Inspection equipment and PTP packaging machines

Abstract

To provide inspection equipment that can enhance inspection accuracy while applying past additional learning.SOLUTION: An additional learning unit 52h generates a trained AI model 200 by causing a neural network of an AI model 200 to additionally learn image data related to conforming products as additional learning data. When a defective product is incorrectly determined to be a conforming product by quality determination using an active AI model 200z used for inspection, the additional learning unit 52h generates a new trained AI model 200 by causing the neural network of the AI model 200 at a stage before obtaining the active AI model 200z to learn data other than additional learning data identified as factors of erroneous determination among the additional learning data used to generate the active AI model 200z from the AI model 200.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The present invention relates to an inspection device for inspecting a predetermined inspection object and a PTP packaging machine for producing PTP sheets having pockets in which tablets are contained. [Background technology]

[0002] A press-through pack (PTP) sheet is known as a blister sheet generally used in the fields of medicines, food, etc. A PTP sheet comprises a container film having a pocket portion in which tablets are placed, and a cover film attached to the container film so as to seal the opening side of the pocket portion.

[0003] Also, a given inspection device may be used to perform visual inspection of inspection objects such as PTP sheets, tablets, etc. In recent years, an inspection device has been proposed that performs visual inspection of inspection objects using an AI model generated by training a neural network using only image data related to inspection objects that are good products (see, for example, Patent Documents 1 and 2). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2022-29241 [Patent Document 2] Patent No. 7046150 Summary of the Invention [Problem to be solved by the invention]

[0005] However, an inspection device using an AI model may erroneously judge a good inspection target to be defective. In this case, the occurrence of erroneous judgments can be reduced by having the neural network of the AI ​​model additionally learn image data related to the erroneous judgment.

[0006] However, as additional learning is repeated, overlearning can occur in the neural network, which can result in situations such as erroneously determining that a defective inspection target is a good product (misjudgment due to overlearning).When such a situation occurs, the conventional approach is to return the AI ​​model to its initial state (for example, the factory default or unlearned state), which can make it impossible to make use of the past additional learning.

[0007] The present invention has been made in consideration of the above circumstances, and has an object to provide an inspection device and the like that can improve inspection accuracy while making use of past additional learning. [Means for solving the problem]

[0008] The following describes each of the means suitable for achieving the above object. Note that, as necessary, the specific effects of the corresponding means will be added.

[0009] Means 1. An inspection device for inspecting a specified inspection object, Image data acquisition means capable of acquiring image data relating to an inspection object; a trained classification means, which is a trained classification means configured to be able to output image data corresponding to the input image data by a neural network having an encoding unit that extracts a feature amount from the input image data and a decoding unit that reconstructs the image data from the feature amount, and which is generated by training the neural network of the classification means with only image data related to inspection objects that are non-defective as training data; an additional learning means for generating a new trained classification means by additionally learning image data related to a non-defective inspection target object as additional learning data to the neural network of the classification means; a reconstructed image data acquisition means for inputting original image data obtained by the image data acquisition means to the learned classification means and acquiring reconstructed image data as reconstructed image data; a determination means for comparing the original image data with the reconstructed image data acquired by inputting the original image data to the learned discrimination means, and determining whether the object to be inspected is good or bad based on the comparison result; a classification means storage means for storing an old classification means which is the classification means at a stage where additional learning by the additional learning means has not been performed in part or in whole; a specific learning data storage means for storing all of the specific learning data, which is the additional learning data used in generating a current identification means, which is the learned identification means used for inspecting an inspection object, from the old identification means, The inspection device is characterized in that, when a defective inspection object is erroneously judged to be a good item based on a pass / fail judgment by the judgment means based on the reconstructed image data reconstructed by inputting the original image data into the current identification means by the reconstructed image data acquisition means, the additional learning means is configured to be able to generate a new learned identification means by having the neural network of the old identification means stored in the identification means memory means learn the specific learning data excluding the error judgment cause data, which is the additional learning data identified as the cause of the error judgment.

[0010] The "image data of good-quality inspection objects" used as "learning data" in the above-mentioned Means 1 may include image data of good-quality inspection objects accumulated in previous inspections, image data of good-quality inspection objects visually selected by an operator, and virtual good-quality image data generated using these image data, etc.

[0011] In addition, since the "trained identification means" in the above-mentioned means 1 is generated by learning only image data related to good inspection objects, the reconstructed image data generated when original image data related to defective inspection objects is input to the trained identification means is almost identical to the original image data from which noise (defective parts) have been removed. In other words, when there is a defective part in the inspection object, virtual image data related to the inspection object on the assumption that there is no defective part is generated as the reconstructed image data.

[0012] The "neural network" includes, for example, a convolutional neural network having multiple convolutional layers. The "learning" includes, for example, deep learning. The "identification means (generative model)" includes, for example, an autoencoder and a convolutional autoencoder (similar to means 6).

[0013] Furthermore, the "old classification means" includes not only classification means in which training has been performed on the neural network, but also classification means in which the neural network is in an untrained state (the same applies to means 6).

[0014] According to the above-mentioned means 1, the original image data is compared with the reconstructed image data obtained by inputting the original image data to the learned identification means and reconstructing the image data, and the quality of the object to be inspected can be judged based on the comparison result. Therefore, both sets of image data to be compared relate to the same object to be inspected. Therefore, the shape and appearance of the object to be inspected are almost the same in both sets of image data to be compared, so that it is not necessary to set relatively loose judgment conditions to prevent erroneous detection due to differences in shape and appearance, and stricter judgment conditions can be set. Furthermore, the conditions for obtaining the image data (for example, the position and angle of the object to be inspected relative to the camera, the light and dark conditions, the angle of view of the camera, etc.) can be made to match in both sets of image data to be compared. As a result, the quality of the object to be inspected can be judged very accurately.

[0015] Further, according to the above-mentioned means 1, when a defective inspection object is erroneously determined to be a good product by a pass / fail judgment based on reconstructed image data obtained by inputting original image data to the current identification means and reconstructing the image data, that is, when it is considered that a misjudgment due to overlearning has occurred, the additional learning means performs learning on the neural network of the old identification means (for example, an initial AI model) which is an identification means at a stage where additional learning has not been performed in part or in whole (an identification means at a stage earlier than the current identification means) to generate a new trained identification means. At this time, the additional learning means trains the neural network of the old identification means with data other than the additional learning data identified as the cause of the misjudgment, among the specific learning data which is the additional learning data used to generate the current identification means from the old identification means. Therefore, in the newly generated trained identification means, the learning state can be corrected so that the occurrence of a misjudgment due to overlearning can be more reliably suppressed while making use of the past additional learning data. As a result, it is possible to more reliably prevent a defective inspection object from being erroneously determined to be a good product, and to improve the inspection accuracy.

[0016] In addition, the additional learning means learns the neural network of the old classification means stored in the classification means storage means, not the neural network of the current classification means used in the inspection. Therefore, the process related to the learning of the neural network and the process related to the extraction (identification) of data on the cause of erroneous judgment, which may be a huge amount of repetitive processing, can be performed without stopping the inspection. This makes it possible to quickly and efficiently inspect a large number of inspection objects.

[0017] Means 2. The identification means storage means stores, in addition to the current identification means, a duplicate current identification means, which is an identification means having the same contents as the current identification means; The inspection device described in means 1 is characterized in that, when a good inspection object is erroneously judged to be defective by the judgment means based on the reconstructed image data reconstructed by inputting the original image data into the current identification means by the reconstructed image data acquisition means, the additional learning means is configured to be able to generate a new learned identification means by additionally learning the original image data that caused the erroneous judgment to the neural network of the duplicate current identification means stored in the identification means memory means.

[0018] According to the above-mentioned means 2, when a good inspection object is erroneously judged to be defective, it is possible to generate a new trained identification means by additionally learning the original image data that caused the erroneous judgment to the neural network of the duplicate current identification means stored in the identification means storage means. Therefore, it is possible to generate a trained identification means that more reliably prevents a good inspection object from being erroneously judged to be defective, and it is possible to further improve the inspection accuracy.

[0019] In addition, according to the above-mentioned means 2, since the duplicate current identification means stored separately from the current identification means used in the inspection are subject to additional learning, it is not necessary to stop the inspection for additional learning. Therefore, it is possible to further improve the speed and efficiency of the inspection.

[0020] Means 3. A defective image data storage means for storing defective image data, which is the original image data related to the erroneous judgment when the defective inspection object is erroneously judged to be a good product by the pass / fail judgment of the judgment means; The inspection device according to the first aspect of the present invention is characterized in that it comprises a factor data identifying means for identifying the erroneous judgment factor data from the specific learning data based on the similarity to the defective image data.

[0021] In addition, the cause data identification means may, for example, have a known similarity judgment means (e.g., an AI generated by training a neural network with at least bad image data) that can calculate the similarity of specific learning data to bad image data (e.g., a score indicating the degree of similarity between both image data based on features extracted from the bad image data and features extracted from the specific learning data), and identify the cause data of the misjudgment based on the results of comparing the calculated similarity with a predetermined threshold value.

[0022] According to the above-mentioned means 3, it is possible to identify misjudgment factor data (additional learning data considered to be a factor of misjudgment) from among specific learning data based on the similarity to the defective image data, which is the original image data related to the misjudgment (original image data that caused the misjudgment). Therefore, it is possible to more accurately identify the misjudgment factor data, and ultimately to more reliably improve the accuracy of the inspection using the newly generated trained discrimination means.

[0023] Means 4. An inspection device as described in Means 1, characterized in that it is provided with a confirmation means for image output AI which confirms the learning state of the newly generated trained identification means by having the judgment means make a pass / fail judgment on the newly generated trained identification means based on the reconstructed image data which is reconstructed by inputting the original image data relating to the erroneous judgment that triggered the generation of the trained identification means.

[0024] According to the above-mentioned means 4, it is possible to easily check whether or not a newly generated learned discrimination means will not generate an erroneous judgment based on the original image data related to the erroneous judgment that triggered the generation of the learned discrimination means, that is, whether or not a better inspection accuracy has been achieved. Furthermore, in the unlikely event that a similar erroneous judgment occurs in an inspection using the newly generated learned discrimination means, it becomes possible to easily deal with such an erroneous judgment (re-identifying the erroneous judgment factor data by a different method, and learning again after removing the re-identified erroneous judgment factor data, or dealing with it other than learning (e.g., reviewing the threshold value used in the stage of comparing the original image data with the reconstructed image data, etc.)).

[0025] Means 5. A display means capable of displaying information related to the learning process by the additional learning means and the confirmation result by the confirmation means for the image output AI; The inspection device described in means 4 is characterized in that it comprises an approval input accepting means capable of accepting input regarding approval of replacing the current identification means with the newly generated learned identification means.

[0026] According to the above-mentioned means 5, it is possible to more reliably prevent the currently used identification means used in the inspection from being unintentionally replaced with the newly generated learned identification means. This makes it possible to more reliably prevent the occurrence of problems due to unintentional replacement.

[0027] Means 6. An inspection device for inspecting a specified inspection object, Image data acquisition means capable of acquiring image data relating to an inspection object; a trained classification means, which is a trained classification means generated by training a neural network of the classification means configured to classify input image data by predetermined items and output the classification result, with image data relating to inspection objects of good and defective items as training data; an additional learning means for additionally learning the neural network of the classification means with image data related to the inspection object as additional learning data, thereby generating a new trained classification means; an identification means storage means for storing an old identification means which is the identification means at a stage where additional learning by the additional learning means has not been performed in part or in whole, and a current identification means which is the learned identification means used for inspecting an inspection object, and a duplicate current identification means which is the identification means having the same contents as the current identification means; a specific learning data storage means for storing all of the specific learning data, which is the additional learning data used in generating the current identification means from the old identification means; The additional learning means is Image data obtained by the image data acquisition means.When an erroneous determination occurs in the classification result outputted by inputting original image data, the original image data causing the erroneous determination can be additionally trained on the neural network of the duplicate current discrimination means stored in the discrimination means storage means, thereby generating a new trained discrimination means; When a confirmation means for a classification output AI that confirms the learning state of the newly generated trained classification means based on the classification result output by inputting the original image data related to the erroneous judgment that triggered the generation of the trained classification means determines that the learning state of the newly generated trained classification means is appropriate, it is determined that the current classification means can be replaced with the newly generated trained classification means, An inspection device characterized in that, when an erroneous judgment occurs in the classification result output by inputting the original image data into the newly generated trained classification means and the confirmation means for the classification output AI judges that the learning state is inappropriate, the inspection device is capable of generating a new trained classification means by having the neural network of the old classification means stored in the classification means storage means learn the specific learning data excluding erroneous judgment cause data, which is data related to an inspection object having the same classification result as the erroneous classification result and is data identified as the cause of the erroneous judgment.

[0028] The "image data of the inspection object" used as the "learning data" in the above-mentioned means 6 includes not only image data of the inspection object of a good product but also image data of the inspection object of a defective product. In addition, the "image data of the inspection object" may include image data obtained during the inspection and virtual image data generated based on the image data.

[0029] Furthermore, the learned identification means may classify the objects to be inspected according to whether they are good or bad, or may classify the objects to be inspected according to the type of defect in the objects to be inspected in addition to whether they are good or bad.

[0030] According to the above-mentioned means 6, when an erroneous determination occurs regarding the classification result of the inspection object, it is possible to generate a new trained discrimination means by additionally learning the original image data that caused the erroneous determination to the neural network of the duplicated current discrimination means stored in the discrimination means storage means. Therefore, a trained discrimination means with better inspection accuracy can be obtained.

[0031] In addition, in the additional learning, learning is performed on the neural network of the duplicated current identification means stored in the identification means storage means, not on the neural network of the current identification means used in the inspection. This allows additional learning to be performed without stopping the inspection, and can further improve the speed and efficiency of the inspection.

[0032] In addition, according to the above-mentioned means 6, the confirmation means for the classification output AI can easily confirm whether or not the newly generated trained classification means will make a repeat erroneous judgment based on the original image data that caused the erroneous judgment, that is, whether or not better inspection accuracy has been achieved.

[0033] Furthermore, only after the confirmation means for the classification output AI has confirmed that a state in which a false positive will not occur again (a state in which better inspection accuracy has been achieved) is it determined that the currently used identification means used in the inspection of the inspection object can be replaced with the newly generated trained identification means. Therefore, it is possible to more reliably prevent an inappropriate trained identification means that may cause a false positive from being used in the inspection.

[0034] On the other hand, if the confirmation means for classification output AI confirms that an erroneous judgment occurs in an inspection using the newly generated trained classification means, that is, if it is considered that an erroneous judgment has occurred due to overlearning, the additional learning means trains the neural network of the old classification means (e.g., initial AI model), which is a classification means at a stage where additional learning has not been performed partially or completely (a classification means at a stage earlier than the current classification means), to generate a new trained classification means. At this time, data that is considered to be the cause of the erroneous judgment is identified as erroneous judgment cause data from data related to the inspection object with the same classification result as the erroneous judgment result, among the specific learning data, which is the additional learning data used to generate the current classification means from the old classification means.

