Model learning device and method for generating model learning data

Abstract

To train a scheduling model capable of outputting highly productive production plans at low cost. [Solution] The model learning device 1 includes an intermediate situation generation unit 12 that generates an intermediate situation located between two selected initial situations in the feature space; a simulation unit 13 that associates each intermediate situation with a dispatching rule suitable for that intermediate situation by performing a simulation of the production line for each combination of the generated intermediate situation and dispatching rule; and a model learning unit 14 that learns a scheduling model using the initial situation corresponding to the dispatching rule and the intermediate situation corresponding to the dispatching rule as learning data.

Description

Technical Field

[0001] The present invention relates to a model learning device and a method for generating model learning data.

Background Art

[0002] When formulating a production plan for executing a plurality of manufacturing processes, an efficient production plan is important for improving the productivity of the manufacturing site. There are multiple options for the manufacturing process to be processed (dispatched) next from the current state, and the productivity may vary significantly depending on which manufacturing process is selected. In addition, if the on-site workers select which manufacturing process to select based on experience and intuition, it takes time to make a decision, and there are also variations in the quality of the decision-making.

[0003] Therefore, Patent Document 1 describes a scheduling device having the following features that can determine the priority rules (dispatching rules, hereinafter sometimes abbreviated as "rules") of the work events to be started next in a short time. · A learning data creation unit that creates learning data by associating a plurality of work event data, a plurality of priority rules applied to each of the plurality of work event data, and a plurality of estimated work times obtained by applying the plurality of priority rules to each of the plurality of work event data. · A learning unit that learns a scheduling model based on a plurality of feature amounts extracted from the plurality of work event data of the learning data, the plurality of priority rules, and an evaluation value obtained from the plurality of estimated work times. · A feature amount extraction unit that extracts feature amounts from the work event data scheduled to start. · A priority rule setting unit that sets an optimal priority rule by inputting the feature amount of the work event data scheduled to start that has been extracted into the scheduling model. A scheduling execution unit that schedules the scheduled work event data according to the optimal priority rules set above. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-39829 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] In the training process of scheduling models, the quantity and quality of the training data used are crucial because they directly impact the model's accuracy. For example, in a production line, it is necessary to recalculate (reschedule) the optimal scheduling for each state, not only in the initial state where no manufacturing processes have started, but also in the work-in-progress state where some manufacturing processes have already begun. To do this, it is necessary to prepare training data for the work-in-progress state in advance in the scheduling model so that it can output the optimal dispatching rules for the work-in-progress state.

[0006] To incorporate the working state, including the in-progress state, into the model's training requires a massive amount of training data that covers an enormous state space. Therefore, instead of trying to cover the entire state space, we will consider prioritizing the preparation of training data for important states that significantly affect the accuracy of the scheduling model. However, conventional scheduling model learning techniques, such as those described in Patent Document 1, do not describe or suggest any approach to improving the accuracy of the model being learned while narrowing down the learning data required for that learning to only the most important data.

[0007] This invention was made in view of these circumstances, and its main objective is to train a scheduling model capable of outputting highly productive production plans at low cost. [Means for solving the problem]

[0008] To solve the above problems, the model learning device of the present invention has the following features. The present invention relates to a model learning device that learns a scheduling model for which dispatching rules are output for determining the next work to be put into work based on the input production status, which indicates the work-in-progress state of each workpiece flowing through a production line. An initial status acquisition unit acquires multiple initial statuses, which are production conditions to which a suitable dispatching rule has been assigned in advance. An intermediate situation generation unit selects two initial situations from the set of initial situations acquired by the initial situation acquisition unit, in which the dispatching rules are different from each other, and generates an intermediate situation located between the two selected initial situations in a feature space formed from the feature quantities extracted from the workpiece of the production situation. A simulation unit that associates each of the intermediate conditions with the dispatching rule suitable for that intermediate condition by performing a simulation of the production line for each combination of the generated intermediate conditions and dispatching rule, The system is characterized by having a model learning unit that learns the scheduling model using the initial state corresponding to the dispatching rule and the intermediate state corresponding to the dispatching rule as learning data. Other features will be described later. [Effects of the Invention]

[0009] According to the present invention, a scheduling model capable of outputting highly productive production plans can be trained at low cost. [Brief explanation of the drawing]

