JP6997964B2 - Discrimination system, discriminator, learning device, discriminant method and program - Google Patents

Description

The present invention relates to a discrimination system, a discrimination device, a learning device, a discrimination method, and a program.

In a device for classifying target data, for example, a defect inspection device for a product, there is a device for classifying between defective and non-defective based on image data obtained by photographing an inspection target. Some of them use a classifier that has acquired classification ability by learning using machine learning. For example, Patent Document 1 describes a classification device that classifies images by a classification unit. In the classification device of Patent Document 1, a sub-classifier is constructed by learning using a teacher data set group showing a teacher image and a teaching result. Then, the classification performance of the classification unit is improved by repeating the learning using a new teacher data set group including a misclassified image whose classification result by the sub-classifier is different from the teaching result.

JP-A-2017-107422

By the way, in the case of the configuration of Patent Document 1, when generating a classifier, it is necessary to prepare training data according to the group to be classified in advance. However, it may not be possible to prepare the training data for all classes before the actual classification. For example, in the case of a classifier (discriminator) that discriminates the presence or absence of defects, it is difficult to grasp all types of defects that may occur before actually inspecting. Therefore, as a result of classification, it may be necessary to add a group other than the set group, but in that case, it was necessary to change the group setting and then re-learn from the beginning. Therefore, there are problems that it takes time to complete learning and that extra computer resources are required. Such a problem is not limited to a defect inspection device using an image, but is a problem widely common to a classification device that classifies some data using machine learning.

Therefore, an object of the present invention is to provide a technique for easily adding a different discrimination function to the configured discriminator.

The discrimination system according to one aspect of the present invention includes a discrimination unit including a first discriminator that classifies input data into one of a plurality of groups based on the feature amount of the input data, and a plurality of discriminators. Generation that generates the second training data in which the input data classified into a specific group among the groups of the above and the information of the specific subgroup among the plurality of subgroups that subdivide the specific group are associated with each other. By training using the unit and the second training data, the input data classified into a specific group is divided into one of a plurality of subgroups based on the feature amount of the input data. A learning unit for generating a second discriminator for classification is provided. In such a discrimination system, the learning unit newly generates a second discriminator by learning based on the input data determined that the first discriminator has the characteristics of a specific group. This makes it possible to add a function capable of discriminating the discrimination result of the first discriminator in more detail without discarding the existing first discriminator and regenerating a new main discriminator. As a result, the discrimination pattern can be expanded without affecting the discrimination ability of the existing first discriminator. In other words, according to the discrimination system according to this aspect, when the discrimination pattern is expanded, it is not necessary to redo the learning of the first discriminator, so that the learning result of the first discriminator can be utilized, and the learning time and effort can be utilized. , Time, computer resources can be reduced.

In the above aspect, the learning unit provides the second discriminator generated by learning to the discriminator, and the discriminator uses the second discriminator for the input data classified into a specific group by the first discriminator. It may be applied to classify the input data into one of a plurality of subgroups.

Further, in the above embodiment, the input data is further provided with a storage unit for storing the first learning data in which the group in which the input data is classified is associated with the input data, and the learning unit is trained using the first learning data. Therefore, a first discriminator for classifying the input data into one of a plurality of groups may be generated, and the generated first discriminator may be provided to the discriminator.

Further, in the above aspect, the generation unit includes input data classified into a specific subgroup among a plurality of subgroups, and a specific small group among a plurality of subgroups that further subdivide the specific subgroup. The third learning data associated with the information of the above is further generated, and the learning unit learns using the third learning data, so that the input data classified into a specific subgroup is the feature of the input data. A third discriminator that classifies into one of a plurality of small groups based on the amount may be generated, and the generated third discriminator may be further provided to the discriminator.

Further, in the above aspect, the input data includes image data of the object to be inspected, and the plurality of groups include a group indicating that the object to be inspected is a non-defective product and a group indicating that the object to be inspected is a defective product. You may do so. Further, the plurality of subgroups may include a plurality of subgroups based on the type of defective product.

The discriminator according to one aspect of the present invention includes a main discriminator including a first discriminator that classifies input data into one of a plurality of groups based on the feature amount of the input data. Generates second training data in which input data classified into a specific group among a plurality of groups and information of a specific subgroup among a plurality of subgroups that subdivide the specific group are associated with each other. By training using the generation unit and the second training data, the input data classified into a specific group is divided into one of a plurality of subgroups based on the feature amount of the input data. For the input data classified into a specific group by the first discriminator using the second discriminator generated so as to classify the input data into one of a plurality of subgroups. It is provided with a generated sub-discrimination unit for classifying into. According to this aspect, a second discriminator is newly generated by learning based on the input data determined by the discriminator to have the characteristics of a specific group. This adds a function to discriminate the discrimination result of the first discriminator in more detail without discarding the existing first discriminator and regenerating the main discriminator in the discriminator. Can be done. As a result, the discrimination pattern of the discrimination device can be expanded without affecting the discrimination ability of the existing first discriminator.