[0035] For example, it is confirmed that a good product is erroneously classified as a defective product in an inspection by a newly generated trained discrimination means. In this case, the misjudgment factor data is identified from "data relating to an inspection target of a defective product" in the specific learning data. Also, it is confirmed that a type of defect occurring in an inspection target is erroneously judged as type Y when it should be judged as type X in an inspection by a newly generated trained discrimination means. In this case, the misjudgment factor data is identified from "data relating to an inspection target in which a defect of type Y occurs" in the specific learning data. This is because, when a misjudgment occurs in a classification result, it is considered that image data related to the same classification result as the misjudged classification result in the data used for additional learning (specific learning data) is the cause of the misjudgment.

[0036] The additional learning means then generates a new trained classification means by having the neural network of the old classification means learn data that excludes the misjudgment factor data from the specific training data. Therefore, when the classification output AI confirmation means judges that the training state is inappropriate, the training state can be corrected so that the occurrence of misjudgments due to over-training can be more reliably suppressed while making use of the past additional training data. As a result, it is possible to more reliably prevent misjudgments regarding the classification of the object to be inspected, and improve the inspection accuracy.

[0037] In addition, if the confirmation means for the classification output AI judges that the learning state is inappropriate, learning is performed on the neural network of the old classification means stored in the classification means storage means, rather than on the neural network of the current classification means used in the inspection. Therefore, the process related to learning the neural network and the process related to extracting (identifying) data on the cause of erroneous judgment, which can be a huge amount of repetitive processing, can be performed without stopping the inspection. This allows the inspection of a large number of inspection objects to be performed more quickly and efficiently.

[0038] Means 7. An erroneous judgment occurrence image data storage means for storing erroneous judgment occurrence image data, which is the original image data related to the erroneous judgment, when the classification result output by inputting the original image data to the newly generated learned classification means occurs and the learning state is judged to be inappropriate by the confirmation means for the classification output AI; The inspection device according to the means 6, further comprising a factor data identifying means for identifying the error judgment factor data from the specific learning data based on the similarity to the error judgment occurrence image data.

[0039] In addition, examples of the factor data identification means include a known similarity determination means (e.g., an AI generated by training a neural network with at least the misjudged image data) that can calculate the similarity of specific learning data to the misjudged image data (e.g., a score indicating the degree of similarity between both image data based on features extracted from the misjudged image data and features extracted from the specific learning data), and identifies the misjudgment factor data based on the results of comparing the calculated similarity with a predetermined threshold value.

[0040] According to the above-mentioned means 7, when the confirmation means for the classification output AI confirms that an erroneous judgment occurs in the trained classification means, it is possible to identify erroneous judgment factor data (additional learning data considered to be a factor of the erroneous judgment) from among the specific learning data based on the similarity to the erroneous judgment-occurring image data, which is the original image data related to the erroneous judgment (the original image data that caused the erroneous judgment). Therefore, it is possible to more accurately identify the erroneous judgment factor data, and ultimately to more reliably improve the accuracy of the inspection using the newly generated trained classification means.

[0041] Means 8. A display means capable of displaying information related to the learning process by the additional learning means and a confirmation result by the confirmation means for the classification output AI; The inspection device described in means 6, further comprising an approval input receiving means capable of receiving input regarding approval for replacing the current identification means with the newly generated learned identification means.

[0042] According to the above-mentioned means 8, the same effects as those of the above-mentioned means 5 are achieved.

[0043] Means 9: A PTP packaging machine equipped with the inspection device according to Means 1 or 6, for inspecting tablets or PTP sheets as inspection objects.

[0044] According to the ninth aspect, the same effects as those of the first or sixth aspect can be achieved.

[0045] The technical matters related to the above-mentioned means may be combined as appropriate. For example, the technical matters related to the above-mentioned means 2 may be combined with at least one of the technical matters related to the above-mentioned means 3 to 5. Also, the technical matters related to the above-mentioned means 7 may be combined with the technical matters related to the above-mentioned means 8. [Brief description of the drawings]

[0046] [Figure 1] FIG. [Diagram 2] FIG. 2 is a partially enlarged cross-sectional view of the PTP sheet. [Diagram 3] FIG. 1 is a perspective view of a PTP film. [Figure 4] FIG. 1 is a schematic configuration diagram of an inspection device according to a first embodiment. [Diagram 5] FIG. 1 is a schematic configuration diagram of a PTP packaging machine in a first embodiment. [Figure 6] 1 is a block diagram showing a functional configuration of an inspection device according to a first embodiment. [Figure 7] 2 is a partially cutaway front view showing a schematic configuration of a pocket forming device and a heating device. FIG. [Figure 8] 10 is a flowchart showing a flow of a pocket portion molding process. [Figure 9] FIG. 2 is a schematic diagram for explaining the structure of a neural network in the first embodiment. [Figure 10] 4 is a flowchart showing the flow of a learning process of a neural network in the first embodiment. [Figure 11] 13 is a flowchart showing a flow of a quality determination process. [Figure 12] 5 is a flowchart showing a flow of additional learning processing in the first embodiment. [Figure 13] 5 is a flowchart showing the flow of a confirmation process in the first embodiment. [Figure 14] 5 is a flowchart showing the flow of additional learning correction processing in the first embodiment. [Figure 15] 4 is a flowchart showing the flow of various processes in the first embodiment. [Figure 16] FIG. 13 is a diagram showing shading patterns and the like that occur in a pocket portion without molding defects. [Figure 17] 17 is a graph showing the luminance values ​​of the pixels along the line AA in FIG. 16. [Figure 18] FIG. 13 is a diagram showing shading patterns and the like that occur in a pocket portion with molding defects. [Figure 19] 19 is a graph showing the luminance values ​​of the pixels along the line BB in FIG. 18. [Figure 20]FIG. 2 is a diagram showing AI models that are sequentially generated and additional learning data used in additional learning processing in the first embodiment. [Figure 21] FIG. 11 is a schematic configuration diagram of a PTP packaging machine in a second embodiment. [Figure 22] FIG. 11 is a block diagram showing the functional configuration of an inspection device according to a second embodiment. [Figure 23] 10 is a flowchart showing the flow of a learning process of a neural network in a second embodiment. [Figure 24] 10 is a flowchart showing a flow of additional learning processing in the second embodiment. [Diagram 25] 13 is a flowchart showing the flow of a confirmation process in the second embodiment. [Figure 26] 13 is a flowchart showing the flow of additional learning correction processing in the second embodiment. [Figure 27] 10 is a flowchart showing the flow of various processes in the second embodiment. [Figure 28] FIG. 11 is a diagram showing AI models that are sequentially generated and additional learning data used in additional learning processing in the second embodiment. [Figure 29] FIG. 11 is a diagram showing an AI model stored in an AI storage unit and additional learning data stored in an image data storage unit in another embodiment. [Diagram 30] FIG. 1 is a perspective view of a blister pack according to another embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0047] Hereinafter, an embodiment will be described with reference to the drawings. [First embodiment] First, we will explain the PTP sheet 1 as the "inspection object." As shown in Figures 1 and 2, the PTP sheet 1 has a container film 3 with multiple pockets 2, and a cover film 4 attached to the container film 3 so as to cover the pockets 2.

[0048] The container film 3 is made of a colorless and transparent thermoplastic resin material such as PP (polypropylene) or PVC (polyvinyl chloride) and has light-transmitting properties. On the other hand, the cover film 4 is made of an opaque material (such as aluminum foil) with a sealant made of polypropylene resin or the like provided on the surface.

[0049] The PTP sheet 1 is formed in a generally rectangular shape in plan view. The PTP sheet 1 has two rows of pockets in the short direction, each row consisting of five pockets 2 arranged along the longitudinal direction. In other words, a total of ten pockets 2 are formed. Each pocket 2 contains one tablet 5 as the contents.

[0050] The pocket portion 2 is composed of a bottom portion 2a which is approximately circular in plan view and is arranged opposite the cover film 4, and a side portion 2b which is approximately cylindrical and connects the bottom portion 2a to the flat portion of the film (non-pocket formed portion) 3b.

[0051] In the present embodiment, the bottom 2a is formed to have a gently curved, generally arc-shaped cross section, but this is not limiting, and the bottom 2a may be formed to have a flat shape. Also, the bottom 2a may be formed to have a larger curvature in cross section, such that the corner 2c where the bottom 2a and the side 2b meet is not clearly defined.

[0052] The PTP sheet 1 is produced by punching out a strip-shaped PTP film 6 (see FIG. 3) formed from a strip-shaped container film 3 and a strip-shaped cover film 4.

[0053] Next, the schematic configuration of a PTP packaging machine 11 for producing the PTP sheet 1 will be described with reference to FIG.

[0054] At the most upstream side of the PTP packaging machine 11, a strip-shaped original web of container film 3 is wound into a roll. The pull-out end of the rolled container film 3 is guided by a guide roll 13. The container film 3 is hung on an intermittent feed roll 14 downstream of the guide roll 13. The intermittent feed roll 14 is connected to a motor that rotates intermittently, and transports the container film 3 intermittently.

[0055] Between the guide roll 13 and the intermittent feed roll 14, a heating device 15 and a pocket forming device 16 are arranged in this order along the transport path of the container film 3. After the container film 3 is heated by the heating device 15 and becomes relatively flexible, a plurality of pockets 2 are formed at once at predetermined positions on the container film 3 by the pocket forming device 16. The pockets 2 are formed during the intervals between the transport operations of the container film 3 by the intermittent feed roll 14. The configurations of the heating device 15 and the pocket forming device 16 will be described in detail later.

[0056] An inspection device 21 is disposed between the guide roll 13 and the intermittent feed roll 14, downstream of the pocket forming device 16. The inspection device 21 inspects the state of the pocket 2 formed by the pocket forming device 16, thereby inspecting the PTP sheet 1 having the pocket 2. The configuration of the inspection device 21 will be described in detail later.

[0057] The container film 3 fed from the intermittent feed roll 14 is wrapped around a tension roll 18, a guide roll 19, and a film receiving roll 20 in this order.

[0058] The film receiving roll 20 is connected to a motor that rotates at a constant speed, so it continuously transports the container film 3 at a constant speed. The tension roll 18 pulls the container film 3 toward the tension side by its elastic force, preventing slack in the container film 3 due to differences in the transport operations of the intermittent feed roll 14 and the film receiving roll 20, and keeping the container film 3 in a constantly tensed state.

[0059] Between the guide roll 19 and the film receiving roll 20, a tablet filling device 22 is disposed along the transport path of the container film 3.

[0060] The tablet filling device 22 functions as a filling means for automatically filling the tablets 5 into the pockets 2. The tablet filling device 22 drops the tablets 5 by opening a shutter at predetermined intervals in synchronization with the conveying operation of the container film 3 by the film receiving roll 20, and the tablets 5 are filled into each pocket 2 in association with this shutter opening operation. The operation of the tablet filling device 22 is controlled by a filling control device 82, which will be described later.

[0061] On the other hand, the original sheet of the cover film 4 formed in a strip shape is wound into a roll on the most upstream side. The drawn end of the cover film 4 wound into a roll shape is guided toward the heating roll 25 by the guide roll 24. The heating roll 25 can be pressed against the film receiving roll 20, and the container film 3 and the cover film 4 are fed between the two rolls 20 and 25. Then, the container film 3 and the cover film 4 pass between the two rolls 20 and 25 in a heated and pressed state, so that the cover film 4 is attached to the container film 3, and the pocket portion 2 is closed with the cover film 4. In this way, the PTP film 6 in which the tablets 5 are accommodated in each pocket portion 2 is manufactured.

[0062] The PTP film 6 sent out from the film receiving roll 20 is hung around a tension roll 27 and an intermittent feed roll 28 in this order.

[0063] The intermittent feed roll 28 is connected to a motor that rotates intermittently, and therefore intermittently transports the PTP film 6. The tension roll 27 is in a state in which the PTP film 6 is pulled toward the tension side by the elastic force, and prevents the PTP film 6 from slackening due to the difference in the transport operation between the film receiving roll 20 and the intermittent feed roll 28, thereby keeping the PTP film 6 in a constantly tensed state.

[0064] The PTP film 6 fed from the intermittent feed roll 28 is wrapped around a tension roll 31 and an intermittent feed roll 32 in this order.

[0065] The intermittent feed roll 32 is connected to a motor that rotates intermittently, and therefore intermittently transports the PTP film 6. The tension roll 31 pulls the PTP film 6 toward the tension side by elastic force, and prevents the PTP film 6 from slackening between the intermittent feed rolls 28 and 32.

[0066] Between the intermittent feed roll 28 and the tension roll 31, a slit forming device 33 and an engraving device 34 are disposed in this order along the transport path of the PTP film 6. The slit forming device 33 has a function of forming a separation slit at a predetermined position on the PTP film 6. The engraving device 34 has a function of applying an engraving at a predetermined position (e.g., a tag portion) on the PTP film 6. Note that the separation slit and the engraving are not shown in FIG. 1 etc.

[0067] The PTP film 6 sent out from the intermittent feed roll 32 is hung around a tension roll 35 and a continuous feed roll 36 in that order on the downstream side.

[0068] Between the intermittent feed roll 32 and the tension roll 35, a sheet punching device 37 is disposed along the transport path of the PTP film 6. The sheet punching device 37 functions as a sheet punching means (separating means) that punches out the outer edge of the PTP film 6 into individual PTP sheets.

[0069] The PTP sheets 1 punched by the sheet punching device 37 are transported by a conveyor 38 and temporarily stored in a finished product hopper 39. However, when a defective product signal is input from a filling control device 82 (described later) to a defective sheet discharge mechanism 40, which can selectively discharge PTP sheets 1, the defective PTP sheets 1 are separately discharged by the defective sheet discharge mechanism 40 and transferred to a defective product hopper (not shown).

[0070] A cutting device 41 is disposed downstream of the continuous feed roll 36. A scrap portion 42 remaining in a strip shape after punching by the sheet punching device 37 is guided to the tension roll 35 and the continuous feed roll 36, and then led to the cutting device 41. The cutting device 41 has a function of cutting the scrap portion 42 to a predetermined size. The cut scrap portion 42 is stored in a scrap hopper 43 and then disposed of separately.

[0071] Next, the configurations of the heating device 15 and the pocket portion forming device 16 will be described with reference to FIG.

[0072] The heating device 15 includes an upper heater plate 15a and a lower heater plate 15b. Both heater plates 15a and 15b are configured to be heatable by a heater (not shown). Both heater plates 15a and 15b are disposed on either side of the transport path of the container film 3, and are each movable in a direction approaching or moving away from the container film 3.

[0073] Further, each of the heater plates 15a, 15b has a plurality of protrusions 15c, 15d at positions corresponding to the portions 3a of the container film 3 to be molded of the pocket portions 2. As shown in FIG.

[0074] The container film 3, which is conveyed intermittently, is sandwiched between the protrusions 15c and 15d as the heater plates 15a and 15b approach each other during a temporary stop, and is partially (spot) heated, and the heated parts are softened. In this embodiment, the contact parts of the protrusions 15c and 15d with the container film 3 are slightly smaller than the planar shape of the pocket portion 2.

[0075] The pocket forming device 16 includes a lower die 61 and an upper die 71. The lower die 61 has a plurality of insertion holes 64 at positions corresponding to the positions of the pockets 2. The lower die 61 is fixed to a support base 63 in a fixed state via a cylindrical lower die chamber 62. The support base 63 has a plurality of through holes, into which a rod-shaped slider 65 is inserted via a bearing mechanism. The slider 65 can be moved up and down by a cam mechanism (not shown).

