JP7564726B2 - Anomaly detection device, anomaly detection method, anomaly detection program, and base model - Google Patents

Description

The present invention relates to an anomaly detection device, an anomaly detection method, an anomaly detection program, and a base model used for anomaly detection, for detecting anomalies occurring in waste treatment facilities.

Waste treatment facilities that treat waste are configured to detect abnormalities in the facility if any occur, in order to incinerate waste stably and appropriately.

For example, Patent Document 1 discloses a combustion device that constantly monitors whether an abnormality such as raw material clogging the pipes has occurred by photographing the inside of the pipes and having a monitor check the photographed images.

Patent document 2 discloses a waste incinerator that detects whether the waste level in a waste storage container is being kept constant by photographing the inside of the container, determining the waste level (surface height of the waste) from the image data, and checking changes in the waste level over time.

Patent Document 3 also discloses a waste incinerator that has a vibration sensor installed in the piping and detects blockages in the piping based on the detection value of the vibration sensor.

JP 2011-75226 A JP 2014-10091 A JP 2011-102785 A

However, the abnormality detection means disclosed in Patent Document 1 requires constant human monitoring, which is time-consuming to detect abnormalities. Furthermore, when detecting abnormalities using images, shadows of fluids flowing through pipes, etc. are cast, making it difficult to distinguish between the fluids (haze) and the shadows, which can make it difficult to detect abnormalities. In particular, when detecting abnormalities automatically from images, the haze can have a significant impact on the detection accuracy.

In addition, while various types of abnormalities occur in waste incineration facilities, the abnormality detection means disclosed in Patent Document 2 is configured to detect abnormalities by detecting whether the amount of waste is being kept constant, and is therefore specialized as a means for detecting abnormalities in the amount of stored waste, making it difficult to apply to detecting various types of abnormalities and low versatility.

Similarly, the abnormality detection means disclosed in Patent Document 3 is configured to detect abnormalities by utilizing changes in pipe vibration caused by blockages in the pipes, making it difficult to apply the method to detect various types of abnormalities and low versatility.

As such, various types of abnormalities occur in waste incineration facilities, and currently, abnormalities are detected using algorithms that correspond to the nature of the abnormality, but there is a need for a more versatile algorithm to detect abnormalities.

The present invention aims to solve the above problems by detecting abnormalities in a highly versatile and accurate manner without much effort.

In order to achieve the above object, an anomaly detection device according to one embodiment of the present invention includes an imaging device that captures images of a predetermined portion of a waste treatment facility, a luminance data generation unit that performs luminance conversion on image data of each pixel of the captured image captured by the imaging device to generate luminance data, a noise removal unit that generates enhanced data for each of the luminance data groups by removing noise using a time series minimum filter from the luminance data generated for each of the captured images captured during a predetermined unit period, a feature calculation unit that divides the enhanced data into a plurality of local regions and calculates a feature for each of the local regions, and a predetermined learning unit that performs a feature calculation for each of the captured images. The device includes a learning input data generation unit that generates learning input data in which the feature amounts calculated for each of the multiple enhancement data generated during a period are associated with the local region, a detection input data generation unit that generates detection input data in which the feature amounts calculated based on a set of the luminance data group corresponding to the multiple captured images captured by the imaging device are associated with the local region, and an abnormality determination unit that determines that an abnormality has been detected in the part when the degree of abnormality of the feature amount related to the detection input data relative to the feature amount related to the learning input data in each of the local regions is equal to or greater than a predetermined threshold value.

The anomaly detection method according to one embodiment of the present invention includes a learning image acquisition step of acquiring a plurality of learning images by photographing a predetermined portion of a waste treatment facility multiple times in each of different predetermined unit periods within a predetermined learning period; a learning luminance data generation step of performing luminance conversion on image data of each pixel of each of the learning images to generate a plurality of learning luminance data corresponding to each of the learning images; a learning enhancement data generation step of treating the plurality of learning luminance data corresponding to the learning images photographed within each of the unit periods as a set of luminance data groups and removing noise using a time series minimum filter to generate a plurality of learning enhancement data from the plurality of sets of the luminance data groups; and a learning input step of dividing each of the learning enhancement data into a plurality of local regions, calculating a learning feature for each of the local regions, and generating a learning input in which the local regions and the learning feature are associated with each other. The method includes a learning input data generation process for generating data for each of the learning enhancement data, a detection image acquisition process for capturing the part as a plurality of detection images, a detection luminance data generation process for converting the luminance of image data for each pixel of each of the detection images to generate a plurality of detection luminance data, a detection enhancement data generation process for generating detection enhancement data from the plurality of detection luminance data by removing noise using a time series minimum filter, a detection input data generation process for dividing the detection enhancement data into the local regions, calculating a detection feature for each of the local regions, and generating detection input data in which the local regions and the detection feature are associated with each other, and an abnormality determination process for determining that an abnormality has been detected in the part when the degree of abnormality of the detection feature with respect to the learning feature in each of the local regions is equal to or greater than a predetermined threshold value.

The anomaly detection program according to one embodiment of the present invention includes a learning image acquisition function that acquires a plurality of learning images of a predetermined portion of a waste treatment facility, the plurality of learning images being photographed multiple times in each of different predetermined unit periods within a predetermined learning period; a learning luminance data generation function that performs luminance conversion on image data of each pixel of each of the learning images to generate a plurality of learning luminance data corresponding to each of the learning images; a learning enhancement data generation function that generates a plurality of learning enhancement data from a plurality of sets of the luminance data groups by treating the plurality of learning luminance data corresponding to the learning images photographed within each of the unit periods as a set of luminance data groups and removing noise using a time series minimum filter; and a learning enhancement data generation function that divides each of the learning enhancement data into a plurality of local regions, calculates a learning feature for each of the local regions, and generates learning input data in which the local regions and the learning feature are associated with each other, the learning enhancement data generation function that generates a plurality of learning enhancement data from a plurality of sets of the luminance data groups by dividing the plurality of learning enhancement data into a plurality of local regions, calculates a learning feature for each of the local regions, and generates learning input data in which the local regions and the learning feature are associated with each other, the learning enhancement data generation function that generates a plurality of learning enhancement data from a plurality of sets of the luminance data groups by using a time series minimum filter to remove noise. The computer is caused to execute a learning input data generation function that generates learning enhancement data for each learning enhancement data, a detection image acquisition function that acquires a plurality of detection images in which the part is photographed, a detection brightness data generation function that performs brightness conversion on image data of each pixel of the plurality of detection images in which the part is photographed to generate a plurality of detection brightness data, a detection enhancement data generation function that generates detection enhancement data from the plurality of detection brightness data by removing noise using a time series minimum filter, a detection input data generation function that divides the detection enhancement data into the local regions, calculates a detection feature for each of the local regions, and generates detection input data in which the local regions and the detection feature are associated, and an abnormality determination function that determines that an abnormality has been detected in the part when the abnormality level of the detection feature with respect to the learning feature in each of the local regions is equal to or greater than a predetermined threshold value.

