Defect occurrence prediction system, learning instruction method for interactive ai, and program

Abstract

To provide a defect occurrence prediction system, a learning instruction method for interactive AI, and a program capable of reducing the occurrence of defective products.SOLUTION: A defect occurrence prediction system 10 includes: a plurality of sensors SEN1, SEN2, ... SENn which measure the state of a manufacturing apparatus that manufactures articles; operation units SW1, SW2, ... SWn which are operated by a user who has recognized defects in the articles; and a transmission unit 50 which transmits learning data to an interactive AI constructed on the basis of a large-scale language model.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to a defect occurrence prediction system, a learning instruction method for a dialogue-type AI, and a program.

Background Art

[0002] Patent Document 1 describes a failure occurrence prediction system. This failure occurrence prediction system includes a situation acquisition unit that acquires data related to the operation status at each timing of an operation device to be monitored in association with timing information, a situation prediction unit that predicts the future operation status according to the operation status acquired by the situation acquisition unit for each of a plurality of periods having different lengths from the present, a third storage unit that stores a first correspondence relationship between the operation status and the occurrence of a predetermined defect in the operation device, and a defect prediction unit that obtains a predetermined index related to the possibility of occurrence of a defect based on the predicted operation status and the first correspondence relationship.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] An object of the present invention is to provide a defect occurrence prediction system, a learning instruction method for a dialogue-type AI, and a program that can reduce the occurrence of defective products.

Means for Solving the Problems

[0005] The invention according to claim 1 is a failure prediction system comprising a plurality of sensors for measuring the state of a manufacturing apparatus for manufacturing an article, an operation unit operated by a user who has grasped a defect occurring in the article, and a transmission unit for transmitting learning data to a dialogue-type AI configured based on a large language model, wherein the learning data includes information A which is the state of the manufacturing apparatus obtained by the plurality of sensors, information B which is the time when a defect occurred in the article, information C regarding the situation in which the defect occurred recorded after the defect occurred, and information D regarding a countermeasure method for the defect recorded after the defect occurred.

[0006] The invention according to claim 2 is the failure prediction system according to claim 1, further comprising a learning instruction unit for instructing a learning method to the dialogue-type AI, wherein the learning instruction unit divides the linguistic feature space into a first space in which tokens extracted from the information C and the information D are vectorized and mapped, and a second space in which the tokens are not mapped, and instructs the dialogue-type AI to generate an answer by regarding the tokens in the first space as more relevant tokens than the tokens in the second space.

[0007] The invention according to claim 3 is the failure prediction system according to claim 2, further comprising a first processing unit having a data processing unit that processes the measurement data output by each of the plurality of sensors and outputs the processed measurement data, and a time specifying unit that specifies the time ts when the operation unit is activated, wherein the learning instruction unit instructs the dialogue-type AI to correct a time deviation that occurred between the time tm when a defect actually occurred in the article and the time ts with respect to the processed measurement data.

[0008] The invention according to claim 4 is the defective occurrence prediction system according to claim 2, further comprising a first processing unit having a data processing unit that processes the measurement data output by each of the plurality of sensors and outputs the processed measurement data, and a time specifying unit that specifies the time ts when the operation unit operates, and further comprising a time correction unit for correcting the time shift that occurred between the time tm when a defect actually occurred in the article and the time ts for the processed measurement data.

[0009] The invention according to claim 5 is the defective occurrence prediction system according to claim 3 or 4, further comprising a second processing unit having a conversion processing unit that converts the defective occurrence time tr when a defect occurred in the article, the information C, and the information D, which were recorded after the defect occurred and were recorded as audio data or video data, into text data.

[0010] The invention according to claim 6 is the defective occurrence prediction system according to claim 5, wherein the second processing unit compares the time ts and the defective occurrence time tr, and outputs an alarm when the time shift that occurred between the time ts and the defective occurrence time tr is greater than a predetermined value.

[0011] The invention according to claim 7 is a program for causing the second processing unit included in the defective occurrence prediction system according to claim 5 to function as a conversion means for converting the information B, information C, and information D recorded as audio data or video data into text data.