[0010] [Figure 1] It is an explanatory diagram of a simple production line related to this embodiment. [Figure 2] It is an explanatory diagram of a complex production line related to this embodiment. [Figure 3] It is a Gantt chart showing a production plan before changing the dispatching rules related to this embodiment. [Figure 4] It is a Gantt chart showing a production plan with the dispatching rules changed from the production plan of FIG. 3 related to this embodiment. [Figure 5] It is a configuration diagram of a model learning device related to this embodiment. [Figure 6] It is a table showing item information and resource information related to this embodiment. [Figure 7] It is an explanatory diagram showing process information related to this embodiment. [Figure 8] It is a table showing order information related to this embodiment. [Figure 9] It is a table showing work-in-progress information related to this embodiment. [Figure 10] It is an explanatory diagram showing work-in-progress information of each work according to the production status related to this embodiment. [Figure 11] It is a table showing rule information and feature quantity information related to this embodiment. [Figure 12] It is a table showing learning data information related to this embodiment. [Figure 13] It is a flowchart showing the processing of a model learning device related to this embodiment. [Figure 14] It is a hardware configuration diagram of a model learning device related to this embodiment. [Figure 15] It is a graph showing a feature space in which icons are distributed for each initial situation related to this embodiment. [Figure 16] It is a graph showing the details from the process of generating an intermediate situation to the process of generating a learned model related to this embodiment. [Figure 17]This is an explanatory diagram illustrating the process of generating an intermediate state from the initial state to the undetermined state according to this embodiment. [Figure 18] This is an explanatory diagram showing the process of correcting an intermediate state of uncertainty related to this embodiment to a first confirmed state. [Figure 19] This is an explanatory diagram showing the process of correcting the intermediate state of the undetermined situation in this embodiment to the second confirmed state. [Figure 20] This is a screen diagram illustrating the process of generating one intermediate state from two initial states in this embodiment. [Figure 21] This is a screen diagram showing the process of generating rule selection model information that has been learned from initial and intermediate state training data information in this embodiment. [Modes for carrying out the invention]

[0011] One embodiment of the present invention will be described below with reference to the drawings.

[0012] Figure 1 is an explanatory diagram of a simple production line 50. On the production line 50, robot 51 retrieves parts stored on parts shelves 52 and sets them in a machining center (M / C) 53 to process the parts. Robot 51 then removes the processed products from the machining center 53 and stores them on a finished product shelf 54, making the products ready for shipment. Here, we define the following terms. "Workpiece" refers to the items being produced, such as parts or finished products. Robot 51 transports the workpiece. "Resources" refer to locations and machines through which workpieces pass in the manufacturing process, such as robots 51, parts shelves 52, processing machines 53, and finished product shelves 54. Parts shelves 52 and finished product shelves 54 are also called stockers or buffers. The production line 50 can be represented as a collection of resources through which workpieces pass sequentially.

[0013] Figure 2 is an explanatory diagram of the complex production line 200. Figure 1 illustrates a simple production line 50 in which one workpiece passes through four resources. Figure 2, on the other hand, illustrates a complex production line 50 in which three workpieces pass through one of eleven resources. The first workpiece, indicated by the thick solid arrow, passes through the following order: parts shelf 201 → robot 231 → processing machine 211 → robot 232 → finished product shelf 221. The second workpiece, indicated by the thin solid arrow, passes through the following sequence: parts shelf 202 → robot 231 → processing machine 212 → robot 232 → finished product shelf 222. The third workpiece, indicated by the thin dashed arrow, passes through the following order: parts shelf 203 → robot 231 → processing machine 213 → robot 232 → finished product shelf 223.

[0014] Here, "production status" is the collection of the current positions of each workpiece at each given time. For example, at a certain time, production situation A was "Workpiece 1 = Robot 232", "Workpiece 2 = Processing Machine 212", and "Workpiece 3 = Robot 231". Subsequently, when Robot 232 moves Workpiece 1 to the finished shelf 221, production situation B changes to "Workpiece 1 = Finished shelf 221", "Workpiece 2 = Processing Machine 212", and "Workpiece 3 = Robot 231". Furthermore, if all workpieces are located at the starting position of production line 200, such as "workpiece 1 = parts shelf 201", "workpiece 2 = parts shelf 202", and "workpiece 3 = parts shelf 203", the production status is pre-work-in-progress. On the other hand, if some workpieces are located beyond the starting position of production line 50 (e.g., processing machine 53 in Figure 1), as in production statuses A and B described above, the production status is post-work-in-progress.

[0015] Next, "dispatching rules" are rules (algorithms) for selecting which workpiece's current position should be advanced to the next resource, based on the input production status. For example, there are the following rules. FIFO (First In, First Out): First-in, first-out. • EDD (Earliest Due Date): Ordered by delivery date (earliest first). • SLACK: Sort by the smallest margin of deadline (deadline - current time - total remaining work time).

[0016] The "production plan" is a plan that, based on the dispatching rules and the current production status, selects the next workpiece to be processed, thereby transitioning the production status from the initial state to the final state. Even when processing the same set of workpieces, production plans can vary in quality and are evaluated using KPIs (Key Performance Indicators) such as the delivery delay rate (the percentage of total workpieces that are delayed) and the make-span (the time from the start to the end of the work). Therefore, in order to improve the KPIs of production planning, it is important to generate (learn) a scheduling model that selects appropriate rules according to the input production status. In other words, the scheduling model is the result of learning using pairs of production status and dispatching rules that performed well in that production status as training data.

