Wind turbine monitoring apparatus and retraining method

Abstract

To provide a wind turbine monitoring apparatus configured to reduce monitoring burden.SOLUTION: A monitoring apparatus 20 includes an image acquisition unit 24A, a storage unit 21, a determination unit 24B, and an input / output unit 22. The image acquisition unit 24A periodically acquires images to be monitored. The storage unit 21 is constructed by using a predetermined network structure and normal image data, and stores information for executing an "anomaly determination model" that determines an anomaly when an input image is abnormal. The determination unit 24B determines, using the anomaly determination model, whether the periodically acquired images to be monitored are abnormal or not. The input / output unit 22 outputs a warning signal when the determination unit 24 determines an anomaly.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a monitoring device and a relearning method for wind power generation.

Background Art

[0002] Conventionally, remote monitoring of power generation facilities and the like has been performed. For example, Patent Document 1 (Japanese Unexamined Patent Application Publication No. 2014-020250) discloses a monitoring device that reduces the work load in the inspection of the inside of a housing that stores a main bearing, a speed increaser, and a generator in a power generation device.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the monitoring device for a power generation device described in Patent Document 1 mentioned above, an image inside the facility is acquired, and the state inside the facility can be visually confirmed from the image.

[0005] However, abnormalities in power generation facilities and the like are often determined from changes over a relatively long period, and it is not always necessary to monitor constantly. Rather, if the monitoring frequency is high, the monitoring burden may increase. The present invention has been made in view of the above circumstances, and an object thereof is to provide a monitoring device capable of reducing the monitoring burden.

Means for Solving the Problems

[0006] The wind power generation monitoring device according to the first aspect includes an image acquisition unit, a storage unit, a determination unit, and an output unit. The image acquisition unit periodically acquires an image of a monitoring target. The storage unit stores information for executing an abnormality determination model that is constructed using a predetermined network structure and normal image data and determines that an input image is abnormal when it is not normal. The determination unit determines whether the image of the monitoring target periodically acquired is abnormal using the abnormality determination model. The output unit outputs a warning signal when it is determined by the determination unit that the image is abnormal.

[0007] Therefore, since the wind power generation monitoring device according to the first aspect determines whether the image of the monitoring target periodically acquired is abnormal using the abnormality determination model and outputs a warning signal when it is determined by the determination unit that the image is abnormal, the monitor does not need to visually monitor at all times, and the monitoring burden can be reduced.

[0008] Furthermore, since it includes a learning unit that constructs an abnormality determination model using normal image data, the construction of the abnormality determination model and the processing using the abnormality determination model can be executed by a single device.

[0009] The wind power generation monitoring device according to the second aspect is the monitoring device according to the first aspect, and the processing by the determination unit is executed when the processing by the learning unit is being executed. In the case of the wind power generation monitoring device according to the second aspect, since the abnormality determination is performed before the learning is completed, the unmonitored time can be shortened.

[0010] The wind power generation monitoring device according to the third aspect is the monitoring device according to the first aspect, and further includes an input / output unit that displays image data and accepts an input of a setting of whether it is normal or not. Therefore, an abnormality determination model reflecting the criteria of normal and abnormal by the monitor can be constructed.

[0011] The wind power generation monitoring device according to the fourth aspect executes re-learning of the anomaly determination model when an attention signal is output a predetermined number of times within a predetermined period. To supplement, in the anomaly determination model used for remotely monitoring an object, images captured by a camera installed at a fixed position are periodically input, and the presence or absence of anomalies is periodically determined. However, when the camera is used for a certain period, the position of the camera may change gradually, and the reliability of the anomaly determination model may decrease. Therefore, when the reliability of the anomaly determination model decreases, it is necessary to reconstruct the model. The monitoring device according to the fifth aspect can execute reconstruction of the anomaly determination model at an appropriate timing with the above configuration.

[0012] The wind power generation monitoring device according to the fifth aspect is such that the anomaly determination model is constructed by learning the network structure using normal image data. In the monitoring device according to the fifth aspect, since there is no need to collect image data other than normal image data, the anomaly determination model can be easily constructed.

[0013] The wind power generation monitoring device according to the sixth aspect is such that the anomaly determination model is constructed by calculating a predetermined statistic from normal image data using the network structure. In the monitoring device according to the sixth aspect, since there is no need to collect image data other than normal image data, the anomaly determination model can be easily constructed.