[0076] A pocket portion forming die 66 is fixed to the upper part of the slider 65, and the pocket portion forming die 66 has a plurality of rod-shaped plugs 66a which can be inserted into the insertion holes 64. The tip shape of the plugs 66a is shaped to correspond to the inner surface of the pocket portion 2. The pocket portion forming die 66 moves up and down in conjunction with the up and down movement of the slider 65 driven by the cam mechanism. The lower die 61, the pocket portion forming die 66, etc. can be appropriately replaced depending on the type of PTP sheet 1 to be produced.

[0077] Furthermore, a circulation path 67 for circulating cooling water (or hot water) is formed inside each of the slider 65 and the pocket portion molding die 66. This makes it possible to suppress variations in the surface temperature of each plug 66a.

[0078] During the molding of the pocket portion 2, the plug 66a is arranged in the initial position, intermediate stop position, and protruding position in this order, and finally returns to the initial position. The initial position is a position where the plug 66a is arranged at the start of the molding process of the pocket portion 2, and the plug 66a arranged at this position is arranged below the insertion hole 64 and outside the insertion hole 64. The intermediate stop position is a position where the plug 66a is arranged at an intermediate stage of the molding process of the pocket portion 2, and the plug 66a arranged at this position is arranged inside the insertion hole 64 and forms a predetermined gap between the plug 66a and the container film 3. The protruding position is a position where the plug 66a is arranged at the final stage of the molding process of the pocket portion 2, and the tip surface of the plug 66a arranged at this position protrudes from the lower die 61 by an amount corresponding to the depth of the pocket portion 2. The operation of the plug 66a is controlled by a molding control device 81 described later.

[0079] On the other hand, the upper die 71 is fixed to an upper plate 73 that is vertically movable via a plate 72, and is movable in a direction approaching or moving away from the lower die 61. The upper die 71 has a gas supply hole 74 at a position opposite to the insertion hole 64 of the lower die 61.

[0080] Furthermore, a gas supply passage 75 communicating with the gas supply hole 74 is formed inside the plate 72 and the upper plate 73, and a predetermined high-pressure gas (inert gas, air in this embodiment) is supplied to the gas supply passage 75 from a gas supply device 76 constituted, for example, by a compressor.

[0081] In this embodiment, a total of 10 pockets 2 corresponding to one PTP sheet 1 are simultaneously formed by one operation of the pocket forming device 16. That is, five pockets 2 are simultaneously formed in the film width direction (Y direction) of the container film 3 and two pockets are simultaneously formed in the film transport direction (X direction).

[0082] 5, a description will be given of a molding control device 81 for controlling the molding of the pocket portion 2 by the heating device 15 and the pocket forming device 16. The molding control device 81 is configured by a computer system having a CPU (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), etc.

[0083] The molding control device 81 stores information about the initial position, intermediate stop position, and protruding position of the plug 66a, and controls the operation of the plug 66a based on this information. The information about the initial position, intermediate stop position, and protruding position of the plug 66a is changed as appropriate depending on the depth of the pocket portion 2 in the PTP sheet 1 to be manufactured.

[0084] Here, the pocket forming process executed under the control of the forming control device 81 will be described with reference to FIG.

[0085] In the pocket molding process, first, an intermediate stop position arranging process in step S101 is performed. In the intermediate stop position arranging process, the slider 65 is moved to move the pocket molding die 66 upward, so that the plug 66a arranged in the initial position moves upward.

[0086] When the plug 66a reaches the set intermediate stop position, the movement of the slider 65 is stopped and the plug 66a is placed at the intermediate stop position. At this time, the tip surface of the plug 66a is spaced a predetermined distance from the container film 3. This predetermined distance is usually smaller than the depth of the pocket portion 2.

[0087] Next, in the clamping process of step S102, the upper die 71 is moved downward to clamp the container film 3 between the lower die 61 and the upper die 71, which are in a fixed state. At this time, an annular portion of the container film 3 located around the to-be-molded portion 3a (see FIG. 7) that will become the pocket portion 2 is clamped between both dies 61, 71. The intermediate stop position arranging process and the clamping process may be performed simultaneously, or the clamping process may be performed prior to the intermediate stop position arranging process.

[0088] In the subsequent expanding process of step S103, gas is supplied from the gas supply device 76 through the gas supply path 75 to the gas supply hole 74, so that compressed air is blown from the front side (upper side in Fig. 7) of the portion to be molded 3a of the pocket portion 2 in the container film 3. By supplying gas, the portion to be molded 3a expands to the side opposite (lower side in Fig. 7) to the protruding side of the pocket portion 2 (upper side in Fig. 7), and is stretched and thinned.

[0089] Then, the portion to be molded 3a expands until it is supported by the tip surface of the plug 66a. When the portion to be molded 3a is expanded by supplying gas, the wall thickness of the portion to be molded 3a after expansion becomes approximately the same overall.

[0090] The amount of stretching of the container film 3 and the thickness of the to-be-molded portion 3a change depending on the intermediate stop position of the plug 66a. When the intermediate stop position of the plug 66a is relatively high, the amount of stretching of the container film 3 is relatively small, so that the to-be-molded portion 3a becomes thick overall. On the other hand, when the intermediate stop position of the plug 66a is relatively low, the amount of stretching of the container film 3 is relatively large, so that the to-be-molded portion 3a becomes thin overall.

[0091] In the subsequent final molding step S104, the plug 66a moves upward and is placed in the protruding position, which reverses the bulging direction of the portion to be molded 3a, and forms a pocket portion 2 having a predetermined depth.

[0092] When the container film 3 is deformed by pressing, the portion of the portion to be molded 3a corresponding to the bottom 2a is cooled by contacting the plug 66a, and is hardly stretched at all. Therefore, if the portion to be molded 3a is made thick overall by setting the intermediate stop position relatively high, the portion corresponding to the bottom 2a will remain thick when pressed by the plug 66a, and as a result, the side portion 2b of the pocket portion 2 to be molded will be relatively thin.

[0093] On the other hand, if the intermediate stop position is set relatively low so that the portion to be molded 3a is thin overall, the portion corresponding to the bottom 2a will be maintained in a thin state when pressed by the plug 66a, and as a result, the side portion 2b of the pocket portion 2 to be molded will be relatively thick.

[0094] In this way, by adjusting the intermediate stop position of the plug 66a and adjusting the thickness of the portion to be molded 3a, it is possible to adjust the balance of the thicknesses of the bottom 2a and side 2b in the pocket portion 2 that is ultimately molded.

[0095] After the final forming step, the plug 66a is placed in the initial position and the container film 3 is released from the clamping by both dies 61, 71, thereby completing the pocket forming step.

[0096] Next, a description will be given of the filling control device 82 for controlling the filling of the tablets 5 by the tablet filling device 22. The filling control device 82 is configured by a computer system having a CPU, RAM, and the like.

[0097] The filling control device 82 can control switching to fill or not fill a predetermined pocket portion 2 with tablets 5 based on the inspection result by the inspection device 21. Specifically, the filling control device 82 receives a pass / fail judgment result for a predetermined PTP sheet 1 (molding state of ten pocket portions 2) from the inspection device 21, and when the pass / fail judgment result is a "good" judgment, controls the tablet filling device 22 to fill all ten pocket portions 2 included in the PTP sheet 1 with tablets 5.

[0098] On the other hand, when the pass / fail judgment result for a given PTP sheet 1 is a "defective" judgment, the filling control device 82 controls the tablet filling device 22 so as not to fill tablets 5 into all ten pockets 2 included in the PTP sheet 1. In addition, the filling control device 82 outputs a defective product signal to the defective sheet discharge mechanism 40. As a result, the defective sheet discharge mechanism 40 discharges the PTP sheet 1 related to the defective product signal (defective sheet).

[0099] Next, a detailed description will be given of the configuration of the inspection device 21. As shown in FIGS.

[0100] The lighting device 50 irradiates a predetermined electromagnetic wave to a predetermined range of the container film 3 from the protruding side of the pocket portion 2 (the lower side in FIG. 4). The lighting device 50 has an electromagnetic wave irradiating device 50a and a diffusion plate 50b covering it, and is configured to be capable of surface emission. The lighting device 50 in this embodiment irradiates the container film 3 with electromagnetic waves including ultraviolet light.

[0101] The camera 51 has sensitivity in the wavelength region of the electromagnetic waves irradiated from the lighting device 50. The camera 51 is provided on the opening side of the pocket portion 2 of the container film 3 (upper side in FIG. 4), and is disposed so that the optical axis OL of the lens is aligned along the vertical direction (Z direction) perpendicular to the film flat portion 3b of the container film 3.

[0102] In addition, a bandpass filter 51a is provided corresponding to the lens of the camera 51. By providing the bandpass filter 51a, only the ultraviolet light transmitted through the container film 3 among the electromagnetic waves irradiated from the lighting device 50 is two-dimensionally imaged by the camera 51. In addition, the transmission image data thus acquired by the camera 51 becomes luminance image data in which the luminance differs for each pixel (each coordinate position) based on the difference in the transmittance of ultraviolet light in the container film 3.

[0103] Particularly in this embodiment, the bandpass filter 51a used is one that passes only ultraviolet light with a wavelength of 253±20 nm, which has a transmittance of approximately 30±10 percent through the container film 3. This is because if the transmittance of the electromagnetic waves passing through the container film 3 is too high or too low, it may be difficult to create a difference in light transmittance between the thin and thick parts of the bottom 2a of the pocket portion 2.

[0104] In addition, the imaging range of the camera 51 in this embodiment is set so as to image at once an area including at least a total of 10 pocket portions 2 corresponding to one PTP sheet 1 formed into the container film 3 in one operation of the pocket portion forming device 16.

[0105] The inspection control device 52 is composed of a computer system including a CPU that executes predetermined arithmetic processing, a ROM that stores various programs and fixed value data, a RAM in which various data is temporarily stored when various arithmetic processing is executed, and peripheral circuits for these.

[0106] The inspection control device 52 functions as various functional units such as a main control unit 52a, an illumination control unit 52b, a camera control unit 52c, an image capture unit 52d, an image processing unit 52e, a reconstructed image data acquisition unit 52f, a determination unit 52g, an additional learning unit 52h, and a storage processing execution unit 52n, which will be described later, by the CPU operating in accordance with various programs. In this embodiment, the reconstructed image data acquisition unit 52f constitutes the "reconstructed image data acquisition means", the determination unit 52g constitutes the "determination means", and the additional learning unit 52h constitutes the "additional learning means".

[0107] However, the various functional units are realized through cooperation of various hardware such as the CPU, ROM, RAM, etc., and there is no need to clearly distinguish between functions realized by hardware and functions realized by software, and some or all of these functions may be realized by hardware circuits such as ICs.

[0108] Furthermore, the inspection control device 52 is provided with an input unit 521 including a keyboard, a mouse, a touch panel, etc., a display unit 522 having a display screen such as a liquid crystal display, a communication unit 528 capable of transmitting and receiving various data to and from the outside, etc. In this embodiment, the display unit 522 constitutes the "display means."

[0109] The inspection control device 52 also includes an inspection information storage unit 523, a used AI storage unit 524, a non-used AI storage unit 525, an image data storage unit 526, and an additional learning information storage unit 527, each of which is configured using a hard disk drive (HDD) or a solid state drive (SSD), etc. In this embodiment, the non-used AI storage unit 525 constitutes the "identification means storage means."

[0110] Prior to describing the various functional units constituting the inspection control device 52, the input unit 521, the display unit 522, the storage units 523 to 527, and the communication unit 528 will be described.

[0111] The input unit 521 is an input means for inputting information to the inspection control device 52. In this embodiment, a predetermined additional learning command or additional learning correction command can be input via the input unit 521.

[0112] The additional learning command is input by an operator or the like, for example, when a good PTP sheet 1 is erroneously determined to be a defective product by the quality judgment of the judgment unit 52g described later. On the other hand, the additional learning correction command is input by an operator or the like, for example, when a defective PTP sheet 1 is erroneously determined to be a good product by the quality judgment of the judgment unit 52g. Of course, when it is determined that an erroneous judgment has occurred by an inspection device provided separately from the inspection device 21, the additional learning command or the additional learning correction command may be automatically input to the inspection control device 52.

[0113] In addition, when an additional learning command is input, an additional learning process described later is performed, and when an additional learning correction command is input, an additional learning correction process described later is performed. In addition, various setting information stored in the test information storage unit 523 can be changed by the input unit 521.

[0114] The display unit 522 is configured to be able to display, for example, various kinds of information stored in each of the storage units 523 to 527. Therefore, the display unit 522 can display image data (transmission image data) obtained by the camera 51, additional learning data 301 and defective image data 302, which will be described later, information used in the inspection (for example, a threshold value, etc.), a pass / fail judgment result of the PTP sheet 1, and the like.

[0115] The inspection information storage unit 523 is for storing various setting information used in the inspection, pass / fail judgment results, etc. In this embodiment, as various setting information, for example, the shapes and dimensions of the PTP sheet 1, the pocket portion 2, and the tablets 5, the shape and dimensions of the pocket frame W for defining the area of ​​the pocket portion 2, the relative positional relationship with the camera 51, the threshold value used for pass / fail judgment and similarity judgment of image data, etc. are stored.

[0116] The AI-in-use storage unit 524 stores the AI ​​model 200 as an "identification means." The AI ​​model 200 has a function of outputting image data corresponding to input image data by using a deep neural network 190 (hereinafter simply referred to as "neural network 190"; see FIG. 9).

[0117] Here, the structure of the neural network 190 will be described with reference to Fig. 9. Fig. 9 is a schematic diagram conceptually showing the structure of the neural network 190. The neural network 190 has a convolutional auto-encoder (CAE) structure including an encoder unit 191 as an "encoding unit" that extracts a feature amount (latent variable) TA from input image data (grayscale pattern data) GA, and a decoder unit 192 as a "decoding unit" that reconstructs image data (grayscale pattern data) GB from the feature amount TA.

[0118] The structure of the convolutional autoencoder is well known, so a detailed description will be omitted. However, the encoder unit 191 has a plurality of convolutional layers 193, and in each of the convolutional layers 193, a result of a convolutional operation performed on input data using a plurality of filters (kernels) 194 is output as input data for the next layer. Similarly, the decoder unit 192 has a plurality of deconvolutional layers 195, and in each of the deconvolutional layers 195, a result of a deconvolutional operation performed on input data using a plurality of filters (kernels) 196 is output as input data for the next layer. Then, in a learning process or additional learning process described later, the weights (parameters) of the filters 194 and 196 are updated.

[0119] 6, the in-use AI storage unit 524 stores, as the AI ​​model 200, a current AI model 200z that is currently being used to inspect the PTP sheet 1. The current AI model 200z will be described in more detail later.

[0120] The unused AI storage unit 525 stores the AI ​​models 200 that are not used for inspection. The AI ​​models 200 stored in the unused AI storage unit 525 include a duplicate current AI model 200z1, which is an AI model 200 with the same content as the current AI model 200z used for inspection of the PTP sheet 1 (stored in the used AI storage unit 524) and is stored separately from the current AI model 200z, and an initial AI model 200a.

