JP7700720B2 - Learning system, defect prediction system, trained model, learning method, and defect prediction method - Google Patents

Description

The present disclosure relates to a learning system, a defect prediction system, a trained model, a learning method, and a defect prediction method.

Patent document 1 discloses a technology that estimates the cause of defects that occur during the manufacturing process from knowledge data expressed in multiple hierarchical structures.

JP 2007-257285 A

However, while the technology described in Patent Document 1 can estimate the cause of defects such as malfunctions that are discovered, it does not take measures to ensure that product defects are not overlooked during the inspection process, and it is inevitable that abnormalities will be overlooked.

In order to avoid overlooking product defects during the inspection process, it is desirable to predict the occurrence and nature of product defects before they are discovered during the inspection process. However, since normal manufacturing conditions are set so that product defects do not occur in the first place, it is difficult to predict the occurrence of product defects from normal manufacturing conditions.

In view of the above-mentioned problems, the objective of the present disclosure is to provide a learning system, a defect prediction system, a trained model, a learning method, and a defect prediction method capable of predicting product defects that may occur on a product production line.

The learning system disclosed herein inputs training data into a learning model, the training data including change point information indicating a change point related to at least one of a worker, production equipment, and work content in a production line that produces a product, and defect information indicating a product defect that has occurred in the production line with the change point, and performs machine learning on the learning model, thereby generating a trained model that inputs the change point information and outputs defect prediction information indicating a product defect that is predicted to occur in response to the change point indicated by the change point information.

In this way, the learning system can perform machine learning on training data including the change point information and the defect information, and generate a trained model that outputs product defects that are predicted to occur in response to a change point. Therefore, the learning system can obtain a trained model that is capable of predicting product defects that may occur in a product production line when a change point occurs. Furthermore, it is thought that product defects often occur when a change point occurs. Therefore, the learning system can obtain a trained model that is capable of predicting product defects that may occur in a production line.

The defect information included in the training data may include information indicating defects discovered after passing through an inspection process included in the production line. This makes it possible for the generated trained model to predict defects that were not discovered during the inspection process.

The defect information included in the training data may include information indicating a product defect that occurred during a period determined according to the content of the change point indicated by the change point information. This allows machine learning to be performed on information indicating a defect that occurred during a period according to the content of the change point in association with the change point, thereby generating a trained model with higher prediction accuracy.

The production line may also be a line that includes assembly work for assembling products. It is expected that change points will frequently occur in production lines that include assembly work, making it easier to execute machine learning that associates change points with defects. Thus, with this configuration, it is possible to obtain a trained model that can accurately predict possible product defects for production lines that include assembly work, where change points frequently occur.

The defect prediction system disclosed herein inputs change point information indicating a change point related to at least one of a worker, production equipment, and work content in a production line that produces a product, and defect information indicating a product defect that has occurred in the production line along with the change point into a trained model trained by machine learning using training data including the change point information, and outputs defect prediction information indicating a product defect that is predicted to occur in response to the change point indicated by the change point information.

In this way, the defect prediction system can use a trained model that has been machine-learned using training data including the change point information and the defect information to output product defects that are predicted to occur in response to a change point. Therefore, the defect prediction system can predict product defects that may occur in a product production line when a change point occurs. Product defects are thought to occur frequently when a change point occurs. Therefore, the defect prediction system can predict product defects that may occur in a production line.

The defect information included in the training data may also include information indicating defects discovered after passing through an inspection process included in the production line. This makes it possible to predict defects that were not discovered during the inspection process.

The defect information included in the training data may include information indicating a product defect that occurred during a period determined according to the content of the change point indicated by the change point information. This allows for accurate prediction of defects using a trained model with higher predictive accuracy that is machine-learned to associate information indicating a defect that occurred during a period according to the content of the change point with the change point.

The defect prediction information may also be notified in a process after the change point indicated by the change point information input to the trained model. This allows the defect prediction results according to the change point to be notified at an appropriate time.

In addition, when the defect prediction information includes information indicating multiple defects, the information to be notified may be determined according to the importance of each of the multiple defects, and the determined information may be notified. In this way, even when multiple defects are predicted, necessary defect prediction information can be notified.

The contents of the inspection to be performed in response to the defect prediction information may be notified in a process after the change point indicated by the change point information input to the trained model. This allows the contents of the inspection to be performed in response to the defect predicted according to the change point to be notified at an appropriate time.

In addition, when there are multiple types of inspections to be performed corresponding to the defect prediction information, the information to be notified may be determined according to the importance of the defect corresponding to each of the multiple types of inspections, and the determined information may be notified. In this way, even when there are multiple types of inspections to be performed for a predicted defect, it is possible to notify the content of the necessary inspection.

In addition, the contents of the inspection to be performed in response to the defect prediction information may be notified, and the trained model may be updated based on the results of the inspection of the notified contents. This can improve the accuracy of the trained model.

The production line may also be a line that includes assembly work for assembling products. It is expected that change points will frequently occur in production lines that include assembly work, making it easier to execute machine learning that associates change points with defects. Thus, with this configuration, it is possible to accurately predict possible product defects for production lines that include assembly work, where change points frequently occur.

The trained model disclosed herein is a machine-learned model that uses training data including change point information indicating change points related to at least one of workers, production equipment, and work content in a production line that produces products, and defect information indicating product defects that have occurred in the production line due to the change points, to input the change point information and output defect prediction information indicating product defects that are predicted to occur in response to the change points indicated by the change point information.

In this way, the trained model is machine-trained on training data including the change point information and the defect information, and can be a model that outputs product defects that are predicted to occur in response to a change point. Therefore, the trained model can predict product defects that may occur in a product production line when a change point occurs. Furthermore, it is thought that product defects often occur when a change point occurs. Therefore, the trained model can predict product defects that may occur in a production line.

