WO2019176354A1 - Learning data collection method, learning data collection device, abnormality detection system, and computer program - Google Patents

Abstract

Provided is a learning data collection method for collecting data for additional learning of a learned auto encoder to which object-to-be-inspected data for detection of an abnormality in an object to be inspected is input and from which feature extraction data indicating an extracted feature of the input object-to-be-inspected data is output, the method comprising: statistically comparing a difference between output feature extraction data and a plurality of first object-to-be-inspected data input to the auto encoder for learning of the auto encoder that has not learned yet with a difference between output feature extraction data and a plurality of second object-to-be-inspected data input to the auto encoder having learned; and when a statistical difference is found, selecting, as data for additional learning, second object-to-be-inspected data related to a normal object to be inspected.

Description

Learning data collection method, learning data collection device, anomaly detection system, and computer program

The present disclosure relates to a learning data collection method, a learning data collection device, an anomaly detection system, and a computer program.
This application claims priority based on Japanese Patent Application No. 2018-45727 filed on Mar. 13, 2018, and incorporates all the content described in the above Japanese application.

The factory production line has a defect detection system that detects product defects. The defect detection system detects a defect of an inspection object using a sensor in a production line of a factory. The sensor is, for example, a camera. In a defect detection system using machine learning such as deep learning, it is possible to learn sensor values and automatically set determination parameters necessary for defect detection.

Non-Patent Document 1 discloses a curriculum learning technique for automatically selecting sample data used for deep neural network learning.
In Non-Patent Document 2, additional learning is performed when abnormal data that does not match any learned category is detected using SOM, which is one of machine learning techniques, and abnormal data is learned as a new category. Technology is disclosed.

The learning data collection method of the present disclosure includes a learned auto encoder that receives inspection target data for detecting an abnormality of an inspection target and outputs feature extraction data obtained by extracting features of the input inspection target data. A learning data collection method for collecting data for additional learning, wherein a plurality of first inspection object data input to the auto encoder and a plurality of output data are output in order to learn the auto encoder before learning. The difference between the first feature extraction data and the plurality of second inspection target data input to the auto encoder after learning and the difference between the plurality of second feature extraction data output after learning are statistically compared, If there is a difference, the second inspection object data related to the normal inspection object is selected as additional learning data.

The learning data collection device of the present disclosure includes a learned auto-encoder that receives inspection target data for detecting an abnormality of the inspection target and outputs feature extraction data obtained by extracting features of the input inspection target data. A learning data collection device for collecting data for additional learning, wherein a plurality of first inspection object data input to the auto encoder and a plurality of output data are output to the auto encoder before learning. A comparison unit that statistically compares the difference between the first feature extraction data and the plurality of second inspection target data input to the auto encoder after learning and the difference between the plurality of second feature extraction data output. And a selection unit that selects the second inspection object data relating to the normal inspection object as data for additional learning when there is a statistical difference. The

The anomaly detection system of the present disclosure includes a learned auto encoder that receives inspection object data for detecting an abnormality of an inspection object, outputs feature extraction data obtained by extracting features of the input inspection object data, and An anomaly detection processing unit for detecting an anomaly of the inspection object by comparing the inspection object data input to the auto encoder and the output feature extraction data, and additional learning of the auto encoder The learning data collecting device for collecting data, and a learning processing unit for additionally learning the auto encoder based on the data collected by the learning data collecting device.

The computer program of the present disclosure adds a learned autoencoder that receives inspection target data for detecting whether there is an abnormality in the inspection target object and outputs feature extraction data obtained by extracting features of the input inspection target data A computer program for causing a computer to collect data for learning, the plurality of first inspection target data input to the auto encoder for causing the computer to learn the auto encoder before learning, and The difference between the plurality of output first feature extraction data and the difference between the plurality of second inspection target data input to the auto encoder after learning and the output plurality of second feature extraction data are statistically calculated. If there is a statistical difference, the second inspection object data relating to the normal inspection object is To execute a process of selecting a data pressurized learning.

It is a schematic diagram which shows the structural example of the anomaly detection system which concerns on Embodiment 1. FIG. It is a block diagram which shows the structural example of the anomaly detection apparatus which concerns on Embodiment 1. FIG. It is a block diagram which shows the structural example of the anomaly detection apparatus which concerns on Embodiment 1. FIG. It is a schematic diagram which shows the neural network and learning method of an auto encoder. It is a schematic diagram which shows the function of the auto encoder which extracts the characteristic of a test target object and outputs characteristic extraction data. It is a schematic diagram which shows the production | generation method of difference data. It is a schematic diagram which shows an example of the test object data obtained by imaging the connector which is a test object, feature extraction data, and difference data. It is a schematic diagram which shows an example of the test object data obtained by imaging the connector which is a test object, feature extraction data, and difference data. It is a schematic diagram which shows an example of the test object data obtained by imaging the connector which is a test object, feature extraction data, and difference data. It is a flowchart which shows the process sequence which concerns on the collection of the data for additional learning, and additional learning. It is a graph which shows the statistical difference of difference amount. It is a contour map which shows the mixed normal distribution obtained based on the test object data for learning obtained by imaging a normal test object. It is a schematic diagram which shows the difference data obtained during the anomaly detection process after learning. It is a schematic diagram which shows the method of discriminate | determining the test object data obtained by imaging the test object with abnormality, and the test object data obtained by imaging a normal test object. It is a schematic diagram which shows the method of discriminate | determining the test object data obtained by imaging the test object with abnormality, and the test object data obtained by imaging a normal test object. 10 is a flowchart illustrating a collection processing procedure of additional learning data according to the second embodiment. It is a schematic diagram which shows the selection reception method of the data for learning. It is a schematic diagram which shows the selection reception method of the data for learning. It is a schematic diagram which shows the structural example of the anomaly detection system which concerns on Embodiment 3. FIG.

[Problems to be solved by the present disclosure]
The sensor value slightly changes when the lighting environment in the factory changes, the sensor becomes dirty, or the location of the production line moves. As a result, the input value to the defect detection system changes. Even a defect detection system using machine learning such as deep learning cannot cope with a change in sensor value after learning, and there is a possibility that the detection accuracy of a defective product is lowered. There is a demand for a defect detection system that can automatically perform additional learning according to changes in the factory environment.

The important point when performing additional learning automatically is the selection of learning data. For example, in the anomaly detection system using an auto encoder, learning data in which inspection object data obtained by imaging a non-defective inspection object and inspection object data obtained by imaging a defective inspection object are mixed. If additional learning is performed using, an abnormality of the inspection object cannot be detected.