[0017] Figure 3 is a Gantt chart showing the production plan before the dispatching rules were changed. The vertical axis of the Gantt chart represents the seven resources, including robot R1, while the horizontal axis represents time. For example, in the current production situation at time t1, workpiece W1 has been removed from buffer B1 by robot R2 (workpiece W1 is in work-in-progress), and workpiece W3 has not yet been removed from buffer B1. In this production plan of Figure 3, the time t2a when machine M2 finishes workpiece W2 is the latest work completion time in the entire production plan, and is the time that determines the makespan.

[0018] Figure 4 is a Gantt chart showing the production plan with the dispatching rules changed from the production plan in Figure 3. In Figure 3, the processing order of robot R2 was "workpiece W1 → W2 → W3" selected according to the FIFO rule, but in Figure 4, it has been changed to "workpiece W2 → W1 → W3" selected according to the SLACK rule. As a result, the workpiece W2, which was the bottleneck determining the makespan in Figure 3, was processed earlier, shortening the time t2b when the processing machine M2 finished working with workpiece W2 in Figure 4 compared to time t2a in Figure 3. Therefore, the production plan in Figure 4 was improved compared to the production plan in Figure 3 due to the shortened makespan. Thus, in order to improve production planning, it is necessary to generate a scheduling model that selects rules appropriate to the production status at each point in time, such as the current time t1. Therefore, we will now explain the model learning device 1 shown in Figure 5.

[0019] Figure 5 is a diagram showing the configuration of the model learning device 1. First, let's explain the application of the model learning device 1. In addition to its application to the production line 50 shown in Figure 1, the model learning device 1 can also be applied to any process system. [Step 1 (Preparation Stage)] The user considers the configuration of the production line 50 and registers a simulation model of the production line 50 with the model learning device 1. [Second Step (Learning Phase)] Before actually operating the production line 50, the model learning device 1 generates a scheduling model that has learned various production conditions by running a simulation using a simulation model of the production line 50.

[0020] [Third Step (Operational Phase)] The model application device 9 uses the scheduling model generated by the model learning device 1 to output dispatching rules that are highly productive (highly rated KPI values) according to the various (current) production conditions that occur after the production line 50 is actually put into operation. Based on the output dispatching rules, the model application device 9 generates a production plan that selects the next workpiece to be processed and applies that production plan to each resource of the production line 50. In this way, the model application device 9 determines dispatching rules using the rule selection model information 29 prepared in advance by the model learning device 1, resulting in a low-load decision process that can handle frequent rescheduling. Therefore, automation of production control makes it easier to automate line operations.

[0021] The model learning device 1 learns a scheduling model that outputs dispatching rules for determining the next workpiece to be processed based on the input production status, which indicates the work-in-progress state of each workpiece flowing through the production line. Therefore, the model learning device 1 has a processing unit 10, a storage unit 20, an input unit 31, and an output unit 32. The memory unit 20 stores item information 21, process information 22, resource information 23, rule information 24, order information 25, work-in-progress information 26, learning data information 27, feature information 28, and rule selection model information 29. The memory unit 20 may be located inside the model learning device 1, or it may be located outside the model learning device 1 and configured to be accessible from the model learning device 1.

[0022] The input unit 31 receives input from the user, including item information 21, process information 22, resource information 23, rule information 24, order information 25, work-in-progress information 26, training data information 27, and feature information 28. The learning data information 27 received from this input unit 31 is referred to below as the "initial status" as the production status initially prepared by the user. Based on the various information received by the input unit 31, the processing unit 10 adds the intermediate status generated from the initial status learning data information 27 to the learning data information 27, and then generates rule selection model information 29 as a scheduling model. An "intermediate status" is an additional production status located between multiple initial statuses. The output unit 32 outputs the rule selection model information 29 generated by the processing unit 10 to the model application device 9. The output unit 32 also outputs various information useful to the user, such as through screen displays.

[0023] The processing unit 10 includes an initial situation acquisition unit 11, an intermediate situation generation unit 12, a simulation unit 13, and a model learning unit 14. The initial status acquisition unit 11 acquires multiple initial statuses, which are production statuses to which suitable dispatching rules have been pre-assigned. The intermediate situation generation unit 12 selects two initial situations with different dispatching rules from the set of initial situations acquired by the initial situation acquisition unit 11, and generates an intermediate situation that is located between the two selected initial situations in the feature space formed from the features extracted from the workpiece of the production situation. The simulation unit 13 runs a simulation of the production line for each combination of generated intermediate conditions and dispatching rules, thereby associating each intermediate condition with a dispatching rule suitable for that condition. The model learning unit 14 learns a scheduling model using the initial state corresponding to the dispatching rule and the intermediate state corresponding to the dispatching rule as training data.

[0024] The output unit 32 outputs, for example, one of the following data: • Output of training data generated by the model learning unit (model learning data generation unit) 14 (in a form that causes the model learning device 1 to execute the model learning data generation method). • The output of the scheduling model trained on that training data. • Output from both the training data and the scheduling model.