The learning method disclosed herein involves inputting training data into a learning model, the training data including change point information indicating a change point relating to at least one of a worker, production equipment, and work content in a production line that produces a product, and defect information indicating a product defect that has occurred in the production line with the change point, and training the learning model through machine learning to generate a trained model that inputs the change point information and outputs defect prediction information indicating a product defect that is predicted to occur in response to the change point indicated by the change point information.

The learning method is expected to produce the same results as the learning system.

The defect prediction method disclosed herein involves inputting change point information indicating a change point relating to at least one of a worker, production equipment, and work content in a production line that produces a product, and defect information indicating a product defect that has occurred in the production line along with the change point, into a trained model trained by machine learning using training data including the change point information, and outputting defect prediction information indicating a product defect that is predicted to occur in response to the change point indicated by the change point information.

The defect prediction method is expected to produce the same effects as the defect prediction system.

The present disclosure provides a learning system, a defect prediction system, a trained model, a learning method, and a defect prediction method that are capable of predicting product defects that may occur on a product production line.

1 is a block diagram showing an example of a configuration of a defect prediction system according to an embodiment; FIG. 1 is a schematic diagram showing an example of a production line. FIG. 1 is a schematic diagram showing an example of classification of product defects. FIG. 2 is a diagram illustrating an example of explanatory variables and target variables of a trained model used in the defect prediction system of FIG. 1 . 2 is a flow chart for explaining an example of a learning process which is a learning stage process in the defect prediction system of FIG. 1; 2 is a flow diagram for explaining an example of a defect prediction process which is a process in an operation stage in the defect prediction system of FIG. 1. 2 is a flow diagram for explaining an example of an overall processing procedure in an operation stage in the defect prediction system of FIG. 1. FIG. 8 is a flow diagram for explaining an example of a defect handling procedure after distribution of a product distributed according to the processing procedure of FIG. 7. A block diagram showing an example configuration of a learning system that generates a trained model used in a defect prediction system according to another embodiment. FIG. 13 is a block diagram showing a configuration example of a defect prediction system according to another embodiment.

The following describes the embodiments in detail with reference to the drawings, but the claimed invention is not limited to the following embodiments. Furthermore, not all of the configurations described in the embodiments are necessarily essential as means for solving the problems. In each drawing, the same elements are given the same reference numerals, and duplicate explanations will be omitted as necessary to clarify the explanation.

(Embodiment)
First, a configuration example of a defect prediction system according to the present embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing a configuration example of a defect prediction system according to the present embodiment. Fig. 2 is a schematic diagram showing an example of a production line, and Fig. 3 is a schematic diagram showing an example of classification of product defects. Fig. 4 is a diagram showing an example of explanatory variables and target variables of a trained model used in the defect prediction system of Fig. 1.

The defect prediction system 100 shown in FIG. 1 is a system that uses a trained model 15 to predict product defects on a production line that produces the products. Below, an example configuration will be described in which the defect prediction system 100 includes a learning system that trains the untrained model 14, but it is possible to predict product defects even without a learning system as long as the trained model 15 is provided. Note that an example in which the learning system and the defect prediction system are configured separately will be described later as another embodiment.

In the following, the production line to be predicted will be described using the example production line 40 shown in Figure 2, but is not limited to this. The example production line 40 has change points 41 to 44 and 46 to 48, which are locations that can be change points, as described below. Of course, there needs to be one or more locations that can be change points. The example production line 40 also has inspection points 45 and 49, which are locations that can be inspection points. Inspection point 49 is the final inspection point, and it is sufficient that there is at least one inspection point on the production line 40.

Furthermore, the type of product and the production process (manufacturing process) are not important. In the following explanation, we will assume that the product is a vehicle, but it may be other types of products such as electronic devices, food, or furniture. Furthermore, a product defect can refer to a defect in the product, a manufacturing defect, or any other problem that could be considered a defect in terms of distributing the product on the market.

As illustrated in FIG. 3, defects that may occur in a vehicle may include fastening-related defects related to fastening and design-related defects related to design. Fastening-related defects may include slanted insertion, where parts are installed at an angle, misaligned holes between parts, insufficient fastening torque, excessive baking during baking paint, and not fastening, where parts are not fastened properly. Design-related defects may include scratches, stains, wrinkles, deformation, loose parts, etc.

As shown in FIG. 1, the defect prediction system 100 according to this embodiment can include an information processing device 10 that predicts product defects, a vehicle production record database (DB) 20, and multiple terminals 31 to 36. The information processing device 10, the vehicle production record DB 20, and the terminals 31 to 36 can be communicatively connected via a network N.

Here, the network N is a communication line network such as the Internet, an intranet, a mobile phone network, or a LAN (Local Area Network). Note that, although an example will be described in which the defect prediction system 100 has six terminals 31 to 36, it is sufficient to provide as many terminals as are capable of inputting change points and defects, as described below, and viewing the prediction results. Therefore, one or more terminals are sufficient.

The information processing device 10 is a device that is the main component of the defect prediction system 100. The information processing device 10 can include a control unit 11 that controls the entire device, a communication unit 12, and a memory unit 13 in which an unlearned model 14 or a learned model 15 is stored.

The information processing device 10 functions as a learning device in the learning stage where the unlearned model 14 is machine-learned, and can function as a defect prediction device in the operation stage where predictions are made using the trained model 15. Note that the learning device and the defect prediction device can be configured independently, and such configuration examples will be described later with reference to Figures 9 and 10.

The control unit 11 can be realized by an integrated circuit, for example a processor such as a CPU (Central Processing Unit), MPU (Micro Processor Unit), or GPU (Graphics Processing Unit), a working memory, and a non-volatile storage device. The control program executed by the processor is stored in this storage device, and the processor reads the program into the working memory and executes it, thereby fulfilling the functions of the control unit 11. This storage device can also be part of the memory unit 13.

