JP2019152968A - Sewage overflow detection device, sewage overflow detection method, program and sewage treatment device - Google Patents

Abstract

To provide a sewage overflow detection device which automatically and surely detects whether sewage overflows or not in sewage treatment within a reaction vessel provided with a partition board.SOLUTION: A sewage overflow detection device 100 is mounted in a sewage treatment device having a reaction vessel 1 for sewage treatment and partition boards 7 partitioning the inside of the reaction vessel 1 into a plurality of regions, to detect whether sewage overflows beyond the partition boards 7. The sewage overflow detection device 100 comprises: image acquisition means which acquires an object image including the partitioning board 7; image processing means which processes the acquired object image into an input image; feature extraction means which extracts a feature quantity by performing prescribed processing on information constituting the input image and finally calculates a specific value; and overflow determination means which determines whether sewage overflows beyond the partitioning boards 7 or not on the basis of the specific value.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a sewage overflow detection apparatus, a sewage overflow detection method, a program, and a sewage treatment apparatus.

  Conventionally, a reaction tank for biologically treating sewage, a membrane separation apparatus for removing solid matter from sewage immersed in the reaction tank and biologically treated, and a membrane separation apparatus installed at the lower part of the membrane separation apparatus There is known a sewage treatment apparatus including a diffuser pipe for supplying a gas such as air. The biological treatment of sewage in the reaction tank is performed based on, for example, an activated sludge method in which sewage is treated with organic sludge containing microorganisms, that is, so-called activated sludge.

  Specifically, in the activated sludge method, a nitrification reaction is performed in which ammonia is converted into nitrous acid or nitric acid in the presence of oxygen (aerobic state). Furthermore, a denitrification reaction is performed in which nitrous acid or nitric acid converted from ammonia by nitrification reaction is converted to nitrogen. Unlike the nitrification reaction performed in the presence of oxygen (aerobic state), the denitrification reaction needs to be performed in an oxygen-free state. The denitrification reaction may be performed in a reaction tank different from the reaction tank in which the nitrification reaction is performed, but in order to save space in the sewage treatment apparatus, the nitrification reaction and the denitrification reaction are performed in a single reaction tank. A sewage treatment apparatus to be performed has been proposed (see, for example, Patent Document 1).

  FIG. 8 is a diagram schematically showing a conventional sewage treatment apparatus. The sewage treatment apparatus of FIG. 8 includes a reaction tank 1 that performs a nitrification reaction in an aerobic state and a denitrification reaction in an oxygen-free state, and a raw water tank 9 for supplying sewage to the reaction tank 1. 1 has a partition plate 7 for partitioning the inside of the reaction tank 1 into a plurality of compartments. Specifically, the reaction tank 1 is partitioned into a sewage area A surrounded by the partition plate 7 and a sewage area B surrounded by the partition plate 7 and the inner wall of the reaction tank 1, and the sewage area A is separated from the membrane separation device 2 and the scatter water. It has a trachea 4. The reaction tank 1 also includes a sewage supply start water level LWL (Low water level) for starting the supply of sewage from the raw water tank 9 and a sewage supply stop water level HWL for stopping the supply of sewage from the raw water tank 9. (High water level). The sewage has a level higher than the upper end of the partition plate 7 (hereinafter referred to as “sewage overflow position”) and a position lower than the upper end of the partition plate 7 (hereinafter referred to as “sewage non-overflow position”). ) To increase or decrease in the reaction tank 1.

  FIG. 9 is a diagram schematically showing the flow of sewage when the level of sewage in the reaction tank 1 in FIG. 8 is at the sewage overflow position. For example, when the sewage water level is at the sewage overflow position, the sewage circulates around the partition plate 7 by the air supplied from the air diffuser 4 to the membrane separation device 2 over the upper end of the partition plate 7. A circulating flow is formed (overflow condition). By this circulation flow, nitric acid in the sewage area A is transferred to the sewage area B and converted into nitrogen gas by a denitrification reaction. On the other hand, when the sewage water level is at the sewage non-overflow position, the circulation of the sewage is divided between the sewage area A and the sewage area B. Therefore, even if the air diffuser 4 supplies air to the membrane separation device 2 A circulation flow circulating around the partition plate 7 is not formed (divided state). As a result, although the nitrification reaction is performed in an aerobic state in the sewage region A, the generated nitric acid does not transfer to the sewage region B, and the denitrification reaction is performed in the sewage region B while maintaining an oxygen-free state. .

JP 2004-261711 A Japanese Patent Laying-Open No. 2006-168046 JP 2017-157138 A

  In the sewage treatment apparatus having such a partition plate, the overflow time during which the sewage flows over the upper end of the partition plate is an important factor for the nitrogen removal capability. The overflow time varies depending not only on the sewage supply start water level LWL and the sewage supply stop water level HWL, but also on the raw water inflow rate, the membrane treatment rate, and the aeration intensity, so it cannot be simply calculated. In order to stably measure such overflow time for a long period of time, it is necessary to accurately grasp whether the sewage level is at the sewage overflow position or the non-sewage overflow position. In 1, the sludge concentration is high and the sewage is extremely dirty, so it is difficult to accurately grasp the water level of the sewage.

