JP7632966B2 - Anomaly detection system, anomaly detection method, and anomaly detection program - Google Patents

Description

The present invention relates to an anomaly detection system, an anomaly detection method, and an anomaly detection program.

Conventionally, anomaly detection systems have been known that detect anomalies (equipment anomalies, product anomalies, etc.) based on various sensor data acquired from a production line (e.g., process data such as temperature data, pressure data, and flow rate data during production).

The anomaly detection system acquires and learns from various sensor data when an anomaly occurs (anomaly data) and various sensor data when no anomaly occurs (normal data), and uses the generated trained model to detect anomalies.

JP 2008-299640 A

However, in the case of the above anomaly detection system, after it is introduced into a production line, the detection performance may deteriorate over time. For example, if a new anomaly that has not been learned occurs, the system may not be able to properly detect the anomaly.

For this reason, it is necessary to build an autonomous anomaly detection system that can adapt to new anomalies that occur after it is introduced to the production line.

In one aspect, the aim is to provide an autonomous anomaly detection system.

According to one aspect, the anomaly detection system comprises:
An extraction unit that extracts feature amounts from sensor data acquired from a production line;
an identification unit that identifies a distribution area in which features extracted from sensor data in a state in which no abnormality occurs are distributed, the distribution area being a predetermined distribution area in a feature space;
a determination unit that determines whether the acquired sensor data is normal data by determining whether the distribution region includes the feature extracted by the extraction unit;
a notification unit that notifies a result of the determination by the determination unit in the production line;
a change unit that changes the distribution area when information regarding the correctness of the judgment result is fed back from the production line in response to the notification of the judgment result ,
The identification unit is
The variational autoencoder is a variational autoencoder in which sensor data acquired when no abnormalities are occurring is input to the variational autoencoder, and the data output from the variational autoencoder is trained to approximate the sensor data acquired when no abnormalities are occurring. By inputting the sensor data acquired when no abnormalities are occurring to an encoder constituting the variational autoencoder in which the training process has been completed, the distribution region is identified based on features output from the encoder constituting the variational encoder in which the training process has been completed .

It is possible to provide an autonomous anomaly detection system.

FIG. 1 is a diagram illustrating an example of a system configuration of an anomaly detection system in a learning phase. FIG. 2 illustrates an example of a hardware configuration of a learning device. FIG. 2 is a diagram illustrating details of each part of the learning device. FIG. 1 is a diagram illustrating an example of a system configuration of an anomaly detection system in an inspection phase. FIG. 2 illustrates an example of a hardware configuration of an inspection apparatus. FIG. 2 is a diagram showing details of each part of the inspection device. 13 is a flowchart showing the flow of a learning process. 13 is a flowchart showing the flow of inspection and autonomy processing.

Each embodiment will be described below with reference to the accompanying drawings. Note that in this specification and the drawings, components having substantially the same functional configuration are designated by the same reference numerals, and redundant description will be omitted.

[First embodiment]
<System configuration of anomaly detection system in the learning phase>
First, a description will be given of a system configuration in a learning phase of an anomaly detection system according to a first embodiment. Fig. 1 is a diagram illustrating an example of a system configuration of an anomaly detection system in a learning phase.

As shown in FIG. 1, the anomaly detection system 100 in the learning phase has a production line 110, an inspection line 120, and a learning device 140.

The production line 110 is a facility for producing a specified product, and on the production line 110, various tasks related to the production process are performed by workers 111 such as operators. In addition, various sensors (not shown) are arranged on the production line 110, and in the learning phase, a sensor data group 130, which is a set of time-series sensor data detected by the various sensors during production on the production line 110, is transmitted to the learning device 140.

In the first embodiment, the sensor data group 130 includes various process data detected by any sensors arranged on the production line 110. For example, if the production line 110 is a facility equipped with rollers, the sensor data group 130 includes roller rotation speed data and torque data, and if the production line 110 is a facility in which products are transported while being stretched between rollers, the sensor data group 130 includes tension data applied to the products. Also, for example, if the production line 110 is a facility equipped with a processing furnace, the sensor data group 130 includes temperature data and pressure data inside the furnace, flow rate data of gas supplied to the furnace, etc.