The learning device according to one aspect of the present invention trains input data using a first storage unit that stores the first training data in which a group in which the input data is classified is associated with the input data, and the first training data. Thereby, the first learning unit that generates the first discriminator that classifies the input data into one of a plurality of groups, the input data that is classified into a specific group among the plurality of groups, and the relevant group. By training using the second training data and the second storage unit that stores the second training data associated with the information of the specific subgroup among the plurality of subgroups that subdivide the specific group. A second learning unit that generates a second discriminator that classifies input data classified into a specific group into one of a plurality of subgroups based on the feature amount of the input data. To prepare for. According to this aspect, a second discriminator is newly generated by learning based on the input data determined that the first discriminator has the characteristics of a specific group. This makes it possible to add a function capable of discriminating the discrimination result of the first discriminator in more detail without discarding the existing first discriminator and regenerating a new main discriminator. As a result, the discrimination pattern can be expanded without affecting the discrimination ability of the existing first discriminator.

The discrimination method according to one aspect of the present invention includes a step in which the computer includes a first discriminator that classifies the input data into one of a plurality of groups based on the feature amount of the input data. , Generates second training data that associates input data classified into a specific group among a plurality of groups with information of a specific subgroup among a plurality of subgroups that subdivide the specific group. The input data classified into a specific group by training using the step to be performed and the second training data is divided into one of a plurality of subgroups based on the feature amount of the input data. And the step of generating the second discriminator to be classified into.

A program according to one aspect of the present invention is a means including a first discriminator for classifying a computer into one of a plurality of groups based on the feature amount of the input data. A means for generating second training data in which input data classified into a specific group among the groups of the above and information of a specific subgroup among a plurality of subgroups subdividing the specific group are associated with each other. , And, by training using the second training data, the input data classified into a specific group is divided into one of a plurality of subgroups based on the feature amount of the input data. It functions as a means for generating a second discriminator for classification.

According to the present invention, different discriminant functions can be easily added to the configured discriminator.

An example of an application scene of the discrimination system according to the present embodiment is schematically illustrated. An example of the functional configuration of the discrimination device and the learning device according to the present embodiment is schematically illustrated. An example of the learning data for the main discriminator according to the present embodiment is schematically illustrated. An example of the learning data for the sub-discriminator according to the present embodiment is schematically illustrated. An example of the hardware configuration of the image processing apparatus according to the present embodiment is schematically shown. To. An example of the processing sequence of the discrimination system according to the present embodiment is schematically illustrated. ..

Preferred embodiments of the present invention will be described with reference to the accompanying drawings. In each figure, those with the same reference numerals have the same or similar configurations.

§1 Application example First, an example of a situation in which the present invention is applied will be described with reference to FIG. FIG. 1 schematically illustrates an example of an application scene of the discrimination system 1 according to the present embodiment. The discrimination system 1 performs an appearance inspection or appearance measurement of the work 4 by executing an image measurement process on the image data obtained by imaging the work 4 to be inspected carried on the belt conveyor 2, for example. , Realize the discrimination between non-defective products and defective products. In the following description, an example in which the image measurement process is applied to the presence / absence of a defect on the surface of the work 4 and the identification of the type of the defect will be described, but the present invention is not limited to this and can be applied to the measurement of the appearance shape and the like. .. Here, the defect is an abnormality possessed by the work, and includes, for example, scratches, unevenness of color and the like, stains, dents, chips, burrs, foreign matter, faint printing, misalignment of printing and the like. Further, the application example of the discrimination system 1 is not limited to the visual inspection by the image, and can be widely applied to the example of classifying the input data such as the failure prediction of the machine and the concentration degree determination of the driver of the automobile.

An image pickup device 30 is arranged on the upper portion of the belt conveyor 2, and the image pickup field of view 6 of the image pickup device 30 is configured to include a predetermined area of the belt conveyor 2. The image data (hereinafter, also referred to as “input data”) generated by the image pickup of the image pickup device 30 is transmitted to the discrimination device 10 and the learning device 20. Imaging by the imaging device is performed periodically or eventually.

The discrimination device 10 and the learning device 20 are connected to the PLC (programmable controller) 70, the database device 80, and the like via the upper network 8. The measurement result in the discrimination device 10 may be transmitted to the PLC 70 and / or the database device 80. In addition to the PLC 70 and the database device 80, any device may be connected to the upper network 8.

The discriminator 10 is set with a main discriminator that classifies the input data into a specific group based on which of the plurality of groups the input data has the characteristics of the group. The discriminator 10 can add a sub-discriminator ex post facto to classify the discriminant result of the main discriminator into a more detailed group. For example, the main discriminator determines whether or not the work in the input data includes features corresponding to defects, so that the input data is a non-defective group (that is, a group of input data relating to the work having no defects. ) And a defective product group (that is, a group of input data relating to a workpiece having a defect). In this case, for example, the sub-discriminator can classify the input data classified into the defective product group into a plurality of groups such as a dent group, a scratch group, and a chip group based on the type of defect. The categories that can be classified by the sub-discriminator are not limited to the types of defects, and are classified based on, for example, the cause of the defect (worker name and process), the imaging date of the input data, the processing date of the work, and the like. But it may be.