[0014] The relearning method according to the seventh aspect is a method for relearning an anomaly determination model used in the wind power generation monitoring device according to the first aspect. In this method, when the relearning mode is on, image data is accumulated at the first time interval until the first number of images is accumulated. Next, using the accumulated image data and the network structure, a first determination model for determining whether the input image is normal is constructed. Then, the first post-determination image data determined to be normal by the first determination model is accumulated at the second time interval. Next, the first post-determination image data is displayed and a setting of whether it is normal is received. Then, when the first post-determination image data set to be normal is accumulated by the second number of images, a second determination model for determining whether the input image is normal is constructed using the first post-determination image data and the network structure. And the second determination model is set as a new anomaly determination model.

[0015] Here, in order to relearn the anomaly determination model, it is necessary to collect a large number of image data. Therefore, it is necessary to set a relatively long time for collecting the image data used for relearning. According to the method according to the seventh aspect, since the second determination model is gradually constructed using the first determination model, a highly reliable anomaly determination model can be constructed quickly.

[0016] In the method according to the eighth aspect, it is the relearning method according to the seventh aspect, and further, the second post-determination image data determined to be normal by the second determination model is accumulated at the third time interval. Then, the first post-determination image data and / or the second post-determination image data is displayed and a setting of whether it is normal is received. Next, the second determination model is reconstructed every time the second post-determination image data set to be normal is accumulated by the fourth number of images until the third number of images is accumulated. And the second determination model after the reconstruction is executed a predetermined number of times is set as a new anomaly determination model. According to the method according to the eighth aspect, since the second determination model is reconstructed a predetermined number of times, a highly reliable anomaly determination model can be provided.

[0017] <Terms> In the present disclosure, the term "learning" means any aspect of constructing the conditions of a machine learning model based on learning data. Therefore, in the present invention, "relearning" not only means resetting the values of the parameters of a predetermined network structure by machine learning, but also includes the concept of reconstructing and setting any conditions related to the machine learning model.

Brief Description of the Drawings

[0018]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Embodiments for Carrying Out the Invention

[0019] (1) Configuration of the Monitoring Device for Wind Power Generation

[0020] FIG. 1 is a schematic diagram showing the configuration of the monitoring device 20 according to the present embodiment. For convenience, in the following description, it is assumed that the monitoring target is a wind power generation facility. More specifically, a camera 15 is installed inside the control room of the wind power generation facility, and the inside of the control room is set as the monitoring target. However, the monitoring target of the monitoring device 20 according to the present embodiment is not limited to this, and it may be any object.

[0021] The monitoring device 20 periodically acquires an image of the monitoring target via the camera 15, and periodically monitors whether an abnormality has occurred in the monitoring target from the image. In addition, the monitoring device 20 transmits the monitoring result to the user terminal device 30 operated by the monitor. Such a monitoring device 20 can be realized by any computer. Here, the monitoring device 20 includes a storage unit 21, an input / output unit 22, a communication unit 23, and a processing unit 24. Note that the monitoring device 20 may be realized as hardware using an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or the like.

[0022] The storage unit 21 stores various kinds of information and is realized by an arbitrary storage device such as a memory and a hard disk. Specifically, the storage unit 21 stores information for executing an abnormality determination model. The "abnormality determination model" is constructed using a predetermined network structure and normal image data, and determines that there is an abnormality when the input image is not normal. Here, the abnormality determination model is constructed by calculating a predetermined statistic from the normal image data using a predetermined network structure. For example, the abnormality determination model can be constructed by a framework such as PaDiM using a network structure such as Wide-ResNet.

[0023] The input / output unit 22 is realized by a keyboard, a mouse, a touch panel, etc., and is used to input various information into the computer or output various information from the computer. Here, the input / output unit 22 displays a screen G1 as shown in FIG. 2 on the display. Further, the input / output unit 22 receives the input of various setting information from the user via this screen G1. For example, via the screen G1, it is possible to set the on / off of the inference mode and the on / off of the automatic relearning mode, which will be described later. Here, the screen G1 includes an image display section d1, a management information display section d2, a setting reception section d3, and a mode setting section d4. An image of the object to be monitored is displayed on the image display section d1. For example, an image of a structure in the control room is displayed. The management information display section d2 displays information for the monitor, such as information on the "current threshold value" used for abnormality determination. The setting reception section d3 is a part that receives the input of various setting information from the monitor. The mode setting section d4 is a part that receives the on / off setting of the "inference mode" and the "automatic relearning mode".