The worker 111 monitors whether the equipment of the production line 110 is normal or not, and notifies the learning device 140 of the monitoring result. As a result, the learning device 140 detects that the sensor data group 130 is
- Whether the data is a set of sensor data (abnormal data) when an abnormality occurs,
- Is the sensor data group in a state where no abnormality is occurring (normal data)?
can be divided into:

In the first embodiment, the abnormalities monitored by the worker 111 include, for example, abnormalities such as abnormal noise or vibration occurring on the production line 110, or abnormalities such as a jig not being installed correctly.

The inspection line 120 is a facility for inspecting the quality of products manufactured on the production line 110, and the inspection is performed by an operator 121 such as an inspector. The operator 121 inspects whether the products are abnormal (whether they have a predetermined quality) and notifies the learning device 140 of the inspection result. As a result, the learning device 140 converts the sensor data group 130 into
- Whether it is a group of sensor data (abnormal data) when an abnormality occurs (when the product does not have the specified quality),
- Is the sensor data group (normal data) in a state where no abnormality has occurred (when the product has a specified quality)?
can be divided into:

In the first embodiment, the abnormalities inspected by the worker 121 include, for example, an abnormality such as the film thickness of the product being above or below a reference value, or an abnormality such as the product containing a defect.

A learning program is installed in the learning device 140, and by executing the program, the learning device 140 functions as a selection unit 141, a learning unit 142, and a feature space generation unit 143.

The selection unit 141 acquires the sensor data group 130, extracts normal data (after excluding abnormal data), and stores the data as learning data in the learning data storage unit 144.

The learning unit 142 has a VAE (Variational Autoencoder). The learning unit 142 performs a learning process on the VAE by inputting learning data to the VAE so that the data output from the VAE approaches the learning data input to the VAE.

The learning unit 142 also inputs learning data to the encoder constituting the VAE for which the learning process has been completed, acquires features (also called latent variables) output from the encoder, and notifies the feature space generation unit 143.

The feature space generation unit 143 is an example of an identification unit, and identifies an area (distribution area) in the feature space (also called latent space) in which the features notified by the learning unit 142 are distributed. The distribution area identified by the feature space generation unit 143 is a distribution area that corresponds to normal data (i.e., the distribution area in which the features extracted from the sensor data group in a state in which no abnormality occurs are distributed).

The method for identifying the distribution region is arbitrary. For example, the center of gravity of all features plotted in the feature space may be calculated, and the distribution region may be identified by setting the distance of 3σ from the center of gravity as the boundary position.

<Hardware configuration of the learning device>
Next, the hardware configuration of the learning device 140 will be described. Fig. 2 is a diagram showing an example of the hardware configuration of the learning device. As shown in Fig. 2, the learning device 140 has a processor 201, a memory 202, an auxiliary storage device 203, an I/F (Interface) device 204, a communication device 205, and a drive device 206. The hardware components of the learning device 140 are connected to each other via a bus 207.

The processor 201 has various computing devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 201 reads various programs (e.g., a learning program, etc.) onto the memory 202 and executes them.

The memory 202 has a primary storage device such as a ROM (Read Only Memory) or a RAM (Random Access Memory). The processor 201 and the memory 202 form a so-called computer, and the processor 201 executes various programs read onto the memory 202, causing the computer to realize, for example, each of the above-mentioned functions (selection unit 141 to feature space generation unit 143).

The auxiliary storage device 203 stores various programs and various data used when the various programs are executed by the processor 201. For example, the learning data storage unit 144 is realized in the auxiliary storage device 203.

The I/F device 204 is a connection device that connects the learning device 140 to an operation device 210 and a display device 211, which are examples of external devices. The I/F device 204 accepts operations on the learning device 140 (for example, input operations of monitoring or inspection results by the workers 111 and 121, or learning operations by an administrator (not shown) of the learning device 140) via the operation device 210. The I/F device 204 also outputs the results of processing by the learning device 140 and displays them to the administrator of the learning device 140 via the display device 211.