  At present, many sensors can be used for detection of overflow, and for example, detection of the water surface level by electrodes or optical sensors is performed. However, in a sewage environment, sludge or dust may get caught between electrodes, or sludge may adhere to the surface of the optical sensor, which may cause the electrode or optical sensor to malfunction. It is enough and not put into practical use. For this reason, overflow is confirmed exclusively by human eyes, but there is also a demand for manpower reduction, and a method for automatically and reliably detecting overflow of sewage is required.

  On the other hand, image recognition technology has made it possible to automatically monitor events that could only be observed visually. In order to recognize the object in the image, a technique called deep learning is used. As a typical technique of deep learning, a convolutional neural network (also called a convolutional neural network, hereinafter referred to as “CNN”) is used. (Abbreviated) is attracting attention. CNN is a technique in which the system automatically extracts and learns the characteristics of a target image by multi-stage arithmetic processing, and recognizes and determines the target with high accuracy. A system for diagnosing a plant disease using this CNN and an apparatus for discriminating a category of a subject in a target image have been reported (see, for example, Patent Documents 2 and 3).

  However, examples of using image recognition technology such as CNN have been reported so far in order to reliably detect the presence or absence of overflow of sewage in sewage treatment in a reaction vessel equipped with a partition plate as described above. Not.

  The present invention relates to a sewage overflow detection device, a sewage overflow detection method, a program, and sewage for automatically and reliably detecting the presence or absence of overflow of sewage in a sewage treatment in a reaction tank provided with a partition plate (partition wall). An object is to provide a processing apparatus.

  In order to achieve the above object, the sewage overflow detection device of the present invention is configured such that the sewage includes a partition disposed between a first region where sewage exists and a second region other than the first region. A sewage overflow detection device that detects whether or not it overflows, an image acquisition unit that acquires a target image including the partition, an image processing unit that processes the acquired target image into an input image, and the input Feature extraction means for extracting a feature value by performing predetermined processing on information constituting the image, and finally calculating a specific value, and whether the sewage overflows the partition based on the specific value Overflow discrimination means for discriminating whether or not.

  In order to achieve the above object, in the method for detecting sewage overflow according to the present invention, the sewage has a partition disposed between a first region where sewage exists and a second region other than the first region. A sewage overflow detection method for detecting whether or not it overflows, an image acquisition step for acquiring a target image including the partition wall, an image processing step for processing the acquired target image into an input image, and the input A feature extraction step of extracting a feature amount by applying predetermined processing to information constituting the image and finally calculating a specific value, and whether the sewage overflows the partition based on the specific value And an overflow determining step for determining whether or not.

  In order to achieve the above object, according to the program of the present invention, is the sewage overflowing a partition wall disposed between the first region where the sewage exists and the second region other than the first region? A program for causing a computer to execute a sewage overflow detection method for detecting whether or not, an image acquisition module for acquiring a target image including the partition wall, an image processing module for processing the acquired target image into an input image, A feature extraction module that extracts a feature amount by performing predetermined processing on information constituting the input image and finally calculates a specific value, and the sewage overflows the partition wall based on the specific value. And an overflow detection module for determining whether or not the vehicle is overflowing.

  In order to achieve the above object, in the sewage treatment apparatus of the present invention, the sewage overflows a partition disposed between a first region where sewage exists and a second region other than the first region. A sewage overflow detection device that detects whether or not the sewage overflow detection device includes an image acquisition unit that acquires a target image including the partition wall, and the acquired target image. Image processing means for processing into an input image, feature extraction means for extracting a feature value by applying predetermined processing to information constituting the input image, and finally calculating a specific value; and And an overflow determining means for determining whether or not the sewage is overflowing the partition.

  ADVANTAGE OF THE INVENTION According to this invention, the presence or absence of the overflow of sewage can be detected automatically and reliably in the sewage treatment in the reaction tank provided with the partition plate (partition wall).

It is a figure which shows roughly a sewage treatment apparatus provided with the sewage overflow detection apparatus which concerns on embodiment of this invention. It is a block diagram which shows the function structure of the sewage overflow detection apparatus in FIG. It is a flowchart which shows the procedure of the intermediate | middle layer process implemented by the feature extraction means in FIG. It is a photograph which shows the input image (overflow state) from which a feature is extracted by the feature extraction means in FIG. It is a photograph which shows the input image (partition state) from which the feature is extracted by the feature extraction means in FIG. It is a figure which shows the outline of the CNN model used in order to demonstrate the predetermined | prescribed process implemented by the feature extraction means in FIG. It is a figure which shows transition of the correct answer rate of learning data and the test data in the sewage overflow detection apparatus of FIG. It is a figure which shows the conventional sewage treatment apparatus schematically. It is a figure which shows roughly the flow of the sewage at the time of the overflow in the reaction tank in FIG.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a diagram schematically showing a sewage treatment apparatus including a sewage overflow detection apparatus 100 according to an embodiment of the present invention.

