JP7625889B2 - Abnormality determination device, abnormality determination method, and abnormality determination program - Google Patents

Description

This invention relates to a technology that uses a variational autoencoder (VAE) to determine abnormalities in an observed object.

Conventionally, there have been devices that use a variational autoencoder (hereinafter referred to as VAE (Variational Autoencoder)) to determine whether the state of an observed object is normal or abnormal (see, for example, Patent Document 1).

The VAE is a model that includes an encoder that calculates latent variables from the observational data, which is the input data, and a decoder that restores the observational data from the latent variables. The VAE is generated by machine learning that mainly uses observational data of the object under normal conditions. Here, the observational data restored from the latent variables is called restored data.

The device described in Patent Document 1 is configured to calculate the difference between observed data and restored data, and determine whether the state of the observed object is normal or abnormal based on this difference.

Patent No. 6740247

In a variety of fields, there is a demand for improved accuracy in determining whether the state of an observed object is normal or abnormal. VAE is attracting attention as a technology that can meet this demand.

The objective of this invention is to provide a technology that utilizes a variational autoencoder to improve the accuracy of determining whether the state of an observed object is normal or abnormal.

To achieve the above objective, the abnormality determination device of the present invention is configured as follows:

The anomaly determination device is equipped with a variational autoencoder (VAE) that is trained using training data including observation data of the object under normal conditions. The observation data is sensing data sensed by a sensor. For example, the observation data may be image data of a frame image capturing an image of the object under observation, sound capture data capturing the sound of the object under observation, or other sensing data.

The storage unit stores, as a reference distribution, the distribution of latent variables extracted by the encoder of the variational autoencoder from the observation data of the observation target under normal conditions. This reference distribution may be, for example, one obtained during training of the variational autoencoder, or one obtained after training of the variational autoencoder. In addition, the reference distribution is, for example, an n-dimensional normal distribution.

The distribution acquisition unit acquires, as a judgment distribution, the distribution of latent variables extracted by the encoder of the variational autoencoder from the observation data for judging the state of the observation target. The first calculation unit calculates the difference between the reference distribution and the judgment distribution. The first calculation unit calculates, for example, the degree of deviation of the judgment distribution from the reference distribution as the difference between the reference distribution and the judgment distribution. For example, the first calculation unit calculates, as the difference between the reference distribution and the judgment distribution, the difference between the distribution mean of the reference distribution and the distribution mean of the judgment distribution.

The determination unit determines whether the observation target is in an abnormal state based on the difference calculated by the first calculation unit.

With this configuration, it is possible to determine whether the state of the observed object is normal or abnormal without being affected by the accuracy of the variational autoencoder decoder.

In addition, when the observation data of the observation target is image data related to a frame image capturing the observation target, the frame image may be divided into each divided region. In this case, the storage unit may store a reference distribution for each divided region, the distribution acquisition unit may acquire a distribution for judgment for each divided region, the first calculation unit may calculate the difference between the reference distribution and the distribution for judgment for each divided region, and the judgment unit may determine whether the observation target is in an abnormal state for each divided region.

With this configuration, it is possible to determine whether or not an abnormality exists for each region of the frame image (i.e., for each part of the observation object captured in that imaging region).

In addition, for example, the configuration may include a second calculation unit that calculates the difference between the observation data for determining the state of the observation object from which the encoder of the variational autoencoder has extracted latent variables and the reconstructed observation data reconstructed from the observation data for determining the state of the observation object by the decoder of the variational autoencoder, and the determination unit may determine whether the observation object is in an abnormal state based on the difference calculated by the first calculation unit and the difference calculated by the second calculation unit.

In this case, for example, the determination unit determines that the observation target is not in an abnormal state when the determination unit determines that the observation target is not in an abnormal state based on the difference calculated by the first calculation unit and determines that the observation target is not in an abnormal state based on the difference calculated by the second calculation unit.

This configuration can reduce the risk of erroneously determining that the observed state is abnormal when it is actually normal.

This invention improves the accuracy of determining whether the state of an observed object is normal or abnormal.