The communication device 205 is a communication device for communicating with other devices (in this embodiment, a gateway device (not shown) that transmits a group of sensor data 130 detected by various sensors arranged on the production line 110).

The drive device 206 is a device for setting the recording medium 212. The recording medium 212 here includes media that record information optically, electrically, or magnetically, such as CD-ROMs, flexible disks, and magneto-optical disks. The recording medium 212 may also include semiconductor memory that records information electrically, such as ROMs and flash memories.

The various programs to be installed in the auxiliary storage device 203 are installed, for example, by setting the distributed recording medium 212 in the drive device 206 and reading the various programs recorded on the recording medium 212 by the drive device 206. Alternatively, the various programs to be installed in the auxiliary storage device 203 may be installed by downloading them from a network via the communication device 205.

<Details of each part of the learning device>
Next, details of each unit of the learning device 140 (here, the learning unit 142 and the feature space generating unit 143) will be described. Fig. 3 is a diagram showing details of each unit of the learning device.

As shown in FIG. 3, the learning unit 142 has an encoder 301 and a decoder 302 that constitute the VAE, and a comparison/change unit 303.

The encoder 301 performs dimensional compression of normal data by inputting the learning data 310 and outputs features. As shown in FIG. 3, the normal data of the learning data 310 input to the encoder 301 is multivariate time series data.

The decoder 302 reproduces the original normal data from the normal data that has been dimensionally compressed by the encoder 301.

The comparison/change unit 303 compares the normal data reproduced by the decoder 302 with the normal data of the learning data 310 input to the encoder 301, and updates the model parameters of the encoder 301 and the decoder 302 so that the two match. This completes the learning process for the encoder 301 and the decoder 302.

In this way, by configuring the learning device 140 to use VAE, it is possible to perform unsupervised learning processing.

When the learning process is completed for the encoder 301, the learning unit 142 again inputs normal data from the learning data 310 and notifies the feature space generation unit 143 of the features output from the encoder 301.

The feature space generation unit 143 plots the features notified by the learning unit 142 in the feature space. In FIG. 3, the feature space 321 is a feature space expressed in two dimensions to simplify the explanation, with the horizontal axis representing feature A and the vertical axis representing feature B. Also, the cross marks indicate the positions in the feature space 321 of the features notified by the learning unit 142.

The feature space generation unit 143 also identifies a distribution region 330 in the feature space 321 in which the features notified by the learning unit 142 are distributed. In FIG. 3, the distribution region 330 corresponds to normal data (i.e., the distribution region in which the features extracted from the sensor data group in a state in which no abnormality occurs are distributed).

In this way, by configuring the system to identify the distribution region corresponding to normal data, even if a new abnormality occurs during the inspection phase, it is possible to respond by changing the distribution region, making it easier to apply the system to the new abnormality.

In the example of FIG. 3, a two-dimensional feature space consisting of two types of features is illustrated, but the feature space is not limited to two dimensions, and a feature space with a number of dimensions according to the type of feature notified by the learning unit 142 is formed.

<System configuration of anomaly detection system in the inspection phase>
Next, a description will be given of a system configuration in the inspection phase of the anomaly detection system according to the first embodiment. Fig. 4 is a diagram showing an example of the system configuration of the anomaly detection system in the inspection phase.

As shown in FIG. 4, the anomaly detection system 400 in the inspection phase includes a production line 110, an inspection line 120, alarm devices 430 and 440, and an inspection device 420.

Of these, the production line 110 and the inspection line 120 have already been explained using Figure 1, so explanations will be omitted here.

An anomaly detection program is installed in the inspection device 420, and by executing the program, the inspection device 420 functions as a feature extraction unit 421, a judgment unit 422, and an autonomy unit 423.

The feature extraction unit 421 is an example of an extraction unit, and in the inspection phase, acquires a sensor data group 410 detected during production on the production line 110. The feature extraction unit 421 also has an encoder for which learning processing has been completed, and by inputting the acquired sensor data group 410, acquires the feature output from the encoder and notifies the determination unit 422.