The purpose of this disclosure is to add a learned auto encoder for anomaly detection to respond to environmental changes when sensor values change due to changes in the lighting environment in the factory, sensor contamination, movement of the installation location of the line, etc. An object is to provide a learning data collection method, a learning data collection device, an anomaly detection system, and a computer program that can collect data for learning.

[Effects of the present disclosure]
According to the present disclosure, when a sensor value changes due to changes in the lighting environment in the factory, contamination of the sensor, movement of the installation location of the line, etc., a learned auto-encoder for anomaly detection is added to respond to the environmental change It is possible to provide a learning data collection method, a learning data collection device, an anomaly detection system, and a computer program that can collect data for learning.
[Description of Embodiment of Present Disclosure]
First, embodiments of the present disclosure will be listed and described. Moreover, you may combine arbitrarily at least one part of embodiment described below.

(1) In the learning data collection method according to this aspect, learning is performed in which inspection target data for detecting an abnormality of the inspection target is input, and feature extraction data obtained by extracting features of the input inspection target data is output. A learning data collection method for collecting data for additionally learning a completed auto encoder, wherein a plurality of first inspection object data and outputs input to the auto encoder for learning the auto encoder before learning The difference between the plurality of first feature extraction data thus obtained and the difference between the plurality of second inspection target data inputted to the auto encoder after learning and the plurality of second feature extraction data outputted If there is a statistical difference, the second inspection object data related to the normal inspection object is selected as additional learning data.

In this aspect, the difference between the difference between the plurality of first inspection object data and the plurality of first feature extraction data and the difference between the plurality of second inspection object data and the plurality of second feature extraction data. Compare differences statistically. The first inspection target data is data used for learning or manufacturing the auto encoder. The second inspection object data is data input to the auto encoder in order to detect an abnormality of the inspection object after learning. By statistically comparing the above differences, it is possible to determine whether or not there has been a change in the anomaly detection environment of the inspection object, such as a change in the lighting environment in the factory, contamination of the sensor, or movement of the installation location of the line.
When there is a statistical difference, second inspection target data related to a normal inspection target object is selected as additional learning data from among the plurality of second inspection target data. If the auto encoder is trained using inspection object data relating to an inspection object having an abnormality, the inspection object having an abnormality cannot be detected. By performing the selection as described above, it is possible to collect learning data that can be adapted to changes in the anomaly detection environment by additional learning.

(2) Add the second inspection object data out of a predetermined confidence interval statistically obtained from the difference between the plurality of first inspection object data and the output first feature extraction data. It is preferable to select the data for learning.

In this aspect, the second inspection target data out of the confidence interval is selected as additional learning data from among the plurality of second inspection target data. Since the inspection target data is data in which the influence of the anomaly detection environment appears, it is possible to effectively additionally learn the auto encoder by selecting such data.

(3) A number of the plurality of first test object data corresponding to a ratio within the confidence interval and a number of the second test object data corresponding to a ratio outside the confidence interval are used for additional learning. A configuration to select as data is preferable.

In this aspect, the first inspection object data and the second inspection object data are mixed at an appropriate ratio and selected as additional learning data. By selecting additional learning data in this way, abnormal learning can be avoided and the auto encoder can be effectively additionally learned.

(4) Additional learning data obtained by randomly mixing the plurality of first inspection object data and the plurality of second inspection object data related to the normal inspection object. The configuration selected as is preferable.

In this aspect, by randomly mixing the first inspection target data and the second inspection target data, it is possible to avoid biased abnormal learning and effectively additionally learn the auto encoder. .

(5) The difference between the plurality of first inspection object data and the outputted first feature extraction data and the difference between the second inspection object data and the outputted feature extraction data. It is preferable that the second inspection object data related to the normal inspection object is selected by comparing the occurrence location.

In this aspect, the second inspection object data relating to the normal inspection object can be selected by comparing the locations where the differences occur. This aspect includes both an invention for manually selecting second inspection object data relating to a normal inspection object and an invention for automatically selecting the second inspection object data.

(6) The difference between the plurality of first inspection object data and the outputted first feature extraction data, and the difference between the second inspection object data and the outputted feature extraction data. In the case where there is no statistical difference between the occurrence location and the occurrence location, it is preferable to select the second inspection target data as the second inspection target data related to the normal inspection target.

In this aspect, whether the second inspection object data is the second inspection object data related to a normal inspection object by comparing the statistical distance of the occurrence point of each difference with a threshold value. It can be determined whether or not. By the statistical distance and threshold comparison processing, the second inspection object data relating to the normal inspection object can be automatically selected.

(7) The difference between the plurality of first inspection object data and the output first feature extraction data, the plurality of second inspection object data and the output feature extraction data. When there is a statistical difference between the occurrence location of the difference and the second inspection object data having a difference in the occurrence location where there is a difference, data indicating the occurrence location is output to the outside, and the normal inspection object It is preferable to accept a selection as to whether or not the data is the second inspection object data.

In this aspect, the data indicating the difference occurrence location is output to the outside together with the second inspection target data having a statistical difference compared to the first inspection target data. That is, the second inspection target data that is a candidate for additional learning is output. The second inspection object data includes inspection object data related to a normal inspection object and inspection object data related to an abnormal inspection object. In view of this, data indicating the location where the difference occurs is externally output so that a person can easily select a normal inspection object. And selection of the 2nd inspection object data concerning a normal inspection object is received.

(8) The difference between the plurality of first inspection object data and the output plurality of first feature extraction data, and the difference between the plurality of second inspection object data and the output feature extraction data. A configuration in which the necessity of additional learning of the auto encoder is determined based on a statistical comparison result is preferable.

In this aspect, between the difference between the plurality of first inspection object data and the plurality of first feature extraction data and the difference between the plurality of second inspection object data and the plurality of second feature extraction data. The necessity of additional learning is determined based on the presence or absence of the statistical difference. Therefore, useless additional learning can be avoided.

(9) The plurality of second inspection target data input to the auto encoder after learning is accumulated, and the second inspection target data has a statistical difference compared to the plurality of first inspection target data. A configuration in which it is determined that additional learning is necessary when the amount of storage exceeds the threshold.

In this aspect, it is determined that additional learning is necessary when a plurality of second inspection target data having statistical differences is sufficiently accumulated as compared with a plurality of first inspection target data. . Therefore, useless additional learning can be avoided and additional learning of the auto encoder can be effectively performed.

(10) It is preferable to periodically determine whether additional learning is necessary.

In this aspect, it is possible to periodically confirm the necessity of additional learning and to perform additional learning of the auto encoder as necessary.

(11) It is preferable that the inspection object data is image data obtained by imaging the inspection object.