In one embodiment of the base model, a specified portion of a waste incineration facility performs luminance conversion on image data of each pixel of multiple learning images taken during different specified unit periods within a specified learning period to generate multiple learning luminance data corresponding to each of the learning images, and generates multiple learning enhancement data from the multiple sets of luminance data groups by treating the multiple learning luminance data corresponding to the multiple learning images taken within the unit period as a set of luminance data groups and removing noise using a time series minimum filter, dividing each of the learning enhancement data into multiple local regions, calculating learning features for each of the local regions, and analyzing the multiple learning input data generated for each of the learning enhancement data by associating the local regions with the learning features to determine the average and variance-covariance matrix of the learning features for each of the local regions.

With the above configuration, by removing noise from the learning input data using a time series minimum filter that is suitable for removing noise when detecting abnormalities in the flow of fluids, it is possible to perform highly versatile and highly accurate abnormality detection of abnormalities in the flow of fluids in various locations.

Furthermore, by detecting anomalies using the degree of anomaly between multiple feature quantities related to the learning input data and feature quantities related to the detection input data, the variance in the feature quantities related to the learning input data is also taken into account, making it possible to detect anomalies with greater accuracy.

It is also preferable that the unit period and the learning period are set according to the type of abnormality detected.

With this configuration, the timing of capturing the captured image can be set as a parameter depending on the type of abnormality, making it possible to detect abnormalities with greater versatility.

The degree of anomaly may also be calculated as the Mahalanobis distance between the feature amount related to the learning input data and the feature amount related to the detection input data.

This configuration allows the degree of anomaly to be calculated appropriately, enabling more accurate detection of anomalies.

The feature amount may also be obtained by dividing each of the local regions into a plurality of blocks by dividing the local regions into equal parts vertically and horizontally in a matrix, determining the difference in brightness between each block and its adjacent blocks and the directional relationship between the blocks as a brightness gradient, and converting the brightness gradient for each of the divided blocks into a unit vector.

This configuration makes it possible to extract features in a manner that emphasizes the features when an abnormality occurs, enabling more accurate detection of abnormalities.

The abnormality detected may also be an abnormality in the flow of fluid through the area.

By using a time series minimum filter to remove noise, it is possible to remove haze, such as the shadows of moving objects, from captured images. This reduces false positives caused by haze, allowing for more accurate detection of abnormalities.

The method may further include a base model generation step for analyzing a plurality of the learning input data during anomaly detection to generate a base model in which the average and variance-covariance matrix of the learning features for each local region are calculated, and the base model may be configured to calculate the degree of anomaly of the detection features relative to the learning features when the detection input data is input, and detect the anomaly based on the degree of anomaly.

This configuration makes it easier to detect abnormalities.

FIG. 1 is a diagram illustrating an example of the configuration of a waste incineration facility. FIG. 1 is a diagram for explaining an overview of anomaly detection. FIG. 2 is a diagram illustrating an example of a functional block configuration of the anomaly detection device. FIG. 11 is a diagram illustrating a flow of performing anomaly detection. 11A and 11B are diagrams illustrating an example of noise removal from a luminance data group. FIG. 11 is a diagram illustrating an example of calculating a feature amount. FIG. 13 is a diagram illustrating an example of a detection result of a blockage abnormality at the supply port of a slag grinder.

[Waste incineration facilities]
As shown in FIG. 1, a waste incineration facility 1 includes a gasification furnace 2, a swirling flow melting furnace 3, a boiler 4, a steam turbine generator 5, an economizer 6, a bag filter 7, and a chimney 8.

The gasification furnace 2 has a sand layer 10 in the lower region, and by introducing forced air into the sand layer 10, the flowing sand is used to gasify waste and other garbage placed on the sand layer 10.

The swirling flow melting furnace 3 melts the gasified waste into molten slag at high temperatures by swirling the waste and reacting it with air. In other words, the waste is incinerated by high-temperature melting, and molten slag, which is the melted and solidified waste, is discharged.

The boiler 4 is connected to the swirling flow melting furnace 3 and generates steam by utilizing the heat of the exhaust gas discharged from the swirling flow melting furnace 3.

The steam turbine generator 5 is supplied with steam generated by the boiler 4, which then rotates a turbine to generate electricity.

The economizer 6 recovers the heat of the exhaust gas discharged from the boiler 4 and transfers it to the feed water of the boiler 4. By heating the feed water supplied to the boiler 4, the boiler efficiency is improved.

The bag filter 7 filters and collects soot and dust from the exhaust gas discharged from the boiler 4 via the economizer 6. In this way, the bag filter 7 purifies the exhaust gas discharged from the chimney 8. Activated carbon is blown into the exhaust gas just before it is introduced into the bag filter 7. By adding activated carbon, dioxins and other substances are adsorbed by the activated carbon and removed from the exhaust gas. The activated carbon is stored just before the bag filter 7, and a predetermined amount is cut out at the cut-out section 15 and blown into the exhaust gas.

The molten slag discharged from the swirling flow melting furnace 3 is crushed and reused. Specifically, the molten slag discharged from the swirling flow melting furnace 3 is solidified and stored in the storage tank 11, and then crushed in the slag grinder 12 installed below the storage tank 11.

The residue containing resources remaining in the sand layer 10 of the gasifier 2 is stored below the gasifier 2 and then discharged and transported by a screw-type discharge device 13. The transported residue is separated into resources such as iron and aluminum and impurities. The resources such as iron and aluminum are recovered and reused after the sand is separated by a sand separation screen 14.

[Outline of abnormality detection]
As described above, the waste incineration facility 1 is configured by combining various devices, and if an abnormality occurs in any of the devices, appropriate processing cannot be performed. Therefore, in this embodiment, an abnormality is detected using a detection method and detection device according to the type of abnormality, even though a common algorithm is used. The abnormality detection algorithm according to this embodiment will be described below using FIG. 2 with reference to FIG. 1.

In the abnormality detection according to this embodiment, first, a predetermined part is continuously photographed and divided into a plurality of image groups each consisting of a plurality of continuously photographed photographed images 20. Each image group is subjected to noise removal using a specific filter, and enhanced data 21 corresponding to the abnormality to be detected is generated for each image group. The filter is selected from a plurality of types of filters according to the type of abnormality to be detected.