The determination unit 422 determines whether the sensor data group 410 is abnormal data or normal data based on the features notified by the feature extraction unit 421. Specifically, the determination unit 422 determines whether the features notified by the feature extraction unit 421 are included in a distribution region corresponding to normal data in the feature space.

Then, if the determination unit 422 determines that the sensor data group 410 is included in the distribution area corresponding to normal data, it determines that the sensor data group 410 is normal data. If the determination unit 422 determines that the sensor data group 410 is not included in the distribution area corresponding to normal data, it determines that the sensor data group 410 is abnormal data. Note that the determination unit 422 reads out the distribution area in the feature space from the feature space storage unit 424 to make the determination. The feature space storage unit 424 stores the distribution area corresponding to the normal data identified in the learning phase.

The determination unit 422 also outputs the determination result to the notification devices 430, 440 (an example of a notification unit). This allows the notification devices 430, 440 to notify the workers 111, 121 of the determination result.

The autonomy unit 423 acquires the monitoring results or inspection results input by the workers 111, 121 to the notification devices 430, 440 in response to the determination unit 422 outputting the determination results to the notification devices 430, 440 and the notification devices 430, 440 notifying the workers 111, 112.

The correspondence between the judgment results output by the judgment unit 422 and the monitoring results or inspection results of the workers 111, 121 is as shown in table 450. In the "Judgment Result" of table 450, "OK" indicates that the judgment unit 422 has judged that the sensor data group 410 is normal data. In the "Judgment Result" of table 450, "NG" indicates that the judgment unit 422 has judged that the sensor data group 410 is abnormal data.

On the other hand, in the "Monitoring or Inspection Results" of table 450, "OK" means
As a result of monitoring by the operator 111, the information that the equipment of the production line 110 is normal is input to the notification device 430, and
The result of the inspection by the worker 121 is that the product is normal (has a predetermined quality), and this is input to the notification device 440.
Shows.

In addition, in the "Monitoring or Inspection Results" of table 450, "NG" means
As a result of monitoring by the worker 111, the occurrence of an abnormality in the equipment of the production line 110 is input to the alarm device 430, or
The worker 121 has inspected the product and input to the notification device 440 that the product is abnormal (does not have a predetermined quality).
Shows.

In other words, the "monitoring result or inspection result" can be thought of as information about whether the judgment result is correct or not. If the "monitoring result or inspection result" is the same as the "judgment result," the "monitoring result or inspection result" is information indicating that the "judgment result" is correct. On the other hand, if the "monitoring result or inspection result" is different from the "judgment result," the "monitoring result or inspection result" is information indicating that the "judgment result" is incorrect.

The autonomy unit 423 is an example of a change unit, and extracts monitoring or inspection results that differ from the judgment result by the judgment unit 422 from among the monitoring or inspection results input by the workers 111 and 121 to the notification devices 430 and 440. The autonomy unit 423 also acquires features corresponding to the extracted monitoring or inspection results from the feature extraction unit 421.

In addition, the autonomy unit 423 plots the features acquired by the feature extraction unit 421 in the distribution area in the feature space read from the feature space storage unit 424, and changes the distribution area based on its positional relationship with the distribution area.

For example, when the judgment result is "OK" and the monitoring result or inspection result is "NG", the autonomy unit 423 changes the distribution region so that the plotted feature value is outside the distribution region. Also, when the judgment result is "NG" and the monitoring result or inspection result is "OK", the autonomy unit 423 changes the distribution region so that the plotted feature value is included in the distribution region.

The notification device 430 notifies the worker 111 of the judgment result output by the judgment unit 422 of the inspection device 420. In response to notifying the judgment result, the notification device 430 also accepts the monitoring result input by the worker 111, associates it with the judgment result, and transmits it to the inspection device 420.

The notification device 440 notifies the worker 121 of the judgment result output by the judgment unit 422 of the inspection device 420. In response to notifying the judgment result, the notification device 440 also accepts the inspection result input by the worker 121, associates it with the judgment result, and transmits it to the inspection device 420.