[0012] The invention according to claim 8 is a method for instructing the learning of a dialogue-type AI configured based on a large language model, which divides a language feature space into a first space in which tokens extracted from learning data are vectorized and a second space in which the tokens are not vectorized, and instructs the dialogue-type AI to generate an answer by using the tokens mapped to the first space as more relevant tokens than the tokens mapped to the second space.

Advantages of the Invention

[0013] According to the present invention, it is possible to provide a defect occurrence prediction system, an interactive AI learning instruction method, and a program that can reduce the occurrence of defective products.

Brief Description of the Drawings

[0014]

Figure 1A

Figure 1B

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Embodiments for Carrying Out the Invention

[0015] Subsequently, with reference to the attached drawings, embodiments embodying the present invention will be described to facilitate understanding of the present invention. In the figures, parts not relevant to the description may be omitted from illustration.

[0016] As shown in FIGS. 1A and 1B, the defect occurrence prediction system 10 according to an embodiment of the present invention can mainly provide two functions by using the interactive AI 20 provided as a cloud service.

[0017] First, based on the state of the manufacturing apparatus 12, the defect occurrence prediction system 10 can present necessary countermeasures to the user before a defect occurs in a machine part (an example of an article) manufactured by the manufacturing apparatus 12. Second, the defect occurrence prediction system 10 can return an appropriate response based on learning data to an inquiry from the user.

[0018] Here, the interactive AI 20 is, for example, ChatGPT provided by OpenAI (「CHATGPT」 is an international registered trademark) and is configured based on large language models. The interactive AI 20 includes not only the interactive AI itself but also means for functioning the interactive AI.

[0019] The manufacturing apparatus 12 is, for example, a manufacturing apparatus for machine parts such as a die-casting machine, an injection molding machine, a blow molding machine, a hydraulic press, and a machining center. The manufacturing apparatus 12 may be a food manufacturing apparatus that manufactures food, and more generally, any manufacturing apparatus for an article may be used.

[0020] As shown in FIG. 2, the defect occurrence prediction system 10 includes a plurality of sensors SEN1, SEN2,... SENn, a plurality of defect occurrence specifying switches SW1, SW2,... SWn, a first processing unit 30, a second processing unit 40, a crawler unit 50, and a learning instruction unit 60. Note that the functions of at least the first processing unit 30, the second processing unit 40, the crawler unit 50, and the learning instruction unit 60 are realized by a computer program.

[0021] A plurality of sensors SEN1, SEN2, ··· SENn are respectively attached to the manufacturing apparatus 12 in plurality, and can measure the state of the manufacturing apparatus 12. Each of the sensors SEN1, SEN2, ··· SENn is, for example, a pressure gauge, a thermometer, a vibration meter, etc., and outputs measurement data.

[0022] Each defect occurrence specifying switch (an example of an operation unit) SW1, SW2, ··· SWn is, for example, a push button switch. Each of the defect occurrence specifying switches SW1, SW2, ··· SWn is operated by a user who has grasped that a defect has occurred in the manufactured machine part, and can generate, for example, a pulse-shaped time specifying signal. Each of the defect occurrence specifying switches SW1, SW2, ··· SWn is provided according to the type of defect. Specifically, there are provided defect occurrence specifying switches that are pressed when a machine part is scratched, defect occurrence specifying switches that are pressed when chipping occurs, and the like.

[0023] The first processing unit 30 is, for example, a terminal used by a user, and receives the measurement data output by each of the sensors SEN1, SEN2, ··· SENn and the time specifying signals output by each of the defect occurrence specifying switches SW1, SW2, ··· SWn. Note that the measurement data output by each of the sensors SEN1, SEN2, ··· SENn is input to the first processing unit 30 at a predetermined period Tc1. This period Tc1 is, for example, 60 seconds.