  The sewage treatment apparatus of FIG. 1 includes a single tank type reaction tank 1 for treating sewage, and the reaction tank 1 includes a membrane separation apparatus 2 that separates contaminants contained in the sewage, and bubbles in the membrane separation apparatus 2. A diffuser pipe 4 for supplying the air-like air and a partition plate 7 (partition wall) for partitioning the inside of the reaction tank 1 into a plurality of regions. The arrangement of the partition plate 7 in the reaction tank 1 is not particularly limited, and the partition plate 7 may surround the entire periphery of the membrane separation device 2, or the partition plate 7 may be membrane-separated together with the tank wall of the reaction tank 1. The periphery of the device 2 may be enclosed. The membrane separation device 2 is connected to a suction pump outside the reaction tank 1. When the suction pump is driven, the biologically treated sewage is filtered by the membrane separation device 2, and the filtered water is taken out of the reaction tank 1.

  The air diffuser 4 is installed at the lower part of the membrane separation device 2 and is connected to a blower outside the reaction tank 1. The blower supplies air to the air diffuser 4. Since the membrane separator 2 filters sewage, sludge substances in the sewage adhere to the membrane surface of the membrane separator 2, and if the sludge substances in the sewage attached to the membrane surface of the membrane separator 2 are left, The membrane separation device 2 is clogged, and the sewage cannot be appropriately filtered. Therefore, the air diffuser 4 supplies air to the membrane surface of the membrane separation device 2 and prevents sludge substances and the like from adhering to the membrane surface of the membrane separation device 2. The reaction tank 1 is connected to a raw water tank 9 for storing sewage through a raw water pump. When the raw water pump is driven, the sewage to be treated is supplied from the raw water tank 9 to the reaction tank 1.

  In the reaction tank 1, sewage has a level higher than the upper end of the partition plate 7 (hereinafter referred to as “sewage overflow position”) and a position lower than the upper end of the partition plate 7 (hereinafter referred to as “sewage non-overflow”). It is increased or decreased so as to come and go between “flow position”. When the level of sewage in the reaction tank 1 is at the sewage overflow position, air is supplied from the air diffuser 4 to the membrane separation device 2, so that the sewage is separated from the inner region A (first region) into the partition plate 7. It overflows the upper end and moves to the external region B (second region) outside the partition plate 7 (overflow state). Thereafter, the sewage descends in the external region B, returns to the internal region A of the partition plate 7 through a region below the partition plate 7. That is, when the sewage water level is at the sewage overflow position and the sewage overflows the partition plate 7, a circulation flow that circulates around the partition plate 7 is formed. When the circulating flow is formed, nitric acid in the inner region A moves to the outer region B, and is converted into nitrogen gas in the outer region B, and a denitrification reaction is performed.

  On the other hand, when the sewage water level is at the sewage non-overflow position, the circulation of the sewage is divided between the sewage area A and the sewage area B. Therefore, even if the air diffuser 4 supplies air to the membrane separation device 2 A circulation flow circulating around the partition plate 7 is not formed (divided state). As a result, the nitrification reaction is performed in the aerobic state in the sewage region A, and the denitrification reaction is performed in the oxygen-free state in the sewage region B. In such a partition-insertion-type membrane separation activated sludge treatment apparatus (B-MBR), the sewage overflow state and the divided state are repeated alternately.

  The sewage overflow detection device 100 of the present invention in FIG. 1 is a device that detects whether sewage overflows the upper end of the partition plate 7 in the reaction tank 1. The imaging device 1011 is image acquisition means 101 (described later) included in the sewage overflow detection device 100.

  FIG. 2 is a block diagram schematically showing the configuration of the sewage overflow detector 100 according to the embodiment of the present invention.

  The sewage overflow detection device 100 in FIG. 2 has a hardware configuration such as a photographing device 1011 such as a camera as an image acquisition unit 101 that acquires a target image including a partition plate 7, and a CPU, ROM, RAM, an external storage device, and the like. Which are connected to each other. A storage area (storage) for storing image data acquired by the imaging apparatus 1011 is connected to the imaging apparatus 1011.

  The CPU expands and executes a program stored in a ROM, an external storage device, or the like on a RAM that is a work memory of the CPU. The external storage device is, for example, an HDD or a memory card, and stores various types of data such as various programs necessary for the CPU to execute processing and data related to thresholds described later.

  The image acquisition unit 101 transfers the image data captured by the imaging apparatus 1011 to the CPU, or stores the image data captured by the imaging apparatus 1011 in a storage area (storage) connected to the imaging apparatus 1011. Transfer the image data to the CPU.

[Outline of Image Recognition Processing by Sewage Overflow Detection Device 100]
Below, the outline of the image recognition process by the sewage overflow detection apparatus 100 of this invention is demonstrated.