Further, the discriminator 10 can add a sub-discriminator that classifies the discriminant result of the sub-discriminator into a more detailed group (subgroup). Specifically, it is assumed that the discriminator 10 is set with a sub-discriminator that classifies input data into four types of defects, such as dents, chips, scratches, and others. In this case, it is possible to add a further sub-discriminator that further classifies the input data classified by the sub-discriminator into the other two types. Further, for example, a further sub-discriminator that classifies the input data classified into the dents into three types of large, medium, and small based on the size of the dent may be added.

Such a sub-discriminator is generated by classifying the input data classified by the main discriminator based on more detailed features and using the classified input data as learning data for additional learning.

As described above, in the discrimination system 1 according to the present embodiment, the learning device 20 newly generates a sub-discriminator by learning based on the input data determined that the discrimination device 10 has the characteristics of a specific group. As a result, in the discrimination device 10, it is possible to add a function capable of discriminating the discrimination result of the main discriminator in more detail without discarding the existing main discriminator and regenerating a new main discriminator. can. As a result, it is difficult to assume all groups in which the input data can be classified before learning (hereinafter, the correspondence between the input data and the group in which the input data is classified is also referred to as a "discrimination pattern"). , Even if it becomes necessary to expand the discrimination pattern in the middle of classification, the classification is expanded by performing additional learning for the sub-discriminator without wasting the previous learning in the main discriminator. It is possible to acquire the ability and create an efficient discriminator. That is, the discrimination pattern of the discrimination device 10 can be expanded without affecting the discrimination ability of the existing main discriminator.

§2 Configuration example An example of the functional configuration of the discrimination device 10 and the learning device 20 will be described with reference to FIG. FIG. 2 is a functional block diagram of the discrimination device 10 and the learning device 20 according to the present embodiment.

"Learning device 20"
First, the functional configuration of the learning device 20 will be described. As shown in FIG. 2, the learning device 20 includes a main discriminator DB 212, a sub-discriminator DB 222, a learning data DB 211 for the main discriminator, a learning data DB 221 for the sub-discriminator, a main discriminator learning unit 210, and a sub-discrimination device. It has a device learning unit 220 and.

In the main discriminator DB 212, a main discriminator (an example of the first discriminator) generated by the main discriminator learning unit 210 is registered. The main discriminator has a function of classifying input data based on the characteristics of the input data. As an example, the main discriminator classifies the input data into a non-defective product group and a defective product group based on whether or not the input data contains features corresponding to defects. The main discriminator may be configured to perform learning using the image data of the non-defective product and discriminate the image data having an outlier of a predetermined value or more from the feature amount of the learned image data of the non-defective product as the image of the defective product. ..

In the sub-discriminator DB 222, a sub-discriminator (an example of a second discriminator) generated by the sub-discriminator learning unit 220 is registered. The sub-discriminator is a discriminator capable of performing more detailed classification than the main discriminator. That is, the sub-discriminator is classified into a specific subgroup among a plurality of groups (subgroups) that further classify the input data that the main discriminator has determined to have the characteristics of a specific group. Input data can be determined.

In the present embodiment, the sub-discriminator classifies the input data discriminated as the defective product group by the main discriminator in more detail. For example, the sub-discriminator can further classify the input data of the defective product group into a dent group, a scratch group, and a chip group based on the type of defect. In this case, the sub-discriminator determines which group the input data is classified into based on whether the feature amount of the learned type of defect and the feature amount of the defect in the input data are within a predetermined range. Determine.

Each discriminator is a discriminator generated by performing a learning process and an evaluation process on an arbitrary machine learning model. The machine learning model has a predetermined model structure and processing parameters that vary depending on the learning process, and the processing parameters are optimized based on the experience obtained from the training data, so that the accuracy of the recognition result is improved. It is a model to do. That is, the main discriminator and the sub-discriminator of the present embodiment include a so-called "trained model" which is a combination of a predetermined model structure and optimized processing parameters. As the algorithm of the machine learning model, for example, a support vector machine, logistic regression, neural network, deep neural network and the like can be used, but the type thereof is not particularly limited.

The learning data DB 211 for the main discriminator (which is an example of the storage unit) stores learning data for the main discriminator learning unit 210 to perform learning. FIG. 3 is a diagram showing an example of learning data registered in the learning data DB 211 for the main discriminator. As shown in FIG. 3, in the learning data DB 211 for the main discriminator, the input data and the discriminant result (presence / absence of defect) that classifies the input data are registered as learning data in association with each other. The learning data registered in the learning data DB 211 for the main discriminator is not limited to the example of FIG. 3, and may be, for example, input data of a plurality of types of non-defective products.

The learning data DB 221 for the sub-discriminator (an example of the storage unit) stores learning data for the sub-discriminator learning unit 220 to perform learning. FIG. 4 is a diagram showing an example of learning data registered in the learning data DB 211 for a sub-discriminator. As shown in FIG. 4, in the learning data DB 221 for the sub-discriminator, the discriminant result regarding the presence or absence of a defect (discrimination result of the main discriminator) in the input data, and when the discriminant result has a defect, more detailed The classification result to the group is linked and registered for each category. In the example of FIG. 4, the categories are the type of defect and the cause of the defect.