[0024] In addition, when it is determined that there is an abnormality in the object to be monitored, the input / output unit 22 outputs an image in a manner that enables the recognition of the abnormal location. For example, when an abnormality occurs in the image of the object to be monitored, the input / output unit 22 outputs a heat map reflecting the degree of abnormality. More specifically, a heat map H colored according to the degree of abnormality in each region is output (see FIG. 3).

[0025] The communication unit 23 is realized by an arbitrary network card or the like, and enables communication with communication devices on the network by wire or wirelessly. When an abnormality occurs in the image of the object to be monitored, the communication unit 23 transmits "caution information" to the user terminal device 30 or the like.

[0026] The processing unit 24 executes various information processes and is realized by a processor such as a CPU or a GPU and a memory. Here, when the programs stored in the storage unit 21 are read into the CPU, GPU, etc. of the computer, the processing unit 24 functions as an image acquisition unit 24A, a determination unit 24B, and a learning unit 24C. The image acquisition unit 24A acquires an image of a monitoring target via the camera 15 or the like, and stores the acquired image in the storage unit 21 at any time.

[0027] The determination unit 24B determines whether an image of a monitoring target acquired regularly is abnormal or not using an abnormality determination model. Specifically, when the "inference mode" is set to the on state, the determination unit 24B executes a determination process described later.

[0028] The learning unit 24C constructs an abnormality determination model. Specifically, the learning unit 24C constructs an abnormality determination model using correct answer image data for learning. Further, when the "automatic relearning mode" is set to the on state, the learning unit 24C executes a relearning process described later.

[0029] (2) Operation of the wind power generation monitoring device (2-1) Inference mode FIG. 4 is a flowchart for explaining the operation of the "inference mode" of the monitoring device 20 according to the present embodiment. First, in the monitoring device 20, the on / off setting of the "inference mode" is performed. When the "inference mode" is set to the on state, the monitoring device 20 periodically acquires image data (R1, R2). The on / off switching of the inference mode is performed via the input / output unit 22 by, for example, clicking the button b1 shown in FIG. 2.

[0030] Subsequently, the monitoring device 20 generates a feature map divided into regions of a predetermined size from the acquired image data. Specifically, the acquired image data is input to a predetermined network structure, and a feature map is generated based on the features extracted from the network structure. For example, when the acquired image is 224×224 pixels, the monitoring device 20 compresses the vertical and horizontal directions by 1 / 8 and generates a feature map of 28×28 pixels. In other words, the monitoring device 20 generates a feature map having features for each region (refer to the broken line in FIG. 5) formed when the vertical and horizontal directions of the acquired image are divided into 28 (R3). Note that Wide-ResNet or the like can be used as the network structure.

[0031] Next, the monitoring device 20 calculates the abnormality degree for each divided area (R4). When calculating the abnormality degree, the average value and covariance matrix of the feature amounts corresponding to each divided area are obtained from N pieces (N is an arbitrary natural number) of normal image data. Then, the abnormality degree is calculated by comparing the N-dimensional normal distribution of the feature amounts of each divided area obtained from the normal image data with the feature amounts of each divided area obtained from the acquired image data. Further, the monitoring device 20 creates a heat map according to the abnormality degree of each area.

[0032] When the abnormality degree in each divided area of the monitoring device 20 is greater than the threshold value, it is determined that an abnormality has occurred, and a caution signal is output (R5-Yes, R6). The monitoring device 20 transmits an e-mail or the like stating that an abnormality has occurred to the user terminal device 30 of the monitor in response to the output of the caution signal.

[0033] On the other hand, when the abnormality degree in each divided area of the monitoring device 20 is less than or equal to the threshold value, it records that no abnormality has occurred in association with the time, and repeats the processing of steps R1 to R5 every time a certain period of time elapses until the inference mode is turned off (R5-No, R1-No).