FIG. 1 is a schematic diagram illustrating a variational autoencoder. 2 is a block diagram showing the configuration of a main part of the abnormality determination device of this example. FIG. 4 is a flowchart showing the operation of the abnormality determination device of this example. 10 is a block diagram showing a configuration of a main part of an abnormality determination device according to a first modified example. FIG. 10 is a flowchart showing the operation of the abnormality determination device of the first modified example. 10 is a flowchart showing the operation of the abnormality determination device of the second modified example. FIG. 7A is a diagram showing an example of observed data, and FIG. 7B is a diagram showing an example in which a frame image is divided into four regions. FIG. 11 is a block diagram showing a configuration of a main part of an abnormality determination device according to a third modified example. 13 is a flowchart showing the operation of the abnormality determination device of the third modified example.

The following describes an embodiment of the invention.

＜1. Application examples＞
FIG. 1 is a schematic diagram for explaining a variational autoencoder. As is well known, the variational autoencoder 1 (hereinafter referred to as VAE1 (Variational Autoencoder)) is composed of a computer. The VAE1 is a model generated by deep learning using training data including normal observation data of an observation target, and has an encoder 2 and a decoder 3. The encoder 2 and the decoder 3 are neural networks. The encoder 2 extracts a latent variable z from the input observation data. The decoder 3 receives the latent variable z as an input and restores the observation data. Here, the observation data restored by the decoder 3 from the latent variable z is called restored data.

The observation data is sensing data sensed by a sensor. For example, the observation data may be image data of a frame image capturing an image of the observation target, sound capture data capturing the sound of the observation target, or other sensing data.

The observation data may be a time series of numerical values, or a frequency image generated by processing such as a Fourier transform.

In deep learning, the encoder 2 learns the conditional probability q(z|x) of extracting a latent variable z from observed data, and the decoder 3 learns the conditional probability p(x|z) of restoring observed data from the latent variable z.

In this example, the encoder 2 and the decoder 3 learn the conditional probability q(z|x) and the conditional probability p(x|z) so that the latent variable z has a specific distribution. The specific distribution may be, for example, an n-dimensional normal distribution, a distribution on the surface of a hypersphere of radius r, or another distribution.

The anomaly determination device of this example includes a VAE1. The anomaly determination device of this example also stores, as a reference distribution, the distribution of latent variables z extracted by the encoder 2 of the VAE1 from the observation data of the object under normal conditions. The reference distribution may be one generated when the encoder 2 and the decoder 3 are learning the conditional probability q(z|x) and the conditional probability p(x|z), or it may be one generated by additionally processing the observation data of the object under normal conditions after the encoder 2 and the decoder 3 have learned the conditional probability q(z|x) and the conditional probability p(x|z).

The anomaly determination device of this example calculates the difference between the distribution of latent variables z extracted from the observation data input by the encoder 2 of the VAE1 and a stored reference distribution. In other words, the anomaly determination device of this example calculates the degree to which the distribution of latent variables z extracted from the observation data input by the encoder 2 of the VAE1 deviates from the reference distribution. Based on the calculated difference, the anomaly determination device of this example determines whether the observed object is in a normal state or an abnormal state.

Therefore, the abnormality determination device of this example can determine whether the observed object is in a normal state or an abnormal state without being affected by the accuracy of the decoder 3 of the VAE 1. This improves the accuracy of determining whether the observed object is in a normal state or an abnormal state.

2. Configuration example
2 is a block diagram showing the configuration of the main parts of the abnormality determination device 10 of this example. The abnormality determination device 10 of this example includes a control unit 11, an input section 12, an output section 13, and a storage section 14.

The control unit 11 of the abnormality determination device 10 is composed of a hardware CPU, memory, and other electronic circuits. The control unit 11 has a VAE1, a distribution acquisition unit 21, a first calculation unit 22, and a determination unit 23.

The input unit 12 receives observation data of an observation target to be determined as to whether it is in a normal state or an abnormal state. As described above, the observation data is sensing data sensed by a sensor. For example, the observation data may be image data of a frame image capturing an image of the observation target, sound pickup data capturing the sound of the observation target, or other sensing data.

The output unit 13 outputs the judgment result (normal state or abnormal state) made by the control unit 11 regarding the observed object to a higher-level device (not shown).