In this aspect, the auto encoder uses the image data obtained by imaging the inspection object as inspection object data to detect an abnormality of the inspection object. According to this aspect, data suitable for additional learning can be collected from a plurality of second inspection object data that is image data.

(12) The learning data collection apparatus according to this aspect receives learning target data for detecting an abnormality of the inspection target object, and outputs feature extraction data obtained by extracting features of the input inspection target data A learning data collection device for collecting data for additionally learning a completed auto encoder, wherein a plurality of first inspection object data and outputs input to the auto encoder for learning the auto encoder before learning The difference between the plurality of first feature extraction data thus obtained and the difference between the plurality of second inspection target data inputted to the auto encoder after learning and the plurality of second feature extraction data outputted If there is a statistical difference between the comparison unit and the comparison unit, the second inspection object data related to the normal inspection object is selected as additional learning data. And a part.

In this aspect, as in aspect 1, the learning data collection device can collect learning data that can be adapted to changes in the anomaly detection environment through additional learning.

(13) The anomaly detection system according to this aspect receives learned object data for detecting an anomaly of an object to be inspected, and outputs feature extraction data obtained by extracting features of the inputted inspection object data. Added an encoder, an abnormality detection processing unit for detecting an abnormality of the inspection object by comparing the inspection object data input to the auto encoder and the output feature extraction data, and the auto encoder A learning data collecting device according to an aspect (12) for collecting data for learning, and a learning processing unit for additionally learning the auto encoder based on the data collected by the learning data collecting device.

In this aspect, the anomaly detection system can detect an anomaly of an inspection object using an auto encoder. Similar to aspect 1, the learning data collection device can collect learning data that can be adapted to changes in the anomaly detection environment by additional learning. The learning processing unit can additionally learn the auto-encoder using the collected learning data, and can adapt the change to the anomaly detection environment.

(14) The computer program according to this aspect has been learned to input inspection target data for detecting whether there is an abnormality in the inspection target and to output feature extraction data obtained by extracting the characteristics of the input inspection target data A computer program for causing a computer to collect data for further learning of an auto encoder, wherein the computer is configured to receive a plurality of first inputs input to the auto encoder so that the computer learns the auto encoder before learning. Differences between inspection object data and output first feature extraction data, differences between a plurality of second inspection object data input to the auto encoder after learning and output second feature extraction data And when there is a statistical difference, the second inspection pair related to the normal inspection object To execute a process of selecting the data as data for the additional learning.

In this aspect, by causing the computer to execute the computer program, the learning data collection device can be adapted to change in the anomaly detection environment by additional learning in the same manner as in aspect 1. Data can be collected.

[Details of Embodiment of the Present Disclosure]
Specific examples of the learning data collection method, the learning data collection device, the anomaly detection system, and the computer program according to the embodiment of the present disclosure will be described below with reference to the drawings. In addition, this indication is not limited to these illustrations, is shown by the claim, and it is intended that all the changes within the meaning and range equivalent to the claim are included.

(Embodiment 1)
FIG. 1 is a schematic diagram illustrating a configuration example of the anomaly detection system according to the first embodiment. The anomaly detection system is an anomaly detection device 1 that detects an anomaly of an inspection object 110 based on an imaging unit 2 that images the inspection object 110 and image data obtained by imaging (hereinafter referred to as inspection object data). And a display unit 3. The inspection object 110 is a connector terminal of the wire harness 100 disposed in the vehicle, for example.

<Hardware configuration of the anomaly detection device 1>
FIG. 2 is a block diagram illustrating a configuration example of the anomaly detection device 1 according to the first embodiment. The anomaly detection device 1 is a computer having an operation unit 1a such as one or a plurality of CPUs (Central Processing Units) and a multi-core CPU. A temporary storage unit 1b, an image input unit 1c, an output unit 1d, an input unit 1e, a clock unit 1f, a storage unit 1g, and a data storage unit 1h are connected to the calculation unit 1a via a bus line.

The calculation unit 1a controls the operation of each component by executing a computer program 5 described later stored in the storage unit 1g, and detects an abnormality of the inspection object 110 using an auto-encoder type neural network. A process of collecting additional learning data according to the present embodiment, a process of additionally learning the neural network using the collected data, and the like. This neural network constitutes an auto encoder 11 (see FIGS. 3 and 4) described later.

The temporary storage unit 1b is a memory such as a DRAM (Dynamic RAM) or an SRAM (Static RAM), and is executed by the computer program 5 read from the storage unit 1g or the arithmetic processing when the arithmetic processing of the arithmetic unit 1a is executed. Temporarily store various data generated.

The storage unit 1g is a non-volatile memory such as a hard disk, an EEPROM (Electrically Erasable Programmable ROM), or a flash memory. The storage unit 1g stores a computer program 5 for executing the anomaly detection process of the inspection object 110, the collection process of additional learning data, and the additional learning process by the operation unit 1a controlling the operation of each component. is doing.
The storage unit 1g may store the computer program 5 read from the recording medium by a reading device (not shown). Recording media are CD (Compact Disc) -ROM, DVD (Digital Versatile Disc) -ROM, optical disc such as BD (Blu-ray (registered trademark) Disc), flexible disc, magnetic disc such as hard disc, magnetic optical disc, semiconductor memory, etc. It is. Further, the computer program 5 according to the first embodiment may be downloaded from an external computer (not shown) connected to a communication network (not shown) and stored in the storage unit 1g.

The image input unit 1c is an interface to which the imaging unit 2 is connected. The image pickup unit 2 converts an image formed by the lens into an electrical signal such as a CCD and a CMOS, and AD converts the electrical signal converted by the image sensor into digital image data. And an image processing unit that outputs image data as inspection target data. The inspection object data output from the imaging unit 2 is input to the anomaly detection device 1 via the image input unit 1c. The inspection target data input to the anomaly detection device 1 is stored in the data storage unit 1h. The imaging unit 2 and the anomaly detection device 1 may be configured to be connected separately by a dedicated cable, or may be configured to be connected via a network such as a LAN (Local Area Network). The inspection target data is digital data in which the pixels arranged in the vertical and horizontal directions are indicated by luminance values of a predetermined gradation. In the present embodiment, description will be made assuming that the image data is monochrome.

The output unit 1d is an interface to which the display unit 3 is connected. The display unit 3 is a liquid crystal panel, an organic EL display, electronic paper, a plasma display, or the like. The display unit 3 displays various information corresponding to the image data given from the calculation unit 1a. For example, the contents of the anomaly detection result, the image of the inspection object 110 having a defect, and the like are displayed. The display unit 3 is an example of an external output device that outputs the anomaly detection result, and may be a buzzer, a speaker, a light emitting element, or another notification device. The factory operator can recognize the state of the inspection object 110 as a result of the anomaly detection from the image displayed on the display unit 3.