Next, each of the emphasis data 21 is divided into local regions 54 (see FIG. 6), and learning features 22 are calculated for each local region.

When detecting an abnormality, the area is continuously photographed, and the multiple photographed images 20 are subjected to noise removal using the same filter as described above to generate enhanced data 21, and detection features 24 are calculated for each local region 54.

Then, the degree of anomaly of the detection feature 24 with respect to the learning feature 22 is calculated for each local region 54. Finally, if the degree of anomaly is equal to or greater than a predetermined threshold value for any local region 54, it is determined that an abnormality has been detected in that area.

In this way, by using a common algorithm and selecting a filter to be used in noise removal according to the anomaly to be detected, it is possible to detect a specific anomaly without using a separate detection algorithm for each anomaly. This makes it possible to detect anomalies easily, with high versatility, and with high accuracy.

First Embodiment
[Specific examples of anomaly detection]
Next, as a first embodiment, a specific configuration example for detecting an anomaly will be described with reference to FIGS. 3 to 6 while also referring to FIGS.

As shown in FIG. 3, the anomaly detection device includes an imaging device 27 that captures an image of a specific part of the waste incineration facility 1, and an anomaly detection unit 28 that detects the occurrence of a specific anomaly in this part.

The abnormality detection unit 28 includes a control unit 31, an image acquisition unit 32, a memory unit 33, a brightness data generation unit 35, a noise removal unit 36, a feature calculation unit 37, a learning input data generation unit 39, a detection input data generation unit 42, and an abnormality determination unit 43.

The control unit 31 controls the operation of each functional block of the abnormality detection unit 28. The control unit 31 includes a processor such as a CPU. The memory unit 33 stores various data and programs. The image acquisition unit 32 acquires the captured image 20 captured by the imaging device 27.

The luminance data generating unit 35 performs luminance conversion on the image data of each pixel 52 of the captured image 20 to generate luminance data. The noise removing unit 36 removes noise using a predetermined filter, and generates the enhancement data 21 from multiple luminance data related to multiple image data captured in time series.

The feature calculation unit 37 divides the emphasis data 21 into a matrix of local regions 54 and calculates the feature for each local region 54, as described below.

The learning input data generation unit 39 generates learning input data 58 by associating the local regions 54 with the learning features 22 calculated for learning.

The detection input data generation unit 42 generates detection input data 59 by associating the local region 54 with the detection feature amount 24 calculated for detection.

The abnormality determination unit 43 calculates the degree of abnormality between the learning feature 22 in the learning input data 58 and the detection feature 24 in the detection input data 59 for each local region 54. The abnormality determination unit 43 then compares the calculated degree of abnormality with a predetermined threshold value, and if the degree of abnormality in any of the local regions 54 is equal to or greater than the threshold value, determines that an abnormality has been detected in the area where the image was captured.

The anomaly detection unit 28 may further include a base model generation unit 41. The base model generation unit 41 analyzes a plurality of learning input data 58 to generate the base model 23. The base model 23 is a model configured to calculate the degree of anomaly between the learning feature 22 and the detection feature 24 in the detection input data 59 in each local region 54 by inputting the detection input data 59, and to determine the presence or absence of an abnormality in the part whose image was captured using a threshold value. When the base model 23 is generated, the anomaly determination unit 43 inputs the detection input data 59 to the base model 23 and performs an anomaly determination.

Furthermore, the functional blocks of the anomaly detection unit 28 are not limited to the above configuration, and two or more functional blocks may be partially or entirely integrated, or any of the functional blocks may be further subdivided. Furthermore, some of the functional blocks may be configured independently of the anomaly detection unit 28.

In addition, some or all of the functions of the anomaly detection unit 28 may be configured not only by hardware but also by software. A program related to the software is stored in the storage unit 33, etc., and executed by a processor built into the control unit 31, etc. Some or all of the functions of the anomaly detection unit 28 may be implemented by any method.

The operation of the anomaly detection device, the anomaly detection method, and the processing of the anomaly detection program are described below using as an example a case where an anomaly is detected in which slag accumulates at the supply port at the top of the slag grinder 12, to which slag is supplied from the storage tank 11, causing the supply port to become blocked. Note that in the following explanation, the anomaly detection method and the anomaly detection program that causes a computer to execute the function are described below using as an example a configuration processed by the anomaly detection device described in Figure 3, but they may be processed by any device configuration, and may be configured without using the base model 23, for example.

First, to generate the learning input data 58, the imaging device 27 captures an image of the inside of the supply port at the top of the slag grinder 12. The images are captured within a number of different unit periods (filter units) during a predetermined learning period (length of the learning data), and a number of learning images 46 are acquired during each unit period. In addition, a number of learning image groups are generated by treating the multiple learning images 46 captured during each unit period as one set of learning image groups (step #1 in FIG. 4).

Specifically, the imaging device 27 continuously captures the inside of the upper supply port of the slag grinder 12 during a learning period (length of learning data) of 30 hours. The imaging is performed 3600 times during a unit period of one hour, and multiple unit periods are provided during the learning period. Among the multiple captured learning images 46, 3600 learning images 46 captured during one hour (filter unit) are used as one set of learning image groups to generate multiple sets of learning image groups. When 30 sets of unit periods are provided during the learning period, a total of 108,000 learning images 46 are captured, and 30 sets of learning image groups are generated by using the 3600 learning images 46 captured during one unit period as one set of learning image groups. The unit periods during the learning period may be provided consecutively, or may be provided at a predetermined interval. The image acquisition unit 32 stores the multiple learning images 46 captured by the imaging device 27 in the memory unit 33 for each learning image group.

Next, the luminance data generation unit 35 performs luminance conversion on the image data of each pixel 52 of each learning image 46 to generate learning luminance data 48. The learning luminance data 48 is data that represents the luminance value of each pixel 52 (step #2 in FIG. 4). The luminance data generation unit 35 generates a luminance data group 50 corresponding to the learning image group from the generated learning luminance data 48 (step #3 in FIG. 4). The luminance data generation unit 35 stores the generated learning luminance data 48 and luminance data group 50 in the memory unit 33.

Next, the noise removal unit 36 removes noise from the multiple learning luminance data 48 for each luminance data group 50 using a time-series median filter to generate learning enhancement data 51 (step #4 in FIG. 4).

Specifically, as shown in FIG. 5, the noise removal unit 36 first obtains 3600 luminance values for each pixel 52 from the 3600 learning luminance data 48 constituting one luminance data group 50. Next, the noise removal unit 36 sets the median value of the 3600 luminance values, that is, the 1800th or 1801st largest luminance value, as the luminance value of that pixel 52. The noise removal unit 36 performs this process on all pixels 52 to generate learning emphasis data 51 in which the median luminance value is associated with each pixel 52. One learning emphasis data 51 is generated for each luminance data group 50, and 30 sets of learning emphasis data 51 are generated. The noise removal unit 36 stores the generated learning emphasis data 51 in the memory unit 33.