The main discriminator learning unit 210 generates a main discriminator based on the learning data stored in the learning data DB 211 for the main discriminator, and registers it in the main discriminator DB 212.

The sub-discriminator learning unit 220 generates a sub-main discriminator based on the learning data stored in the sub-discriminator learning data DB 221 and registers it in the sub-discriminator DB 222. For example, when the learning data of FIG. 4 is given, the sub-discriminator learning unit 220 can generate a sub-discriminator capable of discriminating the type of defect and a sub-discriminator capable of discriminating the cause of the defect. Specifically, the sub-discriminator learning unit 220, as a sub-discriminator capable of discriminating the type of defect, is a sub-discriminator that classifies the defects of the input data based on whether the defects have scratches, chips, or dents. To generate. Further, as a sub-discriminator capable of discriminating the cause of the defect, the defect of the input data has the characteristics of any of the defect generated in the step X, the defect generated by the worker α, and the defect generated in the step y. Generate a sub-discriminator that classifies based on the crab.

"Discrimination device 10"
Next, an example of the functional configuration of the discrimination device 10 will be described. As shown in FIG. 2, the discrimination device 10 includes a main discrimination unit 110, a sub-discrimination unit 112, a display unit 114, an input unit 116, and a generation unit 118.

The main discriminator generated by the learning device 20 is set in the main discriminator 110. The main discriminator 110 determines whether the work in the input image data (input data) is a non-defective product or a defective product (whether or not it has a defect) by using the set main discriminator. The determination result and the input data are registered in, for example, the database device 80.

A sub-discriminator generated by the learning device 20 is set in the sub-discrimination unit 112. The number of sub-discriminators set in the sub-discrimination unit 112 is not limited to one, and may be one or a plurality. Further, the sub-discriminator 112 may be configured to have a plurality of sub-discriminators capable of classifying input data in different categories.

In the present embodiment, the sub-discrimination unit 112 performs more detailed discrimination on the input data determined to be defective by the main discrimination unit 110. Specifically, the sub-discriminator 112 sequentially inputs input data to a plurality of set sub-discriminators. Each sub-discriminator determines whether or not the input data includes the feature amount of the group that can be discriminated by the sub-discriminator. The discrimination result of the sub-discrimination unit 112 is associated with the input data and registered in, for example, the database device 80. The input data for which the group could not be discriminated by any of the sub-discriminators may be discriminated as unknown and registered in the database device 80.

Further, the sub-discriminator 112 may set a sub-discriminator (an example of the third discriminator) that classifies the input data discriminated by the sub-discriminator into more detailed groups (child groups). For example, after the type of defect is discriminated by the sub-discriminator, the input data can be further classified by the size of the defect by the following sub-discriminator. Further, the groups may be classified into detailed groups by the same procedure until the number of data in each group becomes a predetermined number or less.

The display unit 114 is a display that displays the discrimination results of the main discrimination unit 110 and the sub-discrimination unit 112. For example, it is preferable that the display unit 114 displays the input data on the display 104 together with the determination result.

The input unit 116 is a keyboard or mouse that accepts user operations.

The generation unit 118 receives an input of a new subgroup name (label) for classifying the input data from the user via the input unit 116. Specifically, the user classifies the input data determined by the main discrimination unit 110 as a defective product or the input data determined by the sub-discrimination unit 112 to be unknown, based on a predetermined feature. Set a label for. The classification of the input data into groups may be automatically performed by the discrimination device 10 by a method such as clustering based on the similarity of the feature amounts of the defects. Specifically, when a plurality of input data are automatically classified, the plurality of input data are classified into clusters based on the proximity and / or distance of the defect features of each of the plurality of input data. The cluster rig method may be used. Further, for example, a process of arbitrarily extracting N input data from a plurality of input data and classifying the remaining input data into N clusters according to the similarity of the features of the extracted N input data. You may use a non-hierarchical clustering method that is executed multiple times until the optimum N clusters are formed.

Further, as a configuration for automatically classifying, specifically, the input data classified into a specific group (for example, a group having a defect) by the main discriminator is further classified by a machine and classified as the same group. The input data can be the input data belonging to the "new subgroup". At this time, the user may give a name (label) to the new subgroup. Further, the discriminating device 10 may assign a unique identifier to the new subgroup and automatically name the assigned identifier as the name of the group. In addition, as an example of classification by machine, for example, classification by K-means method and other unsupervised learning can be mentioned.

Further, as another example of automatic classification, the following configuration can be considered, especially when defect classification is performed. First, some input data (photographs) having typical defects are prepared. At this time, a plurality of input data of at least one type of defect are required. For example, as input data regarding a defect called "dent", a plurality of photographs of a work having a dent are prepared. An auxiliary discriminator is generated by learning using the prepared input data. At this time, the auxiliary discriminator can discriminate the input data having a defect called "dent". Next, the input data group classified by the main discriminator as having a defect is input to this auxiliary discriminator, and new by the proximity and / or distance from the input data of the type of defect learned by the auxiliary discriminator. Classify into subgroups. The user may give a name to the new subgroup, or the discriminating device 10 may automatically name the new subgroup by automatically numbering an identifier as described above.