[0025] The simulation model of the production line 50 used by the simulation unit 13 is defined by a combination of item information 21, process information 22, and resource information 23. Item information 21 is information that defines the ID of the item that can be produced on the production line. Resource information 23 defines the resources necessary for executing the process, such as equipment and storage space. Process information 22 is information that defines the processes to be performed until the item is completed.

[0026] Figure 6 is a table showing item information 21 and resource information 23. Item information 21 lists the item IDs that will be used for each number (#). Resource information 23 associates the equipment ID representing the resource with its type.

[0027] Figure 7 is an explanatory diagram showing process information 22. Process information 22 is information that models a production line, such as production line 200 in Figure 5. Process information 22 associates the item ID of the workpiece with the process ID indicating the process to be executed, the ID of the preceding process, the equipment ID indicating the resource to execute the process, the destination ID indicating the resource to execute the next process, and the work time of the process.

[0028] The production status of production line 50 is defined by a combination of order information 25 and work-in-progress information 26. The initial status acquisition unit 11 acquires the initial status of the production status by reading data previously input from the input unit 31 or by randomly generating it from a simulation model of production line 50. Order information 25 indicates the order to be put into the production line. Work-in-progress information 26 indicates the progress of each order.

[0029] Figure 8 is a table showing order information 25. Order information 25 associates the order ID, work ID, item ID, quantity, and delivery date.

[0030] Figure 9 is a table showing the work-in-progress information 26. Work-in-progress information 26 associates the production status ID, the work ID included in that production status, the current location ID indicating the resource where the current work exists, the next process ID for the work, and the start date and time for the work.

[0031] Figure 10 is an explanatory diagram showing the work-in-progress information 26 for each workpiece according to its production status. In production status 301, work-in-progress information 26 shows that there are six workpieces, W1 to W6, and their current position IDs are as follows: • Current location ID for workpieces W2 and W6 = parts shelf • Current position ID of workpieces W1 and W3 = buffer • Current position ID of workpieces W4 and W5 = processing machine

[0032] Production status 302 shows the state at a later time than production status 301, and the current position of the workpiece is progressing little by little as follows. • Current position ID of workpieces W2 and W6 = buffer • Current position ID of workpieces W1, W3, W5 = processing machine • Work W4 current location ID = completed shelf

[0033] Rule information 24 is information that shows the dispatching rules for determining the priority of tasks in the production plan. Feature information 28 is information used to define a feature space for defining the distance between production situations. Production situations that are close in distance within the feature space are similar to each other.

[0034] Figure 11 is a table showing rule information 24 and feature information 28. Rule information 24 lists dispatching rules such as FIFO. While Figure 11 illustrates the names of the dispatching rules, in reality, rule information 24 also includes the algorithm of the dispatching rule in addition to its name. Feature information 28 is a list of features extracted from the production status (each workpiece).

[0035] The rule selection model information 29 is information that indicates the scheduling model, and outputs the corresponding highly rated dispatching rule according to the input production status. The training data information 27 is data used to train the rule selection model information 29, and consists of combinations of individual production situations (initial situation, intermediate situation) and highly-rated dispatching rules corresponding to those production situations.

[0036] Figure 12 is a table showing the training data information 27. Tables 27A and 27B associate scenarios with order IDs, production status IDs, and rule IDs, respectively. A scenario is a name assigned to a production status ID, and one scenario corresponds to one production status ID. Table 27A shows the learning data information 27 which stores scenarios "1" to "9" as initial states acquired by the initial state acquisition unit 11. Table 27B shows the learning data information 27 obtained by adding scenarios "12" and "13," which are intermediate situations generated by the intermediate situation generation unit 12 from the initial situations (scenarios "1" to "9") in Table 27A.

[0037] Furthermore, the input data that the rule selection model information 29 associates with the dispatching rule is not limited to production status alone; the following information may also be used as input data in combination with production status. • Combination of workpiece input varieties • Equipment operating status (presence or absence of malfunctions, operating hours) • Delivery date of work This allows for the output of appropriate dispatching rules to prevent productivity declines, depending on the diverse production environments represented by the input data.

[0038] Figure 13 is a flowchart showing the processing of the model learning device 1. First, we will explain the process (S101-S104) for learning the rule selection model information 29 from the initial situation. The initial status acquisition unit 11 acquires item information 21, process information 22, resource information 23, and rule information 24 as master data to be used in the evaluation calculation (S103) of the simulation unit 13 (S101). The initial status acquisition unit 11 acquires information indicating the initial status (order information 25, work-in-progress information 26) (S102).

[0039] The simulation unit 13 evaluates the simulation model of the production line shown in the master data of S101 using a simulator for each initial situation in S102 (S103). In this evaluation process of S103, a KPI is calculated for each combination of rule and initial situation shown in the rule information 24, and the rule ID of the rule that outputs the highest-rated KPI is associated with the production situation ID of the initial situation as learning data information 27. The model learning unit 14 generates initial rule selection model information 29 by machine learning the combination of the initial situation in S102 and the evaluation result (the highest-rated rule) in S103 (S104). The machine learning algorithm used by the model learning unit 14 in the S104 process is preferably an algorithm suitable for feature space partitioning, such as a Decision Tree or a Support Vector Machine (SVM).