As can also be seen from the description of the control unit 11, the information processing device 10 can be configured as a general-purpose computer. When functioning as a learning device, the information processing device 10 can be configured to store the unlearned model 14 in the general-purpose computer, input teacher data to the unlearned model 14, perform machine learning, and generate the trained model 15. When functioning as a defect prediction device, the information processing device 10 can be configured to store the trained model 15 in the general-purpose computer, input information to the trained model 15, and output information indicating the prediction result.

The control unit 11 can also be equipped with a first control unit with high processing power used during learning, and a second control unit used during operation. The second control unit can have a lower processing power than the first control unit.

The communication unit 12 can be equipped with a communication interface that communicates with the vehicle production record DB 20 and terminals 31 to 36 via the network N.

The memory unit 13 is a storage device in which the unlearned model 14 or the trained model 15 is stored. In the learning stage, the unlearned model 14 is stored in the memory unit 13, and the trained model 15 can be generated by inputting teacher data into the unlearned model 14 and performing machine learning, and the trained model 15 can be stored in the memory unit 13. The algorithm of the unlearned model 14, i.e., the algorithm of the trained model 15, can be of any type as long as it performs supervised learning, and hyperparameters, etc., can also be used. The teacher data used for machine learning and the input data and output data during operation will be described later.

Although not shown, the vehicle production record DB 20 can be realized by combining one or more server computers and storage devices connected to the network N, but may also be incorporated into the information processing device 10, for example. The vehicle production record DB 20 reads and writes each piece of data in response to access via the network N, and returns the processing results to the access source.

The vehicle production record DB 20 is a database that manages, in association with each other, product information (i.e., vehicle information) 21, which is information about vehicles produced on the vehicle production line 40, change point information 22 about change points that have occurred, and defect information 23, which indicates defects discovered by product inspection on the production line 40. In the vehicle production record DB 20, change point information 22 and defect information 23 are associated with the vehicle indicated by the product information 21, but since there are cases in which no change points or defects occur during production for some vehicles, there exists product information 21 with which at least one of the change point information 22 and defect information 23 is not associated. Furthermore, the defect information 23 can include not only information about products that have been found to be defective by product inspection on the production line 40, but also information indicating defects discovered after distribution to the market.

The change point information 22 is information indicating a change point regarding at least one of the workers, production equipment, and work content in the production line 40 that produces the product. Examples of changes in workers include a new employee working, or a temporary replacement of a worker while a regular worker is on vacation. Thus, a change point in a worker can refer to a change in a worker, or a change in the rank of a worker performing a task when workers are ranked according to their proficiency. A change point in production equipment refers to a change in one piece of equipment in the production line 40, and for example, replacing a tool from an old one with a new one is also treated as a change in production equipment. A change point in work content can refer to a change in work procedures or work methods for the purpose of improving work efficiency. Another example of a change point in work content is a scene where a mistake occurs in a standard task and the mistake is corrected.

In fact, defects that occur due to these changes include, for example, parts not being assembled correctly when an inexperienced worker takes over for a worker on vacation. Other defects include when a scratch or stain made by an operator is repaired, but another part becomes dirty while repairing. Other defects include when a new employee does not notice that a process has changed and continues working on the same part, resulting in some parts not being repaired.

The terminals 31 to 36 can be information processing devices such as computers, smartphones, or dedicated processing terminals, and can access the vehicle production record DB 20 and the information processing device 10 via the network N.

Users U1, U2, U3, U4, and U5 are responsible for inputting changes in personnel, changes in work, changes in equipment and tools, defects that have been discovered, and defects that have been distributed in the market, for the production line 40 that produces vehicles.

During the learning stage, users U1, U2, U3, U4, and U5 input the corresponding information using terminals 31, 32, 33, 34, and 35, respectively, and each of terminals 31 to 35 can transmit the input information to vehicle production record DB 20 via network N as information for generating training data for learning.

In addition, in the operation phase, users U1, U2, and U3 input corresponding information using terminals 31, 32, and 33, respectively, and each of terminals 31 to 33 can transmit the input information to information processing device 10 as input information for performing defect prediction in the operation phase. Also, in the operation phase, simultaneously with such transmission or as necessary, each of terminals 31 to 33 can transmit the input information to vehicle production record DB 20 via network N as information for updating teacher data for learning.

User U6 is a person who views the prediction results during the operation phase. During the operation phase, terminal 36 can receive notification of the prediction results from information processing device 10, or access information processing device 10 to view the prediction results.

First, we will explain the case where the information processing device 10 functions as a learning device. In this case, the defect prediction system 100 functions as a learning system.

The control unit 11 inputs training data including change point information 22 and defect information 23 that indicates a product defect that occurred on the production line 40 along with a change point indicated by the change point information 22 into the unlearned model 14.

The teacher data can be input by acquiring information from the vehicle performance DB 20 via the communication unit 12. Among the defect information 23, the defect information indicating a product defect that occurred on the production line 40 accompanied by a change point indicated by the change point information 22 can be the defect information 23 associated with the change point information 22 in the vehicle performance DB 20, and will be referred to as defect information 24 below for distinction. The teacher data can also include product information 21.

In other words, the teacher data can include product information 21 and change point information 22 as explanatory variables 201a, and defect information 24 as objective variables 201b, as in the teacher data 201 illustrated in FIG. 4.

However, as described above, product defects that occur on the production line 40 can also include defects in vehicles that are discovered after distribution to the market. In other words, the defect information 24 included in the training data can include information indicating defects that are discovered after passing through the inspection process included in the production line 40. This inspection process includes a finished product inspection process at inspection point 49, where the finished product is inspected, but can also include an intermediate inspection process at inspection point 45, where the product is inspected during production. This makes it possible for the generated trained model 15 to predict defects that could not be discovered in the inspection process used in the production line 40, such as defects discovered after distribution to the market.