[0034] Note that the above-described threshold value can be changed by the monitor in a timely manner. Specifically, on the screen G1 shown in FIG. 2, when the "threshold setting" column c1 is clicked, a screen G2 as shown in FIG. 6 is displayed, and the threshold value can be changed.

[0035] In addition, the monitor can also change the setting of the threshold value for each area of the image data. Specifically, on the screen G1 shown in FIG. 2, when the "watched area setting" column c2 is clicked, an image G3 of the monitoring target as shown in FIG. 7 is displayed. Here, the image G3 shown in FIG. 7 is divided into 28 areas vertically and horizontally. When the image is clicked with a pointer or the like, the color of the area where the pointer is placed changes. Then, a different threshold value can be individually set for the area whose color has changed from the other areas.

[0036] (2-2) Automatic relearning mode FIG. 8 is a flowchart for explaining the operation of the "automatic relearning mode" of the monitoring device 20 according to the present embodiment. When monitoring an object using an image with a fixed shooting area, there may be a case where no abnormality is found in the image but it is determined as an abnormality. This is caused by reasons such as the camera moving from a fixed position or changing its orientation. If the position of the camera changes, unless the anomaly detection model is reconstructed using the image data captured by the camera at the changed position, an attention signal will continue to be output even though no anomaly occurs thereafter.

[0037] Therefore, when the attention signal occurs multiple times within a predetermined period, the monitoring device 20 prompts the execution of the automatic relearning mode in which the anomaly detection model is automatically reconstructed. Specifically, an email stating that the execution of the automatic relearning mode is prompted is sent to the user terminal device 30 of the monitor. Then, when the monitor turns on the "automatic relearning mode", the relearning process is executed (S1). The on / off switching of the relearning mode is performed via the input / output unit 22 by, for example, clicking the button b2 shown in FIG. 2.

[0038] When the "automatic relearning mode" of the monitoring device 20 is turned on, the "inference mode" is turned off (S2). Then, the monitoring device 20 periodically (at a first time interval) acquires image data until 50 sheets (the first number of sheets) of image data of the monitoring target are accumulated (S3 to S5). Here, for example, image data is acquired every 3 hours.

[0039] Next, the monitoring device 20 constructs a "first determination model" indicating whether the input image data is normal based on the 50 sheets of accumulated image data. At this time, all 50 sheets of accumulated image data are regarded as normal image data. Then, the first determination model is constructed by the 50 sheets of accumulated image data and a network structure (for example, Wide-ResNet) for constructing the anomaly detection model.

[0040] Subsequently, the monitoring device 20 periodically acquires image data (at a second time interval) until 100 pieces (second number) of image data of the monitoring target are accumulated (S7 to S9). At this time, since the first determination model is used to remove abnormal image data, the image data can be acquired efficiently.

[0041] Also, at any timing of steps S7 to S9, the monitor can check the image data accumulated so far and correct the label as to whether it is normal or abnormal (S8). Specifically, when the "Outlier Removal" column c3 is pressed on the screen G1 shown in FIG. 2, the screen transitions to a screen G4 on which the images accumulated so far are displayed. As shown in FIG. 9, the screen G4 displays the corresponding image data together with radio boxes r1 and r2 that can alternatively select "Normal" or "Abnormal". Thereby, even if abnormal image data is included in the image data regarded as normal image data in steps S3 to S5, such abnormal image data can be removed, and the reliability of the abnormality determination model can be enhanced. For example, when there is an operator in the control room of the monitoring target and the operator is reflected in the image, the degree of abnormality of that part will be calculated to be large. Therefore, the monitoring device 20 is configured to be able to exclude an image that causes such an outlier.

[0042] Also, when the "False Judgment Correction" column c4 is pressed on the screen G1 shown in FIG. 2, the screen transitions to a screen G5 on which the images determined to be abnormal so far are displayed. As shown in FIG. 10, the screen G5 displays the corresponding image data together with radio boxes r3 and r4 that can alternatively select "Normal" and "Abnormal". Thereby, when normal image data is included in the image data determined to be abnormal by the first determination model in step S7, such misjudged normal image data can be added to the learning image. As a result, the reliability of the abnormality determination model can be improved. Note that the image data corrected as normal image data can be copied to a plurality of pieces instead of one piece and added as a learning image.