[0024] As shown in FIG. 3, the first processing unit 30 has a data processing unit 302 and a time specifying unit 304. The data processing unit 302 preprocesses the measurement data output by each of the sensors SEN1, SEN2, ··· SENn. This preprocessing is, for example, data cleansing including noise removal and the like. The preprocessed measurement data is output as processed measurement data (an example of information A representing the state of the manufacturing apparatus obtained by a plurality of sensors), and is stored in the first storage unit 32 (see FIG. 2).

[0025] The time specifying unit 304 specifies the time ts (an example of the information B indicating the time when a defect occurred in the article) at which the defect occurrence specifying switches SW1, SW2, ··· SWn operated, based on the time specifying signals output from the defect occurrence specifying switches SW1, SW2, ··· SWn. This time ts is also the time when an operator recognized that a defect had occurred in the manufactured machine part. The specified time ts is stored in the first storage unit 32 and transmitted to the second processing unit 40.

[0026] Here, a defect in a machine part occurs due to a malfunction of the manufacturing apparatus 12, and the time ts is the time when an operator recognized that a defect had occurred in the manufactured machine part. Therefore, as shown in FIG. 4, the time ts may be shifted by a time TL1 (delayed by the time TL1) with respect to the time tm (an example of the information B) when a malfunction of the manufacturing apparatus 12 occurred and a defect actually occurred in the machine part. For example, if the manufacturing apparatus 12 is a die-casting machine, there may be a shift of the time TL1 between the time tm when a defect actually occurred in the machine part manufactured by the die-casting machine and the time ts when the defect was recognized.

[0027] The second processing unit 40 is, for example, a terminal used by a user, and can convert defect countermeasure information recorded as audio data or video data into text data. Here, generally, as shown in FIG. 4, after a defect occurs in a machine part, a report (an example of a record) for recording a countermeasure method for the occurred defect is created by a user. The aforementioned defect countermeasure information is the information recorded in this report. The report may be recorded on an electronic medium not only as text data but also as audio data or video data.

[0028] The defect countermeasure information recorded in the report includes at least the following information. (1) Defect occurrence time tr (an example of the information B) The defective occurrence time tr is the time when an operator recognizes that a defect has occurred in the manufactured machine part, and it is the time described in the report. Therefore, the defective occurrence time tr may have a deviation of time TL2 due to factors such as the misunderstanding of the report creator with respect to the time ts when the defective occurrence specific switches SW1, SW2, ··· SWn are activated. (2) Defective occurrence situation (an example of information C regarding the situation where a defect has occurred, recorded after the defect has occurred) The defective occurrence situation is information regarding the situation where a defect has occurred, and includes at least information about the type of defect. (3) Defective countermeasure method (an example of information D regarding the method of dealing with a defect, recorded after the defect has occurred) The defective countermeasure method is information regarding the method of dealing with the occurred defect.

[0029] As shown in FIG. 5, the second processing unit 40 has a conversion processing unit 402 and an alarm output unit 404. The conversion processing unit 402 can convert the defective countermeasure information recorded in the report into text data. The defective countermeasure information converted into text data is stored in the second storage unit 42 (see FIG. 2).

[0030] The alarm output unit 404 (see FIG. 5) outputs an alarm when the result of comparing the defective occurrence time tr included in the above-mentioned defective countermeasure information with the time ts when the defective occurrence specific switches SW1, SW2, ··· SWn are activated shows that the deviation of time TL2 is larger than a predetermined value. After the alarm is output, if the deviation of time TL2 generated between the defective occurrence time tr and the time ts is larger than a predetermined value, the defective occurrence time tr is corrected to the time ts. Details will be described later.

[0031] The crawler unit (an example of a transmission unit) 50 (see FIG. 2) is, for example, a terminal used by a user, can communicate with the first processing unit 30 and the second processing unit 40 respectively, and can acquire the information stored in the first storage unit 32 and the second storage unit 42 respectively. The crawler unit 50 transmits at least the following information to the interactive AI 20 as learning data for fine-tuning after performing necessary processing. (1) Processed measurement data stored in the first storage unit 32 (an example of information A) (2) Defect countermeasure information stored in the second storage unit 42 (an example of information B, information C, and information D)

[0032] The learning instruction unit 60 is, for example, a terminal used by a user, and can instruct the interactive AI 20 about a learning method via an API (Application Programming Interface) provided by the interactive AI 20. As shown in FIG. 6, the API is executed via the cloud computing service 22. Specifically, the learning instruction unit 60 divides the linguistic feature space into a first space in which tokens extracted from the defect countermeasure information are vectorized and mapped, and a second space in which the tokens are not mapped, and can instruct the interactive AI 20 to generate an answer with tokens in the first space as more relevant tokens than the tokens in the second space.