The defect information 24 included in the teacher data may include information indicating product defects that occurred during a period determined according to the content of the change point indicated by the change point information 22. In this way, the period during which product defects occurred to be adopted as the defect information 24 of the teacher data may be determined according to the content of the change point, that is, the period during which defects occurred to be adopted as teacher data may be changed according to the content of the change point. Determining the period to be adopted as the defect information 24 of the teacher data means determining the period of the defect information 24 used for learning, which ultimately means determining the period of the defect information to be predicted (defect prediction information). In fact, depending on the content of the change point, there are cases where it is sufficient to learn product defects at the time when the change point occurred, and cases where it is necessary to learn product defects for a predetermined time after the occurrence of the change point, so such a configuration can be said to be beneficial.

This allows information indicating defects that occurred during a period corresponding to the content of the change point to be machine-learned in association with the change point, and a trained model 15 with higher predictive accuracy can be generated. Then, by using such a trained model 15 to predict product defects, the prediction accuracy can be improved.

The production line 40 may also be a line that includes assembly work for assembling products. Assembly work can refer to the work of assembling parts. It is expected that change points will often occur on a production line that includes assembly work, making it easier to execute machine learning that associates change points with defects. Thus, with this configuration, it is possible to obtain a trained model 15 that can accurately predict possible product defects for a production line 40 that includes assembly work that often generates change points. Then, by predicting product defects using such a trained model 15, it is possible to improve prediction accuracy.

The teacher data may also include at least one of change point information 22 with no associated defect information 23 and defect information 23 with no associated change point information 22. This is because it is possible to learn, for example, that the change point indicated by the change point information 22 does not affect the occurrence of a defect, or that the defect information 23 indicates a defect that occurs independently of the change point.

The control unit 11 then performs machine learning on the unlearned model 14 to which the above-mentioned teacher data has been input, to generate a trained model 15 that, during operation, inputs change point information and outputs defect prediction information indicating product defects that are predicted to occur for the change point indicated by the change point information. The defect prediction information output from the trained model 15 is information equivalent to the defect information 24, and can be information equivalent to the content contained in the defect information 24.

In this way, the information processing device 10 functioning as a learning device can perform machine learning of training data including change point information 22 and defect information 24, and generate a trained model 15 that outputs product defects that are predicted to occur in response to a change point. Therefore, the information processing device 10 can obtain a trained model 15 capable of predicting product defects that may occur in a product production line when a change point occurs. Furthermore, it is considered that product defects occur frequently when a change point occurs. Therefore, it can be said that the information processing device 10 can obtain a trained model 15 capable of predicting product defects that may occur in a production line 40.

In other words, the trained model 15 obtained in this manner is machine-learned using training data including change point information 22 and defect information 24, and can be a model that outputs product defects that are predicted to occur in response to change points. Therefore, the trained model 15 can predict product defects that may occur in the product production line 40 when a change point occurs. Furthermore, it is thought that product defects occur frequently when a change point occurs. Therefore, the trained model 15 can predict product defects that may occur in the production line 40.

Next, a case where the information processing device 10 functions as a defect prediction device will be described.
The control unit 11 inputs change point information to the trained model 15 generated as described above, and outputs defect prediction information indicating product defects predicted to occur for the change point indicated by the change point information. The defect prediction information output from the trained model 15 as described above is information equivalent to the defect information 24, and can be information equivalent to the contents included in the defect information 24. In this way, the control unit 11 can function as a prediction unit that performs defect prediction.

In this way, the trained model 15 can be a model for causing a computer to input change point information 22 and output defect prediction information indicating product defects that are predicted to occur for the change points indicated by the change point information 22.

The control unit 11 can also notify the defect prediction information via the communication unit 12 at a process subsequent to the change point indicated by the change point information 22 input to the trained model 15. The notification destination can be determined as a predetermined notification destination such as the terminal 36. This allows the defect prediction result according to the change point to be notified at an appropriate time.

In addition, when the defect prediction information includes information indicating multiple defects, the control unit 11 may determine the information to be notified according to the importance of each of the multiple defects via the communication unit 12, and notify the determined information. The notification destination here can also be determined as a predetermined notification destination such as the terminal 36. This makes it possible to notify only the necessary defect prediction information, even when multiple defects are predicted. In this case, too, the timing of notification can be a process after the change point.

In addition, when multiple change points are input, the control unit 11 may notify the defect prediction information in a process after the latest change point, or may notify the defect prediction information for at least the corresponding change point in a process after that change point. The notification destination here can also be determined as a predetermined notification destination such as the terminal 36. The control unit 11 may also change the notification destination depending on the importance.

The control unit 11 can also notify the content of the inspection to be performed in response to the defect prediction information at a process subsequent to the change point indicated by the change point information 22 input to the trained model 15. The notification destination here can also be determined as a predetermined notification destination such as the terminal 36. This allows the content of the inspection to be performed in response to the defect predicted according to the change point to be notified at an appropriate time, so that the inspector can receive the notification and perform the inspection.

In addition, when there are multiple types of inspections to be performed corresponding to the defect prediction information, the control unit 11 may determine the information to be notified (the inspection contents) according to the importance of the defect corresponding to each of the multiple inspection contents, and notify the determined information. The notification destination here may also be determined as a predetermined notification destination such as the terminal 36. In this way, even when there are multiple types of inspections to be performed for the predicted defect, it is possible to notify the necessary inspection contents. Even in this case, notification in a process after the change point as described above may be adopted.

