, more accurate analysis becomes possible.

According to the present embodiment described above, by providing the moving body analysis unit using moving images, it is possible to quantitatively evaluate the positions, moving directions, and moving speeds of the flocs and scums, so that the flocs and other particles can be used. Can be discriminated.

Further, the present invention is not limited to the above-described embodiment, and includes various modifications in which components are added, deleted, or converted with respect to the above-mentioned configuration.

100 Water quality analyzer 101 Data storage unit 110 Huatric analysis unit 111 First image processing unit 112 First convolution AE learning unit 113 First convolution AE parameter storage unit 114 First feature quantity extraction unit 115 First clustering unit 116 First clustering parameter storage unit 117 First cluster classification unit 118 First feature space drawing unit 119 First learning necessity determination unit 120 Floating material analysis unit 121 Second image processing unit 122 Second convolution AE Learning unit 123 Second convolution AE parameter storage unit 124 Second feature amount extraction unit 125 Second clustering unit 126 Second clustering parameter storage unit 127 Second cluster classification unit 128 Second feature space drawing unit 129 Second 2 Learning necessity determination unit 130 Moving object analysis unit 131 Third image processing unit 132 Third convolution AE learning unit 133 Third convolution AE parameter storage unit 134 Third feature quantity extraction unit 135 Third clustering unit 136 Third clustering parameter storage unit 137 Third cluster classification unit 138 Third feature Space drawing unit 139 Third learning necessity determination unit 140 Flow estimation unit 201 Data storage unit 210 Moving object analysis unit 211 Image processing unit 212 Flow estimation unit 213 Statistics calculation unit

Claims (2)

The first convolution AE is learned for the image data of the settling pond, the hue feature amount of the settling pond is extracted, and the image data is classified into hue clusters by clustering from the hue feature amount, and the hue feature space. And the hue analysis unit that draws the cluster distribution,
For the image data, a second convolution AE different from the first convolution AE is learned to extract the feature amount of the suspended matter in the settling pond, and the image data is obtained from the feature amount of the suspended matter by clustering. A water quality analyzer equipped with a floating matter analysis unit that classifies the floating matter into clusters and draws the feature space of the floating matter and the cluster distribution. The convolution AE is learned from the image data, and the features of the moving direction and the moving speed obtained by the dense optical flow of the settling pond are extracted, and the image data is classified into a cluster of moving bodies by clustering from the features . The water quality analyzer according to claim 1, further comprising a moving body analysis unit that draws a feature space and a cluster distribution of the moving body.

Images