The operation unit 4 such as a keyboard, a mouse, and a touch sensor is connected to the input unit 1e. A signal indicating the operation state of the operation unit 4 is input to the anomaly detection device 1 via the input unit 1e. The calculation unit 1a can recognize the operation state of the operation unit 4 via the input unit 1e.

The clock unit 1 f keeps timing for additionally learning the neural network of the auto encoder 11. The clock unit 1f outputs a signal at a timing for confirming the necessity of additional learning of the neural network, such as a one-month cycle, for example, and the arithmetic unit 1a determines the necessity of additional learning, and for additional learning as necessary Data collection and additional learning processing are executed.

The data storage unit 1h is a non-volatile memory such as a hard disk, an EEPROM, or a flash memory, like the storage unit 1g. The data storage unit 1h stores the inspection target data for learning, the inspection target data in operation, the difference data, and the inspection target data for additional learning.
Hereinafter, for convenience of explanation, it is necessary to distinguish between the inspection object data for learning used in the learning process at the time of manufacturing the anomaly detection apparatus 1 and the inspection object data input during operation of the anomaly detection apparatus 1 after learning. In some cases, the inspection object data will be described with reference symbols α and β.

The inspection object data α for learning is image data used when learning a neural network for detecting an abnormality of the inspection object 110. That is, the image data used in the manufacturing stage of the anomaly detection device 1.
The inspection object data β in operation is image data that is input to the anomaly detection device 1 when the inspection object 110 is actually inspected on a factory line or the like.
The difference data is data indicating a difference between the inspection target data and the feature extraction data. The feature extraction data is image data obtained by extracting features of the inspection object 110 from the inspection object data in order to detect an abnormality in the inspection object 110. Details of the feature extraction data and the difference data will be described later.
The inspection target data for additional learning is data for additionally learning the neural network. When the illumination environment in the factory changes, the lens of the imaging unit 2 becomes dirty, or the installation location of the line moves, the luminance value of the image obtained by imaging the inspection object 110 changes. If it does so, the abnormality detection precision of the test object 110 may fall. The additional learning data is data for adapting to various environmental changes by additionally learning the neural network.

<Functional part of the anomaly detection device 1>
FIG. 3 is a block diagram illustrating a configuration example of the anomaly detection device 1 according to the first embodiment. The anomaly detection device 1 includes an auto encoder 11, an anomaly detection processing unit 12, a learning data collection device 13, and a learning processing unit 14 as functional units. The abnormality detection processing unit 12 is further configured by a determination data generation unit 12a and an abnormality determination unit 12b. Each functional unit of the anomaly detection device 1 is realized by hardware such as the arithmetic unit 1a and the data storage unit 1h.

The auto encoder 11 is a functional unit that receives the inspection target data output from the imaging unit 2 and outputs feature extraction data obtained by extracting the characteristics of the input inspection target 110. Specifically, when the inspection object data is input, the auto encoder 11 outputs feature extraction data representing the characteristics of the normal inspection object 110. The image related to the inspection target data may include an image of dust, scratches, shadows, anomalous parts of the inspection target 110 itself, and the like. The feature extraction data is image data that reproduces an ideal inspection object 110 that is obtained when these images are removed and a normal inspection object 110 having no abnormality is imaged. Auto-encoding is realized by a neural network. The neural network includes an intermediate layer that dimensionally compresses inspection target data, and inputs and outputs image data having the same number of pixels.

FIG. 4 is a schematic diagram showing a neural network of the auto encoder 11 and a learning method. The auto encoder 11 includes an input layer 11a, a convolution layer (CONV layer), a deconvolution layer (DECONV layer), and an output layer 11b. The input layer 11a is a layer to which data of each pixel value related to inspection target data is input. The convolution layer is a layer that dimensionally compresses the inspection object data α for learning. For example, the convolution layer performs dimensional compression by performing convolution integration. The feature amount of the inspection object 110 is extracted by dimensional compression. The deconvolution layer is a layer that restores data that has been dimensionally compressed in the convolution layer to its original dimension. The deconvolution layer performs deconvolution processing and restores the original dimension. By the restoration, image data representing the original characteristic of the inspection object 110, that is, the characteristic of the normal inspection object 110 is restored. In addition, although the example in which the convolution layer and the deconvolution layer are two layers is shown, one layer or three or more layers may be used. The output layer 11b is a layer that outputs data of each pixel value related to the feature extraction data obtained by extracting the feature of the inspection object 110 in the convolution layer and the deconvolution layer.

As shown in FIG. 4, the auto encoder 11 causes the neural network of the auto encoder 11 to perform machine learning so that the input inspection target data and the output feature extraction data are the same. That is, the neural network is machine-learned so that the input image 110a and the output image 110b are the same. Such machine learning is performed using inspection object data obtained by imaging a normal inspection object 110.

FIG. 5 is a schematic diagram showing the function of the auto encoder 11 that extracts the feature of the inspection object 110 and outputs the feature extraction data. As described above, the auto encoder 11 learned as described above receives either inspection object data obtained by imaging a normal inspection object 110 or inspection object data obtained by imaging the inspection object 110 having an abnormality. Image data obtained by imaging the normal inspection object 110 is output as feature extraction data.
In FIG. 5, the feature extraction data input to the auto encoder 11 from the left side is image data obtained by imaging the inspection object 110 having an abnormality. The image data includes data of the image 110a of the inspection object 110 having the image 111a at the abnormal location. When this inspection object data is input to the auto encoder 11, feature extraction data as shown on the right side in FIG. 5 is output. The feature extraction data is image data obtained by imaging a normal inspection object 110. The image data includes data of the image 110b of the inspection object 110 from which the image 111a of the anomalous portion is removed and the original characteristics of the inspection object 110 appear.

The feature extraction data output from the auto encoder 11 is input to the determination data generation unit 12a shown in FIG. Further, the inspection object data output from the imaging unit 2 is input to the determination data generation unit 12a.
The determination data generation unit 12a calculates the difference between the inspection target data and the feature extraction data, and outputs the difference data obtained by the difference calculation to the anomaly determination unit 12b. That is, the difference between the pixel value of each pixel related to the inspection target data and the pixel value of each corresponding pixel related to the feature extraction data is calculated, and image data having the difference value as the pixel value is output as difference data. The difference data generated by the determination data generation unit 12a is stored in the data storage unit 1h in association with the inspection target data that is the source of the difference data.