In addition, in a configuration in which the control unit 11 notifies the contents of the inspection to be performed in response to the defect prediction information, the control unit 11 may update (re-learn) the trained model 15 based on the results of the inspection of the notified contents. The results of the inspection can be input from the terminal 36 or the like and registered in the vehicle performance DB 20. This can improve the accuracy of the trained model 15.

In this way, the defect prediction system 100 can output product defects that are predicted to occur in response to a change point by using a trained model 15 that has been machine-learned from training data including the change point information 22 and the defect information 24. Therefore, the defect prediction system 100 can exclude defects that are not caused by change points from the data used for prediction, and can predict product defects that may occur in the product production line when a change point occurs (defects that are likely to occur). It is considered that product defects often occur when a change point occurs. Therefore, the defect prediction system 100 can predict product defects that may occur in the production line 40. This can also prevent product defects from being overlooked in an inspection process before they are discovered in the inspection process, for example.

Next, an example of the learning process in the defect prediction system 100 will be described with reference to FIG. 5. FIG. 5 is a flow diagram for explaining an example of the learning stage process (learning process) executed by the information processing device 10 in the defect prediction system 100. Below, an example will be described in which the production line 40 is an assembly line made up of assembly processes, but the present invention is not limited to this.

First, users U1 to U3 output change point information 22 indicating change points in the assembly process from terminals 31 to 33 to vehicle production record DB 20 for recording, and user U4 outputs defect information 24 indicating a defect that has occurred from terminal 34 to vehicle production record DB 20 for recording (step S1). Recording of change points can also be performed, for example, by an assembly worker acting as one of users U1 to U3 when a vehicle is conveyed. Also, recording of defects can also be performed, for example, by an assembly worker acting as user U4 when a vehicle is conveyed.

In addition, step S1 may also include a process in which, after the product is distributed in the market, user U5 outputs defect information 24 indicating defects that occurred in the market from terminal 35 to vehicle production record DB 20 and records the information.

Next, the information processing device 10 refers to the vehicle production record DB 20 and determines whether the vehicle information 21 contains defect information 24 (step S2). If the vehicle information 21 contains defect information 24 (YES in step S2), the information processing device 10 extracts the defect information 24 and presents it on the terminal 36, and classifies the type of defect, the stage at which the defect was discovered by the user U6, and the severity (importance) of the defect from the presented information (step S3).

In step S3, the defects can be classified by type and labeled into each category. For this classification, the classification method illustrated in FIG. 3 can be used. Classifying the defects by type as in step S3 has the effect of, first, improving the accuracy of machine learning to be performed later.

However, the classification method is not limited to using the classification system illustrated in FIG. 3, but the defects that occur can be classified according to the operation in which they occurred, the type of defect, or even by the detailed phenomenon of the defect. By classifying the defects by type, it becomes possible to reflect the tendency of the defects in the trained model 15 and to instruct additional inspection methods for predicted defects at a later time.

In addition, by performing the classification in step S3, the second effect is that the result can be reflected in the ranking of the priority for performing predictive inspections. In other words, the priority of predicted defects that should be found in additional inspections can be determined depending on the stage at which the defect is found and the severity of the defect. As described above, there are two major types of defects: defects found during inspection on the production line 40 (i.e., in the factory) and defects that have leaked into the market. In particular, defects that have leaked into the market cause inconvenience to vehicle users, so they can be said to be defects of particular importance to vehicle manufacturers. Furthermore, among defects that have leaked into the market, the shorter the time between the completion of the vehicle and the discovery, the higher the importance in maintaining the reliability of the vehicle. In contrast, defects found in the factory are often found by basic, predetermined inspections, and can be said to have a lower priority among defects.

Next, the severity of the defect will be explained. The importance of the defect can be broadly categorized into four levels, for example, R1, R2, R3, and R4. Here, R1 can be an accident or fire, or something necessary for safety that has the potential to cause such an accident or fire. R2 can be (1) a major malfunction of the vehicle's functionality that does not fall under R1, or (2) a malfunction that conflicts with safety standards and relevant laws and regulations related to the safety of the country or region, or something that has the potential to cause such an accident or fire, or something that does not fall under R1. R3 can be a malfunction of functionality that has the potential to cause such an accident, and R4 can be a problem that does not fall under any of R1 to R3.

Following step S3, the user U6 uses the terminal 36 to register (record) the information indicating the classified defect as defect information 23 in the vehicle production record DB 20 (step S4). Note that the operations of steps S3 and S4 can be performed as appropriate using a user interface on the terminal 36 that supports the operations.

After processing in step S4 and if the answer is NO in step S2, the information processing device 10 refers to the vehicle production record DB 20 and links the product involved in the change point (the vehicle in this example) with the defect in that vehicle (step S5). That is, in step S5, the change point information 22 and defect information 24 recorded in the vehicle production record DB 20 are associated. This linking (association) can be performed automatically based on the product information 21. Note that no linking will be performed for information for which the answer is NO in step S2.

In step S5, by linking the vehicle involved in the change point with the defect in that vehicle, the causal relationship between them can be learned, and the defect can be predicted. The linking can be performed based on the product information 21. The change point was set as described above, but it is preferable that the linking in step S5 is performed in a beneficial manner, that is, so that a highly accurate prediction can be made.

For example, a defect where a certain part is assembled incorrectly may be caused by the part being held in a different way than usual when cleaning the part that was discovered during assembly, resulting in the part being installed incorrectly. Therefore, in order to address this, the change point can be the fact that the part was cleaned, which is not usually done. Also, a defect where a certain task was not performed may be caused by a change in equipment, and an unskilled worker whose task had changed worked in a hurry, disrupting the procedures and skipping a task. Therefore, in order to address this, the change point can be the fact that a worker (especially a worker with a lower level of skill) or a change in equipment was made.