The storage unit 14 stores the reference distribution of the latent variable z described above. The storage unit 14 is preferably configured with a non-volatile memory, but may be a volatile memory. The storage unit 14 may be a removable memory such as a memory card that is attached to the anomaly determination device 10. The storage unit 14 may utilize part of the storage area of a hard disk drive (HDD) or a solid state drive (SSD).

Next, we will explain the VAE1, distribution acquisition unit 21, first calculation unit 22, and judgment unit 23 that the control unit 11 has.

As described above, VAE1 is a model generated by deep learning using training data that includes observation data of the object under normal conditions.

The distribution acquisition unit 21 acquires the distribution of the latent variable z of the observation data extracted by the encoder 2 of the VAE 1 as a distribution for judgment.

The first calculation unit 22 calculates the difference (degree of deviation) between the judgment distribution acquired by the distribution acquisition unit 21 and the reference distribution stored in the storage unit 14. For example, the first calculation unit 22 calculates the distance between the distribution mean of the reference distribution and the distribution mean of the judgment distribution as the difference. The first calculation unit 22 may also calculate the distance for each element of the latent variable z, and calculate the sum, mean, median, etc. as the difference (degree of deviation) between the judgment distribution and the reference distribution. The first calculation unit 22 may also calculate the feature amount of the judgment distribution and the feature amount of the reference distribution, and calculate the difference between these feature amounts as the difference between the judgment distribution and the reference distribution.

The determination unit 23 determines whether the observed object is in a normal state or an abnormal state based on the difference between the reference distribution and the distribution for judgment calculated by the first calculation unit 22. For example, if the difference between the reference distribution and the distribution for judgment is less than a predetermined judgment threshold, the determination unit 23 determines that the observed object is in a normal state. If the difference between the reference distribution and the distribution for judgment is equal to or greater than the predetermined judgment threshold, the determination unit 23 determines that the observed object is in an abnormal state.

When the hardware CPU constituting the control unit 11 of the abnormality determination device 10 executes the abnormality determination program of the present invention, it operates as the VAE1, the distribution acquisition unit 21, the first calculation unit 22, and the determination unit 23. The memory also has an area for expanding the abnormality determination program of the present invention and an area for temporarily storing data generated during execution of the abnormality determination program. The control unit 11 may be an LSI that integrates the hardware CPU, memory, etc. Also, the hardware CPU is a computer that executes the abnormality determination method of the present invention.

<3. Operation example>
3 is a flowchart showing the operation of the abnormality determination device of this example. The abnormality determination device 10 has an observation data storage unit (not shown) that temporarily stores the observation data of the observation target input to the input unit 12. The observation data may be input at regular time intervals, may be input at the timing when a predetermined specific event is detected, or may be input at other timings. The timing when a specific event is detected is, for example, the timing when a specific operation is performed, or the timing when a state change is detected by a sensor other than the sensor that senses the observation data (for example, the timing when a moving object is detected by an infrared sensor, or the timing when a change in the open/closed state of a door is detected by a door open/close sensor).

The abnormality determination device 10 selects observation data to be processed from the observation data temporarily stored in the observation data storage unit (s1). For example, if the observation data is input to the input unit 12 at 100 msec intervals, the input observation data may be selected as the observation data to be processed in the order in which they were input, or the mth (m>1) observation data from the previously selected observation data to be processed may be selected as the observation data to be processed. For example, if the observation data is a moving image of the observation target captured by an imaging device, the frame images of this moving image may be selected in order as the observation data to be processed, or the mth frame image from the frame image previously selected as the observation data to be processed may be selected as the observation data to be processed.

Note that the above method for selecting the observation data to be processed is just one example, and other methods may be used for selection.

The abnormality determination device 10 encodes the observation data to be processed selected in s1 using the encoder 2 of the VAE1, and extracts the latent variable z of this observation data (s2). At this time, the VAE1 may or may not restore the observation data from the latent variable z extracted by the decoder 3.

The distribution acquisition unit 21 acquires the distribution of the latent variable z extracted from the observed data as a distribution for judgment (s3). The first calculation unit 22 calculates the difference between the distribution for judgment acquired in s3 and a reference distribution of the latent variable z stored in the storage unit 14 as the degree of deviation (s4). For example, the first calculation unit 22 calculates the distance between the distribution mean value of the reference distribution and the distribution mean value of the distribution for judgment as the difference.

