Shape control system

Abstract

To appropriately update a portion of a learning model that need to be updated by relearning while suppressing an unintended effect on the learning model caused by relearning a learning device in a shape control system that uses machine learning to control a shape of an object.SOLUTION: A shape control system according to the present invention corrects a controller model in a data space to which correction data belongs in such a manner that when data sampled within an area at least a threshold distance away from the correction data is input to the controller model, a difference between an output from the controller model before correction and an output from the controller model after correction tends to 0.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a shape control system for controlling the shape of an object. [Background technology]

[0002] A rolling system processes metal into shapes such as plates by passing it between two rotating rolls. One method for controlling a rolling system uses machine learning. In this method, the target shape of the material and the operation amount of the actuator are set as control targets (output from the learning device), and the operating parameters of the rolling system are set as input parameters (input to the learning device). By performing machine learning in advance, the control parameters for obtaining the target shape are calculated.

[0003] The following Patent Document 1 addresses the issue of "learning an optimal operation method for performance data without worsening the state of a controlled plant," and describes the following technology (see abstract): "The control method learning device 21 learns combinations of performance data and control operations for a controlled plant 1, and the control execution device 20 controls the controlled plant in accordance with the learned combination of performance data and control operations, and the control method learning device learns to obtain separate combinations of performance data and control operations for multiple control targets in accordance with the state of the controlled plant, and the obtained combinations of performance data and control operations are used as predetermined combinations of performance data and control operations for the controlled plant in a control rule execution unit 10, and when the state of the controlled plant is similar to the combination of performance data and control operations before correction, the corrected learning data is used to perform control" (see abstract). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2019-028824 Summary of the Invention [Problem to be solved by the invention]

[0005] In the conventional technology for controlling a rolling system using machine learning, when a learner is retrained, additional learning is performed using correction data used for the retraining. This correction data is data that is different from the training data used when the learner was initially trained. The learning results in the learner are updated by the retraining, and the initial learned model becomes a corrected learned model.

[0006] When retraining using correction data, it is desirable to update only the parts of the initial training model that are directly related to the correction data. This is because retraining is usually performed to partially update the initial training model. However, the scope of influence that the correction data has on the initial training model is not necessarily limited to the parts that are being corrected. This is because the data structure inside a training model is generally complex, and internal elements influence each other. In this case, retraining using correction data may update parts of the training model that are not directly related to the retraining, which could unintentionally change the output from the learning module.

[0007] The present invention has been made in consideration of the above-mentioned problems, and aims to appropriately update the parts of the learning model that are to be updated by relearning, while suppressing unintended effects on the learning model caused by relearning of a learning device, in a shape control system that controls the shape of an object using machine learning. [Means for solving the problem]

[0008] The shape control system according to the present invention modifies the controller model so that when data sampled within an area in the data space to which the modified data belongs that is at least a threshold distance away from the modified data is input to the controller model, the difference between the output from the controller model before modification and the output from the controller model after modification approaches zero. [Effects of the Invention]

[0009] According to the shape control system of the present invention, in a shape control system that controls the shape of an object using machine learning, it is possible to appropriately update parts of the learning model that are desired to be updated by relearning while suppressing unintended effects on the learning model caused by relearning of the learning module. Problems, configurations, and effects other than those described above will become clear from the description of the following embodiments. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a schematic diagram showing the configuration of a learning device used by the shape control system according to the first embodiment. FIG. [Figure 2] FIG. 2 is a conceptual diagram of a learning device according to the first embodiment. [Figure 3] FIG. 10 is a schematic diagram illustrating a method for making the output from the differential model B approach 0. [Figure 4] FIG. 10 is a schematic diagram illustrating another method for making the output from the differential model B approach 0. [Figure 5] FIG. 10 is a schematic diagram illustrating another method for making the output from the differential model B approach 0. [Figure 6] FIG. 10 is a configuration diagram of a shape control system 1 according to a second embodiment. [Figure 7] 10 is a flowchart illustrating a procedure in which the model generation computer 11 generates an initial learning model A. [Figure 8] 10 is a flowchart illustrating a procedure in which the model generation computer 11 generates a corrected learning model C. [Figure 9] 10 is a flowchart illustrating a process in which the model generation computer 11 collects sampling data. [Figure 10] 10 is a flowchart illustrating a procedure in which the control computer 12 controls the shape control device 14. [Figure 11] The formulas for calculating the reaction rate and polarity agreement rate are shown below. [Figure 12] The formula for calculating the model change rate is shown below. [Figure 13]1 shows an example of a user interface provided by the shape control system 1. [Figure 14] 10 is an example of an objective function used when a learning device generated by the model generation computer 11 performs machine learning. DETAILED DESCRIPTION OF THE INVENTION