[0033] In addition, the learning instruction unit 60 can instruct the interactive AI 20 to correct a deviation TL1 of the time (see FIG. 4) that occurred between the time tm when a defect actually occurred in the machine part and the time ts for the processed measurement data. However, instead of instructing the learning instruction unit 60 to correct the deviation of the time TL1, the defect occurrence prediction system 10 may include a time correction unit (not shown) for correcting the deviation of the time TL1 that occurred between the time tm when a defect actually occurred in the machine part and the time ts for the processed measurement data.

[0034] Thus, according to the defect occurrence prediction system 10, the processed measurement data indicating the state of the manufacturing apparatus 12 and the defect countermeasure information are transmitted to the dialogue-type AI 20 as learning data. As a result, the dialogue-type AI 20 can learn 1) the relationship between the state of the manufacturing apparatus 12 and the occurrence timing of defects for each type of flaw or chip, etc., and 2) the defect occurrence situation and the defect countermeasure method.

[0035] Next, regarding the operation (defect occurrence prediction method) of the defect occurrence prediction system 10, it will be described separately for the learning period and the execution period of the dialogue-type AI 20.

[0036] During the learning period of the dialogue-type AI 20, each sensor SEN1, SEN2, ··· SENn constantly measures the state of the manufacturing apparatus 12, and the first processing unit 30 stores the processed measurement data in the first storage unit 32.

[0037] Also, when a defect is recognized in the manufactured machine parts, as shown in FIG. 7, the first processing unit 30 stores the time ts for each type of defect in the first storage unit 32, and the user creates a report as text data, voice data, or video data. When a predetermined amount of reports are created by repeating this, based on the user's operation, the defect countermeasure information included in the report created as voice data or video data is converted into text data by the second processing unit 40 and stored in the second storage unit 42.

[0038] Here, the conversion process by the conversion processing unit 402 included in the second processing unit 40 is executed according to the following steps, as shown in FIG. 8.

[0039] (Step SA1) The report is read by the user.

[0040] (Step SA2) The conversion processing unit 402 determines whether the input data is voice data. If the input data is voice data, step SA3 is executed. If the input data is not voice data, step SA4 is executed.

[0041] (Step SA3) The conversion processing unit 402 converts the voice data into text data. After that, step SA6 is executed.

[0042] (Step SA4) The conversion processing unit 402 determines whether the input data is video data. If the input data is video data, step SA5 is executed. If the input data is not video data, step SA6 is executed.

[0043] (Step SA5) The conversion processing unit 402 separates the voice data included in the video data.

[0044] (Step SA6) The conversion processing unit 402 compares the defect occurrence time tr included in the text data converted in step SA3 with the times ts when the defect occurrence identification switches SW1, SW2, ··· SWn are activated.

[0045] (Step SA7) The conversion processing unit 402 determines whether the compared defect occurrence time tr and the times ts when the defect occurrence identification switches SW1, SW2, ··· SWn are activated substantially match. If the deviation of the time TL2 generated between the defect occurrence time tr and the time ts is equal to or less than a predetermined value, step SA8 is executed assuming that the defect occurrence time tr and the time ts substantially match. If it is greater than the predetermined value, step SA9 is executed.

[0046] (Step SA8) The defect handling information is stored in the second storage unit 42 as text data.

[0047] (Step SA9) The alarm output unit 404 (see FIG. 5) outputs an alarm indicating that the defect occurrence time tr and the time ts do not substantially match.