Next, the feature calculation unit 37 calculates SIFT (Scale-Invariant Feature Transform) features from each training enhancement data 51 and sets them as training features 22 (step #5 in FIG. 4).

Specifically, as shown in Fig. 6, the feature calculation unit 37 first divides the learning emphasis data 51 into a number of local regions 54 arranged in a matrix, with each local region 54 being an area with a diameter of 15 pixels. Next, the feature calculation unit 37 divides each local region 54 into four, thus dividing the data into 16 blocks 55. Next, the feature calculation unit 37 calculates the luminance of each block 55 by dividing it proportionately by the area of the pixels 52 that make up each block 55. Next, for each block 55, the feature calculation unit 37 designates the target block 55 as B0, and assumes eight blocks B1 to B8 adjacent to block B0. Blocks B1 to B8 are respectively block B1 oriented at 0° with respect to block B0, block B2 oriented at 45°, block B3 oriented at 90°, block B4 oriented at 135°, block B5 oriented at 180°, block B6 oriented at 225°, block B7 oriented at 270°, and block B8 oriented at 315°.

Next, the feature calculation unit 37 calculates the difference (difference amount) between the luminance of blocks B1 to B8 and the luminance of block B0 as a luminance gradient. Then, for each local region 54, the feature calculation unit 37 sets the luminance gradients associated with eight directions for the 16 blocks 55 as a 128-dimensional vector of SIFT features (feature vector), converts the length of the 128-dimensional feature vector into a unit vector, and calculates it as the learning feature 22. Note that the calculation of the SIFT features can be performed by any method. For example, the SIFT features may be calculated using a SIFT function of an image recognition library such as OpenCV.

Next, the learning input data generating unit 39 generates learning input data 58 by associating the positions of the local regions 54 with the learning features 22 of each local region 54 for each piece of learning emphasis data 51. The learning input data generating unit 39 stores the generated multiple pieces of learning input data 58 in the memory unit 33 (step #6 in FIG. 4).

Next, the base model generation unit 41 analyzes the learning features 22 of each local region 54 from the multiple learning input data 58, and generates a base model 23 for the learning features 22 of each local region 54. Specifically, the base model 23 is a model of the variance-covariance matrix and average of the learning features 22 in each local region 54. The base model 23 may be generated as a model indicating the variance-covariance matrix and average of the learning features 22, but may also be configured to determine the presence or absence of an abnormality by inputting detection data (detection input data 59 described below).

After the base model 23 is generated, for example, after a predetermined time has elapsed since the last learning image 46 was captured, the imaging device 27 captures the inside of the upper supply port of the slag grinder 12 as multiple detection images 47 and stores the detection images 47 in the storage unit 33. For example, after a predetermined time has elapsed since the last learning image 46 was captured, 3600 detection images 47 are captured within a period of one hour (unit of filter) (step #7 in FIG. 4). Next, the brightness data generating unit 35 performs brightness conversion on each detection image 47 to generate detection brightness data 49, and stores the detection brightness data 49 in the storage unit 33 (step #8 in FIG. 4). Next, the noise removing unit 36 removes noise from the multiple detection brightness data 49 using a time-series median filter, and generates detection enhancement data 57. The noise removing unit 36 stores the generated detection enhancement data 57 in the storage unit 33 (step #9 in FIG. 4). Next, the feature calculation unit 37 calculates the detection features 24 for the detection emphasis data 57 in the same manner as for the learning features 22 (step #10 in FIG. 4). Next, the detection input data generation unit 42 generates detection input data 59 by associating the positions of the local regions 54 with the detection features 24 for each local region 54 for the detection emphasis data 57. The detection input data generation unit 42 stores the generated detection input data 59 in the storage unit 33 (step #11 in FIG. 4).

Then, the abnormality determination unit 43 determines whether or not an abnormality has occurred in the upper supply port of the slag grinder 12 based on the detection input data 59 and the base model 23 (step #12 in Figure 4).

となる。 Specifically, the abnormality determination unit 43 calculates, for each local region 54, the degree of abnormality of the detection feature 24 in the detection input data 59 with respect to the learning feature 22 in the learning input data 58. For example, the degree of abnormality is calculated as the Mahalanobis distance between the learning feature 22 in the learning input data 58 and the detection feature 24 in the detection input data 59. More specifically, the average vector and variance-covariance matrix of the learning feature 22 of each local region 54 in the learning input data 58 are calculated, and the Mahalanobis distance between the learning feature 24 in the detection input data 59 and each local region 54 is calculated. If the local region 54 is j (1, 2, . . . ), the detection feature 24 of the local region j is y _j , the average vector of the training feature 22 of the local region j in the training input data 58 is μ _j , and the variance-covariance matrix of the training feature 22 of the local region j in the training input data 58 is Σ _j , then the Mahalanobis distance a(j) in the local region j is given by

It becomes.

The abnormality determination unit 43 then compares the calculated Mahalanobis distance with a predetermined threshold value. The threshold value is, for example, 100 or 120. If the Mahalanobis distance is, for example, 100 or greater in any of the local regions 54, the abnormality determination unit 43 determines that an abnormality has been detected in the upper supply port of the slag grinder 12.

The base model 23 may be generated using learning images 46 taken during a certain period of time, and the base model 23 may be used continuously thereafter. Alternatively, the base model 23 may be updated using learning images 46 taken at any time during an abnormality determination or during a period when an abnormality determination is not being performed, in which no abnormality is detected.

Furthermore, the configuration may be such that, when an abnormality is detected, a predetermined notification is issued. The notification is controlled by the notification unit 62. The notification unit 62 displays, on a display unit (not shown), the location where the abnormality was detected and the fact that an abnormality has occurred. Alternatively, the notification unit 62 may notify the occurrence of an abnormality with a lamp or speaker. Furthermore, the notification unit 62 may be configured to present a method of dealing with an abnormality based on an abnormality degree map created according to a threshold value corresponding to the degree of abnormality when an abnormality is detected.

As described above, the solidified slag is supplied to the slag grinder 12 from the supply port at the top of the slag grinder 12. As a result, slag may accumulate at the supply port of the slag grinder 12, causing the supply port to become clogged, and this blocked state may be detected as an abnormality.