[0040] Next, we will explain the process (S112-S114) for improving the accuracy (retraining) of the rule selection model information 29 based on the intermediate situation generated from the initial situation. The intermediate situation generation unit 12 selects multiple (two) initial situations from the set of initial situations and generates a new production situation that is located between those initial situations as an "intermediate situation" (S112). The simulation unit 13 evaluates the simulation model of the production line shown in the master data of S101 using a simulator for each intermediate situation in S112, in the same manner as the evaluation process in S103 (S113). As a result, new learning data information 27 for each intermediate situation is generated. The model learning unit 14 generates rule selection model information 29 by machine learning the initial situation learning data information 27 generated in S103 and the intermediate situation learning data information 27 generated in S113, in the same manner as the learning process in S104 (S114).

[0041] The model learning unit 14 determines whether the termination criteria, as exemplified below, are met (S115), and if the termination criteria are not met, the process returns to S112. • When the number of intermediate situations generated is less than or equal to a predetermined amount. • When the rule selection model information 29 converges. For example, when the difference between the previous rule selection model information 29 and the current rule selection model information 29 is less than or equal to a predetermined amount. Here, in the second and subsequent S112, the intermediate status generation unit 12 may not only generate an intermediate status from the initial status in S102, but may also consider the intermediate status generated in the previous S112 as the initial status for the current time and generate the current intermediate status.

[0042] Figure 14 is a hardware configuration diagram of the model learning device 1. The model learning device 1 is configured as a computer 900 having a CPU 901, RAM 902, ROM 903, HDD 904, communication I / F 905, input / output I / F 906, and media I / F 907. The communication interface 905 is connected to an external communication device 915. The input / output interface 906 is connected to the input / output device 916. The media interface 907 reads and writes data to the recording medium 917. Furthermore, the CPU 901 controls each processing unit by executing a program (also called an application or app) loaded into the RAM 902. This program can also be distributed via a communication line or by recording it on a recording medium 917 such as a CD-ROM.

[0043] Figure 15 is a graph showing the feature space 310 where icons for each initial situation are distributed. In Figure 15, the feature space 310 is illustrated as a two-dimensional plane with the features C1 and C2 selected from the feature information 28 as its two axes. However, the number of features used to form the feature space 310 is not limited to two; one or more features can be used. For example, one feature can be used to form a one-dimensional number line, or three features can be used to form a three-dimensional space.

[0044] In the feature space 310, the position within the space is determined based on the feature values ​​C1 and C2 of each production status ID (S1 to S9), which represent the initial state. Furthermore, in Figure 15, the same rule ID is enclosed in the same shape, as shown below, by the rule used by each production status ID, that is, the rule ID corresponding to each production status ID in the training data information 27. • The production status IDs (S1-S3) enclosed in circles use FIFO (First-In, First-Out) logic. • Production status IDs (S4-S6) enclosed in a rectangle use EDD (Electronic Data Desk). • Production status IDs (S7-S9) enclosed in triangles use SLACK.

[0045] The feature space 320 shows the state in which the model learning unit 14, using a machine learning algorithm, has determined the boundary surface 324 (boundary in 2D) of each rule, based on the arrangement of each production status ID (S1~S9) present in the feature space 310. The boundary surface 324 demarcates the following regions 321, 322, and 323. Area 321, which includes production status IDs (S1~S3), is the area of ​​production status where FIFO is used. Area 322, which includes production status IDs (S4~S6), is the area of ​​production status where EDD is used. Area 323, which includes the production status IDs (S7~S8), is the area of ​​production status where SLACK is used. This interface 324 is the main information that forms the rule selection model information 29. As a result, the rule selection model information 29 can output an appropriate rule not only when the input is the same production status as the production status ID of the training data information 27, but also when the input is a different production status than the production status ID of the training data information 27.

[0046] In this way, the model learning unit 14 generates rule selection model information 29 by inferring the boundary surface 324 from the set of initial conditions (S104). Therefore, the larger the set of initial conditions obtained in S102 (the more initial conditions prepared), the more accurate the shape of the boundary surface 324 becomes, and the higher the accuracy of the rule selection model information 29. On the other hand, increasing the size of the initial set of conditions increases computation time and memory capacity required for computation, thus raising the computation cost. Therefore, the intermediate situation generation unit 12 of the model learning device 1 improves the accuracy of the interface surface 324 with low computational cost by adding the production conditions near the interface surface 324, which greatly affect the shape of the interface surface 324, to the learning data information 27.