[0121] The replicated current AI model 200z1 is obtained by copying the current AI model 200z, and even if various processes such as update processes are performed on it, it has no effect on the current AI model 200z. In this embodiment, the current AI model 200z, the replicated current AI model 200z1, and the initial AI model 200a each constitute a "learned identification means", and the initial AI model 200a constitutes an "old identification means". In addition, the current AI model 200z constitutes a "current identification means", and the replicated current AI model 200z1 constitutes a "duplicate current identification means".

[0122] The current AI model 200z is generated by the additional learning process and the additional learning correction process by the additional learning unit 52h, which will be described later, and is the AI ​​model 200 that is adopted in the approval process, which will be described later. The inspection device 21 inspects the PTP sheet 1 using the current AI model 200z. On the other hand, the duplicate current AI model 200z1 is not used to inspect the PTP sheet 1.

[0123] The current AI model 200z and the duplicated current AI model 200z1 can be sequentially updated by additional learning processing and additional learning correction processing by the additional learning unit 52h. In this embodiment, the used AI memory unit 524 and the unused AI memory unit 525 store "AI(n)" (see FIG. 20) as the current AI model 200z and the duplicated current AI model 200z1. Note that n represents a natural number.

[0124] The initial AI model 200a is an AI model 200 at a stage where no additional learning process has been performed by the additional learning unit 52h. In this embodiment, the unused AI storage unit 525 stores "AI(0)" (see FIG. 20), which is the AI ​​model 200 at the time of shipment from the factory, as the initial AI model 200a.

[0125] At the time of shipment from the factory, an initial AI model 200a is stored in each of the AI ​​storage units 524 and 525. The initial AI model 200a stored in the in-use AI storage unit 524 is used to inspect the PTP sheet 1 until it is updated to the current AI model 200z by additional learning processing.

[0126] The trained AI models 200, such as the current AI model 200z, the duplicated current AI model 200z1, and the initial AI model 200a, are generation models constructed by deep learning the neural network 190 using only image data related to a non-defective PTP sheet 1 having no molding defects in the pocket portion 2 as training data, and have a so-called autoencoder structure. In this embodiment, the initial AI model 200a is obtained by previously performing a training process, which will be described later, on the neural network 190, and the current AI model 200z and the duplicated current AI model 200z1 are obtained by performing multiple additional training processes on the neural network 190 of the initial AI model 200a.

[0127] Here, the learning process (initial learning process) performed when generating the initial AI model 200a will be described. In this embodiment, the learning process is performed in advance at a location different from the installation location of the inspection device 21 (for example, a manufacturing plant of the inspection device 21) based on the execution of a predetermined learning program. Prior to the learning process, a large number of image data relating to non-defective PTP sheets 1 with no molding defects in the pocket portion 2 are prepared as learning data. This learning data has the same format as the image data (shade pattern data) obtained by the image data acquisition unit 52x described later.

[0128] 10, in the learning process, first, in step S201, an unlearned neural network 190 is prepared. For example, the neural network 190 stored in advance in a predetermined storage device or the like is read out. Alternatively, the neural network 190 is constructed based on network configuration information (for example, the number of layers of the neural network and the number of nodes in each layer) stored in the storage device or the like.

[0129] Next, in step S202, reconstructed image data is acquired. That is, previously prepared learning data is provided as input data to the input layer of the neural network 190. Then, reconstructed image data output from the output layer of the neural network 190 is acquired.

[0130] In the next step S203, the learning data is compared with the reconstructed image data (reconstructed grayscale pattern data) output by the neural network 190 in step S202, and it is determined whether the error is sufficiently small (whether it is equal to or smaller than a predetermined threshold value).

[0131] Here, if the error is sufficiently small, in step S205, it is determined whether or not the termination condition of the learning process is satisfied. For example, if a positive determination is made a predetermined number of times in succession in step S203 without going through the process of step S204 described later, or if learning using all of the prepared learning data is repeated a predetermined number of times, it is determined that the termination condition is satisfied. If the termination condition is satisfied, the neural network 190 and its learning information (updated parameters, etc., described later) are stored in each of the AI ​​storage units 524 and 525 as the initial AI model 200a, and the learning process is terminated.

[0132] On the other hand, if the termination condition is not satisfied in step S205, the process returns to step S202, and the learning of the neural network 190 is performed again.

[0133] If it is determined in step S203 that the error is not sufficiently small, then in step S204, a network update process (learning of the neural network 190) is performed, and then the process returns to step S202 to repeat the above series of processes.

[0134] In the network update process in step S204, the weights (parameters) of the filters 194, 196 in the neural network 190 are updated to more appropriate ones using a known learning algorithm such as backpropagation so that a loss function representing the difference between the training data and the reconstructed image data is minimized. Note that, as the loss function, for example, BCE (Binary Cross-entropy) can be used.

[0135] By repeating the processes of steps S202 to S204 many times, the error between the learning data and the reconstructed image data in the neural network 190 is minimized, and more accurate reconstructed image data is output.

[0136] The initial AI model 200a finally obtained is image data (shade pattern data, sometimes called "original image data") obtained by the image data acquisition unit 52x described later, and when image data related to a good PTP sheet 1 is input, it generates reconstructed image data that is almost identical to the input image data. Also, the initial AI model 200a is image data (original image data) obtained by the image data acquisition unit 52x described later, and when image data related to a defective PTP sheet 1 is input, it generates reconstructed image data that is almost identical to the input image data with noise parts (parts corresponding to defective parts) removed. That is, when the PTP sheet 1 is defective, virtual image data related to the PTP sheet 1 on the assumption that there is no defective part is generated as the reconstructed image data related to the PTP sheet 1.

[0137] Furthermore, it is not necessary to perform the above-mentioned learning process each time an inspection device 21 is manufactured. The neural network 190 and its learning information (updated parameters, etc.) may be acquired in advance and stored as an initial AI model 200a in each AI memory unit 524, 525 of the inspection device 21.

[0138] The image data storage unit 526 stores various image data such as image data (shade pattern data) obtained by the image data acquisition unit 52x described later. The image data (shade pattern data) obtained by the image data acquisition unit 52x corresponds to one PTP sheet 1 and represents a total of 10 shade patterns related to the bottom 2a of the pocket portion 2, etc. In this embodiment, the image data storage unit 526 constitutes a "specific learning data storage means" and a "defective image data storage means."

[0139] The image data stored in the image data storage unit 526 includes additional learning data 301 used in the additional learning process by the additional learning unit 52h. The additional learning data 301 is image data related to a non-defective PTP sheet 1 among a plurality of image data (shade pattern data) obtained by the image data acquisition unit 52x, and satisfies a predetermined condition (e.g., selected by an operator, etc.).

[0140] The additional learning data 301 also includes image data that caused a misjudgment when a good PTP sheet 1 was judged to be defective by the judging unit 52g described later. In other words, the additional learning data includes image data relating to a PTP sheet 1 that was judged to be defective by the judging unit 52g but was actually a good PTP sheet.

[0141] In this embodiment, "IM(1)" to "IM(n)" (see FIG. 20) are stored as the additional learning data 301. For example, "IM(1)" is image data used in the additional learning process for the neural network 190 of the initial AI model 200a, and the AI ​​model 200 generated by this additional learning process is "AI(1)."

[0142] The additional training data 301 also includes specific training data 301a. The specific training data 301a is the additional training data 301 used to generate the current AI model 200z from the initial AI model 200a. Therefore, in this embodiment, all of the additional training data 301 corresponds to the specific training data 301a.

[0143] Furthermore, the image data stored in the image data storage unit 526 includes defective image data 302. The defective image data 302 is shading pattern data (original image data) related to a false determination when a defective PTP sheet 1 is falsely determined to be a good product by a quality determination process by the determination unit 52g described later. In other words, the defective image data 302 is different from the additional learning data 301 and is related to a defective PTP sheet 1.

[0144] In this embodiment, whether or not a defective PTP sheet 1 has been erroneously determined to be a good product by the quality judgment of the judgment unit 52g, or whether or not a good PTP sheet 1 has been erroneously determined to be a defective product by the quality judgment of the judgment unit 52g, is determined by visual inspection by an operator, etc. Of course, an inspection device other than the inspection device 21 may be provided separately, and whether or not an erroneous judgment has occurred may be determined by that inspection device.

[0145] Then, when an operator or the like inputs information using the input unit 521, the image data (shade pattern data) stored in the image data storage unit 526, when a defective PTP sheet 1 is mistakenly judged to be a good product, is identified as defective image data 302.

[0146] The additional learning information storage unit 527 stores information related to the learning of the AI ​​model 200, such as various information related to the additional learning process and the additional learning correction process (i.e., the learning process by the additional learning unit 52h).

[0147] The information stored in the additional learning information storage unit 527 includes, for example, all trained AI models 200 (on which learning processing, additional learning processing, and additional learning correction processing have been performed) generated up to the present time, including the current AI model 200z and the initial AI model 200a, and the additional learning data 301 used to train the trained AI models 200. In this embodiment, the trained AI model 200 constitutes a "trained identification means" like the initial AI model 200a and the current AI model 200z.

[0148] In addition, the information stored in the additional learning information storage unit 527 includes, for example, all operation contents performed on the input unit 521 in various processes (e.g., additional learning process, additional learning correction process, confirmation process and approval process described below), the date and time when the input related to approval was received in the approval process described below, etc.

[0149] Incidentally, all trained AI models 200 generated up to the present time include trained AI models 200 that were rejected in the approval process described below. The rejected trained AI models 200 are given information indicating that they were rejected, and can be distinguished from the adopted trained AI models 200 (i.e., the current or past current AI models 200z). The additional training data 301 includes data used to generate the rejected trained AI models 200. Incidentally, the additional training information storage unit 527 may store information related to all operation contents on the input unit 521 as an audit trail.

[0150] The communication unit 528 includes a wireless communication interface conforming to a communication standard such as a wired LAN (Local Area Network) or a wireless LAN, and is configured to be able to transmit and receive various data to and from the outside. For example, the result of the pass / fail judgment performed by the judgment unit 52g is output to the outside (such as the filling control device 82) via the communication unit 528.

[0151] Next, the above-mentioned various functional units constituting the inspection control device 52 will be described in detail.

[0152] The main control unit 52a is a functional unit that controls the entire inspection device 21, and is configured to be able to transmit and receive various signals to and from other functional units such as the illumination control unit 52b and the camera control unit 52c.

[0153] The lighting control unit 52b is a functional unit that drives and controls the lighting device 50, and controls the lighting timing and the like based on a command signal from the main control unit 52a.

[0154] The camera control unit 52c is a functional unit that drives and controls the camera 51, and controls the timing of imaging based on a command signal from the main control unit 52a. The timing of the illumination and imaging is controlled by the main control unit 52a based on a signal from an encoder (not shown) provided in the PTP packaging machine 11.

[0155] The illumination control unit 52b and the camera control unit 52c execute a process in which, at each interval when the transport of the container film 3 formed in the pocket portion 2 is temporarily stopped, the illumination device 50 irradiates the container film 3 with electromagnetic waves and the camera 51 captures an image of the electromagnetic waves (ultraviolet light) that have passed through the container film 3. The transmission image data captured and generated by the camera 51 is converted into a digital signal (image signal) inside the camera 51 and transferred in the form of a digital signal to the inspection control device 52 (image capture unit 52d).

[0156] The image capture unit 52d is a functional unit for capturing image data captured and acquired by the camera 51. The captured image data (transmission image data) is stored in the image data storage unit 526.

[0157] The image processor 52e is a functional unit that performs a predetermined image processing on the image data captured by the image capture unit 52d to obtain shading pattern data (original image data). In this embodiment, the image processor 52e generates shading pattern data by extracting a portion related to the pocket portion 2 from the image data (transmission image data) captured by the image capture unit 52d.

[0158] More specifically, first, a mask process is executed on the image data (transmission image data) captured by the image capture unit 52d. Specifically, a pocket frame W (see FIG. 16) is set according to the positions of the ten pockets 2 in the image data (transmission image data) captured by the image capture unit 52d, and a mask M is applied to the area other than the pocket area specified by the pocket frame W, i.e., the area corresponding to the film flat portion 3b.

[0159] Next, shading pattern data is generated by extracting images within each pocket frame W from the masked image data (masking image data). The generated shading pattern data corresponds to one PTP sheet 1, and represents a total of ten shading patterns K (see Figs. 16 and 18) at the bottom 2a of the pocket portion 2. The shading pattern data is stored in the image data storage unit 526. In this embodiment, some of the acquired multiple shading pattern data become the additional learning data 301 and the defective image data 302. Note that a total of ten shading pattern data each representing a shading pattern K for one pocket portion 2 may be obtained from the transmission image data.

[0160] Here, FIG. 16 is a diagram showing the shading pattern K that occurs in the bottom 2a of a pocket portion 2 without molding defects in the shading pattern data, and FIG. 17 is a graph showing the brightness values ​​of each pixel along line AA of the shading pattern K shown in FIG. 16.

[0161] Moreover, Figure 18 is a diagram showing a shading pattern K that occurs at the bottom 2a of a pocket portion 2 having a molding defect, and Figure 19 is a graph showing the brightness values ​​of each pixel along line BB of the shading pattern K shown in Figure 18.

[0162] As shown in Figures 16 to 19, the shading pattern K is two-dimensional image information having brightness information for each pixel (for example, any value among 256 gray levels from 0 to 255), and corresponds to an image (intensity distribution image of transmitted electromagnetic waves) showing the two-dimensional distribution of shading occurring at the bottom 2a of the pocket portion 2, etc., based on the relationship between the difference in thickness (thickness distribution) at each position (coordinate position) such as the bottom 2a of the pocket portion 2 and the transmittance of the electromagnetic waves passing therethrough.

[0163] In this embodiment, the pocket frame W is set to match the opening peripheral portion of the pocket portion 2 (the junction between the side portion 2b and the flat portion 3b of the film), so that the shading pattern K includes not only the bottom portion 2a of the pocket portion 2, but also a pattern relating to the side portion 2b of the pocket portion 2, and the corner portion 2c of the pocket portion 2 where the bottom portion 2a and the side portion 2b intersect.

[0164] In the shade pattern K, the range corresponding to the bottom 2a has a shade pattern having a shade distribution (brightness distribution) that generally corresponds to the thickness distribution of the bottom 2a. On the other hand, in the range corresponding to the side 2b or the corner 2c, the brightness information does not correspond to the electromagnetic waves transmitted along the thickness direction (X direction or Y direction) of the side 2b or the like, but corresponds to the electromagnetic waves transmitted along the stretching direction (Z direction) during molding, and therefore has little to do with the thickness of the side 2b or the like.

[0165] In this embodiment, the camera 51, the image capture unit 52d, and the image processing unit 52e constitute an image data acquisition unit 52x, which acquires shading pattern data (original image data) related to the PTP sheet 1. In this embodiment, the image data acquisition unit 52x constitutes an "image data acquisition means."