The base model 23 used to detect anomalies is a model of the learning features 22 of an image (learning image 46) taken of the supply port of the slag grinder 12 when no slag is accumulated at the supply port of the slag grinder 12 or when the amount of slag accumulated is so small that it is not a problem. Anomalies are detected by using the base model 23 to compare the learning features 22 of the learning image 46 with the detection features 24 of the detection image 47 for each local region 54.

Here, the wall surface of the supply port in the photographing area may be discolored due to moisture, dirt, etc. If the color of the wall changes between when the learning image 46 is photographed and when the detection image 47 is photographed, the discolored part of the wall may be erroneously determined to be abnormal. Therefore, when generating the learning enhancement data 51 used to generate the base model 23, noise is removed using a time-series median filter so that the discoloration of the wall becomes normal. Furthermore, parameters such as the filter unit and the length of the learning data in photographing the learning image 46 are optimized according to the abnormality of slag accumulating at the supply port of the slag grinder 12 and the discoloration of the wall, and the learning image 46 is photographed with the filter unit set to 1 hour and the length of the learning data set to 30 hours. In addition, the interval between photographing the learning image 46 and photographing the detection image 47 may be any time, but is set to 10 hours, for example.

Figure 7 shows photographs showing the detection results of deposits that cause blockage abnormalities, with Figure 7(a) being a photograph showing the detection results of deposits when the threshold value is 100, and Figure 7(b) being a photograph showing the detection results of deposits when the threshold value is 120. In the photographs showing the detection results, pixels detected as deposits are surrounded by dashed lines. The notification unit 62 may also be configured to display a photograph such as that shown in Figure 7.

In this way, in order to detect an abnormality in the supply port of the slag grinder 12, where deposits gradually build up and lead to blockage, a time-series median filter that is optimal as a filter is used to remove noise, and parameters such as the filter unit and the length of the learning data are optimized to capture the learning images 46, thereby generating a base model 23 that is optimized for detecting deposits. As a result, abnormalities can be detected with high accuracy and without much effort.

In addition, when capturing the detection image 47, parameters such as the filter unit and the length of the learning data are optional, and the filter used for noise removal and whether or not noise removal is necessary are optional. However, as with the generation of the learning input data 58, by performing noise removal, abnormalities can be detected with greater accuracy.

When determining whether or not there is an abnormality, the distance (degree of abnormality) between the detection features 24 and the learning features 22 for each local region 54 included in the base model 23 is verified. Multiple sets of learning images 46 are captured, and multiple learning features 22 are compared with the detection features 24, so that abnormalities are detected while taking into account the variance of the learning features 22, allowing for accurate detection of abnormalities. Furthermore, by taking into account the mean and variance-covariance matrix of the learning features 22, abnormalities can be detected with greater accuracy. In particular, by using the Mahalanobis distance to verify the degree of abnormality, abnormalities can be detected with greater accuracy.

Second Embodiment
As shown in the first embodiment, when detecting an abnormality in which a blockage occurs due to deposits, it is preferable to perform noise removal using a time series median filter. Furthermore, noise removal can also be performed using an appropriate filter depending on the type of abnormality. For example, when detecting a flow abnormality in which the flow (flow) of a fluid is stagnated, a time series minimum filter is used, which performs noise removal so as to emphasize the fluid.

For example, in an anomaly detection process for detecting flow anomalies such as a decrease in the amount of activated carbon flowing through the cut-out section 15, which supplies activated carbon immediately before the bag filter 7, noise removal may be performed using a time-series minimum filter. In the cut-out section 15, activated carbon is generally constantly flowing, so when noise removal processing is performed to emphasize the activated carbon, the local region 54 corresponding to the inside of the cut-out section 15 in the learning input data 58 has a feature value corresponding to a state where the area is uniformly filled with activated carbon. In contrast, the local region 54 corresponding to the inside of the cut-out section 15 in the detection input data 59 in a state where an abnormality has occurred in the flow of activated carbon has a feature value different from the feature value corresponding to activated carbon. Therefore, flow anomalies in activated carbon can be easily and accurately detected.

When noise is removed using a time series minimum filter, various conditions such as parameters such as the filter unit and length of learning data, and thresholds are optimized according to the type of anomaly and the environment. When detecting flow anomalies in the cutout unit 15, for example, the filter unit is set to 30 seconds, the length of learning data is set to 10 minutes, the threshold for anomaly determination is set to 20, and the interval between capturing the learning image 46 and capturing the detection image 47 is set to 150 seconds. Below, specific configuration examples are described using Figures 3 to 6 with reference to Figures 1 and 2. Note that the configuration of the anomaly detection device shown in Figure 3 is the same as that of the first embodiment, so a detailed description is omitted.

First, to generate the learning input data 58, the imaging device 27 captures an image of the inside of the cutout section 15 that supplies activated carbon. The images are captured within a number of different unit periods (filter units) during a predetermined learning period (length of the learning data), and a number of learning images 46 are acquired during each unit period. In addition, a number of sets of learning images are generated by treating the multiple learning images 46 captured during each unit period as a set of learning image groups (step #1 in FIG. 4).

Specifically, the imaging device 27 continuously captures the inside of the cutout unit 15 during a learning period (length of learning data) of 10 minutes. The capture is performed 300 times during a unit period of 30 seconds, and multiple unit periods are provided during the learning period. Among the captured learning images 46, 300 learning images 46 captured during 30 seconds (filter unit) are used as one learning image group to generate multiple learning image groups. When 20 unit periods are provided during the learning period, a total of 20 learning images 46 are captured, and 20 learning image groups are generated by using the 300 learning images 46 captured during one unit period as one learning image group. The unit periods during the learning period may be provided consecutively, or may be provided at a predetermined interval. The image acquisition unit 32 stores the multiple learning images 46 captured by the imaging device 27 in the memory unit 33 for each learning image group.

Next, the luminance data generation unit 35 performs luminance conversion on the image data of each pixel 52 of each learning image 46 to generate learning luminance data 48. The learning luminance data 48 is data that represents the luminance value of each pixel 52 (step #2 in FIG. 4). The luminance data generation unit 35 generates a luminance data group 50 corresponding to the learning image group from the generated learning luminance data 48 (step #3 in FIG. 4). The luminance data generation unit 35 stores the generated learning luminance data 48 and luminance data group 50 in the memory unit 33.

Next, the noise removal unit 36 performs noise removal on the multiple learning luminance data 48 for each luminance data group 50 using a time series minimum filter to generate learning emphasis data 51 (step #4 in Figure 4).