FIG. 6 is a schematic diagram showing a method for generating difference data. The left figure shows an image 110a related to inspection object data, and the center figure shows an image 110b related to feature extraction data. The image 110a related to the inspection object data includes the image 111a of the anomalous portion. Looking at the image 110b related to the feature extraction data, it can be seen that the image 111a at the anomalous location is removed and the image of the normal inspection object 110 is obtained. The determination data generation unit 12a generates difference data obtained by subtracting the luminance value of each pixel related to the feature extraction data from the luminance value of each pixel related to the inspection target data. The right figure shows difference data, and an image 111a of an abnormal part is extracted as an image of difference data.

7A, 7B, and 7C are schematic diagrams illustrating examples of inspection target data, feature extraction data, and difference data obtained by imaging the connector that is the inspection target 110. FIG. 7, 7B and 7C show a method for generating difference data using inspection object data obtained by imaging the connector of the wire harness 100 as a more specific inspection object 110. FIG. The basic operation is as described in FIG. FIG. 7A shows data to be inspected, and the connector image 110a includes an image 111a of an abnormal location. FIG. 7B shows feature extraction data, which includes an image 111b of a normal inspection object 110. FIG. FIG. 7C is an image obtained by extracting the anomalous location image 111a as an image of difference data.

The anomaly determination unit 12b determines whether there is an anomaly in the inspection object 110 based on the difference data output from the determination data generation unit 12a. The anomaly determination unit 12b is, for example, a learned one-class support vector machine. The one-class support vector machine is a classifier that classifies difference data obtained from normal inspection target data and difference data obtained from abnormal inspection target data. The one class support vector machine is an example of a classifier, and the determination method is not particularly limited as long as difference data obtained from abnormal inspection target data can be statistically determined as an outlier.

<Collecting additional learning data>
Next, the additional learning process of the auto encoder 11 will be described.
FIG. 8 is a flowchart showing a processing procedure related to collection of additional learning data and additional learning. The arithmetic unit 1a executes the following processing at the additional learning confirmation timing of the auto encoder 11 that arrives periodically. First, the calculation unit 1a reads a plurality of learning inspection target data α from the data storage unit 1h, and inputs the read plurality of inspection target data α to the auto encoder 11, thereby extracting feature extraction data of the inspection target data α. Is generated (step S11). And the calculating part 1a produces | generates the frequency distribution of difference amount based on the difference data output from the determination data production | generation part 12a (step S12).

The difference amount is a scalar amount indicating the size of the difference. For example, the arithmetic unit 1a performs threshold processing on the difference data and binarizes it. And the calculating part 1a performs a connection area | region extraction process from the image which concerns on the difference data after noise removal, and calculates the integrated value of the luminance value of each pixel which comprises a connection area | region for every connection area | region. Next, the maximum integrated value among the integrated values of each connected region included in the image related to the difference data is determined as the difference amount of the difference data. That is, the luminance integrated value of the largest difference image portion included in the image related to the difference data is calculated as the difference amount. The difference amount is calculated for each of a plurality of difference data. Finally, the calculation unit 1a calculates a frequency distribution of the difference amount (see FIG. 9).

Next, the calculation unit 1a reads the difference data obtained during operation from the data storage unit 1h (step S13). The difference data is data obtained by calculating a difference between the inspection target data β and the feature extraction data. And the calculating part 1a produces | generates the frequency distribution of difference amount based on the read difference data (step S14). The said difference amount and frequency distribution can be produced | generated by the process similar to step S12.

Next, the calculation unit 1a calculates a test statistic for testing a statistical difference between the frequency distribution calculated in step S12 and the frequency distribution calculated in step S14 (step S15), It is determined whether or not the frequency distribution has a statistically significant difference (step S16).

FIG. 9 is a graph showing the statistical difference of the difference amount. The horizontal axis indicates the difference amount, and the vertical axis indicates the frequency. A in the figure indicates the frequency distribution calculated in step S12, and B in the figure indicates the frequency distribution calculated in step S14. The computing unit 1a calculates a t value as a test statistic, and performs a t test on the difference between the population means of each frequency distribution. The test method is not particularly limited, and an F value may be calculated and a difference in variance may be F-tested. Of course, you may test using other test statistics. It is also possible to determine whether each frequency distribution has a statistically significant difference by calculating the similarity or statistical distance of each frequency distribution and calculating whether the statistical distance is greater than or equal to a threshold value. good.

When it is determined in step S16 that there is no statistically significant difference (step S16: NO), the arithmetic unit 1a finishes the process. Statistics between the frequency distribution of the difference amount obtained from the inspection object data α for learning used at the time of manufacturing the anomaly detection apparatus 1 and the frequency distribution of the difference amount obtained from the inspection object data β captured during operation If there is no significant difference, it is considered that there is no change in the environment that lowers the anomaly detection accuracy, and additional learning of the auto encoder 11 is not necessary.

When it is determined that there is a statistically significant difference (step S16: YES), the calculation unit 1a collects or selects data for additionally learning the auto encoder 11 by executing the processes of steps S17 to S23. Execute the process. The frequency distribution of the difference amount obtained from the inspection object data α for learning used at the time of manufacturing the anomaly detection apparatus 1 and the frequency distribution of the difference amount obtained from the inspection object data β captured during operation are statistically calculated. If they are different, there may be some changes such as changes in the lighting environment in the factory, contamination of the sensor, and movement of the installation location of the line. In this case, the anomaly detection accuracy may be lowered, and it is necessary to additionally learn the auto encoder 11 and adapt it to environmental changes.

The arithmetic unit 1a is configured such that the difference amount of the difference data obtained from the inspection object data β among the plurality of inspection object data β obtained during operation is the difference amount obtained from the inspection object data α for learning. One inspection object data β that is out of the confidence interval of the frequency distribution A is selected (step S17). The confidence interval is, for example, 2σ = about 95%.

The inspection object data β selected in step S17 includes inspection object data β obtained by imaging the inspection object 110 having an abnormality. Such inspection object data β must be excluded from the additional learning data. The processing of step S18 and step S19 following step S17 includes inspection object data β obtained by imaging the normal inspection object 110 and inspection object data β obtained by imaging the inspection object 110 having an abnormality. It is for sorting.

The calculation unit 1a that has finished the process of step S17 calculates a difference occurrence location in the image of the plurality of learning target data α for learning (step S18), and the calculated difference occurrence location is selected in step S17. It is determined whether or not the difference occurrence location in the image of the inspection object data β matches (step S19).
That is, the calculation unit 1a selects whether or not the difference is not an abnormality of the inspection object 110 but a difference caused by an environmental change or the like based on the occurrence location. This will be specifically described below.