(Image acquisition means 101)
The image acquisition unit 101 acquires a target image including the partition plate 7 in the reaction tank 1. Since the sewage overflow detection device 100 of the present invention is a device that detects whether sewage overflows the upper end of the partition plate 7, the target image is at least a part of the upper end of the partition plate 7 in the reaction tank 1. And its surroundings. The target image may be acquired by photographing with a photographing apparatus 1011 such as a camera, or may be obtained from a storage area (storage) in which image data is stored. The target image is a still image or an image of one frame in a moving image. The installation location of the photographing device 1011 such as a camera is not particularly limited as long as it can photograph at least a part of the upper end of the partition plate 7 and the periphery thereof, but usually an image effective for determining the overflow state or the divided state is used. It is fixed at a fixed position where continuous shooting is possible.

  The size of the target image acquired by the image acquisition unit 101 is not particularly limited. For example, vertical pixels × horizontal pixels are 400 to 2200 × 600 to 4000 pixels. Here, the vertical pixel × horizontal pixel is hereinafter also referred to as height × width. The form of the acquired image data is not particularly limited, but the form of RGB luminance data is preferable in the present invention. The acquired target image data is transferred to the image processing means 102. Such a process by the image acquisition unit 101 is referred to as an image acquisition step.

(Image processing means 102)
The image processing unit 102 performs processing necessary for performing arithmetic processing in the feature extraction unit 103 described later on the target image data acquired by the image acquisition unit 101. Examples of the processing include image size change such as image enlargement, reduction, and trimming, and image quality adjustment such as brightness adjustment and color adjustment. Only one type of processing may be performed, or a plurality of processing may be combined. The processed image data is input to the feature extraction unit 103 as an input image. The vertical pixel × horizontal pixel of the input image is preferably 50 to 200 × 100 to 400 pixels. Such a process by the image processing means 102 is referred to as an image processing step.

(Feature extraction means 103)
The feature extraction unit 103 extracts a feature amount of the input image by performing predetermined processing on the information constituting the input image, and finally calculates a new specific value (feature amount). Here, the predetermined process is at least one process selected from a convolution calculation process, a pooling calculation process, and a fully coupled calculation process, and the convolution calculation process and the pooling calculation process include at least information constituting the input image. It is preferable that the treatment is performed once.

  The convolution calculation process is a process of extracting a local feature amount of the input image data by performing a filtering process on the input image data using a weight parameter called kernel C.

  The pooling calculation process is a process for compressing data while leaving important features by taking a maximum value or an average value in an area for image data (feature amount). By this process, the positional deviation of the image can be absorbed, and robustness against minute distortion and parallel movement in the image is ensured.

  The fully connected arithmetic processing is processing for combining all image data into one node, converting the image data into one-dimensional data using an activation function, and outputting the data. This is a process that performs classification based on feature amounts and plays a role of identifying images.

  In the present invention, the predetermined processing in the feature extraction unit 103 is preferably performed according to the CNN method (convolutional neural network). The CNN method has a hierarchical structure having an input layer, an intermediate layer, and an output layer. Input data is received in the input layer, feature quantities are extracted from the data in the intermediate layer, and final results are output in the output layer. In the intermediate layer of the CNN method, multi-stage calculation steps including a convolution layer, a pooling layer, and a fully connected layer are combined. In the CNN method, complicated feature expression can be obtained through such multi-stage feature conversion, so that the image of the upper end portion of the partition plate 7 can be recognized with high accuracy.

  The feature extraction unit 103 combines the convolution calculation process, the pooling calculation process, and the fully coupled calculation process in a specific order as the predetermined process. In the present invention, (1) convolution operation processing, (2) pooling operation processing, (3) multiple convolution operation processing, and (4) It is preferable to perform all combined arithmetic processing in this order. (3) The convolution calculation process of multiple times is normally implemented 2-9 times, Preferably it is implemented 2-7 times. Parameters used for each calculation in the feature extraction unit 103 are learned in advance so that overflow can be determined by using a probability gradient method or the like.

  By performing such predetermined processing step by step on the information constituting the input image, the feature amount of the input image is extracted, and finally a specific value is calculated as one feature amount between 0 and 1. . Such a process by the feature extraction unit 103 is referred to as a feature extraction step.

(Overflow discrimination means 104)
In the overflow determination unit 104, the specific value (feature amount) between 0 and 1 obtained by the feature extraction unit 103 is compared with a predetermined threshold (for example, 0.5), so that the sewage is separated from the partition plate. It is determined whether or not the vehicle overflows 7. Such a process by the overflow determination unit 104 is defined as an overflow determination step.

(Determination result output means 105)
Finally, the discrimination result output means 105 outputs the discrimination result as 0 (non-overflow state) or 1 (overflow state) by a desired method. Such a process by the discrimination result output means 105 is defined as a discrimination result output step.