The first calculation unit 22 may calculate the distance for each element of the latent variable z, and calculate the sum, average, median, etc. of the distance as the difference (degree of deviation) between the judgment distribution and the reference distribution.

The determination unit 23 determines whether the observed object is in a normal state or an abnormal state based on the difference between the reference distribution and the distribution for judgment calculated by the first calculation unit 22 (s5). For example, if the difference between the reference distribution and the distribution for judgment is less than a predetermined judgment threshold, the determination unit 23 determines that the observed object is in a normal state. If the difference between the reference distribution and the distribution for judgment is equal to or greater than the predetermined judgment threshold, the determination unit 23 determines that the observed object is in an abnormal state.

In the abnormality determination device 10, the output unit 13 outputs the determination result in s5 to the higher-level device (s6), and the process returns to s1. In s6, if the determination result is a normal state, the result may not be output to the higher-level device, and may be output to the higher-level device only if the determination result is an abnormal state.

In this way, the abnormality determination device 10 of this example can determine whether the observed object is in a normal state or an abnormal state without being affected by the accuracy of the decoder 3 of the VAE 1. This improves the accuracy of determining whether the observed object is in a normal state or an abnormal state.

4. Modifications
Modification 1
Fig. 4 is a block diagram showing the configuration of the main parts of the abnormality determination device of Modification 1. The abnormality determination device 10A of Modification 1 differs from the abnormality determination device 10 shown in Fig. 2 in that it includes a second calculation unit 24 and a determination unit 25. In addition, the determination unit 25 is configured to determine whether the observation target is in a normal state or an abnormal state, but the process differs from that of the determination unit 23 of the abnormality determination device 10 shown in Fig. 2.

The second calculation unit 24 calculates the difference between the input observation data and the restored data restored by the decoder 3 of the VAE1 from the latent variable z. The second calculation unit 24 may calculate the distance between each element of the input observation data and the corresponding element of the restored data, and calculate the sum of the distances as the difference between the input observation data and the restored data restored by the decoder 3 of the VAE1 from the latent variable z. The second calculation unit 24 may also calculate the feature amount of the input observation data and the feature amount of the restored data, and calculate the difference between these feature amounts as the difference between the observation data and the restored data. As described in the above-mentioned Patent Document 1 (Patent Publication No. 6740247), the second calculation unit 24 may calculate the likelihood that is correlated with the degree of anomaly in each dimension between the input observation data and the restored data restored by the decoder 3 of the VAE1 from the latent variable z, and calculate a score based on this likelihood as the difference between the observation data and the restored data, or may calculate it using other methods.

The determination unit 25 determines whether the observation target is in a normal state or an abnormal state based on the difference between the reference distribution and the determination distribution calculated by the first calculation unit 22 and the difference between the observation data and the restored data calculated by the second calculation unit 24. In this modified example 1, the determination unit 25 determines that the observation target is in a normal state when it determines that the observation target is in a normal state based on the difference between the reference distribution and the determination distribution calculated by the first calculation unit 22 and also determines that the observation target is in a normal state based on the difference between the observation data and the restored data calculated by the second calculation unit 24. In other words, the determination unit 25 determines that the observation target is in an abnormal state when it determines that the observation target is not in a normal state based on one or both of the difference between the reference distribution and the determination distribution calculated by the first calculation unit 22 and the difference between the observation data and the restored data calculated by the second calculation unit 24.

Figure 5 is a flowchart showing the operation of the abnormality determination device of this modified example 1.

The abnormality determination device 10A of this modified example 1 selects observation data to be processed from the observation data temporarily stored in the observation data storage unit (s11). s11 is the same process as s1 in the example above. The abnormality determination device 10A encodes the observation data to be processed selected in s11 using the encoder 2 of the VAE1, and extracts the latent variable z of this observation data (s12). s12 is the same process as s2 in the example above. In addition, the abnormality determination device 10A decodes the latent variable z extracted in s12 using the decoder 3 of the VAE1, and generates restored data of this observation data (s13).