After step S5, the information processing device 10 inputs the linked change point information 22 and defect information 24 as training data into the unlearned model 14, executes machine learning (step S6), saves the generated trained model 15 (step S7), and ends the process.

Next, an example of the defect prediction process in the defect prediction system 100 will be described with reference to FIG. 6. FIG. 6 is a flow diagram for explaining an example of the defect prediction process, which is the operation stage process in the defect prediction system 100.

First, users U1, U2, and U3 use terminals 31, 32, and 33, respectively, to input known change points before or during production on the production line 40 and transmit them to the information processing device 10, which then inputs information indicating the change points (information equivalent to change point information 22) into the trained model 15 (step S11). The change points to be input can be change points that have occurred but have not been input and change points that are known to occur.

Next, the information processing device 10 lists defect prediction information indicating defects that may occur from the change points indicated by the output results from the trained model 15 (step S12), and ranks the list of possible defects based on the importance of the defects and assigns priorities (step S13).

Next, the information processing device 10 selects the necessary number of additional testing methods (contents indicating additional testing) capable of detecting each type of predicted defect from among a number of pre-stored methods (step S14).

In step S14, by selecting an additional inspection method that can find each predicted defect type, an additional inspection corresponding to the type of defect can be output. For example, if the defect is related to fastening, additional inspection methods can be selected such as retightening the target bolt or visually checking whether the bolt is fully fastened and seated. In addition, if the defect is related to design, additional inspection methods can be selected such as actually touching it with your hand, and if it is dirt, visually checking while changing the angle of the light. In addition, during additional inspection, not only the inspection procedure but also the parts and areas of the vehicle where changes have occurred can be clearly displayed on a monitor, etc., allowing the inspector to handle additional inspections for each vehicle.

Following step S14, the information processing device 10 calculates the time required for the additional inspection when the selected additional inspection method is adopted (step S15). The information processing device 10 then selects the predicted defects and additional inspections in order of priority within the specified inspection time, transmits information indicating the contents of the selected additional inspections and the predicted defects to the terminal 36 and presents it to the user U6 (step S16), and ends the process. The user U6 then instructs the worker to inspect, obtains the inspection results, and if a defect occurs, registers them in the vehicle production record DB 20 from the terminal 36 as information for updating the trained model 15 so that they can be used the next time the model is updated.

In this embodiment, as described in step S16, predicted defects and additional inspections are selected in order of priority within the specified inspection time and presented to the worker; that is, the time required for additional inspections is calculated in advance, and only additional inspections that can be performed within the specified inspection time are performed. In this way, additional inspections for predicted defects can be performed during assembly line work, and the burden on workers on the production line 40 can be prevented from increasing.

Next, an example of the overall processing procedure in the operation phase will be described with reference to Figures 7 and 8. Figure 7 is a flow diagram for explaining an example of the overall processing procedure in the operation phase of the defect prediction system 100, and Figure 8 is a flow diagram for explaining an example of a defect response processing procedure after distribution of a product distributed according to the processing procedure in Figure 7.

When assembling a vehicle in an assembly process, the information processing device 10 first obtains information about the assembly work (including information about any change points, if any) from a vehicle assembly management system (not shown) and determines whether or not there are any change points based on that information (step S21).

If the answer is NO in step S21, the assembly process is carried out and the inspection process included therein performs normal, predetermined inspections (step S22), and as described below, at least if a defect occurs, user U4 inputs the inspection results from terminal 34.

After step S22, the terminal 34 refers to the input information and determines whether or not a defect has occurred (step S23). If no defect has occurred (NO in step S23), the information processing device 10 performs processing to bring the product (vehicle) off the line (step S24) and ends the process. Step S24 allows the target vehicle to be distributed in the market. If a defect has occurred (YES in step S23), the user U4 uses the terminal 34 to register the discovered defect in the vehicle production record DB 20 (step S28) and notifies the worker to fix the discovered defect. Upon receiving this notification, the worker makes the fix (step S29). After processing step S29, the process proceeds to step S24 as well.

On the other hand, if the answer is YES in step S21, one of the users U1 to U3 uses the terminal 31 to 33 to input the change point that occurs and transmit it to the information processing device 10, and the information processing device 10 temporarily records the change point information indicating the change point (step S25). At this time, the information processing device 10 or the terminal 31 to 33 can also register the change point information 22 in the vehicle production record DB 20.

After step S25, the information processing device 10 inputs the recorded change point information into the trained model 15, and outputs defect prediction information indicating defects that may occur from the change points that have occurred in real time, thereby predicting defects that may occur due to the change points that have occurred, and presents the corresponding additional inspections to the terminal 36 (step S26). For the processing of step S26, for example, the processing described with reference to FIG. 6 can be used.

Next, the normal predetermined test described in step S22 and the additional test for the predicted defect presented in step S26 are performed, and at least if a defect occurs, the user U4 inputs the test result from the terminal 34 (step S27). Following step S27, the processing of step S23 and the subsequent steps are executed.

After the target vehicle is distributed on the market in step S24, user U5 collects information on defects (defective products that have been released onto the market) obtained from vehicle users or vehicle sales, etc., on terminal 356. User U5 checks whether there have been any defective products released for each vehicle (step S31), and ends the process for vehicles that have not had any defective products released. In this case, the vehicle user can continue to use the vehicle.

If step S31 is YES, that is, for a vehicle that has had an outflow defect, user U5 registers (records) defect information 23 in vehicle production record DB 20 (step S32), notifies the worker to fix the defect that was found, and the worker makes the fix (step S33). The process of step S32 may be performed by obtaining defect information from the seller and automatically registering it. Also, the worker in step S33 may be a repair shop. Also, by referring to vehicle information 21 about the target vehicle, the defect information 23 registered in step S32 can be linked to change point information 24 if a change point exists.