In this way, in the anomaly detection system 400, the distribution area is changed by repeating the dialogue between the inspection device 420 and the workers 111, 121 (notification of the judgment result by the inspection device 420 and feedback of the monitoring result or inspection result by the workers 111, 121). As a result, according to the first embodiment, it is possible to realize an autonomous anomaly detection system 400 that can adapt to new anomalies that occur after being introduced into the production line 110.

<Hardware configuration of the inspection device>
Next, the hardware configuration of the inspection device 420 will be described. Fig. 5 is a diagram showing an example of the hardware configuration of the inspection device. As shown in Fig. 5, the hardware configuration of the inspection device 420 is generally the same as the hardware configuration of the learning device 140, so the following description will focus on the differences from the hardware configuration of the learning device 140.

The processor 501 reads various programs (e.g., an anomaly detection program) onto the memory 502 and executes them. The processor 501 executes the various programs read onto the memory 502, and the computer formed by the processor 501 and the memory 202 realizes, for example, the above-mentioned functions (feature extraction unit 421, determination unit 422, and autonomy unit 423).

The auxiliary storage device 503, for example, realizes the feature space storage unit 424.

The I/F device 204 is a connection device that connects the notification devices 430 and 440, which are examples of external devices, with the inspection device 420.

The communication device 505 is a communication device for communicating with other devices (in this embodiment, a gateway device (not shown) that transmits a group of sensor data 410 detected by various sensors arranged on the production line 110).

<Details of each part of the inspection device>
Next, the details of each unit (feature amount extraction unit 421, determination unit 422, and autonomy unit 423) of the inspection device 420 will be described. Fig. 6 is a diagram showing the details of each unit of the inspection device.

As shown in FIG. 6, the feature extraction unit 421 has an encoder 610 for which the learning process has been completed. The feature extraction unit 421 acquires the features output from the encoder 610 by inputting the sensor data groups 410 and 411 detected at times T1 and T2, respectively, during the inspection phase, and notifies the determination unit 422.

In the example of FIG. 6, a sensor data group 410 at time T1 and a sensor data group 411 at time T2 are input to an encoder,
Feature amount 621_T1 (feature amount A, feature amount B at time T1),
Feature amount 622_T1 (feature amount C, feature amount D at time T1),
Feature amount 621_T2 (feature amount A, feature amount B at time T2),
Feature amount 622_T2 (feature amount C, feature amount D at time T2),
and notifying the determination unit 422.

The determination unit 422 has a distribution region in the feature space read from the feature space storage unit 424. The example in FIG. 6 shows that the determination unit 422 has a distribution region 330 in the feature space formed by feature A and feature B, and a distribution region 630 in the feature space formed by feature C and feature D.

Note that in the example of FIG. 6, a two-dimensional feature space is shown for simplicity of explanation, but the feature space is not limited to being two-dimensional. Also, in the example of FIG. 6, for simplicity of explanation, the distribution region is identified by dividing the feature space into multiple (two) feature spaces, but it is also possible to identify one distribution region without dividing the feature space. Alternatively, the distribution region may be identified by dividing the feature space into multiple (three or more) feature spaces.

The determination unit 422 plots the feature amounts notified by the feature amount extraction unit 421 in the respective feature amount spaces to determine whether or not the feature amounts are included in the distribution areas 330 and 630. In the example of FIG. 6, the feature amounts at time T1 are as follows:
Since the feature amount 621_T1 is not included in the distribution region 330, the sensor data group 410 at the time T1 is determined to be abnormal data,
Since the feature amount 622_T1 is included in the distribution area 630, the sensor data group 410 at the time T1 is determined to be normal data.
It shows the situation.

In addition, in the example of FIG. 6, the feature amount at time T2 is as follows:
Since the feature amount 621_T2 is not included in the distribution region 330, the sensor data group 411 at the time T2 is determined to be abnormal data,
Since the feature amount 622_T2 is included in the distribution area 630, the sensor data group 411 at the time T2 is determined to be normal data.
It shows the situation.