The distribution acquisition unit 21 acquires the distribution of the latent variable z extracted from the observation data as a distribution for judgment (s14). S14 is the same process as S3 in the above example. The first calculation unit 22 calculates the difference between the distribution for judgment acquired in S14 and the reference distribution of the latent variable z stored in the storage unit 14 as the degree of deviation (s15). S15 is the same process as S4 in the above example. The judgment unit 25 performs a first judgment process to judge whether the observation target is in a normal state or an abnormal state based on the difference between the reference distribution and the distribution for judgment calculated by the first calculation unit 22 (s16). S16 is the same process as S5 in the above example.

When the abnormality determination device 10A determines in the first determination process of s16 that the observation target is in an abnormal state, the output unit 13 outputs the abnormal state of the observation target to the higher-level device (s17, s22) and the process returns to s11.

When the abnormality determination device 10A determines that the observation target is in a normal state in the first determination process of s16, the second calculation unit 24 calculates the difference between the observation data and the restored data generated in the decoding process of s3 (s17, s18). The determination unit 25 performs a second determination process (s19) to determine whether the observation target is in a normal state or an abnormal state based on the difference between the observation data calculated in s18 and the restored data.

When the abnormality determination device 10A determines in the second determination process of s19 that the observation target is in an abnormal state, the output unit 13 outputs the abnormal state of the observation target to the higher-level device (s20, s22) and returns to s11. Also, when the abnormality determination device 10A determines in the second determination process of s19 that the observation target is in a normal state, the output unit 13 outputs the normal state of the observation target to the higher-level device (s20, s21) and returns to s11.

In this way, the abnormality determination device 10A of this modified example 1 determines whether the observed object is in a normal state or an abnormal state based on the difference between the determination distribution of the latent variable z obtained from the observed data and the reference distribution of the latent variable z, and the difference between the observed data and the restored data. Therefore, the accuracy of determining whether the observed object is in a normal state or an abnormal state can be improved.

In addition, in this modified example 1, if the first judgment process (judgment based on the difference between the judgment distribution of the latent variable z obtained from the observation data and the reference distribution of the latent variable z) judges that the observation target is not in a normal state, the processes from s18 to s21 are not performed. Therefore, the processing load of the abnormality judgment device 10A is also reduced.

In addition, in the above explanation, the first judgment process (judgment based on the difference between the judgment distribution of the latent variable z obtained from the observed data and the reference distribution of the latent variable z) is configured to be performed before the second judgment process (judgment based on the difference between the observed data and the restored data), but the second judgment process may be performed before the first judgment process. In this case, if the observed object is judged to be not in a normal state in the second judgment process executed first, the first judgment process may not be executed.

As in the above example, the abnormality determination device 10A may be configured such that the output unit 13 does not execute the process of outputting the normal state of the observed object to the higher-level device (the process related to s21).

Modification 2
The abnormality determination device of the second modification has a configuration similar to that of the abnormality determination device 10A of the first modification shown in FIG.

Figure 6 is a flowchart showing the operation of the abnormality determination device of this modified example 2. The abnormality determination device 10A of this modified example executes the processes of s11 to s16 described in modified example 1 (s31 to s36). Furthermore, after completing the process of s36, it executes the processes of s18 and s19 described in modified example 1 (s37, s38).

When the abnormality determination device 10A of this modification 2 determines that the observation target is in a normal state in both the first determination process and the second determination process, the output unit 13 outputs the normal state of the observation target to the higher-level device (s40) and returns to s31.

In addition, when the abnormality determination device 10A of this modified example 2 determines that the observed object is in a normal state in either the first determination process or the second determination process, and determines that the observed object is in an abnormal state in the other process, the output unit 13 outputs a caution-requiring state of the observed object to the higher-level device (s41, s42) and returns to s31.

In addition, in the abnormality determination device 10A of this modified example 2, if the observation target is determined to be in an abnormal state in both the first determination process and the second determination process, the output unit 13 outputs the abnormal state of the observation target to the higher-level device (s41, s43), and the process returns to s31.

The abnormality determination device 10A of the second modification can determine the state of the object of observation in three stages, including a normal state, an abnormal state, and a state requiring attention. This allows more detailed state management of the object of observation.

As in the above example, the abnormality determination device 10A may be configured such that the output unit 13 does not execute the process of outputting the normal state of the observed object to the higher-level device (the process related to s40).

Modification 3
The abnormality determination device 10B of the third modification receives frame images of moving images captured by an imaging device such as a video camera as observation data. Fig. 7A is a diagram showing an example of the observation data. The observation data shown in Fig. 7 is an example in which the observation target is a road on which a vehicle travels.