Further, the generation unit 118 associates the label set with the input data and registers it in the learning DB 221 for the sub-discriminator. As a result, in the learning device 20, a new sub-discriminator can be generated based on the result of the visual inspection performed by the discriminating device 10.

§3. Operation Flow Next, the processing flow of the discrimination system 1 configured as described above will be described with reference to FIG. FIG. 6 is a sequence diagram showing an example of the processing flow of the discrimination system 1 according to the present embodiment.

First, prior to the visual inspection in the discriminator 10, a learning data set for learning of the main discriminator is registered in the learning data DB 211 for the main discriminator of the learning device 20 (S21), and the main discriminator learning unit 210 performs the main discriminating. A vessel is generated (S22).

The generated main discriminator is set in the main discriminator 110 of the discriminator 0 (S23). Next, in the visual inspection, when the input data is input, the main discriminator 110 discriminates the input data by using the set main discriminator (S11). The determination result is displayed on the display unit 114 together with the input data (S12).

The generation unit 118 sets a label for a group that further classifies the input data determined to be a defective product (that is, a workpiece having a defect) among the input data displayed on the display unit 114 in more detail based on a predetermined feature. Is received from the user (S13). In the case of a configuration in which classification is automatically performed and naming is automatically performed, the process of S13 may be skipped.

The input data and the set label are registered in the sub-discriminator learning data DB 221 as a learning data set for learning the sub-discriminator (S14).

The learning device 20 generates a sub-discriminator using the learning data newly registered in the sub-discriminator learning data DB 221 (S24), and sets the generated sub-discriminator in the sub-discriminator 112 of the discriminator 10. (S25).

When the sub-discriminator is newly set, the discriminating device 10 discriminates the input data by using the main discriminating unit 110 and the sub-discriminating unit 112 (S15). For example, the sub-discrimination unit 112 can discriminate the type of defective product from the input data determined as a defective product by the main discriminating unit 110.

The discrimination device 10 can also return to S12 after performing discrimination using the sub-discrimination unit 112 and accept further settings of the sub-discriminator.

§4. Advantages In the discrimination system 1 according to the present embodiment, the learning device 20 newly generates a sub-discriminator based on the input data determined that the discrimination device 10 has the characteristics of a specific group. As a result, in the discrimination device 10, it is possible to add a function capable of discriminating the discrimination result of the main discriminator in more detail without discarding the existing main discriminator and regenerating a new main discriminator. can. As a result, the discrimination pattern of the discrimination device 10 can be expanded without affecting the discrimination ability of the existing main discriminator.

§5. Hardware Configuration Next, with reference to FIG. 5 , an example of the hardware configuration in the case where the discrimination device 10 and the learning device 20 described above are realized by the computer 800 will be described. The function of each device can be realized by dividing it into a plurality of devices.

As shown in FIG. 5 , the computer 800 includes a processor 801 and a memory 803, a storage device 805, an input I / F unit 807, a data I / F unit 809, a communication I / F unit 811 and a display device 813.

The processor 801 controls various processes in the computer 800 by executing a program stored in the memory 803. For example, the main discriminator 110 of the discriminator 10, the sub-discrimination unit 112, the display unit 114, the input unit 116, the generation unit 118, the main discriminator learning unit 210 of the learning device 20, the sub-discriminator learning unit 220, and the like may be used. It can be realized as a program that is temporarily stored in the memory 803 and then mainly operates on the processor 801. That is, the processor 801 interprets and executes the program temporarily stored in the memory 803, so that the main discriminator 110, the sub-discrimination unit 112, the display unit 114, the input unit 116, the generation unit 118, and the learning device 20 are the main discriminators. The functions of the learning unit 210 and the sub-discriminator learning unit 220 are realized.

The memory 803 is a storage medium such as, for example, a RAM (Random Access Memory). The memory 803 temporarily stores the program code of the program executed by the processor 801 and the data required when the program is executed.

The storage device 805 is a non-volatile storage medium such as a hard disk drive (HDD) or a flash memory. The storage device 805 stores an operating system and various programs for realizing each of the above configurations. In addition, the storage device 805 can also store the learning data DB 211 for the main discriminator, the main discriminator DB 212, the learning data DB 221 for the sub discriminator, and the sub-discriminator DB 222. Such programs and data are referenced from the processor 801 by being loaded into the memory 803 as needed.

The input I / F unit 807 is a device for receiving input from the user. Specific examples of the input I / F unit 807 include a keyboard, a mouse, a touch panel, various sensors, a wearable device, and the like. The input I / F unit 807 may be connected to the computer 800 via an interface such as USB (Universal Serial Bus).

The data I / F unit 809 is a device for inputting data from the outside of the computer 800. Specific examples of the data I / F unit 809 include a drive device for reading data stored in various storage media. It is also conceivable that the data I / F unit 809 is provided outside the computer 800. In that case, the data I / F unit 809 is connected to the computer 800 via an interface such as USB.