Table 641 shows the correspondence between the judgment result by the judgment unit 422 at time T1 and the monitoring result or inspection result by the worker 111 or 121. In the example of FIG. 6, for the sensor data group 410 at time T1, the judgment result based on the distribution area 630 matches the monitoring result or inspection result, but the judgment result based on the distribution area 330 differs from the monitoring result or inspection result.

Furthermore, table 642 shows the correspondence between the judgment result by judgment unit 422 at time T2 and the monitoring result or inspection result by operator 111 or 121. In the example of FIG. 6, for sensor data group 411 at time T2, the judgment result based on distribution area 330 matches the monitoring result or inspection result, but the judgment result based on distribution area 630 differs from the monitoring result or inspection result.

The autonomy unit 423 extracts, for each feature space, from tables 641 and 642, monitoring results or inspection results that differ from the judgment results by the judgment unit 422. In addition, the autonomy unit 423 acquires features corresponding to the extracted monitoring results or inspection results from the judgment unit 422.

The autonomy unit 423 also plots the feature (feature 621_T1) acquired by the determination unit 422 in distribution area 330 in the feature space (feature space formed by feature A and feature B) read from the feature space storage unit 424. Furthermore, the autonomy unit 423 changes distribution area 330 based on the positional relationship between distribution area 330 and feature 621_T1, and generates distribution area 651. Thereby, it is possible to avoid overdetecting normal data as abnormal data when a feature equal to feature 621_T1 is calculated thereafter.

Similarly, the autonomy unit 423 plots the feature (feature 622_T2) acquired by the determination unit 422 in a distribution area 630 in the feature space (feature space formed by feature C and feature D) read from the feature space storage unit 424. Furthermore, the autonomy unit 423 changes the distribution area 630 based on the positional relationship between the distribution area 630 and feature 622_T2 to generate distribution area 652. Thereby, it is possible to avoid abnormal data going undetected when a feature equal to feature 622_T2 is calculated thereafter.

The autonomy unit 423 stores the changed distribution area 651 in the feature space storage unit 424 as a new distribution area in the feature space formed by feature A and feature B. The autonomy unit 423 also stores the changed distribution area 652 in the feature space storage unit 424 as a new distribution area in the feature space formed by feature C and feature D.

In this way, when the monitoring results or judgment results of the workers 111 and 121 are fed back, the autonomy unit 423 appropriately changes the distribution area, making it possible to adapt to new abnormalities and improving the accuracy of detecting abnormal data.

<Learning process flow>
Next, a description will be given of the flow of the learning process performed by the learning device 140. Fig. 7 is a flowchart showing the flow of the learning process.

In step S701, the learning device 140 acquires a group of sensor data detected during production on the production line 110.

In step S702, the learning device 140 extracts normal data from the acquired sensor data group and stores it as learning data in the learning data storage unit 144.

In step S703, the learning device 140 performs a learning process on the VAE using the learning data.

In step S704, the learning device 140 extracts features by inputting the learning data to the encoder that constitutes the VAE for which the learning process has been completed.

In step S705, the learning device 140 plots the extracted features in feature space and identifies the distribution region.

In step S706, the learning device 140 determines whether or not to end the learning phase, and if it determines not to end the learning phase (No in step S706), it returns to step S701.

On the other hand, if it is determined in step S706 that the learning phase is to end, the learning process ends.

<Inspection and autonomy processing flow>
Next, a description will be given of the flow of inspection and autonomous processing by the inspection device 420. Fig. 8 is a flowchart showing the flow of inspection and autonomous processing.

In step S801, the inspection device 420 acquires a group of sensor data at a specified time detected during production on the production line 110.

In step S802, the inspection device 420 extracts features by inputting the acquired sensor data group to the encoder that constitutes the VAE for which the learning process has been completed.

In step S803, the inspection device 420 determines whether the sensor data group at a given time is normal or abnormal based on whether the acquired feature is included in the distribution area.

In step S804, the inspection device 420 notifies the determination result via the notification device.

In step S805, the inspection device 420 acquires the monitoring or inspection results entered by the operator in response to the notification of the judgment result.