[0166] The reconstructed image data acquisition unit 52f inputs the shading pattern data (original image data) obtained by the image data acquisition unit 52x into a trained AI model 200 such as the current AI model 200z and acquires the reconstructed image data as reconstructed image data. More specifically, the reconstructed image data acquisition unit 52f provides the shading pattern data (original image data) obtained by the image data acquisition unit 52x as input data to the input layer of the neural network 190. Then, the reconstructed image data acquisition unit 52f acquires the reconstructed shading pattern data output from the output layer of the neural network 190 as reconstructed image data. Note that, when inspecting the PTP sheet 1, the reconstructed image data acquisition unit 52f inputs the original image data into the current AI model 200z stored in the AI ​​storage unit 524. On the other hand, during a confirmation process described later, the reconstructed image data acquisition unit 52f inputs the original image data into a trained AI model 200 different from the current AI model 200z, which is newly generated by an additional learning process or the like.

[0167] The judgment unit 52g is a functional unit that performs a quality judgment process for the molding state of the pocket portion 2. The judgment unit 52g compares the shading pattern data (original image data) obtained by the image data acquisition unit 52x with reconstructed image data that is reconstructed by inputting the shading pattern data into the trained AI model 200, and judges the quality of the pocket portion 2 and therefore the PTP sheet 1 based on the comparison result.

[0168] The pass / fail judgment process by the judgment unit 52g will be described in more detail with reference to Fig. 11. In the pass / fail judgment process, first, in step S301, the reconstructed image data acquisition unit 52f acquires reconstructed image data by inputting the shading pattern data (original image data) acquired by the image data acquisition unit 52x to the trained AI model 200. Note that the input target of the original image data is the current AI model 200z when inspecting the PTP sheet 1, and is the newly generated trained AI model 200 when performing the confirmation process described later.

[0169] Next, in step S302, the shading pattern data (original image data) obtained by the image data acquisition unit 52x is compared with the reconstructed image data related to the shading pattern data obtained by the reconstructed image data acquisition unit 52f, and the difference in luminance between each pixel of both data is calculated. Then, pixels whose difference is not within a predetermined allowable range are identified as defective pixels.

[0170] Next, in step S303, the area of ​​the connected components of the defective pixels (defective area) is calculated.

[0171] Next, in step S304, it is determined whether the maximum area among the calculated areas is equal to or less than a predetermined threshold value set in advance. In other words, it is determined whether the defective area is within an allowable range. If the maximum area is equal to or more than the threshold value, it is determined that at least one pocket portion 2 is defective, and in step S305, the PTP sheet 1 corresponding to the shading pattern data (original image data) is determined to be a "defective product."

[0172] On the other hand, if the maximum area is below the threshold value, in step S306, the PTP sheet 1 corresponding to the shading pattern data (original image data) is judged to be a "good product." The result of the quality judgment is stored in the inspection information storage unit 523. Furthermore, when the PTP sheet 1 is inspected, the quality judgment result is output to the filling control device 82, etc.

[0173] The above-mentioned method of determining pass / fail is merely an example, and the pass / fail determination may be performed by other methods, such as a method of determining the degree of dispersion (distribution state) of the connected components of the defective pixels. Of course, if there is even one defective pixel, regardless of the size of the defective pixel, the product may be determined to be "defective."

[0174] The additional learning unit 52h performs additional learning of the neural network 190 using the additional learning data 301, thereby sequentially generating new trained AI models 200. When the additional learning command is input by an operator or the like, the additional learning unit 52h performs the additional learning process.

[0175] The additional learning process will be described in more detail with reference to Fig. 12 and Fig. 20. In the additional learning process, first, in step S401, reconstructed image data is acquired. For example, when "IM(1)" is newly stored as unlearned additional learning data 301, and an additional learning command is input, the additional learning unit 52h provides "IM(1)", which is the additional learning data 301, as input data to the input layer of the neural network 190 of "AI(0)", which is the duplicated current AI model 200z1 (however, the duplicated current AI model 200z1 at this stage is the initial AI model 200a) stored in the unused AI storage unit 525. Then, reconstructed image data output from the output layer of the neural network 190 is acquired.

[0176] In the following step S402, the additional learning data 301 is compared with the reconstructed image data output by the neural network 190 in step S401, and it is determined whether or not the error therebetween is sufficiently small (whether or not it is equal to or smaller than a predetermined threshold value).

[0177] If the error is sufficiently small, in step S404, it is determined whether or not the termination condition for the additional learning is satisfied. For example, if the additional learning using the stored additional learning data 301 is repeated a predetermined number of times, it is determined that the termination condition is satisfied. If the termination condition is satisfied, the neural network 190 and its learning information (updated parameters, etc.) are stored in the unused AI storage unit 525 as a new trained AI model 200, and the additional learning process is terminated.

[0178] On the other hand, if the error is not sufficiently small in step S402, a network update process (learning of neural network 190) is performed in step S403, and then the process returns to step S401 to repeat the above series of processes. The network update process is the same as the process of step S204 in the above learning process. By repeating the processes of steps S401 to S403, in neural network 190, the error between additional learning data 301 and reconstructed image data is minimized, and more accurate reconstructed image data is output.

[0179] The trained AI model 200 newly generated by the additional learning process is shading pattern data (original image data) obtained by the image data acquisition unit 52x, similar to the current AI model 200z and the initial AI model 200a, and when image data related to a non-defective PTP sheet 1 is input, generates reconstructed image data that is almost identical to the input image data. The trained AI model 200 newly generated is shading pattern data (original image data) obtained by the image data acquisition unit 52x, similar to the current AI model 200z and the initial AI model 200a, and when image data related to a defective PTP sheet 1 is input, generates reconstructed image data that is almost identical to the input image data from which noise parts (parts corresponding to defective parts) have been removed.

[0180] Furthermore, the additional learning unit 52h includes a factor data identification unit 52i and an image output AI confirmation unit 52j as functional units. In this embodiment, the factor data identification unit 52i constitutes the "factor data identification means", and the image output AI confirmation unit 52j constitutes the "image output AI confirmation means".

[0181] The factor data identifying unit 52i functions particularly in the additional learning correction process. When the defective PTP sheet 1 is erroneously determined to be a good product by the quality determination process of the determining unit 52g, the factor data identifying unit 52i identifies the additional learning data 301 that caused the erroneous determination as erroneous determination factor data from the specific learning data 301a based on the similarity to the defective image data 302 related to the erroneous determination. In this embodiment, the factor data identifying unit 52i has a known similarity determining means (e.g., an AI generated by having a neural network learn at least the defective image data 302, etc.) capable of calculating the similarity of the specific learning data 301a to the defective image data 302 (e.g., a score indicating the similarity between both data 301a, 302 based on the feature amount extracted from the defective image data 302 and the feature amount extracted from the specific learning data 301a). Then, the factor data identifying unit 52i identifies the erroneous determination factor data based on the result of comparing the similarity calculated by the similarity determining means with a preset threshold value. The calculated similarity is, for example, 0 to 100, and one or more specific learning data 301a whose similarity is equal to or greater than the threshold value is identified as erroneous judgment factor data.

[0182] The confirmation unit 52j for image output AI performs a confirmation process to confirm the learning state of the trained AI model 200 newly generated by the additional learning process or the additional learning correction process. The confirmation process is performed when a new trained AI model 200 is generated by the additional learning process or the additional learning correction process. The confirmation process is performed as follows.

[0183] As shown in FIG. 13, first, in step S501, it is determined whether the learning state of the newly generated AI model 200 is appropriate.

[0184] Specifically, the reconstructed image data acquisition unit 52f inputs the additional learning data 301 or the defective image data 302 to the newly generated trained AI model 200 to acquire the reconstructed image data. Here, if the new trained AI model 200 has been generated by an additional learning process, the additional learning data 301 (original image data related to the erroneous determination that triggered the additional learning process) is input to the AI ​​model 200. On the other hand, if the new trained AI model 200 has been generated by an additional learning correction process, the defective image data 302 (original image data related to the erroneous determination that triggered the additional learning correction process) is input to the AI ​​model 200.

[0185] Next, the judgment unit 52g obtains a pass / fail judgment result based on a comparison between the additional learning data 301 or the defective image data 302 and the reconstructed image data.

[0186] Then, it is judged whether the obtained pass / fail judgment result is correct or not. Here, since the additional learning data 301 relates to a pass / fail PTP sheet 1, if the judgment result is "pass / fail" when the additional learning data 301 is input to a new trained AI model 200, the pass / fail judgment result can be said to be correct. Also, since the defective image data 302 relates to a defective PTP sheet 1, if the judgment result is "defective" when the defective image data 302 is input to a new trained AI model 200, the pass / fail judgment result can be said to be correct.

[0187] If the pass / fail judgment result is correct, i.e., if the learning state is appropriate, then in step S502 it is judged that the current AI model 200z can be replaced with the newly generated trained AI model 200, and the confirmation process is terminated. On the other hand, if the pass / fail judgment result is incorrect, i.e., if the learning state is inappropriate, then in step S503 it is judged that the current AI model 200z cannot be replaced with the newly generated trained AI model 200, and the confirmation process is terminated.

[0188] In addition, in the confirmation process, additional learning data 301 other than the original image data related to the erroneous judgment that triggered the additional learning process or the additional learning correction process may be input to the newly generated trained AI model 200 to determine whether the obtained pass / fail judgment result is correct.

[0189] Furthermore, in response to the input of the additional learning correction command, the additional learning unit 52h cooperates with the factor data specification unit 52i and the confirmation unit 52j for image output AI to execute additional learning correction processing.

[0190] 14, in the additional learning and correction process, first, in step S551, the factor data identifying unit 52i identifies one or more pieces of erroneous determination factor data. As a result, the erroneous determination factor data is extracted from the identified learning data 301a.

[0191] Next, in step S552, a re-learning process is executed. In the re-learning process, the neural network 190 of the initial AI model 200a stored in the unused AI storage unit 525 is trained with the specific learning data 301a excluding the misjudgment factor data identified in step S551, to generate a new trained AI model 200. For example, in step S551, it is assumed that "IM(3)" is identified as the misjudgment factor data from among the specific learning data 301a "IM(0)" to "IM(n)" stored in the image data storage unit 526. In this case, in step S552, the initial AI model 200a is subjected to the additional learning process using the specific learning data 301a other than "IM(3)", namely "IM(0)" to "IM(2)" and "IM(4)" to "IM(n)". As a result, a new trained AI model 200 is generated that has learned the specific learning data 301a other than "IM(3)", namely "IM(3)".

[0192] Next, in step S553, it is determined whether the learning state of the new trained AI model 200 is appropriate. This process is the same as the process in step S501 in the confirmation process. If the learning state is appropriate, it is determined that the defective image data 302 that caused the erroneous judgment has been corrected so that an appropriate pass / fail judgment is performed, and the additional learning correction process is terminated.

[0193] On the other hand, if the learning state is inappropriate (step S553: ​​NO), the process returns to step S551, and the misjudgment factor data is identified again. At this time, the factor data identifying unit 52i appropriately adjusts the threshold value to be compared with the similarity score, identifies only one misjudgment factor data with the highest similarity, and then identifies the misjudgment factor data again. Note that, if the learning state is determined to be inappropriate in step S553, a process of changing the threshold value used in the pass / fail judgment process may also be performed at the same time.

[0194] Then, by repeating the above process until the learning state becomes correct, a trained AI model 200 is finally obtained that can appropriately deal with the defective image data 302 that caused the erroneous judgment.

[0195] Further, the additional learning unit 52h includes, as functional units, an approval process execution unit 52k, an approval input acceptance unit 52l, and a replacement unit 52m. In this embodiment, the approval input acceptance unit 52l constitutes "approval input acceptance means."

[0196] The approval process execution unit 52k executes the approval process when a new trained AI model 200 is generated by executing the additional learning process or the additional learning correction process, and the confirmation process determines that the current AI model 200z can be replaced with the trained AI model 200. By executing the approval process, at least various information related to the learning process (additional learning process and additional learning correction process) by the additional learning unit 52h stored in the additional learning information storage unit 527 and the confirmation result by the confirmation process are displayed on the display unit 522. For example, an image showing the additional learning data 301 used to generate the new trained AI model 200 is displayed.

[0197] Furthermore, by executing the approval process, the display unit 522 displays an approval button that can be selected by an operator or the like when approving the replacement of the current AI model 200z (stored in the AI-in-use storage unit 524) used in the inspection of the PTP sheet 1 with the newly generated AI model 200. The display unit 522 also displays a non-approval button that can be selected by an operator or the like when not approving the replacement.

[0198] When the approval button is selected using the input unit 521, the approval input receiving unit 52l receives an input related to approval for replacing the current AI model 200z (stored in the used AI storage unit 524) used for inspecting the PTP sheet 1 with the newly generated AI model 200. When the approval input receiving unit 52l receives the input related to approval, it outputs a predetermined replacement execution command to the replacement unit 52m and ends the approval process.

[0199] On the other hand, when the non-approval button is selected using the input unit 521, the approval input receiving unit 52l ends the approval process without outputting the replacement execution command.

[0200] Note that a configuration may be adopted in which, upon execution of the approval process, a confirmation process execution button is displayed, and an operator or the like selects this button to execute the confirmation process again on the newly generated trained AI model 200. At this time, the additional training data 301 used in the confirmation process may be selected manually or automatically.

[0201] When the replacement execution command is input, the replacement unit 52m executes a replacement process to replace the current AI model 200z used for inspecting the PTP sheet 1 (stored in the used AI memory unit 524) with the newly generated trained AI model 200.

[0202] In the replacement process, first, the production of the PTP sheet 1 by the PTP packaging machine 11 is temporarily stopped. Then, the current AI model 200z stored in the used AI storage unit 524 is replaced with the new trained AI model 200 stored in the unused AI storage unit 525. Then, after the current AI model 200z is replaced, the production of the PTP sheet 1 is resumed. As a result, inspection of the PTP sheet 1 using the new current AI model 200z is started. Note that, as a result of the replacement process, the new trained AI model 200 stored in the unused AI storage unit 525 becomes the duplicated current AI model 200z1.

[0203] The save process execution unit 52n stores information related to the learning of the AI ​​model 200, such as various information related to the additional learning process and the additional learning correction process (i.e., the learning process by the additional learning unit 52h), in the additional learning information storage unit 527. The save process execution unit 52n stores, in the additional learning information storage unit 527, all trained AI models 200 (including AI models 200 not adopted in the approval process) generated up to the present time, including the current AI model 200z and the initial AI model 200a, the additional learning data 301 used to generate the trained AI model 200, all operation contents performed on the input unit 521 in various processes (e.g., the additional learning process, the additional learning correction process, the confirmation process, and the approval process), the date and time when the input related to approval was received in the approval process, and the date and time when the replacement of the current AI model 200z was performed.

[0204] The flow of the various processes (additional learning process, additional learning correction process, etc.) performed by the additional learning unit 52h described in detail above is, in brief, as shown in FIG.

[0205] That is, when a defective PTP sheet 1 is erroneously judged as a good product in the inspection (step S11: YES), an additional learning and correction command is input by an operator or the like, and an additional learning and correction process (step S12-A) is executed. That is, a new trained AI model 200 is generated by training the neural network 190 of the initial AI model 200a on the specific training data 301a excluding the misjudgment factor data.

[0206] On the other hand, if the inspection erroneously determines that a good PTP sheet 1 is defective (step S11: NO), an additional learning command is input by an operator or the like, and additional learning processing (step S12-B) is executed. That is, a new trained AI model 200 is generated by additionally learning the image data related to the erroneous determination for the duplicated current AI model 200z1.