[0043] Next, the monitoring device 20 constructs a "second determination model" indicating whether the input image data is normal or not based on the accumulated 100 pieces (second number of pieces) of image data (S10). The second determination model is constructed using a network structure (for example, Wide-Res Net) for constructing an anomaly determination model and the accumulated 100 pieces of image data. Also, when the second determination model is constructed, the monitoring device 20 turns on the inference mode (S11). This quickly restores the monitoring of the object.

[0044] Subsequently, the monitoring device 20 periodically (at a third time interval) acquires image data until 1000 pieces (for the third number of pieces) of image data of the monitoring target are accumulated (S12 to S16). At this time, by using the second determination model to remove abnormal image data, image data can be acquired efficiently. Also, at any timing of steps S12 to S14, the accumulated image data can be checked, and the label indicating whether it is normal or abnormal can be corrected (S13). The process in step S13 is the same as that in step S8.

[0045] In addition, in the monitoring device 20, each time 100 pieces (fourth number of pieces) of image data are accumulated while the processes in steps S12 to S16 are being performed, the second determination model is reconstructed using the accumulated data (steps S14-Yes, S15, S16-No). This can enhance the reliability of the second determination model.

[0046] Then, when 1000 pieces of image data of the monitoring target are accumulated, the monitoring device 20 ends the automatic re-learning mode and sets the second determination model at that time as a "new anomaly determination model" (S16-Yes, S17).

[0047] In the above description, the number of accumulated pieces of image data of the monitoring target is set to 1000 pieces, but this numerical value is only set for the convenience of explanation. That is, the number of accumulated pieces of image data of the monitoring target is appropriately set to the number of pieces necessary and sufficient for constructing a "new anomaly determination model".

[0048] Also, in the above description, although the process of reconstructing the second determination model is executed several times, it is not always necessary to execute such a process, and steps S12 to S16 may be omitted as appropriate. When steps S12 to S16 are not executed, the second determination model constructed in step S10 is set as the "new anomaly determination model".

[0049] (3) Features of the wind power generation monitoring device As described above, the monitoring device 20 according to the present embodiment uses an anomaly determination model to determine whether the periodically acquired image of the monitoring target is abnormal, and outputs a warning signal when it is determined to be abnormal by the determination unit 24B. Therefore, the monitor does not need to constantly visually monitor the target object, and the monitoring burden can be reduced.

[0050] In particular, the monitoring device 20 is suitable for monitoring power generation facilities provided in remote locations. For example, this monitoring device 20 applied to the nacelle internal equipment of a wind turbine with severe shaking and movement, or the blades of a wind turbine, etc., collects images and automatically reconstructs the anomaly determination model. Therefore, even when the camera's angle of view or the like changes, it has a remarkable feature that it does not require maintenance work such as correcting the camera's position.

[0051] In addition, when the monitoring device 20 outputs a warning signal a predetermined number of times within a predetermined period, it executes the reconstruction of the anomaly determination model, so that the anomaly determination model can be reconstructed at an appropriate timing.

[0052] In addition, in the monitoring device 20, an anomaly determination model constructed by calculating a predetermined statistic from normal image data using a predetermined network structure is used. Since it is not necessary to collect image data other than normal image data, the anomaly determination model can be easily constructed.

[0053] In addition, in the monitoring device 20 according to the present embodiment, the anomaly determination model is automatically reconstructed. Here, in order to reconstruct the anomaly determination model, it is necessary to collect a large number of image data. Since the anomaly determination model cannot be used while collecting the image data required for learning, it becomes impossible to execute the anomaly determination of the monitoring target during that time. On the other hand, according to the method executed by the automatic relearning mode described above, the first determination model and the second determination model are constructed step by step, and the inference mode is turned on without waiting for the reconstruction of the anomaly determination model to be completed, so that the determination process can be quickly restored.

[0054] In addition, according to the method executed by the automatic relearning mode described above, since the second determination model is reconstructed a predetermined number of times, a highly reliable anomaly determination model can be provided.