In step S806, the inspection device 420 determines whether or not the distribution area needs to be changed. If the determination result matches the monitoring result and the inspection result in step S806, it determines that the distribution area does not need to be changed (No in step S806), and proceeds to step S808.

On the other hand, if the judgment result does not match the monitoring result or inspection result in step S806, it is determined that the distribution area needs to be changed (Yes in step S806), and the process proceeds to step S807.

In step S807, the inspection device 420 changes the distribution area.

In step S808, the inspection device 420 determines whether or not to end the inspection and autonomy processing. If it is determined in step S808 that the inspection and autonomy processing should not be ended (No in step S808), the process returns to step S801.

On the other hand, if it is determined in step S808 that the inspection and autonomy processing should be terminated, the inspection and autonomy processing is terminated.

<Summary>
As is apparent from the above description, the anomaly detection system according to the first embodiment includes:
- Extract features from sensor data acquired from the production line.
- Determine whether the sensor data is normal data by determining whether the features extracted from the sensor data are included in a predetermined distribution region in the feature space, in which the features extracted from the normal data are distributed.
The judgment results are notified on the production line.
In response to notification of the judgment result, when information regarding the success or failure of the judgment result is fed back from the production line, the distribution area is changed.

In this way, by configuring the distribution area to change based on feedback from the production line, the first embodiment makes it possible to realize an autonomous anomaly detection system that can adapt to new anomalies that occur after being introduced to the production line.

Second Embodiment
In the above first embodiment, the autonomy unit 423 has been described as changing the distribution region each time a group of sensor data for a predetermined time range at times T1 and T2 is acquired. However, the timing at which the autonomy unit 423 changes the distribution region is arbitrary, and the distribution region may be changed when a certain number of feature amounts calculated for each predetermined time range that are necessary for changing the distribution region have been accumulated.

In the above first embodiment, the case has been described where the autonomy unit 423 changes the distribution area to realize an autonomous anomaly detection system. However, the method for realizing an autonomous anomaly detection system is not limited to this. For example, in addition to the autonomy unit 423 sequentially changing the distribution area,
- Sensor data whose features are distributed near the boundary of the distribution region, or
- Sensor data whose features were used to change the distribution area,
When a certain number of sensor data are accumulated, the learning unit 142 may perform a re-learning process on the VAE using the sensor data (however, normal data), and the feature space generation unit 143 may be configured to re-identify the distribution area.

In the above first embodiment, a case where monitoring results or inspection results are fed back has been described. However, the information regarding the correctness of the judgment result is not limited to this. For example, information directly indicating the correctness of the judgment result (for example, "0" if the judgment result is correct, and "1" if the judgment result is incorrect) may be fed back.

In addition, in the first embodiment, the learning device 140 and the inspection device 420 are described as being configured separately, but the learning device 140 and the inspection device 420 may be configured as an integrated device.

The present invention is not limited to the configurations shown here, including combinations of the configurations and other elements in the above-mentioned embodiments. These points can be changed without departing from the spirit of the present invention, and can be appropriately determined according to the application form.

100: Anomaly detection system 110: Production line 120: Inspection line 130: Sensor data group 140: Learning device 141: Selection unit 142: Learning unit 143: Feature space generation unit 301: Encoder 302: Decoder 321: Feature space 330: Distribution area 410, 411: Sensor data group 420: Inspection device 421: Feature extraction unit 422: Determination unit 423: Autonomization unit 430: Notification device 440: Notification device

Claims (8)