[0047] Figure 16 is a graph showing the details from the process of generating intermediate states (S112) to the process of generating the trained model (S114). In the feature space 330, production situations with different rules (production situation IDs = S3, S4) are arranged as initial states. The model learning unit 14 then calculated an initial interface 331 that reflects the arrangement of these production situation IDs = S3, S4. The intermediate status generation unit 12 selects production status IDs S3 and S4, which have different rules, and generates a new production status (production status ID = S11) that is located between them as an intermediate status.

[0048] Note that Figure 15 also shows other production situations (production situation IDs = S1, S5) that follow different rules. However, it is desirable for the intermediate situation generation unit 12 to select the two initial situations (production situation IDs = S3, S4) that are closest in distance in the feature space to generate intermediate situations. As a result, the intermediate situations generated from the selected production situations IDs = S3, S4 are expected to be located near the boundary surface 324, contributing to the efficient modification of the boundary surface 324.

[0049] In Figure 16, for the sake of clarity, the midpoint of the dashed line connecting the locations of production status ID=S3 and S4 is designated as production status ID=S11. On the other hand, the intermediate situation generation unit 12 may generate a production situation that is similar to both initial situations, rather than being limited to the midpoint of two initial situations, as an intermediate situation located between two initial situations with different rules. Production situations are considered similar if the distance between them in the feature space is less than a predetermined value.

[0050] Furthermore, although the intermediate situations generated by the intermediate situation generation unit 12 are placed at arbitrary positions within the feature space, they may sometimes be data that cannot be used in a production plan in reality, such as "the current position of the workpiece is about 30% closer to the position of the finished shelf from the position of the processing machine." Production situations in which the current position of the workpiece is not determined in this way are called "undetermined situations," and production situations in which the current position of all workpieces is determined (possible values) are called "determined situations." Therefore, the intermediate status generation unit 12 approximates the undetermined status (production status ID = S11) to a similar confirmed status (production status ID = S12, S13). This approximation process is part of the process for generating the intermediate status (S112). Thus, if the intermediate position in the intermediate state does not exist on a resource that processes workpieces within the production line (in the case of an undetermined state), the intermediate state generation unit 12 generates an intermediate state for a determined state by moving that intermediate position to the location of a nearby resource.

[0051] The feature space 340 shows the interface 341 learned from the training data information 27 to which intermediate situations (production status IDs = S12, S13) have been added. Based on the evaluation process in the simulator (S113), the same EDD rules as for production status ID = S4 are suitable for production status ID = S12 and S13.

[0052] Therefore, the interface 341 of the new rule selection model information 29 generated by the process of generating the learned model (S114) shows the following changes compared to the initial interface 331. The EDD area, which includes production status IDs S4, S12, and S13 where the EDD rules are suitable, will be slightly expanded. The area of ​​the FIFO containing production status ID=S3 will be slightly reduced. In this way, the intermediate state generation unit 12 concentrates on placing intermediate states near the initial boundary surface 331, thereby making a significant contribution to correcting the boundary surface even with a small number of intermediate states.

[0053] The following details of the process by which the intermediate status generation unit 12 approximates an undetermined status (production status ID = S11) to a similar confirmed status (production status ID = S12, S13) will be explained with reference to Figures 17 to 19.

[0054] Figure 17 is an explanatory diagram illustrating the process of generating an intermediate state from the initial state to an undetermined state. Here, production status 303 in Figure 17 is an intermediate state, with production status 301 and production status 302 in Figure 10 as the initial states (for example, production status ID=S11 in Figure 16). The intermediate status generation unit 12 sequentially selects each workpiece in production status 301 and production status 302, and for the selected workpiece, it sets the average value (midpoint) of the current position in production status 301 and the current position in production status 302 as the current position in production status 303. For example, when workpiece W2 is selected, the midpoint between the current position W2B in production status 301 and the current position W2A in production status 302 is the current position of workpiece W2 in production status 303.

[0055] Figure 18 is an explanatory diagram showing the process of correcting the intermediate, undetermined situation (production status 303 in Figure 17) to the first confirmed situation. In production status 304, the current position of workpiece W5 has already been determined by the processing machine, and no correction is necessary. Other workpieces W1-W4 and W6 do not currently reside on any specific resource (undetermined locations enclosed by dashed lines), so their current locations need to be determined to one of the nearby resources (determined locations enclosed by solid lines). The intermediate status generation unit 12 moves the workpieces W1 to W4 and W6 to this determined position, thereby changing the production status to a determined state without significantly altering the intermediate status of the undetermined state (for example, production status ID = S12 in Figure 16).

[0056] Figure 19 is an explanatory diagram showing the process of correcting the intermediate, undetermined situation (production status 303 in Figure 17) to the second, confirmed situation. The current position of workpieces W1, W3-W6 is the same in both production status 304 in Figure 18 and production status 305 in Figure 19.