[0207] Then, when a new trained AI model 200 is generated, a confirmation process is performed in step S14. In the confirmation process, first, in step S14-1, the learning state of the generated AI model 200 is confirmed. If the learning state is appropriate (step S14-1: YES), it is determined that the current AI model 200z can be replaced with the new trained AI model 200 (step S14-2-A). On the other hand, if the learning state is inappropriate (step S14-1: NO), it is determined that the current AI model 200z cannot be replaced with the new trained AI model 200 (step S14-2-B).

[0208] If it is determined that replacement is possible, an approval process (step S15) is executed, and information required by the operator or the like to make the replacement is displayed on the display unit 522. Then, when the operator or the like selects the approval button (step S16: YES), the PTP packaging machine 11 is temporarily stopped, and the current AI model 200z is replaced with the new trained AI model 200 (step S17). On the other hand, when the operator or the like selects the non-approval button (step S16: No), the replacement is not made, and the current AI model 200z is used as is.

[0209] Furthermore, if a new erroneous judgment occurs before the confirmation process or approval process is completed, the execution of a new additional learning process or additional learning correction process may be suspended until approval or disapproval of the replacement is made.

[0210] As described above in detail, according to this embodiment, the original image data (shade pattern data) acquired by the image data acquisition unit 52x is compared with the reconstructed image data reconstructed by inputting the original image data into the current AI model 200z, and the quality of the PTP sheet 1 is judged based on the comparison result. Therefore, both image data to be compared relate to the same PTP sheet 1. Therefore, in both image data to be compared, the shape and appearance of the PTP sheet 1 are almost the same, so there is no need to set relatively loose judgment conditions to prevent erroneous detection due to differences in shape and appearance, and stricter judgment conditions can be set. Furthermore, in both image data to be compared, the conditions when obtaining image data (for example, the position and angle of the PTP sheet 1 relative to the camera 51, the light and dark state, the angle of view of the camera 51, etc.) can be matched. As a result, the quality judgment can be performed with high accuracy.

[0211] In addition, when the defective PTP sheet 1 is erroneously determined to be a good product by the pass / fail judgment based on the reconstructed image data reconstructed by inputting the original image data into the current AI model 200z, that is, when it is considered that the erroneous judgment has occurred due to overlearning, the additional learning unit 52h performs learning on the neural network 190 of the initial AI model 200a, which is the AI ​​model 200 at a stage where additional learning has not been performed (the AI ​​model 200 at a stage earlier than the current AI model 200z), to generate a new trained AI model 200. At this time, the additional learning unit 52h makes the neural network 190 of the initial AI model 200a learn data other than the erroneous judgment factor data of the specific learning data 301a. Therefore, in the newly generated AI model 200, the learning state can be corrected so that the occurrence of erroneous judgment due to overlearning can be more reliably suppressed while making use of the past additional learning data 301 (specific learning data 301a). As a result, it is possible to more reliably prevent the defective PTP sheet 1 from being erroneously determined to be a good product, and to improve the inspection accuracy.

[0212] In addition, the additional learning unit 52h learns the neural network 190 of the duplicated current AI model 200z1, not the neural network 190 of the current AI model 200z used for the inspection. Therefore, the process related to the learning of the neural network 190 and the process related to the extraction (identification) of the misjudgment factor data, which may be a huge amount of repetitive processing, can be performed without stopping the inspection. This allows the inspection of a large number of PTP sheets 1 to be performed quickly and efficiently.

[0213] In addition, in this embodiment, when a good PTP sheet 1 is erroneously determined to be defective, the original image data that caused the erroneous determination is additionally learned by the neural network 190 of the duplicated current AI model 200z1, thereby generating a new trained AI model 200. Therefore, it is possible to more reliably prevent a good PTP sheet 1 from being erroneously determined to be defective, and to further improve the inspection accuracy.

[0214] In addition, since this additional learning is performed on the replicated current AI model 200z1, there is no need to stop testing for the additional learning. Therefore, testing of a large number of PTP sheets 1 can be performed more quickly and efficiently.

[0215] Furthermore, the factor data identifying unit 52i can identify misjudgment factor data (additional learning data 301 considered to be a factor of the misjudgment) from the specific learning data 301a based on the similarity to the defective image data 302, which is the original image data related to the misjudgment (original image data that caused the misjudgment). Therefore, the misjudgment factor data can be identified more accurately, and the accuracy of the inspection using the newly generated AI model 200 can be more reliably improved.

[0216] Furthermore, the image output AI confirmation unit 52j can easily confirm whether or not the newly generated AI model 200 will not generate an erroneous judgment based on the original image data related to the erroneous judgment that triggered the generation of the AI ​​model 200, that is, whether or not better inspection accuracy has been achieved. Furthermore, in the unlikely event that it is confirmed that a similar erroneous judgment occurs in an inspection using the newly generated AI model 200, it becomes possible to easily deal with such an erroneous judgment (for example, re-identifying the erroneous judgment factor data and re-learning after removing the re-identified erroneous judgment factor data, changing the threshold value used in the stage of comparing the original image data with the reconstructed image data, etc.).

[0217] In addition, only after approval by an operator or the like, the current AI model 200z is replaced with the newly generated trained AI model 200. This makes it possible to more reliably prevent the current AI model 200z from being unintentionally replaced with the newly generated trained AI model 200. This makes it possible to more reliably prevent the occurrence of malfunctions due to unintentional replacement. Second Embodiment Next, the second embodiment will be described with reference to Fig. 21 etc., focusing on the differences from the first embodiment. The main differences between the first embodiment and the second embodiment, relating to the AI ​​model and its learning, are as follows: (First embodiment above) · Trained AI model 200... Generated by learning only image data related to a non-defective PTP sheet 1, when image data is input, it outputs the image data. Additional learning process: This is executed when a good PTP sheet 1 is erroneously judged as defective. Additional learning correction process: This is executed when a defective PTP sheet 1 is mistakenly determined to be a good product. (Second embodiment) Trained AI model 400: Generated by learning image data relating to PTP sheets 1 of non-defective and defective products, when image data is input, it outputs a classification result. Additional learning process: This process is executed not only when a good PTP sheet 1 is erroneously determined to be defective, but also when a defective PTP sheet 1 is erroneously determined to be good, or when an erroneous classification occurs. Additional learning correction process: This is executed if an incorrect judgment occurs in the confirmation process that follows the additional learning process.

[0218] The second embodiment will be described in detail below.

[0219] In the first embodiment, the inspection device 21 is disposed downstream of the pocket forming device 16, and judges the quality of the PTP sheet 1 by inspecting the formed state of the pocket portion 2. In contrast, as shown in FIG. 21, the inspection device 23 in the second embodiment is disposed downstream of the tablet filling device 22, and judges the quality of the PTP sheet 1 by inspecting the state of the tablets 5 filled in the pocket portion 2 and the presence or absence of foreign matter mixed into the pocket portion 2. The inspection device 23 also judges the type of defect of the PTP sheet 1 in addition to judging the quality of the PTP sheet 1. In the second embodiment, the defective sheet discharge mechanism 40 discharges the defective PTP sheet 1 when a defective product signal is input from the inspection device 23.

[0220] The inspection device 23 is a transmitted light type inspection device that performs inspection from the protruding portion side of the pocket portion 2 of the container film 3 before sealing. The inspection device 23 has a lighting device 53, a camera 54, and an inspection control device 55.

[0221] The lighting device 53 is disposed downstream of the tablet filling device 22 on the opening side of the pocket portion 2 of the container film 3, and irradiates a predetermined area of ​​the container film 3 with predetermined light (near-infrared light, etc.).

[0222] The camera 54 is provided on the protruding side of the pocket portion 2 of the container film 3, and has sensitivity to the wavelength region of the light irradiated from the lighting device 53. The camera 54 captures the light (near-infrared light) irradiated from the lighting device 53 that has passed through the container film 3. The camera 54 captures an image of an area including a total of 10 pocket portions 2 corresponding to at least one PTP sheet 1 at a time.

[0223] The inspection control device 55 is configured by a computer system having a CPU and the like, similar to the inspection control device 52 in the first embodiment. As shown in Fig. 22, the CPU operates in accordance with various programs, so that the inspection control device 55 functions as various functional units such as a main control unit 55a, a lighting control unit 55b, a camera control unit 55c, an image capture unit 55d, an image processing unit 55e, a judgment processing unit 55g, an additional learning unit 55h, and a storage processing execution unit 55n. In the second embodiment, the additional learning unit 55h constitutes the "additional learning means."

[0224] In addition, the main control unit 55a, the lighting control unit 55b, the camera control unit 55c, the image capture unit 55d, and the image processing unit 55e function in the same manner as the main control unit 52a, the lighting control unit 52b, the camera control unit 52c, the image capture unit 52d, and the image processing unit 52e in the first embodiment described above.

[0225] Furthermore, the inspection control device 55 is provided with an input unit 551, a display unit 552, and a communication unit (not shown) capable of transmitting and receiving various data to and from the outside. The communication unit functions in the same manner as the communication unit 528 in the first embodiment. In the second embodiment, the display unit 552 constitutes a "display means."

[0226] The inspection control device 55 also includes an inspection information storage unit 553, a used AI storage unit 554, a non-used AI storage unit 555, an image data storage unit 556, and an additional learning information storage unit 557, each of which is configured with a storage device such as an HDD. In this embodiment, the non-used AI storage unit 555 constitutes the "identification means storage means."

[0227] The input unit 551 is an input device used for inputting information to the inspection control device 55 and changing the information, and a predetermined additional learning command can be input via the input unit 551. The additional learning command is input by an operator or the like when an erroneous determination is made in the classification result by the trained AI model 400 described later. Of course, a configuration may be adopted in which an additional learning command is automatically input to the inspection control device 55 when it is determined that an erroneous determination has occurred by an inspection device provided separately from the inspection device 23. Note that the input of the additional learning command causes an additional learning process described later to be performed.

[0228] The display unit 552 is a display device capable of displaying various pieces of information stored in the respective storage units 553-557.

[0229] The inspection information storage unit 553 is for storing various setting information used in the inspection, pass / fail judgment results, etc. In the second embodiment, a threshold value for pass / fail judgment and a threshold value for classification are stored in advance as various setting information. These threshold values ​​are used for classification by the trained AI model 400.

[0230] The AI-in-use storage unit 554 stores the AI ​​model 400 as a "classification means." The AI ​​model 400 classifies input original image data by predetermined items using a deep neural network (hereinafter simply referred to as a "neural network"), and outputs the classification result. In other words, the AI ​​model 400 in the second embodiment is a classification type AI model.

[0231] The in-use AI storage unit 554 stores, as the AI ​​model 400, a current AI model 400z that is used to inspect the PTP sheet 1. The current AI model 400z will be described in more detail later.

[0232] The unused AI storage unit 555 stores the AI ​​models 400 that are not used for inspection. The AI ​​models 400 stored in the unused AI storage unit 555 include a duplicated current AI model 400z1 that has the same content as the current AI model 400z that is used for inspecting the PTP sheet 1 and is stored separately from the current AI model 400z, and an initial AI model 400a.

[0233] In the second embodiment, trained AI models 400, such as the current AI model 400z, the duplicated current AI model 400z1, and the initial AI model 400a, constitute a "trained identification means," and the initial AI model 400a constitutes an "old identification means." In addition, the current AI model 400z constitutes a "current identification means," and the duplicated current AI model 400z1 constitutes a "duplicate current identification means."

[0234] The current AI model 400z is generated by the additional learning process and the additional learning correction process by the additional learning unit 55h, and is the AI ​​model 400 that is adopted in the approval process described below. The inspection device 23 inspects the PTP sheet 1 using the current AI model 400z.

[0235] The current AI model 400z and the duplicated current AI model 400z1 can be updated sequentially by additional learning processing and additional learning correction processing by the additional learning unit 55h. In this embodiment, each AI memory unit 554, 555 stores "AI2(n)" (see FIG. 28) as the current AI model 400z and the duplicated current AI model 400z1. Note that n represents a natural number.

[0236] The initial AI model 400a is an AI model 400 at a stage where a learning process, which will be described later, has been performed, but no additional learning process has been performed by the additional learning unit 55h. In this embodiment, the AI ​​storage unit 554 stores "AI2(0)" (see FIG. 28), which is the AI ​​model 400 at the time of shipment from the factory, as the initial AI model 400a.

[0237] The trained AI models 400, such as the current AI model 400z and the initial AI model 400a, are generated models constructed by deep learning a neural network using not only image data related to a good PTP sheet 1 but also image data related to a defective PTP sheet 1 as training data. In other words, the trained AI model 400 is generated by training image data related to good and defective PTP sheets 1. In the second embodiment, the initial AI model 400a is a model in which a training process, which will be described later, has been performed in advance on its neural network, and the current AI model 400z and the duplicated current AI model 400z1 are models in which additional training processes, such as additional training processes, are performed multiple times by the additional training unit 55h on the neural network of the initial AI model 400a.

[0238] When original image data is input, the trained AI model 400 calculates the probability that the PTP sheet 1 related to the original image data is a "good product", the probability that the PTP sheet 1 is a "defective product", and the probability related to the defect type. The trained AI model 400 then compares the calculated probability of being a "good product" or a "defective product" with a threshold value for determining whether the PTP sheet 1 related to the input original image data is good or bad, and outputs the result of the determination.

[0239] In addition, the trained AI model 400 compares the calculated probability of the defect type with the classification threshold stored in the inspection information storage unit 553 to determine the type of defect in the PTP sheet 1 related to the input original image data, and outputs the determination result. The types of defects in this second embodiment are "missing tablet", "foreign matter contamination", and "other defects".

[0240] To explain the judgment in more detail, the trained AI model 400 calculates a score for "missing tablet" and a score for "foreign matter contamination", and if the calculated score exceeds the classification threshold, outputs the type of defect related to the score ("missing tablet" or "foreign matter contamination") as the judgment result. On the other hand, when the trained AI model 400 judges the PTP sheet 1 to be defective and each calculated score is equal to or less than the classification threshold, it outputs "other defects" as the judgment result.

[0241] In addition, "tablet chipping" means that chipping has occurred in the tablet 5, "foreign matter contamination" means that foreign matter other than the tablet 5, such as fragments of the tablet 5, is present in the pocket portion 2, and "other defects" means that defects other than "tablet chipping" and "foreign matter contamination" have occurred. Of course, other defect types than "tablet chipping" and "foreign matter contamination" may be adopted. Examples of other defect types include "tablet cracking" (cracks have occurred in the tablet 5), "missing tablets" (tablets 5 are not filled in the pocket portion 2), "tablet shape / size differences" (the shape and size of the tablet 5 are not compatible with the manufactured variety), and "tablet surface peeling" (the surface layer of the tablet 5 is peeled off).

[0242] Here, the learning process (initial learning process) performed when generating the initial AI model 400a will be described. Prior to this learning process, a large number of image data related to good PTP sheets 1 and image data related to defective PTP sheets 1 are prepared as learning data. In the second embodiment, the image data related to defective PTP sheets 1 include images of different types of defects. In addition, information on the correct answer of the classification result is associated with each learning data. Therefore, for example, learning data related to defective products is associated with information on the defective product and information on the type of defect.