The communication I / F unit 811 is a device for performing data communication via the Internet N by wire or wirelessly with an external device of the computer 800. It is also conceivable that the communication I / F unit 811 is provided outside the computer 800. In that case, the communication I / F unit 811 is connected to the computer 800 via an interface such as USB.

The display device 813 is a device for displaying various information. Specific examples of the display device 813 include a liquid crystal display, an organic EL (Electro-Luminescence) display, a display of a wearable device, and the like. The display device 813 may be provided outside the computer 800. In that case, the display device 813 is connected to the computer 800 via, for example, a display cable or the like.

Although the embodiments of the present invention have been described in detail above, the above description is merely an example of the present invention in all respects. Needless to say, various improvements and modifications can be made without departing from the scope of the present invention. For example, in the above-described embodiment, an example in which the discrimination device 10 is applied to a defect inspection using an image has been described, but the present invention is not limited thereto. The discrimination device 10 can be widely applied to a discrimination device for classifying input data as well as failure prediction of a machine (for example, a manufacturing machine such as a packaging machine).

As an example, the discrimination device 10 can be applied to a failure prediction device of a machine. For example, the failure prediction device can acquire operating state data indicating the operating state of the target device from the device to be monitored (target device). The operating state data is, for example, time-series data such as torque fluctuation data of the motor of the target device, vibration data, and sound emitted by the target device. The failure prediction device has, for example, sensors such as a voice sensor, an image sensor, a motor rotation angle sensor, and an acceleration sensor, and may be configured to acquire operating state data from the target device, or the target device may have such a configuration. It may have a sensor and may be configured to provide the acquired operating state data to the failure prediction device.

The failure prediction device can detect a failure (or a sign of failure) of the target device based on the feature amount of the acquired operating state data. For example, the failure prediction device learns the feature amount of the operating state data of the target device during normal operation (that is, during normal operation) as normal data. As a result, when the operating state data having the feature amount deviating from the normal data by a certain amount or more is acquired, the main discriminator for determining the failure or the sign of the failure is generated.

Next, by learning based on the feature amount of the operating state data acquired when the cause of the failure such as the missing gear, the tooth spill of the belt, and the abnormal wear of the bearing is observed in the target device, it is subordinated. Generate a discriminator. As a result, the failure prediction device can classify the operating state data determined to be abnormal by the main discriminator into subgroups based on the proximity and / or distance from the feature amount for each failure factor. can.
According to such a failure prediction device, by continuously monitoring the operating state of the target device, it is possible to detect a sign of failure at an early stage and prompt the operator for maintenance in advance. As a result, it is possible to prevent inconveniences such as damage to the parts being worked on due to a failure occurring during the operation of the target device and stopping. Further, the failure prediction device classifies the operation status data in the sub-discriminator into subgroups for each cause of the failure, so that the operator can take a more appropriate response at the time of maintenance of the target device.

Further, the above-mentioned discrimination device 10 can also be applied to a concentration level discriminating device for discriminating the concentration level of a driver of an automobile. In this case, for example, the concentration ratio determination device can acquire the driver's state from a DMS (Drive Monitoring System) or a vehicle sensor. Specifically, the DMS is a sensor that detects the state of the driver of the automobile, and may be composed of a seat sensor, a camera that captures the upper body of the driver while the driver of the automobile is sitting in the driver's seat, and the like. Further, the vehicle sensor is a sensor that acquires information on the traveling state of the vehicle, and may be composed of, for example, a vehicle speed sensor, an acceleration sensor, a yaw rate sensor, or the like. The concentration determination device determines whether or not the driver is concentrated from the state of the driver of the vehicle and the running state of the vehicle acquired from these sensors by the main discriminator, and the group determined to be in a state of no concentration. A sub-discriminator that discriminates the factors (sleepy, eating, smartphone operation, etc.) may be generated for the input image group).