[0057] On the other hand, workpieces W2 and W6 were in the same current position (an intermediate position between the parts shelf and the buffer) before the correction. In this case, when multiple workpieces are in the same current position, the intermediate status generation unit 12 determines which workpiece to prioritize based on the production status dispatching rules. For example, in production status 304 in Figure 18, the intermediate status generation unit 12 adopted a rule (e.g., FIFO) to prioritize workpiece W6 over workpiece W2 and proceed to the next process (buffer). On the other hand, in production status 305 in Figure 19, the intermediate status generation unit 12 adopts a different rule (e.g., SLACK) to prioritize workpiece W2 over workpiece W6.

[0058] The following describes the screen diagrams output by the output unit 32 of the model learning device 1 to the display device, with reference to Figures 20 and 21. Figure 20 is a screen diagram illustrating the process of generating one intermediate state from two initial states. Of the two initial statuses, production status display 301G corresponds to production status ID=S3 in Figure 16 and production status 301 in Figure 10, and displays work-in-progress information 26 for each of the six workpieces arranged vertically. For example, in work W1, the production line passes through four resources "R1→R2→R3→R4" in order. Resource "R1," indicated by the white circle, has been executed, while resources "R2, R3, and R4," indicated by the black circles, have not yet been executed, and resource "R2" has a triangle mark indicating its current position. Similarly, production status display 302G, one of the two initial states, corresponds to production status ID=S4 in Figure 16 and production status 302 in Figure 10. Production status display 303G, which is an intermediate state between these two initial states, corresponds to production status ID=S11 in Figure 16 and production status 303 in Figure 17.

[0059] Therefore, the intermediate status generation unit 12 creates work-in-progress information 26 (Figure 9) for intermediate statuses such as production status 303 using the following procedure. [Step 1] Assign a new production status ID to the intermediate stage. [Step 2] Substitute the work IDs from the initial status (works W1 to W6) into the work IDs for the intermediate status. [Step 3] The current location ID of the intermediate situation is the midpoint (e.g., the average) of the current location IDs of the two initial situations. [Step 4] The next process ID for the intermediate status will be the transport from the resource with the current position ID to the next resource on the production line. [Step 5] The start date and time for the intermediate situation will be the midpoint (e.g., the average) of the start dates and times for the two initial situations.

[0060] As described in these steps, the intermediate status generation unit 12 obtains the current position indicating the work-in-progress state within the production line for each workpiece in the initial status and for the same workpiece included in multiple initial statuses, and generates an intermediate status by taking the intermediate position in the feature space of each current position as the workpiece's position in the intermediate status. The output unit 32 then displays information on the screen, as shown in Figure 20, indicating the production line for each workpiece and its current position within that production line, along with the intermediate status generated by the intermediate status generation unit 12 and the initial status used to generate it. This allows the user to intuitively understand what kind of intermediate status has been generated.

[0061] Figure 21 is a screen diagram showing the process of generating rule selection model information 29, which is trained using initial and intermediate state training data information 27. The output unit 32 displays side by side the feature space 410 of the rule selection model information 29 generated from the initial situation in S104 and the feature space 420 of the rule selection model information 29 generated from the initial and intermediate situations in S114. In this way, the output unit 32 places the intermediate and initial states within the feature space, and also displays the information of the scheduling model boundary, which divides the region to which the same dispatching rule is applied within the feature space, on the screen as shown in Figure 21.

[0062] This allows the user to intuitively understand that the initial boundary surface 411, indicated by a circle in the feature space 410, has been modified (improved) to the boundary surface 421, which incorporates the intermediate situation indicated by a square in the feature space 420. Furthermore, each region demarcated by the boundary surface 421 may also be displayed in a visually distinguishable manner, such as by changing the background color.

[0063] In the embodiment described above, the intermediate situation generation unit 12 creates an intermediate situation from two initial situations with different rules, thereby causing the model learning unit 14 to generate rule selection model information 29 in which the boundary surface has been modified according to a rule suitable for the intermediate situation. This allows for concentrated expansion of production data near the conditions under which rules should be switched, thereby reducing computational costs compared to comprehensively generating production data while achieving equivalent model accuracy. Therefore, the model application device 9, using highly accurate rule selection model information 29, can reflect appropriate rules in future production plans according to various situations, not only in the early stages of work-in-progress production, but also in the middle and late stages.

[0064] Furthermore, the present invention is not limited to the embodiments described above, and it goes without saying that various other applications and modifications can be taken as long as they do not depart from the gist of the invention as described in the claims. For example, the embodiments described above are detailed and specific descriptions of the configuration of the model learning device 1 in order to explain the present invention in an easy-to-understand manner, and are not necessarily limited to those comprising all the components described. Also, it is possible to replace a part of the configuration of one embodiment with a component of another embodiment. It is also possible to add a component of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, replace, or delete other components for a part of the configuration of each embodiment.