[0048] (Step SA10) The user corrects the defect occurrence time tr (an example of information B) included in the text data to the time ts. Thereafter, step SA6 is executed again.

[0049] By repeating the above steps, the defect handling information converted into text data is accumulated in the second storage unit 42. Thereafter, as described above, the crawler unit 50 transmits the following information to the dialogue type AI 20 as learning data. (1) Processed measurement data (an example of information A) stored in the first storage unit 32 (2) Defect handling information (an example of information B, information C, and information D) stored in the second storage unit 42

[0050] When the dialogue type AI 20 receives the learning data from the crawler unit 50, as shown in FIG. 9, loop processing for performing deep learning is executed. Note that steps SB1, SB2 and steps SC1 to SC3 are executed in parallel.

[0051] (Step SB1) The dialogue type AI 20 receives the processed measurement data stored in the first storage unit 32.

[0052] (Step SB2) The received processed measurement data is corrected for the deviation of the time TL1 that occurred between the time tm when a defect actually occurred in the machine part and the time ts based on the instruction of the learning instruction unit 60, and then recorded in a storage unit (not shown).

[0053] However, as described above, when the defect prediction system 10 includes a time correction unit (not shown) for correcting the deviation of the time TL1 that occurred between the time tm when a defect actually occurred in the machine part and the time ts for the processed measurement data, since the processed measurement data for which the deviation of the time TL1 has already been corrected by this time correction unit is used, the processed measurement data with the corrected time is directly recorded in the storage unit (not shown).

[0054] (Step SC1) The dialogue-type AI 20 receives defect countermeasure information. Incidentally, since the above-described step SA10 is executed, the defect occurrence time tr included in the defect countermeasure information substantially coincides with the time ts when the defect occurrence specifying switches SW1, SW2, ··· SWn are activated.

[0055] (Step SC2) Based on the instruction of the learning instruction unit 60, the above-described first space is created in the language feature space. That is, the language feature space is divided into a first space and a second space.

[0056] When the dialogue-type AI 20 receives all the learning data from the crawler unit 50 and the loop process is completed, the dialogue-type AI 20 completes the learning. Thereafter, as shown in FIG. 7, the accuracy of the dialogue-type AI 20 predicting the occurrence of a defect is verified.

[0057] In this way, by repeating a series of procedures of defect occurrence, countermeasure, report creation, text data conversion of defect countermeasure information, learning of the dialogue-type AI 20 (transmission of learning data by the crawler unit 50), and accuracy verification, the dialogue-type AI 20 learns 1) the relationship between the state of the manufacturing apparatus 12 and the timing of occurrence of defects for each type of scratch or chip, etc., and 2) the defect occurrence situation and the defect countermeasure method, and the fine-tuning of the dialogue-type AI 20 during the learning period is completed.

[0058] During the execution period of the dialogue-type AI 20, as shown in FIG. 10, the prediction of the occurrence of a defect is executed at any time. The crawler unit 50 transmits the processed measurement data to the interactive AI 20 at a predetermined cycle Tc2. This cycle Tc2 is, for example, 60 seconds, and it is preferably equal to or shorter than the length of the cycle Tc1 at which the measurement data output by each sensor SEN1, SEN2, ··· SENn is acquired by the first processing unit 30.

[0059] On the other hand, a prompt for inquiring about the possibility of occurrence of defects at a predetermined cycle Tc3 is input, and the interactive AI 20 returns a response based on the learned information to the user terminal 80 used by the user or the user's smartphone 82. This cycle Tc3 is, for example, 60 seconds, and it is preferably equal to or shorter than the length of the cycle Tc1.

[0060] Also, when a prompt for inquiring about, for example, a method for preventing the occurrence of defects is input from the user terminal 80 or the smartphone 82, the interactive AI 20 returns a response based on the learned defect countermeasure information.

[0061] As described above, since the interactive AI 20 generates an answer by regarding the token in the first space as a more relevant token than the token in the second space, the possibility of returning an answer different from the fact is reduced. Note that the user terminal 80 and the smartphone 82 are preferably connected to the interactive AI 20 via a VPN (Virtual Private Network) connection.