(Appendix 1)
A discriminator (110, 112) including a first discriminator that classifies the input data into one of a plurality of groups based on the feature amount of the input data.
A second learning data is generated in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are associated with each other. Generation unit (118) and
By training using the second learning data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. The learning unit (210, 220) that generates the second discriminator to be classified, and
A discrimination system (1).
(Appendix 2)
The learning unit (210, 220) provides the second discriminator generated by learning to the discriminating unit (110, 112).
The discriminator (110, 112) applies the second discriminator to the input data classified into a specific group by the first discriminator, and applies the input data to the input data among a plurality of subgroups. Classify into one of the subgroups,
The discrimination system (1) described in Appendix 1.
(Appendix 3)
A storage unit (212) for storing the first learning data, in which the input data is associated with a group in which the input data is classified, is further provided.
The learning unit (210, 220) generated and generated a first discriminator that classifies the input data into one of the plurality of groups by learning using the first learning data. The first discriminator is provided to the discriminator (110, 112).
The discrimination system (1) according to Appendix 1 or 2.
(Appendix 4)
The generation unit (118) includes input data classified into a specific subgroup among the plurality of subgroups, and a specific subgroup among a plurality of subgroups that further subdivide the specific subgroup. Further generate the third learning data associated with the information,
The learning unit (210, 220) makes the input data classified into the specific subgroup by learning using the third learning data, based on the feature amount of the input data, the plurality of input data. A third discriminator for classifying into any of the small groups is generated, and the generated third discriminator is further provided to the discriminator (110, 112).
The discrimination system (1) according to any one of Supplementary note 1 to 3.
(Appendix 5)
The input data includes image data of the object to be inspected.
The plurality of groups include a group indicating that the object to be inspected is a non-defective product and a group indicating that the object to be inspected is a defective product.
The discrimination system (1) according to any one of Supplementary note 1 to 4.
(Appendix 6)
The plurality of subgroups include a plurality of subgroups based on the type of defective product.
The discrimination system (1) according to Appendix 5.
(Appendix 7)
A main discriminator (110) including a first discriminator that classifies the input data into one of a plurality of groups based on the feature amount of the input data.
A second learning data is generated in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are associated with each other. Generation unit (118) and
By training using the second training data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. Using the second discriminator generated to classify, for the input data classified into a specific group by the first discriminator, the input data is subjected to one of a plurality of subgroups. Sub-discrimination unit (112) to generate and classify into
(10).
(Appendix 8)
A first storage unit (211) for storing the first learning data, in which the input data is associated with a group in which the input data is classified, and
A first learning unit (210) that generates a first discriminator that classifies input data into one of a plurality of groups by training using the first learning data.
The second learning data in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are stored is stored. Second storage unit (221) and
By training using the second learning data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. The second learning unit (220) that generates the second discriminator to be classified, and
A learning device (20).
(Appendix 9)
The computer (800)
A step including a first discriminator that classifies the input data into one of a plurality of groups based on the feature amount of the input data.
A second learning data is generated in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are associated with each other. Steps to do and
By training using the second learning data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. Steps to generate a second discriminator to classify,
How to determine what to do.
(Appendix 10)
Computer (800),
A means including a first discriminator for classifying input data into one of a plurality of groups based on the feature amount of the input data.
Generates second training data in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups that subdivide the specific group are associated with each other. And the input data classified into the specific group by training using the second training data, any one of the plurality of subgroups based on the feature amount of the input data. A means of generating a second discriminator that classifies into subgroups of
A program that makes it work.
(Appendix 11)
The generator is
Among the plurality of groups, a plurality of input data classified into a specific group are classified based on the similarity of the feature amounts of the plurality of input data to generate the specific subgroup.
The discrimination system described in Appendix 1.
(Appendix 12)
It is a defective product inspection system using the discrimination system described in Appendix 1.
Image data of the object to be inspected is applied as the input data,
The plurality of groups include a non-defective product group indicating that the object to be inspected is a non-defective product and a defective product group indicating that the object to be inspected is a defective product.
The specific group is the defective product group.
The subgroup is a subdivision of the defective product group based on the type of defective product.
Defective product inspection system.
(Appendix 13)
A failure determination system using the determination system described in Appendix 1.
As the input data, the operating status data indicating the operating status of the device is applied, and the operating status data is applied.
The plurality of groups include a normal group indicating that the operating state of the device is normal, and an abnormal group indicating that the operating state of the device is abnormal.
The specific group is the abnormal group.
The subgroup is a subdivision of the abnormal group based on the type of feature amount in the operating state.
Failure judgment system.

1 Discrimination system 2 Belt conveyor 4 Work 6 Imaging field 8 Upper network 10 Discriminator 20 Learning device 30 Imaging device 80 Database device 104 Display 110 Main discrimination unit 112 Sub-discrimination unit 114 Display unit 116 Input unit 118 Generation unit 210 Main discriminator learning Unit 220 Sub-discriminator Learning unit 800 Computer 801 Processor 803 Memory 805 Storage device 813 Display device

Claims (13)