The abnormality determination device 10B of this variant example 3 determines that the road being observed is in an abnormal state if, for example, there is an obstacle (fallen object) on the road.

The abnormality determination device 10B of this modification 3 divides a frame image, which is observation data, into a number of predetermined divided regions. FIG. 7(B) is a diagram showing an example in which a frame image is divided into four divided regions A to D (divided into two vertically and two horizontally).

Figure 8 is a block diagram showing the configuration of the main parts of the abnormality determination device of this modified example 3. In Figure 8, the same components as those in the abnormality determination device 10 shown in Figure 2 are given the same reference numerals.

The abnormality determination device 10B of this modified example 3 differs in that the control unit 11 has a division unit 30, and that VAEs 1a to 1d are provided for each of the divided regions A to D shown in FIG. 7. In addition, the storage unit 14a stores a reference distribution of the latent variable z for each of the divided regions A to D.

The division unit 30 divides the frame image, which is the observation data, into four division areas A to D shown in FIG. 7(B).

VAE1a is a model generated by deep learning using partitioned region A of the observation data of the object when it is normal. VAE1b is a model generated by deep learning using partitioned region B of the observation data of the object when it is normal. VAE1c is a model generated by deep learning using partitioned region C of the observation data of the object when it is normal. VAE1d is a model generated by deep learning using partitioned region D of the observation data of the object when it is normal.

Figure 9 is a flowchart showing the operation of the abnormality determination device of this modified example 3. The abnormality determination device 10B selects the observation data to be processed (s51). s51 is the same process as s1 and the like described above.

The division unit 30 divides the frame image, which is the selected observation data, into four division areas A to D (s52). The abnormality determination device 10B selects one division area from among the unprocessed division areas A to D that have not been subjected to the processing from s54 onwards (s53). The abnormality determination device 10B performs the same processing as s2 to s5 described above on the division areas A to D selected in s53 (s54 to s57).

At this time, in the encoding process in s54, the abnormality determination device 10B uses VAEs 1a to 1d corresponding to the divided region selected in s53. In addition, in the deviation degree calculation process in s56, the first calculation unit 22 reads out the reference distribution of the latent variable z corresponding to the divided region selected in s53 from the storage unit 14 and calculates the deviation degree.

When the abnormality determination device 10B completes the processes in s54 to s57 for the divided regions A to D selected in s53, it determines whether or not there are any unprocessed divided regions A to D (s58). If there are any unprocessed divided regions A to D, the abnormality determination device 10B returns to s53 and repeats the above processes.

In s58, the abnormality determination device 10B determines that there are no unprocessed divided regions A to D, outputs the determination result in s57 for each divided region A to D (s59), and returns to s51.

In this way, the abnormality determination device 10B of the third modified example can output whether each divided area of the frame image, which is the observation data, is in a normal state or an abnormal state. Therefore, it is possible to notify staff etc. of the locations where abnormalities exist in detail. In other words, staff etc. can respond quickly and appropriately to detected abnormal states.

In addition, in this modification example 3, an example has been described in which a frame image, which is observation data, is divided into four divided regions A to D, but the number of divided regions into which the frame image is divided is not limited to four and may be any number. Also, in this modification example 3, an example has been described in which each divided region A to D is divided so as not to overlap with other adjacent divided regions A to D, but each divided region may be divided so as to partially overlap with other adjacent divided regions A to D.

In addition, when the observed data is a frame image, the location of the abnormal state may be obtained by using a latent variable z that holds the location information through convolution processing such as CNN (Convolution Neural Network).

Note that this invention is not limited to the above-described embodiment as it is, and in the implementation stage, the components can be modified to the extent that does not deviate from the gist of the invention. In addition, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, components from different embodiments may be appropriately combined.