FIG. 10 is a contour diagram showing a mixed normal distribution obtained based on learning inspection object data α obtained by imaging a normal inspection object 110. The computing unit 1a identifies the position where the difference occurs based on the difference data calculated in step S11 and step S12. In FIG. 10, the code | symbol 6 has shown the generation | occurrence | production location of a difference. The position of the difference occurrence position may be estimated using, for example, a mixed normal distribution, where the coordinates of the image related to the difference data are (x, y) and the magnitude of the difference is p (x, y). A mixed normal distribution is represented by a superposition of a plurality of normal distributions. The computing unit 1a estimates a plurality of normal distributions i that reproduce p (x, y) using an EM algorithm or the like. The computing unit 1a calculates the number N of normal distributions i to be mixed, the center coordinates μi = (μix, μii) of each normal distribution, and the mixing ratio πi. In the example shown in FIG. 10, the center of each mountain is the central coordinate μi. i (= 1, 2,..., N) is an integer indicating a normal distribution to be mixed.
In the example shown in FIG. 10, six peaks appear, and in the simplest case, the magnitude p (x, y) of the difference can be expressed by a mixed normal distribution obtained by superimposing six normal distributions i. . Note that the number of peaks and the number of normal distributions i to be superimposed do not necessarily match.

FIG. 11 is a schematic diagram showing difference data obtained during anomaly detection processing after learning.
The calculation unit 1a calculates the difference center coordinates μk = (μkx, μky) based on the difference data selected in step S17 using the k-means or k-means ++ method (k = 1, 2). , ...). In the example illustrated in FIG. 11, the portion of the mark X where the difference occurs is the difference center coordinate μk, and the number of the difference centers is “1”.

12A and 12B show a method for discriminating between inspection object data β obtained by imaging an inspection object 110 having an abnormality and inspection object data β obtained by imaging a normal inspection object 110. It is a schematic diagram. The computing unit 1a calculates a statistical distance between the center coordinate μi and the center coordinate μk, and specifies the center coordinate μi closest to the center coordinate μk. As the statistical distance, for example, the L2 norm, that is, the Euclidean distance may be calculated. Then, the calculation unit 1a statistically determines the center coordinate μi that is a difference occurrence location related to the inspection target data α for learning and the center coordinate μk that is a difference occurrence location related to the inspection target data β obtained during operation. It is determined whether it is close to.
Specifically, the peak of the mixed normal distribution, that is, the normal distribution j, which is closest to the center coordinate μk of the difference occurrence location related to the inspection target data β obtained during operation, is specified. Then, when the value obtained by multiplying the value pj (μk) of the normal distribution j by the reciprocal of the mixture ratio πj is equal to or greater than a predetermined threshold value, the calculation unit 1a determines that the difference occurrence points match. To do.
In the example shown in FIG. 12A, since the closest center coordinate μi and the center coordinate μk are statistically separated, the inspection target data β selected in step S17 images the inspection target 110 having an abnormality. It can be estimated that the obtained image data.
In the example shown in FIG. 12B, since the closest center coordinate μi and the center coordinate μk are statistically close, the inspection object data β selected in step S17 was obtained by imaging the normal inspection object 110. It is image data, and it can be estimated that the data is affected by environmental changes.

When it is determined in step S19 that the difference occurrence locations match (step S19: YES), the calculation unit 1a selects the inspection object data β selected in step S17 as data for additional learning (step S20). ). When it is determined that the difference occurrence points do not match (step S19: NO), the calculation unit 1a excludes the inspection target data β selected in step S17 from the additional learning data (step S21).

Next, the arithmetic unit 1a determines whether or not the selection of whether or not all the inspection object data β accumulated during operation is suitable as additional learning data has been completed (step S22). If it is determined that the sorting has not been completed (step S22: NO), the computing unit 1a returns the process to step S17.
In addition, although the example which selects all the test object data (beta) accumulate | stored in the data storage part 1h was demonstrated, you may use a part of stored test object data (beta) as a candidate of the data for learning.

If it is determined that the selection has been completed (step S22: YES), the learning inspection target data α stored in the data storage unit 1h and the inspection target data β selected in the processing of steps S17 to S21 are randomly selected. (Step S23). The inspection object data α and β obtained by mixing are preferably stored in the data storage unit 1h as learning data.
The mixing ratio of the inspection object data α for learning and the selected inspection object data β is not particularly limited, but may be mixed at a ratio related to the confidence interval used in step S17. That is, it is preferable to mix the number of inspection object data α corresponding to the ratio within the confidence interval and the number of the selected inspection object data β corresponding to the proportion outside the confidence interval.
For example, if the confidence interval is 2σ = 95% and the number of inspection object data β selected in step S20 is 500, 9500 inspection object data α may be read from the data storage unit 1h and mixed.

Next, the calculation unit 1a performs additional learning of the auto encoder 11 using the inspection object data α and β mixed in step S23 as data for additional learning (step S24), and ends the process. The additional learning method is the same as the learning method of the auto encoder 11 at the time of manufacture. That is, the arithmetic unit 1a causes the neural network constituting the auto encoder 11 to machine-learn so that the inspection object data α and β input to the auto encoder 11 and the output feature extraction data are the same.

According to the learning data collection method, the learning data collection device 13, the anomaly detection system, and the computer program 5 configured as described above, the lighting environment in the factory changes, the dirt of the sensor, the movement of the installation location of the line, etc. When the sensor value changes, it is possible to automatically collect learning data for additionally learning the auto encoder 11 so as to respond to environmental changes.

Further, since the test object data β deviating from the confidence interval (see FIG. 9) obtained based on the test object data α for learning is selected as data for additional learning, the auto encoder 11 is effectively used. Can be additionally learned.

Further, the inspection object data α for learning and the inspection object data β accumulated during operation are randomly mixed at an appropriate ratio, and the mixed inspection object data α and β are used as additional learning data. Therefore, abnormal learning can be avoided and the auto encoder 11 can be additionally learned effectively.

Furthermore, since the inspection object data α and β are mixed according to the ratio of the confidence interval, biased abnormal learning can be avoided and the auto encoder 11 can be effectively additionally learned.

Furthermore, the inspection object data β obtained by imaging the normal inspection object 110 can be selected by comparing the difference occurrence points related to the inspection object data α, β, and the auto encoder 11 can be appropriately selected. Additional learning can be done.

Furthermore, the inspection object data β obtained by imaging the normal inspection object 110 can be automatically selected, and the auto encoder 11 can be additionally learned.