[0055] In addition, the monitoring device 20 displays the accumulated image data and accepts the setting of whether it is normal or not. Thereby, even when the image data is automatically collected and the anomaly determination model is constructed, a highly reliable anomaly determination model can be provided. In short, in the relearning method according to the present embodiment, instead of leaving a slight room for improvement in reliability, the anomaly determination model is constructed early and the reliability can be increased later. In particular, in the case of a monitoring target with little change over time, since the acquired image data is often normal, such a relearning method can be preferably used.

[0056] (4) Modification example In the above description, the anomaly determination model is constructed by calculating a predetermined statistic from the normal image data using a predetermined network structure. However, the anomaly determination model is not limited to this. For example, as the anomaly determination model, one constructed by learning using the network structure with the normal image data may be adopted. Such an anomaly determination model can be constructed by an Auto Encoder, a GAN, or the like. Also in this case, since only the normal image data needs to be collected, the anomaly determination model can be easily constructed.

[0057] Also, in the above description, the user terminal device 30 and the monitoring device 20 were described as separate devices, but they may be integrally configured. Furthermore, the monitoring device 20 may be constructed as a system with an arbitrary configuration. For example, the monitoring device 20 may be provided inside the control room, or may be provided on the cloud that can communicate via a communication device provided inside the control room. Also, the monitoring device 20 may be configured such that parts other than the learning unit are provided inside the control room and only the learning unit is provided on the cloud.

[0058] <Other Embodiments> The present disclosure is not limited to the above-described embodiments as they are. The present disclosure can be embodied by modifying the components without departing from the gist thereof at the implementation stage. Also, the present disclosure can form various disclosures by appropriately combining a plurality of components disclosed in the above-described embodiments. For example, some components may be deleted from all the components shown in the embodiments. Furthermore, components may be appropriately combined from different embodiments.

Description of Reference Numerals

[0059] 10 Control Room (Object) 15 Camera 20 Monitoring Device 21 Storage Unit 22 Input / Output Unit 23 Communication Unit 24 Processing Unit 24A Image Acquisition Unit 24B Determination Unit 24C Learning Unit

Claims

1. An image acquisition unit that periodically acquires images of a monitoring target; A storage unit that stores information for executing an abnormality determination model that is constructed using a predetermined network structure and normal image data and determines an abnormality when the input image is abnormal; A determination unit that determines whether an image of a monitoring target periodically acquired is abnormal using the abnormality determination model; An output unit that outputs a warning signal when determined to be abnormal by the determination unit; A learning unit that constructs the abnormality determination model using the normal image data; A monitoring device for wind power generation, comprising the above components.

2. The monitoring device for wind power generation according to claim 1, wherein the process by the determination unit is executed when the process by the learning unit is being executed.

3. The monitoring device for wind power generation according to claim 1, further comprising an input / output unit that displays image data and accepts input of a setting for whether it is normal or not. The monitoring device for wind power generation according to claim 1.

4. When the warning signal is output a predetermined number of times within a predetermined period, re-learning of the abnormality determination model is executed. The monitoring device for wind power generation according to claim 1.

5. The abnormality determination model is constructed by learning the network structure using normal image data. The monitoring device for wind power generation according to claim 1.

6. The abnormality determination model is constructed by calculating a predetermined statistic from normal image data using the network structure. The monitoring device for wind power generation according to claim 1.

7. A re-learning method for re-learning the abnormality determination model used in the monitoring device for wind power generation according to claim 1, comprising: When the re-learning mode is on, accumulating image data at a first time interval until a first number of images is accumulated; Constructing a first determination model that determines whether an input image is normal using the accumulated image data and the network structure; Accumulating the first determination post-image data determined to be normal by the first determination model at a second time interval, displaying the first determination post-image data, and accepting a setting for whether it is normal or not; When the first determination post-image data set to be normal is accumulated for a second number of images, constructing a second determination model that determines whether an input image is normal using the first determination post-image data and the network structure; Setting the second determination model as a new abnormality determination model.

8. Accumulate the second post-judgment image data determined to be normal by the second determination model at a third time interval, display the first post-judgment image data and / or the second post-judgment image data, and accept a setting of whether it is normal. Until the third number of sheets is accumulated, reconstruct the second determination model every time the second post-judgment image data set to be normal is accumulated by the fourth number of sheets. Set the second determination model after reconstruction has been performed a predetermined number of times as a new anomaly determination model. The relearning method according to claim 7.

Images