Furthermore, the correspondence between the configuration according to the present invention and the configuration according to the above-mentioned embodiment can be described as follows.
<Additional Notes>
In an abnormality determination device (10) including a variational autoencoder (1) trained using training data including observation data of an object under normal conditions,
a storage unit (14) for storing, as a reference distribution, a distribution of a latent variable (z) extracted by the encoder (2) of the variational autoencoder (1) from the observation data of the object under normal conditions;
a distribution acquisition unit (21) that acquires, as a judgment distribution, a distribution of a latent variable (z) extracted by an encoder (2) of the variational autoencoder (1) from observation data for judging a state of the observation target;
A first calculation unit (22) that calculates a difference between the reference distribution and the judgment distribution;
and a determination unit (23) that determines whether the observation target is in an abnormal state based on the difference calculated by the first calculation unit (22).

1, 1a to 1d...variational autoencoder (VAE)
2... Encoder 3... Decoder 10, 10A, 10B... Abnormality determination device 11... Control unit 12... Input section 13... Output section 14, 14a... Storage section 21... Distribution acquisition section 22... First calculation section 23, 25... Determination section 24... second calculation unit 30... division unit

Claims (9)

1. An anomaly detection device including a variational autoencoder trained using training data including observation data of an object under normal conditions,
a storage unit that stores a distribution of latent variables extracted by an encoder of the variational autoencoder from normal observation data of the object to be observed as a reference distribution;
a distribution acquisition unit that acquires, as a judgment distribution, a distribution of latent variables extracted by an encoder of the variational autoencoder from observation data for judging a state of the observation target;
a first calculation unit that calculates a difference between the reference distribution and the judgment distribution;
a determination unit that determines that the observation target is in an abnormal state if the difference calculated by the first calculation unit is equal to or greater than a determination threshold . The anomaly determination device according to claim 1, wherein the reference distribution is an n-dimensional normal distribution. The anomaly determination device according to claim 1 or 2, wherein the first calculation unit calculates the degree of deviation of the determination distribution from the reference distribution as a difference between the reference distribution and the determination distribution. The anomaly determination device according to any one of claims 1 to 3, wherein the first calculation unit calculates the difference between the distribution average value of the reference distribution and the distribution for determination as the difference between the reference distribution and the distribution for determination. The observation data of the observation target is each divided area obtained by dividing a frame image of the observation target,
the storage unit stores the reference distribution for each of the divided regions;
the distribution acquisition unit acquires the judgment distribution for each of the divided regions;
the first calculation unit calculates a difference between the reference distribution and the judgment distribution for each of the divided regions;
5. The abnormality determination device according to claim 1, wherein the determination unit determines whether the observation target is in an abnormal state for each of the divided regions. a second calculation unit that calculates a difference between observed data for state determination of the object from which an encoder of the variational autoencoder extracts latent variables and reconstructed observed data reconstructed from the observed data for state determination of the object by a decoder of the variational autoencoder;
The abnormality determination device according to any one of claims 1 to 5, wherein the determination unit determines whether the observation target is in an abnormal state based on a determination based on a difference calculated by the second calculation unit in addition to a determination based on the difference calculated by the first calculation unit. 7. The abnormality determination device according to claim 6, wherein the determination unit determines that the observed object is not in an abnormal state when the determination based on the difference calculated by the first calculation unit determines that the observed object is not in an abnormal state and the determination based on the difference calculated by the second calculation unit determines that the observed object is not in an abnormal state. a computer of an anomaly determination device including an encoder of a variational autoencoder trained using training data including observation data of an object under normal conditions, the encoder including a memory unit storing a distribution of latent variables extracted from the observation data of the object under normal conditions as a reference distribution,
a distribution acquisition step in which an encoder of the variational autoencoder acquires a distribution of latent variables extracted from observation data for determining a state of the observation target as a determination distribution;
a first calculation step of calculating a difference between the reference distribution and the judgment distribution;
a determination step of determining that the observation target is in an abnormal state if the difference calculated in the first calculation step is equal to or greater than a determination threshold . A computer of an anomaly determination device including a storage unit in which an encoder of a variational autoencoder trained using training data including observation data of an object under normal conditions stores a distribution of latent variables extracted from the observation data of the object under normal conditions as a reference distribution,
a distribution acquisition step in which an encoder of the variational autoencoder acquires a distribution of latent variables extracted from observation data for determining a state of the observation target as a determination distribution;
a first calculation step of calculating a difference between the reference distribution and the judgment distribution;
and a determination step of determining that the observation target is in an abnormal state if the difference calculated in the first calculation step is equal to or greater than a determination threshold .

Images