[0062] Thus, according to the defect occurrence prediction system 10 according to the present embodiment, the possibility of occurrence of defects in the machine parts manufactured by the manufacturing apparatus 12 is reduced.

[0063] As described above, the embodiments of the present invention have been described. However, the present invention is not limited to the above-described embodiments, and all changes and the like that do not deviate from the gist are within the scope of application of the present invention. In the foregoing embodiments, the time tm when a defect actually occurred in the machine part, the time ts when the defect occurrence specifying switches SW1, SW2, ··· SWn actuated, and the defect occurrence time tr are all examples of the information B.

Explanation of Signs

[0064] 10 Defect occurrence prediction system 12 Manufacturing apparatus 20 Interactive AI 30 First processing unit 32 First storage unit 40 Second processing unit 42 Second storage unit 50 Crawler unit 60 Learning instruction unit 80 User terminal 82 Smartphone 302 Data processing unit 304 Time specifying unit 402 Conversion processing unit 404 Alarm output unit SEN1, SEN2, ··· SENn Sensors SW1, SW2, ··· SWn Defect occurrence specifying switches

Claims

1. A plurality of sensors for measuring the state of a manufacturing apparatus for manufacturing an article; An operation unit operated by a user who has grasped a defect occurring in the article; A transmission unit that transmits learning data to a dialogue-type AI configured based on a large language model, and comprising: A defect occurrence prediction system in which the learning data includes information A on the state of the manufacturing apparatus obtained by the plurality of sensors, information B on the time when a defect occurred in the article, information C on the situation in which the defect occurred recorded after the defect occurred, and information D on the countermeasure method for the defect recorded after the defect occurred.

2. In the defect occurrence prediction system according to Claim 1, further comprising a learning instruction unit that instructs a learning method for the dialogue-type AI, wherein the learning instruction unit divides a linguistic feature space into a first space in which tokens extracted from the information C and the information D are vectorized and mapped, and a second space in which the tokens are not mapped, and instructs the dialogue-type AI to generate an answer with tokens in the first space being more relevant tokens than tokens in the second space. A defect occurrence prediction system.

3. In the defect occurrence prediction system according to Claim 2, further comprising a first processing unit having a data processing unit that processes measurement data output from each of the plurality of sensors and outputs the processed measurement data, and a time specifying unit that specifies a time ts when the operation unit operates, wherein the learning instruction unit instructs the dialogue-type AI to correct a time deviation that occurred between a time tm when a defect actually occurred in the article and the time ts with respect to the processed measurement data. A defect occurrence prediction system.

4. In the defect occurrence prediction system according to Claim 2, further comprising a first processing unit having a data processing unit that processes measurement data output from each of the plurality of sensors and outputs the processed measurement data, and a time specifying unit that specifies a time ts when the operation unit operates, A defect occurrence prediction system further comprising a time correction unit for correcting a time deviation that occurred between a time tm when a defect actually occurred in the article and the time ts with respect to the processed measurement data.

5. In the defect occurrence prediction system according to Claim 3 or 4, A failure prediction system further comprising a second processing unit having a conversion processing unit that converts the failure occurrence time tr, the information C, and the information D of the article recorded after a failure occurred and recorded as audio data or video data into text data.

6. In the failure prediction system according to claim 5, the second processing unit compares the time ts and the failure occurrence time tr, and outputs an alarm when the time deviation that occurred between the time ts and the failure occurrence time tr is greater than a predetermined value. A failure prediction system having an alarm output unit.

7. A second processing unit included in the failure prediction system according to claim 5, A program for causing the conversion means for converting the information B, information C, and information D recorded as audio data or video data into text data to function.

8. For a dialogue AI configured based on a large language model, divide the language feature space into a first space in which tokens extracted from the learning data are vectorized and a second space in which the tokens are not vectorized, A method for instructing learning of a dialogue AI that instructs the dialogue AI to generate an answer by using tokens mapped to the first space as more relevant tokens than tokens mapped to the second space.

Images