  The above is the outline of the image recognition processing by the sewage overflow detection apparatus 100 according to the present embodiment.

[Details of Image Recognition Processing by Sewage Overflow Detection Device 100]
Next, a detailed flow of the image recognition processing by the sewage overflow detection apparatus 100 according to the embodiment of the present invention will be described using an example of a specific processing method.

  First, one target image including at least a part of the upper end of the partition plate 7 in the reaction tank 1 and its periphery is acquired from the camera 1011 in the form of RGB luminance data (image acquisition step). The size of the target image is 480 × height 640. One piece of image data has three height × width determinants of red (Red), green (Green), and blue (Blue).

Next, the acquired target image is trimmed to change the height to 350 × width 400, and then reduced to obtain the input image G _{0 in} which the image size is changed to height 116 × width 133 (image processing step). ). FIG. 4 shows an example of the input image in the overflow state. FIG. 5 shows an example of an input image in a divided state. In image data, a large amount of information such as color component information is given to one pixel position, and such a degree of freedom is called a channel. The number of channels of the input image G ₀ is a 3 channel of red, green and blue.

Next, a predetermined process using the CNN method is performed on the input image G ₀ , and an arithmetic process with a hierarchical structure having an input layer, an intermediate layer, and an output layer is performed (feature extraction step). FIG. 3 is a flowchart showing a combination of arithmetic processing in the intermediate layer of the CNN method. FIG. 6 is a diagram showing an outline of processing of the input layer, the intermediate layer, and the output layer in the CNN method.

Specifically, the input image G ₀ (height 116 × width 133, channel number 3) obtained in the image processing step is input as two-dimensional data to the input layer of the feature extraction unit 103, and then input to the intermediate layer. Is done. In the intermediate layer, as shown in FIG. 3, the convolution operation 1 (step 1031), the pooling operation (step 1032), the convolution operations 2 to 7 (steps 1033 to 1038), and the fully coupled operation (step 1039) are performed in this order. Will be implemented.

In step 1031, a convolution operation 1 is performed. In the convolution operation, filter processing is performed on the input image data using a weight parameter called kernel C _{i, and} a local feature amount of the input image data is extracted. Here, the kernel C _i represents the value of the filter in the i-th convolution operation. Specifically, the numerical value of the input image data is convolved by multiplying the numerical value of the input image data by the filter value (kernel). The convolution operation is expressed by the following equation (1).

G _i = G _i−1 * C _i (1)
(In Expression (1), G represents a feature value, C represents a kernel, and i represents the number of convolution operations.)

In convolution operation 1, first the value of the weight parameters of the kernel C ₁ convolution is read from the external storage device. The size of the convolution kernel _{C 1} may be selected from a two-dimensional matrix of 8 to 15 × 8 to 15, but here, the 11 × 11. The stride of the convolution operation 1 (the width for shifting the application position of the kernel) can be selected from 1 to 4, but is 2 here. The kernel size and stride may be changed according to the size of the input image and the size of the image after convolution. If after convolution desired to obtain an image of a multi-channel (M-channel) it can be used M types of kernel, the convolution operation 1, the convolution kernel C ₁ 32 type is used. Red performs convolution multiplies the filter (kernel), respectively for the three-channel input image G ₀ of the green and blue pixel values of subsequent same position bring added over the entire channel. This process is performed for 32 types of kernels.

  If the size after convolution is desired to be increased to some extent, padding may be performed to add a wrinkle of a predetermined thickness around the input image. For example, zero padding that fills the input image with zeros can be performed.

By such a convolution operation 1, the feature amount G ₀ is reduced by convolving the matrix while retaining the feature, and finally, the activation function f is applied to the convolved image data in a non-linear manner. Convert. For example, a sigmoid function, a hyperbolic function (tanh), a ReLU function, or the like is used as the activation function f. Result of convolution operation 1, the feature amount _{G 1} of the channel number 32 × height 53 × width 62 is obtained. The obtained feature value G ₁ is defined as a convolution layer 1. Details of the convolution operation are described, for example, in Deep Learning (Machine Learning Professional Series) Okaya Takayuki (Author).

In the next step 1032, a pooling operation is performed, and the image data (feature value G ₁ ) obtained by the convolution operation 1 is compressed to reduce the number of pixels. The pooling process includes a maximum pooling process that takes a maximum value in a predetermined area and an average pooling process that takes an average value. In the present invention, it is preferable to perform a maximum pooling operation. The size of the predetermined area can be selected from 1 to 6 × 1 to 6, and the stride (the width for shifting the predetermined area) can be selected from 1 to 5, but these are set appropriately according to the number of pixels after compression, etc. it can. In step 1032, the predetermined area is 3 × 3 and the stride is 3. In addition, when it is desired to increase the size after the pooling process to some extent, padding that adds wrinkles of a predetermined thickness around the input image may be performed. For example, zero padding that fills the input image with zeros can be performed. By this pooling processing, image data (feature amount G ₁ ′) having 32 channels × 17 height × 20 width is obtained. The obtained feature amount G ₁ ′ is used as a pooling layer.