As described above, the defect prediction system 100 can predict product defects that may occur on the production line 40. This can also prevent product defects from being overlooked in the inspection process, for example, before they are discovered in the inspection process.

For example, in existing methods for identifying the cause of defects, when a defect occurs, the cause is identified by referring to the record of change points, and measures are then taken to address the cause. In contrast, in this embodiment, possible defects are predicted from the learning results based on change points input before or during work, making it possible to prevent defects from leaking out.

Furthermore, many of the defects that occur during product production processes are caused by changes in people, work, tools, or equipment in the process, but even if it is empirically known that a certain defect is likely to occur when a certain change point occurs, there is no known inspection method that corresponds to the change point, so the defect ends up on the market. This problem is particularly noticeable when there are many change points in the process, and is also evident in products such as vehicles, where many people share the work of assembling a single product, because unpredictable defects occur when the work is done by machines.

In contrast, in this embodiment, by using machine learning to learn change points in past processes and the defects that occurred as a result of those change points, it is possible to predict what type of defects will occur by inputting change points occurring in the current process. Then, by instructing workers to carry out an inspection in addition to the general inspection for each defect predicted in this way, it becomes possible to discover defects that could end up on the market.

Other Embodiments
Next, a configuration example in which the learning device and the defect prediction device are configured independently will be described with reference to Fig. 9 and Fig. 10, but various application examples of the above-mentioned embodiment can be applied to this embodiment as well. Fig. 9 is a block diagram showing an example of a configuration of a learning system that generates a trained model used in the defect prediction system according to this embodiment, and Fig. 10 is a block diagram showing an example of a configuration of the defect prediction system according to this embodiment.

As shown in FIG. 9, the learning system 50 can include a control unit 51, an input unit 52, and a memory unit 53. The learning system 50 can be constructed, for example, using a computer such as a PC for AI (Artificial Intelligence) learning. However, the learning system 50 can be configured as a single device or with functions distributed across multiple devices.

The control unit 51 controls the entire learning system 50. The control unit 51 can be realized, for example, by an integrated circuit, and can be realized, for example, by a processor such as a CPU, MPU, or GPU, a working memory, and a non-volatile storage device. The control program executed by the processor is stored in this storage device, and the processor reads the program into the working memory and executes it, thereby fulfilling the functions of the control unit 51. The control program can include a learning program that executes learning. Note that the storage device can also utilize the memory unit 53.

The input unit 52 acquires and inputs a data set of teacher data 54 required for learning, for example from the vehicle production record DB 20 in FIG. 1, and stores it in the memory unit 53 so that it can be referenced during learning. The input unit 52 can be configured with a communication unit. The memory unit 53 can store this teacher data 54, and can store a learning model 55 as an unlearned model.

The processing by the learning system 50 is as described with reference to FIG. 1 etc., and the control unit 51 performs machine learning on the learning model 55 as an unlearned model based on the teacher data 54, and the learning model 55 becomes a trained model. Furthermore, if re-learning is required, the learning model 55 can be re-trained based on a newly prepared data set.

As shown in FIG. 10, the defect prediction device 60 can include a control unit 61, a communication unit 62, and a memory unit 63 that stores a trained model 65 trained by the learning system 50. The defect prediction device 60 can be constructed as an information processing device such as a computer, as described in FIG. 1. However, the defect prediction device 60 may be configured as a single device or may have its functions distributed across multiple devices.

The control unit 61 controls the entire defect prediction device 60. The control unit 61 can be realized, for example, by an integrated circuit, and can be realized, for example, by a processor such as a CPU, MPU, or GPU, a working memory, and a non-volatile storage device. A control program executed by the processor is stored in this storage device, and the processor reads the program into the working memory and executes it, thereby fulfilling the functions of the control unit 61. This control program can include a prediction program that inputs information including change point information that serves as an input parameter to the trained model 65, and obtains defect prediction information that is an output result. Note that the storage device can also utilize the memory unit 63.

The communication unit 62 is an example of an input unit that receives and inputs change point information from one or more of the terminals 31 to 35, and is also an example of an output unit that outputs defect prediction information (and information indicating the contents of additional testing) by transmitting it to the terminal 36.

The processing by the defect prediction device 60 is as described with reference to FIG. 1 etc., and the control unit 61 receives information including change point information via the communication unit 62, inputs it to the trained model 65, obtains defect prediction information from the trained model 65, and presents it to the terminal 36. The control unit 61 can also present information indicating the additional inspection contents corresponding to the defect indicated by the defect prediction information to the terminal 36 together with the defect prediction information.

(Alternative examples, etc.)
In the above-mentioned embodiments, various change points in the process that have a large influence as a cause of defects are used as explanatory variables, but the explanatory variables are not limited to this. As explanatory variables, other types of information can be applied, such as the workability of each production process (difficulty of work, working posture, etc.), characteristics of the worker (age, experience, personality, etc.), and the time period during which the work is performed (before consecutive holidays when workers are likely to be careless, time periods when they become tired, etc.). Workability, or workability and time period, can be expressed, for example, by an index related to ergonomics. In this way, by increasing the amount of information used as explanatory variables, it becomes possible to more accurately predict defects.

In addition, in each of the above embodiments, the product is assumed to be a vehicle, but other products may also be used, and the systems and devices described are not limited to the configurations shown as examples, as long as they perform their functions.

Furthermore, the systems and devices described in each of the above embodiments can be constructed as a system in which the functions are distributed among multiple devices, or as a single device. These multiple devices or single devices can be configured with general-purpose or dedicated circuitry, or as a device equipped with a processor, memory, and interface.

This processor can be a CPU, MPU, GPU, FPGA (Field-Programmable Gate Array), quantum processor (quantum computer control chip), etc. This interface can include an interface with an operating device for accepting user operations, an interface with a display device for displaying information, a communication interface for sending and receiving information, etc. The functions of the above-mentioned multiple devices or a single device are realized by the processor reading a control program stored in memory and executing it while exchanging the necessary information via the interface.