[0011] <First Embodiment> Fig. 1 is a schematic diagram showing the configuration of a learning device used by a shape control system according to a first embodiment of the present invention. The shape control system is a system that controls the shape of an object (e.g., a metal material) by, for example, a rolling process. The shape control system performs machine learning in advance, using the target shape of the material as a control target (output from the learning device) and operation parameters of the rolling system as control parameters (input to the learning device). The learning device is sometimes called a controller model.

[0012] The following can be used as inputs to the learning device: (a) shape deviation (the difference between the final shape of the object and its current shape), (b) speed (the speed at which the object passes through the rolls), (c) pass number (described later), and (d) the current positions of the upper and lower intermediate rolls. Upon receiving the above input parameters, the learning device is configured to output the manipulated variables. The learning device performs machine learning in advance to achieve the target shape by optimizing the manipulated variables.

[0013] Depending on the rolling system, the target shape may be achieved by repeatedly passing the target object (metal material) through the rolls multiple times. In this case, the final target shape is determined, but the shape formed by each pass is not necessarily strictly determined, and the operator repeatedly adjusts the shape of the target object each time while observing its shape. This repetition number is called the pass number. Because the repetition number correlates with the shape formed in each pass, it was used as input to the learning machine.

[0014] The manipulated variables include the shift amount of each intermediate roll. Furthermore, if there is a member that presses the upper intermediate roll toward the object from above the upper intermediate roll, the number and shift amount of that member also become manipulated variables. The same applies to the lower intermediate roll and the member below it. If the rolls can be tilted, the tilt (leveling) also becomes a manipulated variable. The learning device outputs these manipulated variables. These manipulated variables are used to control the shape of the object (e.g., the thickness of the object, the position and degree of bending of the object, etc.).

[0015] FIG. 2 is a conceptual diagram of a learning device in the first embodiment. The shape control system creates an initial learning model A by performing initial learning using training data. The manipulated variable output by the initial learning model A may need to be changed during the process of operating the rolling system. Therefore, the shape control system updates the initial learning model A by performing re-learning to create a corrected learning model C. The learning data used in re-learning is called correction data. Re-learning is performed to re-optimize the learning device by updating suboptimal parts of the initial learning model A.

[0016] However, during the relearning process, parts of the initial learning model A that are not directly related to the purpose of the relearning may be updated. This is because the internal structure of a machine learning device is complex, and the internal components are interrelated. Therefore, in embodiment 1, when relearning is performed, a differential model B that is differentially applied to the initial learning model A is created, and this is applied to the initial learning model A to create a corrected learning model C. Therefore, the correction data is for creating the differential model B.

[0017] In order to minimize the impact on parts of the initial learning model A that are not directly related to relearning, it is desirable to make the output from the differential model B approach 0 in areas far from the correction data. Therefore, below we will explain a method for making the output from the differential model B approach 0.

[0018] Fig. 3 is a schematic diagram illustrating a method for making the output from differential model B approach 0. For convenience of description, the data space to which the training data belongs is represented as a two-dimensional space, and each data point in the data space is represented by a mark in Fig. 3.

[0019] In the left diagram of Figure 3, the training data used to train initial learning model A is placed in the data space. The correction data used to create differential model B is newly added to the data space. When new data samples are collected within a certain range of proximity to the correction data in the data space and re-learning is performed using these data samples, the proximity of the training data and the correction data may affect the learning results of the training data (i.e., initial learning model A). This is indicated by the fact that part of the training data is included in the circle around the correction data in the left diagram of Figure 3.

[0020] Therefore, in the first embodiment, the training data near the correction data is removed, and re-learning is performed using the remaining training data as a reference. The center diagram of Fig. 3 shows the training data remaining after the removal.

[0021] The right diagram of Figure 3 shows an example of new sample data. In relearning, a new data sample is input to a learning device, and the correct answer data for that data sample is also provided to the learning device. This process is repeated to perform the learning process. It is desirable that the new data sample have as little influence as possible on the initial learning model A. Therefore, in the first embodiment, in addition to (a) learning the corrected data itself, (b) new data samples are collected from a peripheral area of ​​the post-removal training data shown in the center diagram of Figure 3 as a reference. These new data samples are collected from an area that has little influence on the initial learning model A that is different from the intention of relearning, so that relearning can be performed appropriately while suppressing the influence on the initial learning model A.