A discriminator including a first discriminator that classifies the input data into one of a plurality of groups based on the feature amount of the input data.
A second learning data is generated in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are associated with each other. And the generator
By training using the second learning data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. A learning unit that generates a second discriminator for classification,
The first discriminator uses an auxiliary discriminator capable of discriminating the input data having the second feature, which is generated by learning using the input data having the second feature, which is a subdivision of the first feature. The input data classified into the groups related to the first feature are subgrouped or other subgroups related to the second feature based on the proximity and / or distance from the input data having the second feature. Auxiliary discriminator to classify into groups and
Equipped with
The generation unit is a discrimination system that acquires information on the specific subgroup based on the result of classifying the input data classified into the specific group by the auxiliary discrimination unit into subgroups . The learning unit provides the second discriminator generated by learning to the discriminating unit.
The discriminator applies the second discriminator to the input data classified into a specific group by the first discriminator, and applies the input data to one of a plurality of subgroups. Classify into
The discrimination system according to claim 1. Further, a storage unit for storing the first learning data, in which the input data is associated with the group in which the input data is classified, is provided.
The learning unit generates a first discriminator that classifies the input data into one of the plurality of groups by training using the first learning data, and the generated first discriminator. To the discriminating unit,
The discrimination system according to claim 1 or 2. The generation unit obtains input data classified into a specific subgroup among the plurality of subgroups and information on a specific subgroup among a plurality of subgroups that further subdivides the specific subgroup. Further generate the associated third training data,
The learning unit uses the third learning data to train the input data classified into the specific subgroup, based on the feature amount of the input data, among the plurality of small groups. A third discriminator to be classified into any of the small groups is generated, and the generated third discriminator is further provided to the discriminator.
The discrimination system according to any one of claims 1 to 3. The input data includes image data of the object to be inspected.
The plurality of groups include a group indicating that the object to be inspected is a non-defective product and a group indicating that the object to be inspected is a defective product.
The discrimination system according to any one of claims 1 to 4. The plurality of subgroups include a plurality of subgroups based on the type of defective product.
The discrimination system according to claim 5. The generator is
Among the plurality of groups, a plurality of input data classified into a specific group are classified based on the similarity of the feature amounts of the plurality of input data to generate the specific subgroup.
The discrimination system according to claim 1. A defective product inspection system using the discrimination system according to claim 1.
Image data of the object to be inspected is applied as the input data,
The plurality of groups include a non-defective product group indicating that the object to be inspected is a non-defective product and a defective product group indicating that the object to be inspected is a defective product.
The specific group is the defective product group.
The subgroup is a subdivision of the defective product group based on the type of defective product.
Defective product inspection system. A failure determination system using the determination system according to claim 1.
As the input data, the operating status data indicating the operating status of the device is applied, and the operating status data is applied.
The plurality of groups include a normal group indicating that the operating state of the device is normal, and an abnormal group indicating that the operating state of the device is abnormal.
The specific group is the abnormal group.
The subgroup is a subdivision of the abnormal group based on the type of feature amount in the operating state.
Failure judgment system. A main discriminator including a first discriminator that classifies the input data into one of a plurality of groups based on the feature amount of the input data.
A second learning data is generated in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are associated with each other. And the generator
By training using the second training data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. Using the second discriminator generated to classify, for the input data classified into a specific group by the first discriminator, the input data is subjected to one of a plurality of subgroups. Sub-discrimination unit to classify into
The first discriminator uses an auxiliary discriminator capable of discriminating the input data having the second feature, which is generated by learning using the input data having the second feature, which is a subdivision of the first feature. The input data classified into the groups related to the first feature are subgrouped or other subgroups related to the second feature based on the proximity and / or distance from the input data having the second feature. Auxiliary discriminator to classify into groups and
Equipped with
The generation unit is a discrimination device that acquires information on the specific subgroup based on the result of classifying the input data classified into the specific group by the auxiliary discrimination unit into subgroups . A first storage unit that stores the first learning data, in which the input data is associated with a group in which the input data is classified,
A first learning unit that generates a first discriminator that classifies input data into one of a plurality of groups by training using the first learning data.
The second learning data in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are stored is stored. The second storage part to do
By training using the second learning data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. The second learning unit that generates the second discriminator to classify,
The first discriminator uses an auxiliary discriminator capable of discriminating the input data having the second feature, which is generated by learning using the input data having the second feature, which is a subdivision of the first feature. The input data classified into the groups related to the first feature are subgrouped or other subgroups related to the second feature based on the proximity and / or distance from the input data having the second feature. Auxiliary discriminator to classify into groups and
Equipped with
The learning device in which the information of the specific subgroup is acquired based on the result of classifying the input data classified into the specific group into the subgroup by the auxiliary discrimination unit . The computer
A step of generating a first discriminator that classifies the input data into one of a plurality of groups based on the feature amount of the input data.
A second learning data is generated in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups subdividing the specific group are associated with each other. Steps to do and
By training using the second learning data, the input data classified into the specific group is divided into one of the plurality of subgroups based on the feature amount of the input data. Steps to generate a second discriminator to classify,
The first discriminator uses an auxiliary discriminator capable of discriminating the input data having the second feature, which is generated by learning using the input data having the second feature, which is a subdivision of the first feature. The input data classified into the groups related to the first feature are subgrouped or other subgroups related to the second feature based on the proximity and / or distance from the input data having the second feature. Steps to classify into groups and
And run
In the step of generating the second learning data, the information of the specific subgroup is acquired based on the result of classifying the input data classified into the specific group into the subgroups by using the auxiliary discriminator. Discrimination method. Computer,
A means for generating a first discriminator that classifies input data into one of a plurality of groups based on the feature amount of the input data.
Generates second training data in which the input data classified into a specific group among the plurality of groups and the information of the specific subgroup among the plurality of subgroups that subdivide the specific group are associated with each other. And the input data classified into the specific group by training using the second training data, any one of the plurality of subgroups based on the feature amount of the input data. A means of generating a second discriminator that classifies into subgroups of
The first discriminator uses an auxiliary discriminator capable of discriminating the input data having the second feature, which is generated by learning using the input data having the second feature, which is a subdivision of the first feature. The input data classified into the groups related to the first feature are subgrouped or other subgroups related to the second feature based on the proximity and / or distance from the input data having the second feature. How to classify into groups and
To make it work
The means for generating the second learning data acquires the information of the specific subgroup based on the result of classifying the input data classified into the specific group into the subgroups by using the auxiliary discriminator. program.

Images