Specifically, as shown in FIG. 5, the noise removal unit 36 first obtains 300 luminance values for each pixel 52 from the five learning luminance data 48 that constitute one luminance data group 50. Next, the noise removal unit 36 sets the minimum of the 300 luminance values as the luminance value of that pixel 52. The noise removal unit 36 performs this process on all pixels 52 to generate learning emphasis data 51 in which each pixel 52 is associated with the minimum luminance value. One learning emphasis data 51 is generated for each luminance data group 50, and twenty sets of learning emphasis data 51 are generated. The noise removal unit 36 stores the generated learning emphasis data 51 in the memory unit 33.

Next, the feature calculation unit 37 calculates SIFT (Scale-Invariant Feature Transform) features from each training enhancement data 51 and sets them as training features 22 (step #5 in FIG. 4).

Specifically, as shown in Fig. 6, the feature calculation unit 37 first divides the learning enhancement data 51 into a plurality of local regions 54 arranged in a matrix, with each local region 54 being an area with a diameter of 15 pixels. Next, the feature calculation unit 37 divides each local region 54 vertically and horizontally into four, thus dividing it into 16 blocks 55. Next, the feature calculation unit 37 calculates the luminance of each block 55 by dividing it proportionately by the area of the pixels 52 that make up each block 55. Next, for each block 55, the feature calculation unit 37 designates the target block 55 as B0, and assumes eight blocks B1 to B8 adjacent to block B0. Blocks B1 to B8 are respectively block B1 oriented at 0° with respect to block B0, block B2 oriented at 45°, block B3 oriented at 90°, block B4 oriented at 135°, block B5 oriented at 180°, block B6 oriented at 225°, block B7 oriented at 270°, and block B8 oriented at 315°.

Next, the feature calculation unit 37 calculates the difference (difference amount) between the luminance of blocks B1 to B8 relative to the luminance of block B0 as a luminance gradient. Then, for each local region 54, the feature calculation unit 37 sets the luminance gradient associated with eight directions for the 16 blocks 55 as a 128-dimensional vector of SIFT features (feature vector), converts the length of the 128-dimensional feature vector into a single vector, and calculates it as the learning feature 22. Note that the calculation of the SIFT features can be performed by any method. For example, the SIFT features may be calculated using a SIFT function of an image recognition library such as OpenCV.

Next, the learning input data generating unit 39 generates learning input data 58 by associating the positions of the local regions 54 with the learning features 22 of each local region 54 for each piece of learning emphasis data 51. The learning input data generating unit 39 stores the generated multiple pieces of learning input data 58 in the memory unit 33 (step #6 in FIG. 4).

Next, the base model generation unit 41 analyzes the learning features 22 of each local region 54 from the multiple learning input data 58, and generates a base model 23 for the learning features 22 of each local region 54. Specifically, the base model 23 is a model of the variance-covariance matrix and average of the learning features 22 in each local region 54. The base model 23 may be generated as a model indicating the variance-covariance matrix and average of the learning features 22, but may also be configured to determine the presence or absence of an abnormality by inputting detection data (detection input data 59 described below).

After the base model 23 is generated, for example, after a predetermined time has elapsed since the last learning image 46 was captured, the imaging device 27 captures the inside of the cutout unit 15 as multiple detection images 47 and stores the detection images 47 in the storage unit 33. For example, after a predetermined time has elapsed since the last learning image 46 was captured, 300 detection images 47 are captured within a period of 30 seconds (unit of filter) (step #7 in FIG. 4). Next, the luminance data generating unit 35 performs luminance conversion on each detection image 47 to generate detection luminance data 49, and stores the detection luminance data 49 in the storage unit 33 (step #8 in FIG. 4). Next, the noise removing unit 36 removes noise from the multiple detection luminance data 49 using a time series minimum filter, and generates detection emphasis data 57. The noise removing unit 36 stores the generated detection emphasis data 57 in the storage unit 33 (step #9 in FIG. 4). Next, the feature calculation unit 37 calculates the detection features 24 for the detection emphasis data 57 in a manner similar to that for the learning features 22 (step #10 in FIG. 4). Next, the detection input data generation unit 42 generates detection input data 59 by associating the positions of the local regions 54 with the detection features 24 for each local region 54 for the detection emphasis data 57. The detection input data generation unit 42 stores the generated detection input data 59 in the storage unit 33 (step #11 in FIG. 4).

Then, the abnormality determination unit 43 determines whether or not an abnormality has occurred in the cut-out unit 15 based on the detection input data 59 and the base model 23 (step #12 in Figure 4).

となる。 Specifically, the abnormality determination unit 43 calculates, for each local region 54, the degree of abnormality of the detection feature 24 in the detection input data 59 with respect to the learning feature 22 in the learning input data 58. For example, the degree of abnormality is calculated as the Mahalanobis distance between the learning feature 22 in the learning input data 58 and the detection feature 24 in the detection input data 59. Specifically, the average vector and variance-covariance matrix of the learning feature 22 of each local region 54 in the learning input data 58 are calculated, and the Mahalanobis distance between the learning feature 24 in the detection input data 59 and each local region 54 is calculated. If the local region 54 is j (1, 2, . . . ), the detection feature 24 of the local region j is y _j , the average vector of the training feature 22 of the local region j in the training input data 58 is μ _j , and the variance-covariance matrix of the training feature 22 of the local region j in the training input data 58 is Σ _j , then the Mahalanobis distance a(j) in the local region j is given by

It becomes.

Next, the abnormality determination unit 43 compares the calculated Mahalanobis distance with a predetermined threshold value. The threshold value is, for example, 20. Then, if the Mahalanobis distance is 20 or more in any of the local regions 54, the abnormality determination unit 43 determines that an abnormality has been detected in the cut-out unit 15.

The base model 23 may be generated using learning images 46 taken during a certain period of time, and the base model 23 may be used continuously thereafter. Alternatively, the base model 23 may be updated using learning images 46 taken at any time during an abnormality determination or during a period when an abnormality determination is not being performed, and in which no abnormality is detected.

Furthermore, the configuration may be such that, when an abnormality is detected, a predetermined notification is issued. The notification is controlled by the notification unit 62. The notification unit 62 displays, on a display unit (not shown), the location where the abnormality was detected and the fact that an abnormality has occurred. Alternatively, the notification unit 62 may notify the occurrence of an abnormality with a lamp or speaker. Furthermore, the notification unit 62 may be configured to present a method of dealing with an abnormality based on an abnormality degree map created according to a threshold value corresponding to the degree of abnormality when an abnormality is detected.

As mentioned above, the flow of activated carbon through the cutout section 15 may be stagnated, and this flow anomaly is detected as an anomaly.

The base model 23 used to detect anomalies is a model of the learning features 22 of the captured image (learning image 46) of the cutout section 15 when the flow of activated carbon through the cutout section 15 is normal. Anomalies are detected by using the base model 23 to compare the learning features 22 of the learning image 46 with the detection features 24 of the detection image 47 for each local region 54.