[0022] Figure 4 is a schematic diagram explaining another method for bringing the output from differential model B closer to 0. As in Figure 3, correction data is placed in a two-dimensional data space. Areas that are relatively close to the correction data (within the solid circle in Figure 4) are thought to have a large impact on the initial training model A. On the other hand, areas that are far away from the correction data may not be very useful for re-training.

[0023] In view of the above, in FIG. 4, new data samples for re-learning are not collected from areas close to the modified data, but new sample data is collected from areas that are a certain distance (but not too far) from the modified data. In other words, new sample data is collected from areas in the data space that are at least a certain distance (at least a first threshold distance) and at most a certain distance (at most a third threshold distance) from the modified data. The modified data itself is also used for re-learning. This allows re-learning to be performed appropriately while suppressing the influence of the modified data on the initial training model A.

[0024] FIG. 5 is a schematic diagram illustrating another method for bringing the output from differential model B closer to 0. As in FIG. 3, correction data is arranged in a two-dimensional data space. In FIGS. 3 and 4, a method for suppressing the influence of correction data during re-learning was explained. Instead, in FIG. 5, the influence of correction data is suppressed when obtaining output from a learning device that has completed learning.

[0025] Specifically, when the learning device performs inference using data input from a region in the data space that is close to the modified data (a region whose distance from the modified data is within a first threshold), the output from differential model B is made sufficiently small (for example, set to 0). This makes it possible to perform re-learning using the modified data while suppressing the influence of the modified data on the initial learning model A.

[0026] <Embodiment 2> 6 is a configuration diagram of a shape control system 1 according to embodiment 2 of the present invention. The shape control system 1 is a system that implements the method described in embodiment 1. The shape control system 1 includes a model generation computer 11, a control computer 12, an operation terminal 131, an operation console 132, and a shape control device 14.

[0027] The shape control device 14 is a device that controls the shape of an object, for example, a metal material as the object, by a rolling process. The control computer 12 is a computer that controls the shape control device 14 using control parameters output by the learning model generated by the model generation computer 11. The control computer 12 includes a control unit 121 and a memory unit 122. The control unit 121 controls the shape control device 14. The memory unit 122 is a device that stores data (e.g., control parameters, etc.) used by the control unit 121.

[0028] The model generation computer 11 is a computer that generates learning models (initial learning model A, differential model B, and corrected learning model C) using the method described in embodiment 1. The model generation computer 11 includes a sample generation unit 111, a model generation unit 112, a display management unit 113, and a storage unit 114. The storage unit 114 is a device that stores data used by each unit. Other functional units will be described later.

[0029] The operation terminal 131 is a terminal through which an operator inputs instructions to the model generation computer 11 or displays data output by the model generation computer 11. The operation console 132 is a terminal through which an operator inputs instructions to the shape control device 14.

[0030] 7 is a flowchart illustrating the procedure for the model generation computer 11 to generate an initial learning model A. The model generation unit 112 acquires training data and creates appropriate learning data by dividing the training data into batch sizes (S701). The model generation unit 112 updates the internal parameters of the initial learning model A using the created learning data (S702). If a termination condition is met (e.g., learning data of a specified size or more has been learned), the flowchart ends (S703: YES); if not, the process returns to S701 (S703: NO).

[0031] 8 is a flowchart illustrating the procedure by which the model generation computer 11 generates the corrected learning model C. Each step in FIG. 8 will be described below.

[0032] (Figure 8: Step S801) The model generation unit 112 generates data for each appropriate processing unit by dividing the correction data into batch sizes, etc. This step is performed for convenience of processing and may be omitted as appropriate.

[0033] (Figure 8: Step S802) The sample generation unit 111 excludes the training data that is close to the modified data. The model generation unit 112 collects sampling data to be used for re-learning from the vicinity of the training data after the exclusion process. This step corresponds to the method described in FIG. 3.

[0034] (Figure 8: Step S803) The sample generation unit 111 collects sample data to be used for re-learning from an area that is somewhat distant (but not too distant) from the correction data. This step corresponds to the method described in FIG. 4.