[0243] As shown in Fig. 23, in the learning process, first, in step S601, an unlearned neural network is prepared. Next, in step S602, a classification result is obtained. That is, previously prepared learning data is provided as input data to the input layer of the neural network. Then, the classification result output from the output layer of the neural network is obtained.

[0244] In the next step S603, the output classification result is compared with the correct answer of the classification result associated with the learning data, and it is determined whether the output classification result is correct.

[0245] Here, if the output classification result is correct, in step S605, it is determined whether or not the termination condition of the learning process is satisfied. For example, if a positive determination is made a predetermined number of times in succession in step S603 without going through the processing of step S604 described later, or if learning using all of the prepared learning data is repeated a predetermined number of times, it is determined that the termination condition is satisfied. If the termination condition is satisfied, the neural network and its learning information (updated parameters, etc., described later) are stored in each of the AI-in-use storage units 554, 555 as the initial AI model 400a, and the learning process is terminated. Note that the initial AI model 400a stored in the AI-in-use storage unit 554 is updated to the current AI model 400z in accordance with the additional learning process and the additional learning correction process.

[0246] On the other hand, if the termination condition is not met in step S605, the process returns to step S602, and the neural network learning is performed again.

[0247] If the output classification result is not correct in step S603, a network update process (neural network learning) is performed in step S604, and then the process returns to step S602 to repeat the above series of processes.

[0248] In the network update process in step S604, a known learning algorithm is used to update the weights (parameters) in the neural network to more appropriate ones so that the output classification result is correct. Then, by repeating the processes in steps S602 to S604 many times, the neural network will output a more accurate classification result.

[0249] The image data storage unit 556 stores various image data such as image data (shade pattern data) obtained by the image data acquisition unit 55x described later. The image data (shade pattern data) obtained by the image data acquisition unit 55x corresponds to one PTP sheet 1 and represents a total of 10 shade patterns related to the pocket portion 2 and tablets 5. In the second embodiment, the image data storage unit 556 constitutes a "specific learning data storage means" and an "erroneous judgment occurrence image data storage means."

[0250] The image data stored in the image data storage unit 556 includes additional learning data 501 used in the additional learning process by the additional learning unit 55h. The additional learning data 501 is image data that, among multiple image data (shade pattern data) obtained by the image data acquisition unit 55x, has caused an erroneous determination in the classification result by the current AI model 400z and satisfies a predetermined condition (e.g., selected by an operator, etc.). Therefore, the additional learning data 501 includes not only image data related to a good PTP sheet 1, but also image data related to a defective PTP sheet 1. Note that the additional learning data 501 is associated with information on the correct answer of the classification result.

[0251] In this embodiment, "IM2(1)" to "IM2(n)" (see FIG. 28) are stored as the additional learning data 501. For example, "IM2(1)" is image data used in the additional learning process for the neural network of the initial AI model 400a, and the AI ​​model 400 generated by this additional learning process is "AI2(1)."

[0252] The additional learning data 501 also includes specific learning data 501a. The specific learning data 501a is the additional learning data 501 used to generate the current AI model 400z from the initial AI model 400a. Therefore, in the second embodiment, all of the additional learning data 501 corresponds to the specific learning data 501a.

[0253] Furthermore, the image data stored in the image data storage unit 556 includes erroneous judgment image data 502. The erroneous judgment image data 502 is shading pattern data (original image data) related to an erroneous judgment when an erroneous judgment of the classification result occurs in the confirmation process described later.

[0254] In the second embodiment, whether or not an erroneous determination has occurred in the classification result is determined by visual inspection by an operator, etc. Of course, whether or not an erroneous determination has occurred may be determined by an inspection device provided separately from the inspection device 23.

[0255] Then, an operator or the like inputs information using the input unit 551, and erroneous judgment occurring image data 502 is specified from the image data (shade pattern data) stored in the image data storage unit 556.

[0256] The additional learning information storage unit 557 stores information related to the learning of the AI ​​model 400, such as various information related to the additional learning process and the additional learning correction process (i.e., the learning process by the additional learning unit 55h). In the second embodiment, in addition to the information listed in the first embodiment, for example, the pass / fail judgment threshold value and the classification threshold value used by the trained AI model 400 are stored in association with the trained AI model 400. In addition, in the second embodiment, a heat map (data in which pixels that are the basis of judgment in the image data are highlighted according to importance) is stored for each trained AI model 400.

[0257] Next, the various functional units constituting the inspection control device 55 will be described. However, the control units 55a-55c, image acquisition unit 55d, and image processing unit 55e function in the same manner as the control units 52a-52c in the first embodiment, and therefore their description will be omitted. In the second embodiment, the camera 54, image acquisition unit 55d, and image processing unit 55e constitute an image data acquisition unit 55x, which acquires shading pattern data (original image data) related to the PTP sheet 1. In the second embodiment, the image data acquisition unit 55x constitutes the "image data acquisition means."

[0258] The acquired shading pattern data (original image data) corresponds to one PTP sheet 1, and represents a total of 10 areas related to the pockets 2 and tablets 5. The shading pattern data is stored in the image data storage unit 556. In the second embodiment, a part of the acquired multiple shading pattern data becomes the additional learning data 501 and the erroneous judgment occurring image data 502.

[0259] The judgment processing unit 55g inputs the shading pattern data (original image data) related to the PTP sheet 1 to be inspected to the current AI model 400z, and obtains the classification result output from the current AI model 400z. The obtained classification result is the inspection result of the PTP sheet 1, and indicates the type of the PTP sheet 1, whether it is good or bad, or whether it is defective.

[0260] The additional learning unit 55h performs additional learning processing of the neural network using the additional learning data 501, thereby sequentially generating trained AI models 400. When the additional learning command is input, the additional learning unit 55h performs the additional learning processing.

[0261] In the additional learning process, as shown in FIG. 24, in step S701, a classification result is obtained. That is, for example, "IM2(1)", which is the newly stored unlearned additional learning data 501, is provided to the input layer of the neural network of "AI2(0)", which is the duplicated current AI model 400z1 (however, the duplicated current AI model 400z1 at this stage is the initial AI model 400a) stored in the unused AI storage unit 555. Then, a classification result output from the output layer of the neural network is obtained. Note that the additional learning data 501 such as "IM2(1)" may be image data related to a good PTP sheet 1 or image data related to a defective PTP sheet 1. In addition, when the additional learning data 501 is image data related to a defective PTP sheet 1, the additional learning data 501 relates to at least one type of defect.

[0262] In the next step S702, it is determined whether the classification result is correct. If the classification result is correct, it is determined in step S704 whether a termination condition for the additional learning process is satisfied. For example, if additional learning using the stored additional learning data 501 has been repeated a predetermined number of times, it is determined that the termination condition is satisfied. If the termination condition is satisfied, the neural network and its learning information (updated parameters, etc.) are stored in the unused AI storage unit 555 as a new trained AI model 400.

[0263] On the other hand, if the classification result is incorrect in step S702, a network update process (neural network learning) is performed in step S704, and then the process returns to step S701 to repeat the above series of processes. The network update process is the same as the process in step S604 in the above learning process. By repeating the processes in steps S701 to S704, the neural network will output a more accurate classification result. When the end condition for the additional learning is met, the additional learning process is terminated.

[0264] Furthermore, the additional learning unit 55h includes, as functional units, a factor data identification unit 55i and a confirmation unit 55j for classification output AI. In the second embodiment, the factor data identification unit 55i constitutes a "factor data identification means", and the confirmation unit 55j for classification output AI constitutes a "confirmation means for classification output AI".

[0265] When it is confirmed by a confirmation process described later that an erroneous judgment has occurred in the classification result, the factor data identifying unit 55i identifies the additional learning data 501 that caused the erroneous judgment as erroneous judgment factor data from the specific learning data 501a based on the similarity to the erroneous judgment-causing image data 502 related to the erroneous judgment (causing the erroneous judgment). In the second embodiment, the factor data identifying unit 55i identifies erroneous judgment factor data from the specific learning data 501a related to the PTP sheet 1 of the same classification result as the erroneous judgment. For example, when an erroneous judgment of "foreign matter contamination" occurs in the confirmation process, the factor data identifying unit 55i identifies erroneous judgment factor data from the specific learning data 501a related to the PTP sheet 1 of "foreign matter contamination". Also, when an erroneous judgment of "good product" occurs in the confirmation process, the factor data identifying unit 55i identifies erroneous judgment factor data from the specific learning data 501a related to the PTP sheet 1 of "good product". This is because, when an erroneous classification result occurs, it is considered that image data related to the same classification result as the erroneous classification result among the data used for additional learning (specific learning data 501a) is the cause of the erroneous classification. Note that the erroneous classification factor data can be specified by the same method as in the first embodiment.

[0266] The confirmation unit 55j for classification output AI executes confirmation processing to confirm the learning state of the trained AI model 400 newly generated by additional learning processing or the like.

[0267] The confirmation process is performed as follows: First, as shown in Fig. 25, in step S801, when a new trained AI model 400 is generated by the additional learning process, it is confirmed whether the learning state of the AI ​​model 400 is appropriate.

[0268] Specifically, the original image data related to the misjudgment that triggered the generation of the newly generated trained AI model 400 is input to the newly generated trained AI model 400 to obtain a classification result.

[0269] Next, the acquired classification result is compared with the correct classification result associated with the input original image data to determine whether the classification result is correct.

[0270] If the classification result is correct (step S801: YES), in step S802, it is determined that the current AI model 400z can be replaced with the newly generated trained AI model 400, and the confirmation process is terminated. On the other hand, if the classification result is incorrect (step S801: NO), an additional learning correction process described below is executed, and the confirmation process is terminated.

[0271] In addition, in the confirmation process, additional learning data 501 other than the original image data related to the erroneous judgment that triggered the additional learning process may be input to the newly generated trained AI model 400 to determine whether the obtained classification result is correct.

[0272] Furthermore, if the confirmation process determines that the classification result is incorrect (step S801: No), that is, if an erroneous judgment occurs in the trained AI model 400 newly generated by the additional learning process, the additional learning unit 55h cooperates with the factor data identification unit 55i and the confirmation unit 55j for image output AI, etc., to perform additional learning correction processing.

[0273] In the additional learning and correction process, as shown in FIG. 26, first, in step S851, the factor data identifying unit 55i identifies one or more pieces of erroneous determination factor data.

[0274] Next, in step S852, a re-learning process is executed. In the re-learning process, the neural network of the initial AI model 400a stored in the unused AI storage unit 555 is trained with the specific learning data 501a excluding the misjudgment factor data identified in step S851, to generate a new trained AI model 400. For example, in step S851, it is assumed that "IM2(3)" is identified as the misjudgment factor data from among "IM2(0)" to "IM2(n)" which are the specific learning data 501a stored in the image data storage unit 556. In this case, in step S852, the above-mentioned additional learning process is performed on the initial AI model 400a using "IM2(0)" to "IM2(2)" and "IM2(4)" to "IM2(n)" which are the specific learning data 501a other than "IM2(3)". This generates a new trained AI model 400 that has learned specific training data 501a other than the misclassification factor data "IM2(3)".

[0275] Next, in step S853, it is determined whether the learning state of the new trained AI model 400 is appropriate using the misjudgment-causing image data 502. This process is the same as the process in step S801 in the confirmation process. If the learning state is appropriate, the new trained AI model 400 is stored in the unused AI storage unit 555 as having been corrected so that an appropriate pass / fail judgment is made on the misjudgment-causing image data 502 that caused the misjudgment, and the additional learning correction process is terminated.

[0276] On the other hand, if the learning state is inappropriate, the process returns to step S851, and the identification of the misjudgment factor data is performed again. Note that, if the learning state is determined to be inappropriate in step S853, a process of changing the threshold value used for classification may also be performed.

[0277] Then, by repeating the above process until the learning state becomes correct, a trained AI model 400 that can properly handle the erroneous judgment-causing image data 502 that caused the erroneous judgment is finally generated. The generated trained AI model 400 is stored in the unused AI storage unit 555.

[0278] Further, the additional learning unit 55h includes, as functional units, an approval process execution unit 55k, an approval input acceptance unit 55l, and a replacement unit 55m. In the second embodiment, the approval input acceptance unit 55l constitutes "approval input acceptance means".

[0279] The approval process execution unit 55k, the approval input receiving unit 55l, and the replacement unit 55m function in the same manner as the approval process execution unit 52k, the approval input receiving unit 52l, and the replacement unit 52m in the first embodiment. Therefore, when the replacement unit 55m executes the replacement process, the current AI model 400z is replaced with the newly generated trained AI model 400, and the inspection of the PTP sheet 1 is started using the new current AI model 400z.

[0280] The save process execution unit 55n basically functions in the same manner as the save process execution unit 52n in the first embodiment. Therefore, the save process execution unit 55n stores information related to the learning of the AI ​​model 400, such as various information related to the additional learning process and the additional learning correction process (i.e., the learning process by the additional learning unit 55h), in the additional learning information storage unit 557. However, the save process execution unit 55n in the second embodiment stores the pass / fail judgment threshold value and the classification threshold value in association with the trained AI models 400, together with the trained AI models 400 adopted or rejected in the approval process, in the additional learning information storage unit 557. In addition, the save process execution unit 55n stores a heat map for each trained AI model 400.

[0281] The flow of the various processes (additional learning process, etc.) performed by the additional learning unit 55h described in detail above is, in brief, as shown in FIG.

[0282] That is, if an erroneous judgment occurs in the classification result output by the inspection device 23 (current AI model 400z) (step S21: YES), an additional learning command is input by an operator or the like, and an additional learning process (step S22) is executed. That is, a new trained AI model 400 is generated by having the duplicated current AI model 400z1 stored in the unused AI storage unit 555 learn image data of a non-defective or defective product related to the erroneous judgment.

[0283] Then, in a confirmation process (step S23), the learning state of the newly generated trained AI model 400 is confirmed. If the learning state is determined to be appropriate (step S23-1: YES), then in step S23-2-A, it is determined that the current AI model 400z can be replaced with the new trained AI model 400.

[0284] On the other hand, if the learning state is determined to be inappropriate (step S23-1: NO), additional learning and correction processing is performed in step S23-2-B. The additional learning and correction processing (step S23-2-B) is then repeated until the classification result in the processing of step S23-1 becomes correct, ultimately resulting in a trained AI model 400 that can appropriately handle the erroneous judgment-causing image data 502 that caused the erroneous judgment. In the repeated additional learning and correction processing, processing is repeatedly performed to re-identify the erroneous judgment factor data and to learn the specific learning data 501a excluding the re-identified erroneous judgment factor data.

[0285] Then, when it is determined in step S23-2-A that replacement with a new trained AI model 400 is possible, an approval process (step S24) is executed, and information required by an operator or the like to make the replacement is displayed on the display unit 551. Then, when the operator or the like selects the approval button (step S25: YES), the PTP packaging machine 11 is temporarily stopped, and the current AI model 400z is replaced with the new trained AI model 400 (step S26). On the other hand, when the operator or the like selects the non-approval button (step S25: No), the replacement is not made, and the current AI model 400z is used as is.

[0286] As described above in detail, according to the second embodiment, when an erroneous determination occurs regarding the classification result of the PTP sheet 1, it is possible to generate a new trained AI model 400 by additionally training the original image data that caused the erroneous determination on the neural network of the duplicated current AI model 400z1. Therefore, it is possible to obtain a trained AI model 400 with better inspection accuracy.