  Steps 1033 to 1038 are processes for repeatedly performing a convolution operation. In the arithmetic processing of FIG. 3, the convolution operations 2 to 7 are performed six times. The processing methods of the convolution operations 2 to 7 are the same as those of the convolution operation 1 of step 1031 except that the parameters such as the feature amount (number of channels × height × width), kernel size, stride and kernel type of the input image are different. This is the same as the processing method. Table 1 below shows parameters that can be used in the convolution operations 2 to 7.

By the convolution operations 2 to 7, the feature amount G ₁ obtained by the convolution operation 1 is successively reduced while the features are retained, and the convolution layer has the feature amounts as shown in Table 2 below. 2-7 are obtained. In addition, you may implement batch normalization (Batch Normalization) after each convolution layer.

In the next step 1036, a fully coupled operation is performed. Here, non-linear calculation processing using the activation function θ is performed on the feature amount G ₇ obtained by the convolution calculation 7. As the activation function θ, various forms such as a sigmoid function, a half-wave rectification function, a piecewise linear convex function, and a ReLU function (normalized linear function) can be used. In the present invention, the activation function θ is expressed by the following formula (2). It is preferable to use a sigmoid function.

G ₈ = θ (G ₇ (p)),
θ (x) = 1 / (1 + e ^−X ) (2)
(Wherein, G represents a feature quantity, theta represents a sigmoid function, p is representing each pixel of the G _7.)

With this fully coupled calculation process, 64-node image data (feature value G ₈ ) is obtained. The characteristic quantity G ₈ obtained here and the total binding layer. In the fully connected arithmetic processing, dropout may be performed to invalidate some of the nodes in the target layer. By performing dropout, it is possible to forcibly reduce the degree of freedom of the network during learning to improve generalization performance and avoid overlearning.

Feature quantity G ₈ obtained by total bound calculation process, then the output layer, is converted into a one-dimensional real number of 0 to 1 by the activation function. As a result, a specific value is finally output as one feature amount between 0 and 1. Examples of the activation function used in the output layer include a sigmoid function, a hyperbolic function (tanh), and a ReLU function.

  Next, based on a specific value between 0 and 1 finally calculated in the feature extraction step, it is determined whether or not sewage overflows the partition plate 7 (overflow determination step). As a specific determination method, if the specific value is 0.5 or more, it is determined that the overflow state (1), and if the specific value is less than 0.5, the divided state (0) is determined. To do. The determination result (overflow state or divided state) determined in the overflow determination step is then output in the determination result output step.

As described above, the procedure for outputting the overflow determination result from the single target image data by the sewage overflow detector 100 has been described. In the actual sewage overflow detection method, first, the CNN model (number of channels × height × width, kernel size, and stride) is set for the input layer, intermediate layer (each processing layer) and output layer of the feature extraction means 103. For example, the table below is created in advance. Next, a number of photographs taken at the time of operation of the sewage treatment apparatus (upper end of the partition plate and its surroundings) are taken. The person in charge determines whether each photograph is overflowing or divided (labeled correctly) and classifies the photographed photographs into learning data and test data. Data is transmitted in the forward direction by applying a CNN model to the learning data for calculation. Although a parameter learning method will be described later, parameters for each calculation process (kernel C _i weighting parameters, bias, etc.) so that the error between the correct data of the learning data and the output data of the CNN model is minimized by an error back propagation method or the like. Is updated. By performing test data overflow detection using the updated parameters, it is possible to detect overflow of sewage with high accuracy.

(Method of learning the weight parameter kernel _{C i)}
Here, a learning method of the weight parameter of the convolution kernel C _i used in the feature conversion unit 103 will be described. In deep learning, the cross entropy (an error function for obtaining an error between correct data based on learning data and output data based on the CNN model) represented by the following equation (3) is defined as a loss function L = H (q, q ′). A method of adjusting the weight value by updating the weight value so that the error is minimized is widely known.

H (q, q ') = -Σ x q (x) · Log q' (x) ··· (3)
In equation (3), q (x) is a true probability distribution of category x (in the case of overflow), and q ′ (x) is category x (in the overflow state) estimated by the CNN model (recognition system). Distribution). Here, it is assumed that the CNN model includes the computation of the kernel C _i as a part.