Here, the flowing matter is captured as a haze in the captured image 20. In order to extract only the flowing matter, that is, the haze, noise is removed using a time series minimum filter to emphasize the flowing matter when generating the learning enhancement data 51 used to generate the base model 23. Furthermore, parameters such as the filter unit and the length of the learning data in capturing the learning image 46 are optimized according to the occurrence state of the situation in which the flow abnormality occurring in the cutout section 15 is easily mistaken for a haze, and the learning image 46 is captured with the filter unit set to 30 seconds and the length of the learning data set to 10 minutes. In addition, the interval between capturing the learning image 46 and capturing the detection image 47 may be any time, but is set to 150 seconds, for example.

In addition, when capturing the detection image 47, parameters such as the filter unit and the length of the learning data are optional, and the filter used for noise removal and whether or not noise removal is necessary are optional. However, as with the generation of the learning input data 58, by performing noise removal, abnormalities can be detected with greater accuracy.

When determining whether or not there is an abnormality, the distance (degree of abnormality) between the detection features 24 and the learning features 22 for each local region 54 included in the base model 23 is verified. Multiple sets of learning images 46 are captured, and multiple learning features 22 are compared with the detection features 24, so that abnormalities are detected while taking into account the variability of the learning features 22, allowing for accurate detection of abnormalities. In particular, by using the Mahalanobis distance to verify the degree of abnormality, abnormalities can be detected with greater accuracy.

[Another embodiment]
(1) In the first embodiment, detection of an abnormality at the supply inlet of the slag grinder 12 was described as an example, but such abnormality detection can be applied to detection of abnormalities in various parts of the waste incineration facility 1.

In addition to selecting a filter, various conditions such as parameters such as filter unit and length of learning data, thresholds, etc. are set according to the type of abnormality to be detected and the type of object to be removed, such as discoloration of a wall. This allows for accurate detection of abnormalities according to the type of abnormality and the environment, and by adjusting parameters while detecting anomalies using the same algorithm, versatility can be improved without hassle.

The abnormality detection according to the first embodiment can be used, for example, to detect an abnormality in which residue accumulates in the discharge device 13, which transports and discharges the residue stored below the gasification furnace 2, causing the discharge device 13 to become clogged. The abnormality detection according to the first embodiment can also be used to detect an abnormality in which foreign matter accumulates on the sand separation screen 14, causing the sand separation screen 14 to become clogged.

When detecting a blockage abnormality in the discharge device 13, for example, the filter unit is 3 seconds, the length of the learning data is 60 seconds, 25 images 20 are taken per second, the threshold value for abnormality determination is set to 35, and the interval between taking the learning image 46 and taking the detection image 47 is set to 15 seconds. The image capture position is the outlet of a non-combustible storage tank (not shown) located downstream of the discharge device 13.

When detecting a blockage abnormality in the sand separation screen 14, for example, the filter unit is 3 minutes, the length of the learning data is 30 minutes, 5 images 20 are taken per second, the threshold value for determining an abnormality is set to 15, and the interval between taking the learning image 46 and the detection image 47 is set to 6 minutes.

(2) The degree of anomaly between the learning features 22 and the detection features 24 is not limited to being calculated using the Mahalanobis distance, but may be calculated using a similarity calculation method using various vectors such as Euclidean distance depending on the type of anomaly, etc. Furthermore, the features are not limited to being SIFT features, but may be calculated using any method.

(3) As described above, the filters used for noise removal are a time series median filter when detecting blockage anomalies caused by deposits, and a time series minimum filter when detecting flow anomalies in fluids. However, different filters may be used depending on the type of anomaly. This makes it possible to detect anomalies with greater versatility.

(4) As described above, when removing noise, different filters are used depending on the type of abnormality. However, the anomaly detection unit 28 may be configured to be provided with multiple filters in advance and to be able to select the filter to be used when removing noise. This allows the anomaly detection device to detect various abnormalities in various parts, making it even more versatile.

For example, the anomaly detection device may further include a filter selection unit 63, and may be configured to perform noise removal using a filter selected by the filter selection unit 63.

The anomaly detection device may further include a parameter setting unit 64. The parameters are optimized according to the type of anomaly, etc. The anomaly detection device may be configured to allow any parameter to be set through the parameter setting unit 64, and more preferably, to allow filter selection and parameter setting, thereby making it possible to easily respond to newly required anomaly detection and to perform anomaly detection with greater versatility.

The anomaly detection method and the anomaly detection program may also be configured to allow for filter selection and parameter setting.

(5) The size of the local region 54 is also not limited to a diameter of 15 pixels, but is optimized according to the type of abnormality to be detected and the size of the detection target. For example, when detecting a blockage abnormality in the discharge device 13, a diameter of 15 pixels is used as one local region 54, when detecting a blockage abnormality in the sand separation screen 14, a diameter of 30 pixels is used as one local region 54, and when detecting a flow abnormality in activated carbon, a diameter of 15 pixels is used as one local region 54. This allows for accurate detection of abnormalities.

(6) The local region 54 is not limited to being divided into 16 parts, and may be divided into any number of blocks 55. In addition, the number of adjacent blocks 55 for which the luminance gradient is to be calculated, i.e., the directions of the adjacent blocks 55, is not limited to eight directions, and the luminance gradient for the blocks 55 in any number of directions, such as four directions, may be calculated.

As a result, in the above example, brightness gradients in eight directions were found for each of the 16 divided blocks 55, and features were calculated from a 16 x 8 = 128-dimensional feature vector, but features may be calculated from a feature vector with a number of dimensions determined by the number of blocks 55 to be divided and the number of directions for which the brightness gradients are found.

The number of divided blocks 55 and the number of directions for which the brightness gradient is calculated are optimized according to the type of abnormality, etc. Therefore, various abnormalities can be detected with high accuracy.

(7) The waste incineration facility 1 is not limited to the fluidized bed gasification and melting furnace illustrated in FIG. 1, and the above abnormality detection can be applied to various incineration facilities. For example, it can be applied to a stoker type incinerator, a fluidized bed gasification and combustion furnace, etc.

In a stoker-type incinerator, a time-series median filter is used to remove noise in the discharge path of the dust ash and bottom ash discharged from the stoker furnace, making it possible to detect blockage abnormalities caused by the accumulation of dust ash and bottom ash. Also, a time-series minimum filter is used to remove noise in the discharge path of the fly ash discharged from the economizer or bag filter, and in the cut-out section that supplies activated carbon, making it possible to detect flow abnormalities in the fly ash and activated carbon.