[0287] In addition, in the additional learning, learning is performed not on the neural network of the current AI model 400z used in the inspection, but on the neural network of the duplicated current AI model 400z1 stored in the unused AI storage unit 555. This allows additional learning to be performed without stopping the inspection, thereby further improving the speed and efficiency of the inspection.

[0288] Furthermore, the confirmation unit 55j for the classification output AI makes it possible to easily confirm whether or not the newly generated trained AI model 400 will produce a repeat erroneous judgment based on the original image data that caused the erroneous judgment, that is, whether or not better inspection accuracy has been achieved.

[0289] In addition, only when the classification output AI confirmation unit 55j confirms that a state in which a false positive will not occur again (a state in which better inspection accuracy has been achieved) is it determined that the current AI model 400z used in the inspection can be replaced with the newly generated trained AI model 400. Therefore, it is possible to more reliably prevent an inappropriate AI model 400 that may cause a false positive from being used in the inspection.

[0290] On the other hand, when the confirmation unit 55j for classification output AI confirms that an erroneous judgment occurs in the test using the newly generated trained AI model 400, that is, when it is considered that an erroneous judgment occurs due to overlearning, the additional learning unit 55h performs learning on the neural network of the initial AI model 400a, which is the AI ​​model 400 at a stage where additional learning has not been performed, to generate a new trained AI model 400. At this time, data considered to be the cause of the erroneous judgment is identified as erroneous judgment cause data from among the data related to the PTP sheet 1 of the same classification result as the erroneous judgment result, among the specific learning data 501a, which is the additional learning data 501 used in generating the current AI model 400z from the initial AI model 400a.

[0291] The additional learning unit 55h then generates a new trained AI model 400 by having the neural network of the initial AI model 400a learn data that is obtained by excluding the misjudgment factor data from the specific training data 501a. Therefore, when the classification output AI confirmation unit 55j judges that the learning state is inappropriate, the learning state can be corrected so that the occurrence of misjudgment due to overlearning can be more reliably suppressed while making use of the past additional training data 501 (specific training data 501a). As a result, it is possible to more reliably prevent the occurrence of misjudgment regarding the classification of the PTP sheet 1, and improve the inspection accuracy.

[0292] In addition, if the classification output AI confirmation unit 55j determines that the learning state is inappropriate, learning is performed on the neural network of the initial AI model 400a stored in the unused AI storage unit 555, instead of the neural network of the current AI model 400z used in the inspection. Therefore, the process related to the learning of the neural network and the process related to the extraction (identification) of the misjudgment factor data, which may be a huge repetitive process, can be performed without stopping the inspection. This allows the inspection of a large number of PTP sheets 1 to be performed more quickly and efficiently.

[0293] Furthermore, when the confirmation unit 55j for classification output AI confirms that an erroneous judgment will occur, the erroneous judgment cause data can be identified from the specific learning data 501a based on the similarity to the erroneous judgment occurring image data 502, which is the original image data related to the erroneous judgment (the original image data that caused the erroneous judgment). Therefore, the erroneous judgment cause data can be identified more accurately, and the accuracy of the inspection using the newly generated trained AI model 400 can be more reliably improved.

[0294] Furthermore, only after approval by an operator or the like is received, the current AI model 400z is replaced with the newly generated trained AI model 400. This makes it possible to more reliably prevent the current AI model 400z from being unintentionally replaced with the newly generated trained AI model 400. This makes it possible to more reliably prevent the occurrence of malfunctions due to unintentional replacement.

[0295] The present invention is not limited to the above-described embodiment, and may be implemented as follows: Of course, other applications and modifications not exemplified below are also possible.

[0296] (a) In the above embodiment, the initial AI models 200a, 400a constitute the "old identification means." In contrast, the "old identification means" may be constituted by any of "AI(1)" to "AI(n-1)" or "AI2(1)" to "AI2(n-1)," which are the AI ​​models 200, 400 at a stage where some additional learning processes have not been performed. Therefore, for example, "AI(99)" may constitute the "old identification means" (see FIG. 29). Furthermore, in such a configuration, if the data relating to "AI(0)" to "AI(98)" and the additional learning data 301 ("IM(1)" to "IM(99)") used for the additional learning of these AI models 200 ("AI(0)" to "AI(98)") are not stored in the image data storage unit 526, when a defective PTP sheet 1 is erroneously determined to be a good product, for example, it is possible to identify data that caused the misjudgment from the additional learning data 301 after "IM(100)" and perform a re-learning process for "AI(99)" using the additional learning data 301 after "IM(100)" excluding this data that caused the misjudgment.

[0297] In addition, all of the AI ​​models 200, 400 at a stage where some or all of the additional learning processes have not been performed, such as "AI(0)" to "AI(n-1)" and "AI2(0)" to "AI2(n-1)", may constitute the "old identification means", and the unused AI storage units 525, 555 may store all of these AI models 200, 400. In this case, the re-learning process may be performed on the AI ​​model 200 immediately before additional learning of the misjudgment factor data. For example, when the misjudgment factor data is identified as "IM(3)", the re-learning process may be performed on "AI(2)", which is the AI ​​model 200 immediately before additional learning of the misjudgment factor data.

[0298] Furthermore, the "old identification means" may be appropriately changed in relation to the current AI model 200z. For example, the current AI model 200z may be used as a reference, and the AI ​​model 200 at a stage before a predetermined number of recent (e.g., 10) additional learning processes may be performed may be used as the "old identification means." In this case, the AI ​​model 200 that becomes the "old identification means" is changed as the current AI model 200z is updated.

[0299] (b) In the above embodiment, the initial AI models 200a, 400a are assumed to have undergone a learning process in advance, but the initial AI models 200a, 400a may not have undergone a learning process. In this case, the additional learning units 52h, 55h may perform the learning process described in the above embodiment on the initial AI models 200a, 400a to generate trained AI models 200, 400. Therefore, the additional learning units 52h, 55h may be capable of executing both the learning process and the additional learning process.

[0300] (c) In the above embodiment, the erroneous judgment factor data is identified based on the similarity with the defective image data 302 and the erroneous judgment occurring image data 502, but the identification method may be changed as appropriate. For example, the additional learning data 301, 501 used in the most recent additional learning process may be identified as the erroneous judgment factor data.

[0301] (d) In the above embodiment, the inspection devices 21, 23 are used to inspect the PTP sheet 1, but the objects to be inspected by the inspection devices 21, 23 are not limited to the PTP sheet 1. Therefore, for example, the inspection devices 21, 23 may be configured to inspect tablets 5.

[0302] The type and shape of the tablet are not limited to the above embodiment. For example, the tablet includes not only medicine but also tablets used for eating and drinking. In addition, the tablet includes not only plain tablets, but also sugar-coated tablets, film-coated tablets, oral disintegrating tablets, enteric tablets, gelatin-coated tablets, and various capsule tablets such as hard capsules and soft capsules.

[0303] Also, the inspection object may be a blister pack 800 other than a PTP sheet, for example, as shown in Fig. 30. The blister pack 800 has a pocket 801 for containing a predetermined content.

[0304] (e) The configuration of the AI ​​models 200, 400 as the "identification means" and the learning method thereof are not limited to the above embodiment. For example, when performing the learning process of the neural network or the acquisition process of the reconstructed image data, a configuration may be adopted in which normalization or other processing is performed on various data as necessary. In addition, the structure of the neural network 190 is not limited to that shown in FIG. 9, and may be a configuration in which, for example, a pooling layer is provided after the convolution layer 193. Of course, the number of layers of the neural network 190, the number of nodes in each layer, the connection structure of each node, etc. may be different.

[0305] Furthermore, in the above embodiment, the AI ​​model 200 (neural network 190) is a generative model having a convolutional autoencoder (CAE) structure, but is not limited to this. For example, the AI ​​model 200 may be a generative model having a different type of autoencoder structure, such as a variational autoencoder (VAE).

[0306] In the above embodiment, the neural network 190 is configured to learn using the error backpropagation method, but the present invention is not limited to this, and various other learning algorithms may be used for learning.

[0307] In addition, the neural network may be configured by a dedicated AI processing circuit such as an AI chip. In that case, only learning information such as parameters is stored in the AI ​​storage units 524, 525 (554, 555), and the dedicated AI processing circuit reads this information and sets it in the neural network to configure the AI ​​models 200, 400.

[0308] (f) In the second embodiment, the trained AI model 400 is capable of classifying the inputted original image data into a good product or a defective product and various types of defects, but it is sufficient that the classification items at least include a good product / defective product. Therefore, the trained AI model 400 may classify into a good product or a defective product, but may not classify by type of defect. [Explanation of symbols]

[0309] 1... PTP sheet (inspection object), 11... PTP packaging machine, 21, 23... inspection device, 52f... reconstructed image data acquisition unit (reconstructed image data acquisition means), 52g... judgment unit (judgment means), 52h, 55h... additional learning unit (additional learning means), 52i, 55i... factor data identification unit (factor data identification means), 52j... confirmation unit for image output AI (confirmation means for image output AI), 52x, 55x... image data acquisition unit (image data acquisition means), 55j... confirmation unit for classification output AI (confirmation means for classification output AI), 190... neural network , 191...encoder unit (encoding unit), 192...decoder unit (decoding unit), 200, 400...AI model (identification means), 200a, 400a...initial AI model (learned identification means, old identification means), 200z, 400z...current AI model (learned identification means, current identification means), 200z1, 400z1...duplicate current AI model (learned identification means, duplicate current identification means), 525, 555...unused AI model memory unit (identification means memory means), 526, 556...image data memory unit (specific learned data memory means, defective image data memory means).

Claims

1. An inspection device for inspecting a predetermined inspection object, Image data acquisition means capable of acquiring image data relating to an inspection object; a trained classification means, which is a trained classification means configured to be able to output image data corresponding to the input image data by a neural network having an encoding unit that extracts a feature amount from the input image data and a decoding unit that reconstructs the image data from the feature amount, and which is generated by training only image data related to inspection objects that are non-defective as training data for the neural network of the classification means; an additional learning means for generating a new trained classification means by additionally learning image data related to a non-defective inspection target object as additional learning data to the neural network of the classification means; a reconstructed image data acquisition means for inputting original image data obtained by the image data acquisition means to the learned classification means and acquiring reconstructed image data as reconstructed image data; a determination means for comparing the original image data with the reconstructed image data acquired by inputting the original image data to the learned discrimination means, and determining whether the object to be inspected is good or bad based on the comparison result; a classification means storage means for storing an old classification means which is the classification means at a stage where additional learning by the additional learning means has not been performed in part or in whole; a specific learning data storage means for storing all of the specific learning data, which is the additional learning data used in generating a current identification means, which is the learned identification means used for inspecting an inspection object, from the old identification means, The inspection device is characterized in that, when a defective inspection object is erroneously judged to be a good item based on a pass / fail judgment by the judgment means based on the reconstructed image data reconstructed by inputting the original image data into the current identification means by the reconstructed image data acquisition means, the additional learning means is configured to be able to generate a new learned identification means by having the neural network of the old identification means stored in the identification means memory means learn the specific learning data excluding the error judgment cause data, which is the additional learning data identified as the cause of the error judgment.

2. the identification means storage means stores, separately from the current identification means, a duplicate current identification means which is an identification means having the same contents as the current identification means, The inspection device according to claim 1, characterized in that, when a good inspection object is erroneously judged to be defective by the judgment means based on the reconstructed image data reconstructed by inputting the original image data into the current identification means by the reconstructed image data acquisition means, the additional learning means is configured to additionally learn the original image data that caused the erroneous judgment to generate a new learned identification means.

3. a defective image data storage means for storing defective image data, which is the original image data related to a misjudgment made by the judgment means when the inspection object is misjudged as a defective product as a non-defective product by the judgment means; 2. The inspection apparatus according to claim 1, further comprising a factor data specifying means for specifying the erroneous judgment factor data from the specific learning data based on a similarity to the defective image data.

4. The inspection device described in claim 1, characterized in that it is further characterized by a confirmation means for image output AI that confirms the learning state of the newly generated trained identification means by having the judgment means make a pass / fail judgment on the newly generated trained identification means based on the reconstructed image data that is reconstructed by inputting the original image data related to the erroneous judgment that triggered the generation of the trained identification means.

5. A display means capable of displaying information related to the learning process by the additional learning means and a confirmation result by the confirmation means for the image output AI; 5. The inspection device according to claim 4, further comprising an approval input receiving means capable of receiving an input relating to approval of replacing the currently used identification means with the newly generated learned identification means.

6. An inspection device for inspecting a predetermined inspection object, Image data acquisition means capable of acquiring image data relating to an inspection object; a trained classification means, which is a trained classification means generated by training a neural network of the classification means configured to classify input image data by predetermined items and output the classification result, with image data relating to inspection objects of good and defective items as training data; an additional learning means for additionally learning the neural network of the classification means with image data related to the inspection object as additional learning data, thereby generating a new trained classification means; an identification means storage means for storing an old identification means which is the identification means at a stage where additional learning by the additional learning means has not been performed in part or in whole, and a current identification means which is the learned identification means used for inspecting an inspection object, and a duplicate current identification means which is the identification means having the same contents as the current identification means; a specific learning data storage means for storing all of the specific learning data, which is the additional learning data used in generating the current identification means from the old identification means; the additional learning means is capable of generating a new trained classification means by additionally learning the original image data that caused the misjudgment to the neural network of the duplicate current classification means stored in the classification means storage means when an erroneous judgment occurs in the classification result outputted by inputting original image data, which is image data obtained by the image data acquisition means, to the current classification means; When a confirmation means for a classification output AI that confirms the learning state of the newly generated trained classification means based on the classification result output by inputting the original image data related to the erroneous judgment that triggered the generation of the trained classification means determines that the learning state of the newly generated trained classification means is appropriate, the current classification means is determined to be replaceable with the newly generated trained classification means, This inspection device is characterized in that, when an erroneous classification result occurs when the original image data is input to the newly generated trained classification means and the classification output AI determines that the learning state is inappropriate, the neural network of the old classification means stored in the classification means storage means is trained with the specific learning data excluding erroneous classification cause data, which is data related to an inspection object having the same classification result as the erroneous classification result and is data identified as the cause of the erroneous classification, thereby generating a new trained classification means.

7. an erroneous judgment occurrence image data storage means for storing erroneous judgment occurrence image data, which is the original image data related to the erroneous judgment, when an erroneous judgment occurs in the classification result output by inputting the original image data to the newly generated learned identification means and the learning state is judged to be inappropriate by the confirmation means for the classification output AI; 7. The inspection apparatus according to claim 6, further comprising a factor data specifying means for specifying the error judgment factor data from the specific learning data based on a similarity to the error judgment occurrence image data.

8. A display means capable of displaying information related to the learning process by the additional learning means and a confirmation result by the confirmation means for the classification output AI; 7. The inspection device according to claim 6, further comprising an approval input receiving means capable of receiving an input relating to approval of replacing the currently used identification means with the newly generated learned identification means.

9. 10. A PTP packaging machine equipped with the inspection device according to claim 1 or 6, which inspects tablets or PTP sheets as inspection objects.

Images