An extraction unit that extracts feature amounts from sensor data acquired from a production line;
an identification unit that identifies a distribution area in which features extracted from sensor data in a state in which no abnormality occurs are distributed, the distribution area being a predetermined distribution area in a feature space;
a determination unit that determines whether the acquired sensor data is normal data by determining whether the distribution region includes the feature extracted by the extraction unit;
a notification unit that notifies a result of the determination by the determination unit in the production line;
a change unit that changes the distribution area when information regarding the correctness of the judgment result is fed back from the production line in response to the notification of the judgment result ,
The identification unit is
A variational autoencoder is a variational autoencoder in which sensor data acquired in a state where no abnormality occurs is input to the variational autoencoder, and the data output from the variational autoencoder is subjected to a learning process so as to approximate the sensor data acquired in a state where no abnormality occurs, and the sensor data acquired in a state where no abnormality occurs is input to an encoder constituting the variational autoencoder in which the learning process has been completed, and the distribution region is identified based on a feature quantity output from the encoder constituting the variational encoder in which the learning process has been completed.
Anomaly detection system. The change unit is
The anomaly detection system according to claim 1 , further comprising: a distribution region that changes, when feedback is received indicating that the determination result is incorrect, such that the determination result becomes correct. The change unit is
When the feature quantity extracted by the extraction unit is not included in the distribution region, the feature quantity extracted by the extraction unit is changed so as to be included in the distribution region;
3. The anomaly detection system according to claim 2, wherein, when the feature extracted by the extraction unit is included in the distribution region, the feature extracted by the extraction unit is changed so that it is not included in the distribution region. The determination unit is
2. The anomaly detection system according to claim 1, wherein when the distribution area is changed by the change unit, the system determines whether the acquired sensor data is normal data by determining whether the feature extracted by the extraction unit is included in the changed distribution area. The anomaly detection system according to claim 1, wherein the notification unit, in response to notifying the judgment result, receives information on the correctness or incorrectness of the judgment result input by an operator of the production line and transmits the information to the change unit. When sensor data determined to be normal or not by the determination unit is stored, the sensor data has a feature amount distributed near a boundary position of the distribution region, or the sensor data has a feature amount used to change the distribution region, a re-learning process is performed on the variational autoencoder using the stored sensor data;
6. The anomaly detection system according to claim 5, wherein the identification unit inputs the accumulated sensor data to an encoder constituting a variational autoencoder for which the re-learning process has been completed, and re-identifies the distribution region based on features output from the encoder constituting the variational autoencoder for which the re-learning process has been completed. An extraction step of extracting feature amounts from sensor data acquired from a production line;
a step of identifying a distribution region in which the feature values extracted from the sensor data in a state where no abnormality occurs are distributed, the distribution region being a predetermined distribution region in the feature value space;
a determination step of determining whether the acquired sensor data is normal data by determining whether the distribution region includes the feature extracted in the extraction step;
a notification step of notifying the result of the determination step in the production line;
and a change step of changing the distribution area when information on the correctness of the judgment result is fed back from the production line in response to the notification of the judgment result ,
The identifying step includes:
A variational autoencoder is a variational autoencoder in which sensor data acquired in a state where no abnormality occurs is input to the variational autoencoder, and the data output from the variational autoencoder is subjected to a learning process so as to approximate the sensor data acquired in a state where no abnormality occurs, and the sensor data acquired in a state where no abnormality occurs is input to an encoder constituting the variational autoencoder in which the learning process has been completed, and the distribution region is identified based on a feature quantity output from the encoder constituting the variational encoder in which the learning process has been completed.
Anomaly detection methods. An extraction step of extracting feature amounts from sensor data acquired from a production line;
a step of identifying a distribution region in which the feature values extracted from the sensor data in a state where no abnormality occurs are distributed, the distribution region being a predetermined distribution region in the feature value space;
a determination step of determining whether the acquired sensor data is normal data by determining whether the distribution region includes the feature extracted in the extraction step;
a notification step of notifying the result of the determination step in the production line;
a change step of changing the distribution area when information on the correctness of the determination result is fed back from the production line in response to the notification of the determination result ,
The identifying step includes:
A variational autoencoder is a variational autoencoder in which sensor data acquired in a state where no abnormality occurs is input to the variational autoencoder, and the data output from the variational autoencoder is subjected to a learning process so as to approximate the sensor data acquired in a state where no abnormality occurs, and the sensor data acquired in a state where no abnormality occurs is input to an encoder constituting the variational autoencoder in which the learning process has been completed, and the distribution region is identified based on a feature quantity output from the encoder constituting the variational encoder in which the learning process has been completed.
Anomaly detection program.

Images