In addition, in a fluidized bed gasification and combustion furnace, noise can be removed using a time-series median filter in the exhaust device 13, making it possible to detect blockage abnormalities caused by accumulation of residues. Also, noise can be removed using a time-series minimum filter in the discharge path for fly ash discharged from the economizer or bag filter, and in the cut-out section that supplies activated carbon, making it possible to detect flow abnormalities in fly ash or activated carbon.

In addition to the above, in a fluidized bed gasification and melting furnace, abnormalities in the flow of fly ash or activated carbon can be detected by using a time series minimum filter to remove noise in the discharge path of the fly ash discharged from the economizer or bag filter.

The present invention can be applied to detecting various abnormalities in various parts of waste treatment facilities such as fluidized bed gasification and melting furnaces, stoker incinerators, and fluidized bed gasification and combustion furnaces.

REFERENCE SIGNS LIST 1 Waste incineration facility 20 Captured image 21 Enhanced data 22 Learning features 23 Base model 24 Detection features 27 Imaging device 35 Luminance data generation unit 36 Noise removal unit 37 Features calculation unit 39 Learning input data generation unit 41 Base model generation unit 42 Detection input data generation unit 43 Abnormality determination unit 46 Learning image 47 Detection image 48 Learning luminance data 49 Detection luminance data 50 Luminance data group 51 Learning enhancement data 52 Pixel 54 Local region 55 Block 57 Detection enhancement data 58 Learning input data 59 Detection input data

Claims (8)

An imaging device for photographing a predetermined portion of a waste treatment facility;
a luminance data generating unit that performs luminance conversion on image data of each pixel of an image captured by the imaging device to generate luminance data;
a noise removal unit that removes noise from the luminance data generated for each of the plurality of captured images captured during a predetermined unit period as a set of luminance data groups using a time-series minimum filter, thereby generating enhancement data for each of the luminance data groups;
a feature amount calculation unit that divides the emphasis data into a plurality of local regions and calculates a feature amount for each of the local regions;
a learning input data generating unit that generates learning input data in which the feature amounts calculated for each of the plurality of pieces of emphasis data generated during a predetermined learning period are associated with the local region;
a detection input data generation unit that generates detection input data in which the feature amount calculated based on a set of the luminance data group corresponding to a plurality of the captured images captured by the imaging device is associated with the local region;
and an abnormality determination unit that determines that an abnormality has been detected in the portion when a degree of abnormality of the feature amount related to the detection input data with respect to the feature amount related to the learning input data in each of the local regions is equal to or greater than a predetermined threshold. The anomaly detection device according to claim 1, wherein the unit period and the learning period are set according to the type of the anomaly to be detected. The anomaly detection device according to claim 1 or 2, wherein the degree of anomaly is calculated as a Mahalanobis distance between the feature amount related to the learning input data and the feature amount related to the detection input data. The anomaly detection device according to any one of claims 1 to 3, wherein the feature amount is obtained by dividing each of the local regions into a plurality of blocks by dividing each of the local regions into equal parts vertically and horizontally in a matrix, determining the difference in brightness between each of the blocks and adjacent blocks and the directional relationship between the blocks as a brightness gradient, and converting the brightness gradient for each of the divided blocks into a unit vector. The abnormality detection device according to any one of claims 1 to 4, wherein the abnormality detected is an abnormality in the flow of a fluid flowing through the part. a learning image acquisition step of acquiring a plurality of learning images by photographing a predetermined portion of a waste treatment facility a plurality of times in each of different predetermined unit periods within a predetermined learning period;
a learning luminance data generating step of performing luminance conversion on image data of each pixel of each of the learning images to generate a plurality of learning luminance data corresponding to each of the learning images;
a learning enhancement data generating step of generating a plurality of learning enhancement data from the plurality of sets of the luminance data groups by removing noise using a time series minimum filter, the plurality of learning luminance data corresponding to the learning images captured within each of the unit periods being treated as a set of luminance data groups;
a learning input data generating step of dividing each of the learning enhancement data into a plurality of local regions, calculating a learning feature for each of the local regions, and generating, for each of the learning enhancement data, learning input data in which the local regions and the learning feature are associated with each other;
a detection image acquisition step of photographing the site as a plurality of detection images;
a detection luminance data generating step of performing luminance conversion on image data of each pixel of each of the detection images to generate a plurality of detection luminance data;
a detection enhancement data generating step of generating detection enhancement data from the plurality of detection luminance data by removing noise using a time series minimum filter;
a detection input data generating step of dividing the detection enhancement data into the local regions, calculating a detection feature for each of the local regions, and generating detection input data in which the local regions are associated with the detection feature;
and an anomaly determination step of determining that an anomaly has been detected in the portion if the degree of anomaly of the detection feature with respect to the learning feature in each of the local regions is equal to or greater than a predetermined threshold. a base model generating step of analyzing a plurality of the learning input data to generate a base model in which an average and a variance-covariance matrix of the learning feature for each of the local regions are obtained;
7. The anomaly detection method according to claim 6, wherein the base model is configured to, when the detection input data is input, calculate the degree of anomaly of the detection feature with respect to the learning feature, and detect the anomaly based on the degree of anomaly. A learning image acquisition function that acquires a plurality of learning images of a predetermined portion of a waste treatment facility, the learning image being captured a plurality of times in each of different predetermined unit periods within a predetermined learning period;
a learning luminance data generating function for performing luminance conversion on image data of each pixel of each of the learning images to generate a plurality of learning luminance data corresponding to each of the learning images;
a learning enhancement data generating function for generating a plurality of learning enhancement data from the plurality of sets of the luminance data groups by removing noise using a time series minimum filter, the plurality of learning luminance data corresponding to the learning images captured within each of the unit periods being treated as a set of luminance data groups;
a learning input data generation function that divides each of the learning enhancement data into a plurality of local regions, calculates a learning feature for each of the local regions, and generates, for each of the learning enhancement data, learning input data in which the local regions and the learning feature are associated with each other;
A detection image acquisition function that acquires a plurality of detection images of the part;
a detection luminance data generating function for performing luminance conversion on image data of each pixel of a plurality of detection images obtained by photographing the part, to generate a plurality of detection luminance data;
a detection enhancement data generating function for generating detection enhancement data from the plurality of detection luminance data by removing noise using a time series minimum filter;
a detection input data generation function that divides the detection enhancement data into the local regions, calculates a detection feature for each of the local regions, and generates detection input data in which the local regions and the detection feature are associated with each other;
and an abnormality determination function that determines that an abnormality has been detected in the area if the degree of abnormality of the detection feature relative to the learning feature in each of the local regions is equal to or greater than a predetermined threshold.

Images