Furthermore, in step S16, the difference between the difference data of the inspection target data α and β can be statistically determined to determine whether additional learning is necessary. Accordingly, useless additional learning can be avoided, and the auto encoder 11 can be additionally learned as necessary.

Furthermore, it is possible to periodically check the necessity of additional learning using the clock unit 1f and to perform additional learning of the auto encoder 11 as necessary.

Furthermore, the anomaly detection apparatus 1 according to the first embodiment detects an anomaly of the inspection object 110 using image data obtained by imaging the inspection object 110, and the additional learning data collection apparatus is Data suitable for additional learning can be selected by image analysis of the inspection object data β.

In the first embodiment, the inspection target data of the image is taken as an example, and anomaly detection of the inspection target 110, data collection for additional learning of the auto encoder 11, and additional learning have been described. The format and contents are not particularly limited. For example, the inspection target data may be voice data detected by a microphone, vibration data detected by a vibration sensor, voltage data detected by a voltmeter or an ammeter, current data, and the like.

In the first embodiment, the processing shown in FIG. 8 is periodically executed, and there is a statistical difference between the difference amount distribution related to the inspection target data α and the difference amount distribution of the inspection target data β. In this case, the example in which the additional learning process is executed has been described. In step S16, the calculation unit 1a determines whether or not the accumulated amount of the inspection target data β having a statistical difference exceeds a predetermined threshold value. If the threshold value is exceeded, the processing in step S17 and subsequent steps may be executed.

In this case, the inspection target data β, which is a candidate for additional learning data, is sufficiently accumulated and then the process related to additional learning is performed. Therefore, unnecessary additional learning can be avoided and the auto encoder can be additionally learned effectively. it can.

(Embodiment 2)
In the second embodiment, the operator determines whether the inspection object data β outside the confidence interval obtained based on the inspection object data α for learning is inspection object data suitable as additional learning data. Is different from the first embodiment in that it is selected semi-automatically. In the following, mainly such differences will be described. Since other configurations and operational effects are the same as those of the first embodiment, the corresponding portions are denoted by the same reference numerals, and detailed description thereof is omitted.

FIG. 13 is a flowchart showing a collection processing procedure of additional learning data according to the second embodiment. The arithmetic unit 1a that has executed the processing of step S11 to step S16 of the first embodiment has the difference amount of the difference data obtained from the inspection target data β among the plurality of inspection target data β obtained during operation. The inspection object data β that is out of the confidence interval of the frequency distribution A of the difference amount obtained from the inspection object data α for learning is selected (step S217). In step S217 of the second embodiment, the calculation unit 1a selects all the inspection target data β that is outside the confidence interval.

And the calculating part 1a calculates the difference generation | occurrence | production location in the image of the some test object data (alpha) for learning (step S218). Specifically, the computing unit 1a estimates the difference occurrence location in the mixed normal distribution in the same manner as in step S18 of the first embodiment.

Next, the calculation unit 1a calculates a difference occurrence point in the images of the plurality of pieces of inspection target data β selected in step S217 (step S219). Specifically, the computing unit 1a estimates the difference occurrence location in the mixed normal distribution in the same manner as in step S218.

Next, the calculation unit 1a compares the difference occurrence location calculated in step S218 with the difference occurrence location calculated in step S219, and the inspection target data β at a position deviating from the difference occurrence location related to the inspection target data α. The difference occurrence location is extracted (step S220). That is, the inspection object data β having a difference in a difference occurrence place that is normally found in the learning inspection object data α and a difference occurrence place is extracted.
Since it is highly likely that the inspection object data β is abnormal data due to an abnormality of the inspection object 110, the operator determines whether or not the inspection object data is appropriate as learning data.

And the calculating part 1a selects the some test object data (beta) which has a difference in the difference generation location extracted by step S220 (step S221).

Next, the calculation unit 1a outputs the inspection target data β selected in step S221 and the difference occurrence location display image 7 indicating the difference occurrence location extracted in step S220 to the display unit 3 via the output unit 1d. (Step S222). And the calculating part 1a receives the suitability as data for additional learning in the operation part 4 (step S223).

14A and 14B are schematic diagrams showing a method for accepting selection of learning data. In FIG. 14A, the difference occurrence location display image 7 is superimposed and displayed on the image 110 a of the inspection object 110 related to the inspection object data β. In the example illustrated in FIG. 14A, there is an abnormality in the inspection object 110, a difference is generated due to the abnormality, and the difference occurrence location is indicated by the difference occurrence location display image 7. In the example shown in FIG. 14B, the inspection object 110 is normal, not caused by an abnormality, and a difference occurrence place display image 7 shows a difference occurrence place that normally occurs.
The operator confirms the video as shown in FIG. 14A or FIG. 14B displayed on the display unit 3, and selects the suitability as additional learning data using the operation unit 4.

The calculation unit 1a performs selection related to the additional learning data based on the selection result received by the operation unit 4 (step S224). When the operation unit 1a receives an operation indicating that it is appropriate as additional learning data, the calculation unit 1a selects, as additional learning data, inspection target data β having a difference at the difference occurrence location extracted in step S220. Conversely, when an operation indicating that it is inappropriate as additional learning data is received, the inspection object data β having a difference at the difference occurrence location extracted in step S220 is excluded from the additional learning data.
When the difference occurrence location calculated in step S218 and the difference occurrence location calculated in step S219 are statistically coincident, the calculation unit 1a also determines the inspection target data β having a difference in the difference occurrence location. Automatically selected as additional learning data.

According to Embodiment 2, semi-automatic collection of additional image data taking into account the judgment by the operator is possible by checking the suitability as the learning data for the inspection target data β that may be abnormal. Thus, the auto-encoder 11 can be additionally learned more effectively.

(Embodiment 3)
FIG. 15 is a schematic diagram illustrating a configuration example of the anomaly detection system according to the third embodiment.
The anomaly detection system according to Embodiment 3 is different from that of Embodiment 1 in that the learning data collection device 301 and the anomaly detection device 310 that performs anomaly detection are configured separately and connected via a communication line L. Different. In the following, mainly such differences will be described. Since other configurations and operational effects are the same as those of the first embodiment, the corresponding portions are denoted by the same reference numerals, and detailed description thereof is omitted.

The hardware configuration of the learning data collection device 301 is the same as that of the anomaly detection device 310 of the first embodiment, and further includes a communication unit 1i that communicates with the anomaly detection device 310. The hardware configuration on the anomaly detection device 310 side is the same, and further includes a communication unit 313 that communicates with the learning data collection device 301.