[0065] Furthermore, some or all of the above configurations, functions, and processing units may be implemented in hardware, for example, by designing them as integrated circuits. Broadly defined processor devices such as FPGAs (Field Programmable Gate Arrays) and ASICs (Application Specific Integrated Circuits) may be used as hardware. Furthermore, each component of the model learning device 1 according to the above-described embodiment may be implemented on any hardware, as long as the respective hardware can send and receive information from each other via a network. Also, the processing performed by a certain processing unit may be implemented by a single piece of hardware, or by distributed processing using multiple pieces of hardware. [Explanation of Symbols]

[0066] 1. Model learning device (computer) 9 Models Applicable Devices 10 Processing Unit 11 Initial Status Acquisition Unit 12 Intermediate Situation Generation Unit 13. Simulation Department 14. Model Learning Unit (Model Training Data Generation Unit) 20 Memory section 21 Item Information 22 Process information 23 Resource Information 24. Rule Information 25 Order Information 26. Work in Progress Information 27. Learning Data Information 28 Feature Information 29. Rule Selection Model Information 31 Input section 32 Output section 50,200 production lines

Claims

1. A model learning device that learns a scheduling model for which, based on the production status indicating the work-in-progress state of each workpiece flowing through a production line, dispatching a dispatch rule for determining the next workpiece to be put into production based on the input production status, An initial status acquisition unit acquires multiple initial statuses, which are production conditions to which a suitable dispatching rule has been assigned in advance. An intermediate situation generation unit selects two initial situations from the set of initial situations acquired by the initial situation acquisition unit, in which the dispatching rules are different from each other, and generates an intermediate situation located between the two selected initial situations in a feature space formed from the feature quantities extracted from the workpiece of the production situation. A simulation unit that associates each intermediate situation with the dispatching rule appropriate for that intermediate situation by performing a simulation of the production line for each combination of the generated intermediate situation and the dispatching rule, The system is characterized by having a model learning unit that learns the scheduling model using the initial state corresponding to the dispatching rule and the intermediate state corresponding to the dispatching rule as learning data. Model learning device.

2. The intermediate state generation unit is characterized by selecting the two initial states that are closest in distance to each other in the feature space in order to generate the intermediate state. The model learning device according to claim 1.

3. The intermediate status generation unit is characterized by generating the intermediate status by obtaining the current position indicating the work-in-progress state within the production line for each workpiece in the initial status and for the same workpiece included in multiple initial statuses, and by setting the midpoint of each current position in the feature space as the position of the workpiece in the intermediate status. The model learning device according to claim 1.

4. The intermediate status generation unit is characterized in that, if the intermediate position in the intermediate status does not exist on a resource that processes workpieces within the production line, it moves that intermediate position to the location of a nearby resource to generate the intermediate status. The model learning device according to claim 3.

5. The aforementioned model learning device further includes an output unit that outputs information to be displayed on the screen. The output unit is characterized by displaying on the screen information indicating the production line for each workpiece and its current position within that production line, based on the intermediate status generated by the intermediate status generation unit and the initial status used to generate it. The model learning device according to claim 3.

6. The aforementioned model learning device further includes an output unit that outputs information to be displayed on the screen. The output unit is characterized in that it places the intermediate and initial states within the feature space, and also displays on the screen the information placed within the feature space, including the boundary surface of the scheduling model that demarcates the region to which the same dispatching rule is applied within the feature space. The model learning device according to claim 3.

7. A method for generating model training data to train a scheduling model in which a computer outputs dispatching rules for determining the next work to be put into work based on the input production status, which indicates the work-in-progress state of each workpiece flowing through a production line, the computer generates training data for training a scheduling model, the computer outputs dispatching rules for determining the next workpiece to be put into work based on the input production status, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model in which a computer outputs dispatching rules for determining the next workpiece to be put into work based on the production status which indicates the work-in-progress state of each workpiece flowing through a production line, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model in which a generates training data for determining the next workpiece to be put into work based on the production status which indicates the work-in-progress state of each workpiece flowing through a production line, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model in which a computer generates training data for training a scheduling model in which a computer generates training data for determining the next workpiece to be put into work based on the production status which indicates the work-in-progress state of each workpiece flowing through a production line, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model in which a computer generates training data for determining the next workpiece to be put into work based on the production status which indicates the work-in-progress state of each workpiece flowing through a production line, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model, the computer generates training data for training a scheduling model, The aforementioned computer includes an initial state acquisition unit, an intermediate state generation unit, a simulation unit, and a model training data generation unit. The initial status acquisition unit acquires multiple initial statuses, which are production statuses to which a suitable dispatching rule has been assigned in advance. The intermediate situation generation unit selects two initial situations from the set of initial situations acquired by the initial situation acquisition unit, in which the dispatching rules are different from each other, and generates an intermediate situation located between the two selected initial situations in the feature space formed from the feature quantities extracted from the workpiece of the production situation. The simulation unit performs a simulation of the production line for each combination of the generated intermediate conditions and dispatching rules, thereby associating each intermediate condition with the dispatching rule appropriate for that intermediate condition. The model training data generation unit is characterized by generating the initial state corresponding to the dispatching rule and the intermediate state corresponding to the dispatching rule as training data. Method for generating data for model training.

Images