Specifically, first, the weights W _i of all the convolution kernels C _i are initialized with random numbers. Here, W _i is c _i × d _{i + 1} × d _i weight variables, c _i is the number of channels, d _i and d _{i + 1} are the number of dimensions of the feature quantity before and after the convolution operation, respectively. is there. Next, an estimated distribution q ′ (x) of category x is calculated for each pixel of each learning image from the output of the CNN model obtained by inputting the learning image data. Then, the value of the j-th element w _nj of the weight W _n is updated according to the following equation (4).

w _nj (t + 1) = w _nj (t) −η∂L / ∂w _nj (t),
L = Σ _i Σ _p L _ip (4)
In Expression (4), L _ip is a loss function related to the pixel p of the learning image i. η is a learning coefficient having a value smaller than 1. The weight parameter W _i other than the final layer may be updated sequentially calculated for each layer by the error back propagation method is a common technique in neural net. Also present various forms of derivative such as the type plus a term which in the above update equations (4) called the attenuation term of the inertia term and the weight w _n. The individual elements of the learning calculation shown here are widely known as deep learning techniques (see, for example, Deep Learning (Machine Learning Professional Series) Takayuki Okaya).

  Here, a form of a learning method called supervised learning that uses data including correct data as learning data as input data has been described. However, in the present invention, parameters are learned in various forms, such as a form in which only the middle layer performs unsupervised learning or a form in which supervised learning is added step by step from a layer close to the input layer. Can do.

(program)
In the sewage overflow detection device 100 of the present invention, the overflow of sewage is detected using a program that causes a computer to detect whether or not sewage is overflowing the partition plate 7 as described below. .

That is, the program of the present invention is
An image acquisition module for acquiring a target image including the partition plate 7;
An image processing module that processes the acquired target image into an input image;
A feature extraction module that extracts a feature amount by applying predetermined processing to information constituting the input image, and finally calculates a specific value;
An overflow determination module that determines whether sewage overflows the partition plate 7 based on a specific value.

  The source code of the image acquisition module, image processing module, feature extraction module, and overflow detection module constituting the program of the present invention is the image acquisition means 101 and image processing means 102 that constitute the sewage overflow detection device 100 of the present invention. It can be easily created by a known method using the above-described image processing conditions used in the feature extraction unit 103 and the overflow detection unit 104 (for example, deep learning (machine learning professional series) Takayuki Okaya (Author)) reference).

  The sewage overflow detection device 100 of the present invention may be installed as a separate device from the sewage treatment device itself, or may be configured as one of the components of the sewage treatment device in the same manner as the reaction tank 1 and the partition plate 7. It may be. If the sewage overflow detection device 100 of the present invention is used in a sewage treatment device including the reaction tank 1 and the partition plate 7, it is not necessary to be directly joined to the sewage treatment device.

  Hereinafter, although an example of the present invention will be specifically described based on examples, the present invention is not limited to the following examples.

1. B-MBR treatment system In the partition-insertion-type membrane separation activated sludge treatment apparatus (B-MBR) shown in Fig. 1, sewage was continuously treated using the following treatment conditions.
(Sewage treatment conditions)
Volume of membrane reactor: 3.7 m ³
HRT: 5.9 hours SRT: 26.6 days MLSS concentration: 9000 to 11000 mg / L
^Two modules of PTFE membrane (membrane pore diameter 0.2 μm) (area 30 m ² ) are used. Membrane flux: 0.25 m ³ / m ² / day membrane aeration air amount: 18 m ³ / hour (SAD _m : 0.30 _m ³⁾ / M ² / hour equivalent)

2. Creation of CNN model The CNN model shown in Table 3 below was created. Specifically, a CNN model is defined by using the deep learning library Torch, and (1) a convolution layer, (2) a maximum pooling layer, (3) a plurality of convolution layers, and (4) a fully connected layer. A CNN model that sequentially configures the above is set. Table 3 shows the number of channels × height × width, kernel size and stride of each layer. The following CNN model is a deep learning model with a total of 11 layers including an input layer and an output layer.

3. Creation of learning data In B-MBR, a camera was fixed at a position where a photograph including at least a part of the upper end of the partition plate in the reaction tank and its periphery could be taken. Next, during the operation of the B-MBR, 1000 photographs were taken every 10 seconds, and a target image of height 480 × width 640 was obtained in the form of RGB luminance data (image acquisition step). The acquired target image was trimmed and changed to a pixel having a height of 350 × width 400, and then reduced to obtain an input image having a height 116 × width 133 (image processing step).

The person in charge determined whether each photograph was overflow (1) or divided (0) and labeled it (correct data). Since the denitrification reaction proceeds and nitrogen is generated in the outer region of the partition plate 7, the surface of the sewage is foamed. Therefore, it is difficult to clearly confirm the upper end of the partition plate 7 in the overflow state. Therefore, the determination criterion of whether the person in charge is overflowing is a divided state when the upper end portion of the partition plate 7 can be clearly confirmed, and an overflow state when the upper end portion of the partition plate 7 cannot be clearly confirmed. . FIG. 4 shows one of the input images determined as the overflow condition (1). FIG. 5 shows one of the input images determined to be in the divided state (0).
4). Learning and overflow detection Among the 1000 photographs taken, 900 pictures were taken as learning data, and the remaining 100 pictures were taken as destination data. The CNN model shown in Table 1 was applied to the input image of learning data to perform sequential calculation processing (feature extraction step). After each convolution layer, batch normalization was performed, and 50% dropout was performed in all tie layers. In the output layer, 0 to 1 one-dimensional real numbers were output. When the output value of the output layer is 0.5 or more, the overflow state (1) is determined, and when it is less than 0.5, the overflow state is determined (overflow determination step). In the learning, for each of 900 photos, each weight parameter is updated so as to minimize the loss function based on the correct answer data attached to each photo. This series of work was set as 1 epoch, and learning was performed up to 50 epochs. After each epoch, the correct answer rate was obtained for 100 pieces of test data not used for learning, and the transition was confirmed as shown in FIG. 7 for learning. The updated parameters were used to determine the overflow of the test data image. FIG. 7 shows the transition of the accuracy (Accuracy) of each data with respect to the number of epochs (number of learning), that is, the correct answer rate.