The anomaly detection device 310 transmits the inspection object data β and the difference data to the learning data collection device 301 through the communication unit 313 during operation. The learning data collection device 301 receives the inspection object data β and the difference data transmitted from the anomaly detection device 310 by the communication unit 1i. The learning data collection method is the same as that in the first embodiment, and the learning data collection device 301 collects the necessity for additional learning of the anomaly detection device 310 and data for additional learning. The learning data collection device 301 transmits the collected learning data to the anomaly detection device 310 via the communication unit 1i. The anomaly detection device 310 receives learning data at the communication unit 313 and performs additional learning based on the received data.
It should be noted that various parameters defining the weighting factor and layer structure for configuring the neural network of the auto encoder 311 are exchanged between the learning data collection device 301 and the anomaly detection device 310, and additional learning is also performed on the learning data collection device 301 side. You may comprise so that it may be performed.

It should be noted that the processing related to data collection and additional learning that causes the auto-encoder 311 to perform additional learning can be distributed as appropriate according to the performance of the hardware, and there is no particular problem where the processing is executed.

In the anomaly detection system according to the third embodiment, as in the first embodiment, when the sensor value changes due to changes in the lighting environment in the factory, dirt on the sensor, movement of the installation location of the line, etc. For this purpose, learning data for additionally learning the auto encoder 311 can be collected.

DESCRIPTION OF SYMBOLS 1 Abnormality detection apparatus 1a Arithmetic unit 1b Temporary storage unit 1c Image input unit 1d Output unit 1e Input unit 1f Clock unit 1g Storage unit 1h Data storage unit 2 Imaging unit 3 Display unit 4 Operation unit 5 Computer program 7 Difference occurrence location display image 11 Auto encoder 12 Anomaly detection processing unit 12a Determination data generation unit 12b Anomaly determination unit 13 Learning data collection device 14 Learning processing unit 100 Wire harness 110 Inspection object 301 Learning data collection device

Claims (14)

1. Learning that collects data for additionally learning a learned auto-encoder that outputs feature extraction data obtained by extracting features of the inputted inspection target data, when inspection target data for detecting an abnormality of the inspection target is input Data collection method,
   In order to learn the auto encoder before learning, the difference between the plurality of first inspection target data input to the auto encoder and the plurality of output first feature extraction data is input to the auto encoder after learning. Statistically comparing the plurality of second inspection target data and the difference between the plurality of output second feature extraction data,
   A learning data collection method for selecting the second inspection object data relating to the normal inspection object as data for additional learning when there is a statistical difference.
2. The second inspection object data that deviates from a predetermined confidence interval statistically obtained from the difference between the plurality of first inspection object data and the output plurality of first feature extraction data. The learning data collection method according to claim 1, wherein the learning data collection method is selected as data.
3. The plurality of pieces of first inspection object data corresponding to the ratio within the confidence interval and the plurality of pieces of second inspection object data corresponding to the proportion outside the confidence interval are selected as additional learning data. The learning data collection method according to claim 2.
4. The learning inspection object data obtained by randomly mixing the plurality of first inspection object data and the plurality of second inspection object data related to the normal inspection object is selected as additional learning data. The learning data collection method according to any one of claims 1 to 3.
5. Differences between the plurality of first inspection object data and the output first feature extraction data, and differences between the second inspection object data and the output feature extraction data. The learning data collection method according to any one of claims 1 to 4, wherein the second inspection object data relating to the normal inspection object is selected by comparing.
6. Differences between the plurality of first inspection object data and the output first feature extraction data, and differences between the second inspection object data and the output feature extraction data. 6. The learning data collection method according to claim 5, wherein if there is no statistical difference between the second inspection target data, the second inspection target data is selected as the second inspection target data related to the normal inspection target.
7. Occurrence point of difference between the plurality of first inspection object data and the output plurality of first feature extraction data, and generation of difference between the plurality of second inspection object data and the output feature extraction data If there is a statistical difference between the location and the second inspection object data having a difference in the occurrence location where there is a difference, externally output data indicating the occurrence location,
   The learning data collection method according to claim 5, wherein selection of whether or not the second inspection object data relating to the normal inspection object is received.
8. The difference between the plurality of first inspection object data and the output first feature extraction data and the difference between the plurality of second inspection object data and the output feature extraction data are statistically calculated. The learning data collection method according to any one of claims 1 to 7, wherein the necessity of additional learning of the auto encoder is determined based on the comparison result.
9. Accumulating the plurality of second inspection target data input to the auto encoder after learning,
   9. The learning according to claim 8, wherein additional learning is determined when an accumulation amount of the second inspection target data having a statistical difference compared to the plurality of first inspection target data exceeds a threshold. Data collection method.
10. The learning data collection method according to claim 8 or 9, wherein the necessity of additional learning is periodically determined.
11. The learning data collection method according to any one of claims 1 to 10, wherein the inspection object data is image data obtained by imaging the inspection object.
12. Learning that collects data for additionally learning a learned auto-encoder that outputs feature extraction data obtained by extracting features of the inputted inspection target data, when inspection target data for detecting an abnormality of the inspection target is input Data collection device,
    In order to learn the auto encoder before learning, the difference between the plurality of first inspection target data input to the auto encoder and the plurality of output first feature extraction data is input to the auto encoder after learning. A comparison unit that statistically compares the plurality of second inspection target data and the difference between the plurality of output second feature extraction data;
    A learning data collection device comprising: a selection unit that selects the second inspection object data related to the normal inspection object as data for additional learning when there is a statistical difference.
13. Inspection target data for detecting an abnormality of the inspection target is input, and a learned auto encoder that outputs feature extraction data obtained by extracting the characteristics of the input inspection target data;
    An anomaly detection processing unit that detects an anomaly of the inspection object by comparing the inspection object data input to the auto encoder and the output feature extraction data;
    The learning data collection device according to claim 12, which collects data for additionally learning the auto encoder;
    An anomaly detection system comprising: a learning processing unit that additionally learns the auto-encoder based on data collected by the learning data collection device.
14. Inspection object data for detecting whether there is an abnormality in the inspection object is input, and data for additionally learning a learned auto encoder that outputs feature extraction data obtained by extracting features of the input inspection object data, A computer program for causing a computer to collect,
    In the computer,
    In order to learn the auto encoder before learning, the difference between the plurality of first inspection target data input to the auto encoder and the plurality of output first feature extraction data is input to the auto encoder after learning. Statistically comparing the plurality of second inspection target data and the difference between the plurality of output second feature extraction data,
    A computer program for executing a process of selecting the second inspection object data related to the normal inspection object as data for additional learning when there is a statistical difference.
    

Images