The program includes instructions (or software code) that, when loaded into a computer, cause the computer to perform one or more functions described in the embodiments. The program may be stored on a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable media or tangible storage media include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technology, CD-ROM, digital versatile disc (DVD), Blu-ray (registered trademark) disk or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The program may be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, a transitory computer-readable medium or communication medium includes electrical, optical, acoustic, or other form of propagated signal.

The present invention is not limited to the above embodiment, and can be modified as appropriate without departing from the spirit and scope of the invention.

U1 to U6 users N network 10 information processing device 11 control unit 12 communication unit 13 storage unit 14 unlearned model 15 learned model 20 vehicle production record DB
21 Product information 22 Change point information 23, 24 Defect information 31 to 36 Terminal 40 Production line 41 to 44, 46 to 48 Change point 45, 49 Inspection point 50 Learning system 51 Control unit 52 Input unit 53 Memory unit 54 Teacher data 55 Learning model 60 Defect prediction device 61 Control unit 62 Communication unit 63 Memory unit 65 Trained model 100 Defect prediction system 201 Teacher data 201a Explanatory variable 201b Objective variable

Claims (14)

A learning model is provided with training data including change point information indicating a change point in a production line that produces a product, the change point information indicating a change point in a production line that produces a product, the change point information indicating a defect in a product that has occurred in the production line along with the change point, and the defect information indicating a defect in a product that has occurred in the production line along with the change point.
By subjecting the learning model to machine learning, a trained model is generated that inputs the change point information and outputs defect prediction information indicating a product defect that is predicted to occur for the change point indicated by the change point information;
The defect information included in the teacher data includes information indicating a defect of a product that occurred during a period determined according to the content of the change point indicated by the change point information,
The teacher data includes the defect information to which the change point information is not associated .
Learning system. The defect information to be included in the teacher data includes information indicating a defect found after passing through an inspection process included in the production line.
The learning system according to claim 1 . The production line is a line including an assembly operation for assembling a product.
A learning system according to claim 1 or 2. inputting the change point information into a trained model trained by machine learning using training data including change point information indicating change points in a production line that produces products, related to a worker, related to a worker and production equipment, related to a worker and work content, or related to a worker, production equipment, and work content, and defect information indicating a defect in a product that has occurred in the production line along with the change point;
outputting defect prediction information indicating product defects predicted to occur for the change point indicated by the change point information;
The defect information included in the teacher data includes information indicating a defect of a product that occurred during a period determined according to the content of the change point indicated by the change point information,
The teacher data includes the defect information to which the change point information is not associated .
Defect prediction system. The defect information to be included in the teacher data includes information indicating a defect found after passing through an inspection process included in the production line.
The failure prediction system according to claim 4 . The defect prediction information is notified at a process subsequent to the change point indicated by the change point information inputted to the trained model.
The failure prediction system according to claim 4 or 5 . When the defect prediction information includes information indicating a plurality of defects, information to be notified is determined according to the importance of each of the plurality of defects, and the determined information is notified.
The failure prediction system according to claim 4 or 5 . Notifying the contents of the inspection to be performed in response to the defect prediction information at a process subsequent to the change point indicated by the change point information inputted to the trained model.
The failure prediction system according to claim 4 or 5 . When there are a plurality of contents of the inspection to be performed corresponding to the defect prediction information, information to be notified is determined according to the importance of the defect corresponding to each of the contents of the plurality of inspections, and the determined information is notified.
The failure prediction system according to claim 4 or 5 . notifying the contents of the inspection to be performed in response to the defect prediction information;
Updating the trained model based on the results of inspecting the notified content.
The failure prediction system according to claim 4 or 5 . The production line is a line including an assembly operation for assembling a product.
The failure prediction system according to claim 4 or 5 . A trained model that has been machine-learned using training data including change point information indicating a change point in a production line that produces a product, related to a worker, related to a worker and production equipment, related to a worker and work content, or related to a worker, production equipment, and work content, and defect information indicating a product defect that has occurred in the production line with the change point, is inputted, and defect prediction information indicating a product defect that is predicted to occur in response to the change point indicated by the change point information is outputted,
a trained model for causing a computer to function so as to input the change point information and output the defect prediction information ;
The defect information included in the teacher data includes information indicating a defect of a product that occurred during a period determined according to the content of the change point indicated by the change point information,
The teacher data includes the defect information to which the change point information is not associated.
Trained model . The computer
A learning model is provided with training data including change point information indicating a change point in a production line that produces a product, the change point information indicating a change point in a production line that produces a product, the change point information indicating a defect in a product that has occurred in the production line along with the change point, and the defect information indicating a defect in a product that has occurred in the production line along with the change point.
By subjecting the learning model to machine learning, a trained model is generated that inputs the change point information and outputs defect prediction information indicating a product defect that is predicted to occur for the change point indicated by the change point information;
The defect information included in the teacher data includes information indicating a defect of a product that occurred during a period determined according to the content of the change point indicated by the change point information,
The teacher data includes the defect information to which the change point information is not associated .
How to learn. The computer
inputting the change point information into a trained model trained by machine learning using training data including change point information indicating change points in a production line that produces products, related to a worker, related to a worker and production equipment, related to a worker and work content, or related to a worker, production equipment, and work content, and defect information indicating a defect in a product that has occurred in the production line along with the change point;
outputting defect prediction information indicating product defects predicted to occur for the change point indicated by the change point information;
The defect information included in the teacher data includes information indicating a defect of a product that occurred during a period determined according to the content of the change point indicated by the change point information,
The teacher data includes the defect information to which the change point information is not associated .
Defect prediction methods.

Images