5. Results From the results shown in FIG. 7, the correct answer rate was 100% for the learning model, and the correct answer rate was about 95% for the test data. The boundary between the overflow state and the divided state is not always clear, and the person in charge often makes a mistake. Therefore, the correct answer rate of 95% reproduces the judgment of the person in charge with high accuracy. Therefore, according to the present invention, the overflow time representing the nitrogen removal performance of the B-MBR treatment system can be measured.

(Other embodiments)
The present invention supplies software (program) that realizes the functions of the above-described embodiments to a system or apparatus via a network or various storage media, and a computer (CPU, MPU, etc.) of the system or apparatus reads and executes the program. It is processing to do. The present invention may be applied to a system composed of a plurality of devices, or may be applied to an apparatus composed of one device. Further, the present invention is not limited to the above-described embodiments, and various modifications can be made based on the spirit of the present invention, and they are not excluded from the scope of the present invention. That is, all the structures which combined the Example mentioned above and its modification are also included in this invention, for example, it has the sewage overflow detection apparatus 100 for discriminating whether at least sewage overflows the partition plate 7 or not. The configuration is included in the present invention.

  The present invention can provide a sewage treatment apparatus and method that can stably perform sewage treatment.

DESCRIPTION OF SYMBOLS 1 Reaction tank 2 Membrane separation apparatus 4 Aeration pipe 7 Partition plate 9 Raw water tank 100 Sewage overflow detection device 101 Image acquisition means 102 Image processing means 103 Feature extraction means 104 Overflow discrimination means 105 Discrimination result output means 1011 Imaging device (camera)
1031 Convolution operation 1
1032 Pooling operation 1033 Convolution operation 2
1034 Convolution operation 3
1035 Convolution 4-7
1036 Fully combined operation

Claims (7)

A sewage overflow detector for detecting whether or not the sewage overflows a partition wall disposed between a first region where sewage exists and a second region other than the first region,
Image acquisition means for acquiring a target image including the partition;
Image processing means for processing the acquired target image into an input image;
A feature extraction unit that extracts a feature amount by performing a predetermined process on information constituting the input image, and finally calculates a specific value;
And a sewage overflow detection device, comprising: overflow determining means for determining whether or not the sewage overflows the partition wall based on the specific value.   The sewage overflow detection device according to claim 1, wherein the predetermined process is at least one process selected from a convolution calculation process, a pooling calculation process, and a fully coupled calculation process.   The sewage overflow detection device according to claim 2, wherein the convolution calculation process and the pooling calculation process are performed at least once on information constituting the input image.   Information constituting the input image includes (1) the convolution operation process, (2) the pooling operation process, (3) a plurality of convolution operation processes, and (4) the fully coupled operation process in this order. The sewage overflow detector according to claim 2 or 3, wherein the sewage overflow detector is provided. A sewage overflow detection method for detecting whether or not the sewage overflows a partition disposed between a first region where sewage exists and a second region other than the first region,
An image acquisition step of acquiring a target image including the partition;
An image processing step of processing the acquired target image into an input image;
A feature extraction step of extracting a feature amount by applying a predetermined process to information constituting the input image, and finally calculating a specific value;
An overflow determination step for determining whether or not the sewage overflows the partition wall based on the specific value. The computer executes a sewage overflow detection method for detecting whether or not the sewage overflows a partition wall disposed between a first region where sewage exists and a second region other than the first region. A program to
An image acquisition module for acquiring a target image including the partition;
An image processing module for processing the acquired target image into an input image;
A feature extraction module that extracts a feature amount by applying a predetermined process to information constituting the input image, and finally calculates a specific value;
An overflow detection module that determines whether or not the sewage overflows the partition based on the specific value. Sewage provided with a sewage overflow detection device for detecting whether or not the sewage overflows a partition disposed between a first region where sewage exists and a second region other than the first region. A processing device comprising:
The sewage overflow detector is
Image acquisition means for acquiring a target image including the partition;
Image processing means for processing the acquired target image into an input image;
A feature extraction unit that extracts a feature amount by performing a predetermined process on information constituting the input image, and finally calculates a specific value;
Overflow discrimination means for discriminating whether or not the sewage overflows the partition wall based on the specific value.