[0035] (Figure 8: Steps S802 to S803: Supplement) In S802 and S803, the methods described in Figures 3 and 4 are used in combination, but it should be noted that by using at least one of these methods, re-learning can be performed while suppressing the influence of the correction data on the initial learning model A. The same applies to the method described in Figure 5, which will be described later. In other words, it is sufficient to use at least one of these three methods.

[0036] (Figure 8: Steps S804 to S806) The model generation unit 112 creates appropriate training data by dividing the sampling data by batch size, for example (S804). The model generation unit 112 updates the internal parameters of the differential model B using the created training data (the corrected data in S801 and the sampling data in S804) (S805). If the termination condition (e.g., training data of a specified size or more has been trained) is met, this flowchart ends (S806: YES); if not met, the process returns to S801 (S806: NO).

[0037] 9 is a flowchart illustrating the process by which the model generation computer 11 collects sampling data. This flowchart can be used when the sample generation unit 111 collects sampling data related to shape deviation in S802 or S803. Each step in FIG. 9 will be described below.

[0038] (Figure 9: Step S901) The sample generation unit 111 samples the radius value from a predetermined range (r1, r2). This step corresponds to obtaining the radius value of the solid circle in FIG. 3 in S802, and to obtaining the radius value of the region between the solid circle and the dotted circle in FIG. 4 in S803.

[0039] (Figure 9: Step S902) The sample generation unit 111 adjusts the radius according to the dimension of the shape deviation. This step is for adjusting the radius sampled in S901 so that it is not biased (for example, so that it is not too close to either r1 or r2).

[0040] (Figure 9: Steps S903 to S904) The sample generation unit 111 acquires a unit vector in an arbitrary direction in the data space to which the training data belongs (S903). The sample generation unit 111 calculates a perturbation vector by multiplying the radius calculated in S902 by the unit vector (S904). The perturbation vector is used to generate new sampling data by arbitrarily changing the reference data (training data in the right diagram of FIG. 3 or correction data in FIG. 4) used when collecting sampling data for re-learning.

[0041] (Figure 9: Steps S905 to S906) The sample generation unit 111 samples one piece of data from the data space (S905). This corresponds to sampling one piece of training data in S802, and corresponds to sampling one piece of correction data in S803. The sample generation unit 111 obtains new sampled data by adding the perturbation vector of S904 to the sampled data as a reference (S906).

[0042] 10 is a flowchart illustrating a procedure in which the control computer 12 controls the shape control device 14. Before carrying out this flowchart, it is assumed that the model generation computer 11 has already generated the corrected learning model C.

[0043] The control unit 121 receives data to be input to the learning device (each input illustrated in FIG. 1) (S1001). The control unit 121 determines whether the received input data is within the cutoff region described in FIG. 5 (i.e., determines whether the input data is within the cutoff distance from the modified data used to generate the modified learning model C) (S1002). If the input data is within the cutoff region, the control unit 121 overwrites the output from the differential model B to a value equal to or lower than a threshold (e.g., 0) (S1003). If the input data is not within the cutoff region, the process skips to S1004. The control unit 121 controls the shape control device 14 by transmitting to the shape control device 14 an operation amount obtained by inputting the input data to the modified learning model C (S1004).

[0044] <Third Embodiment> In the third embodiment of the present invention, an example of a configuration will be described in which the shape control system 1 described in the first and second embodiments evaluates the results of re-learning. As evaluation criteria, for example, the response rate, polarity agreement rate, model change rate, etc., which will be described below, can be used. The other configurations of the shape control system 1 are the same as those of the first and second embodiments.

[0045] FIG. 11 shows the calculation formulas for the response rate and polarity agreement rate. The response rate indicates whether the actuator of the shape control device 14 has responded to the correct answer data input to the controller model (learner), and whether the estimated value by the controller model and its correct answer value match. The polarity agreement rate indicates the degree to which the sign of the actuator's operation amount for the correct answer data input to the controller model matches the estimated value by the controller model and its correct answer value. The calculation formulas are just examples, and other calculation formulas may be used as long as they can calculate similar evaluation indicators.

[0046] Figure 12 shows the formula for calculating the model change degree. The model change degree represents the degree to which the learning module has changed before and after re-learning. The calculation formula is just an example, and other calculation formulas may be used as long as they can calculate similar evaluation indicators.

[0047] FIG. 13 shows an example of a user interface provided by the shape control system 1. The shape control system 1 (e.g., the model generation computer 11) calculates the evaluation indices described in FIGS. 11 and 12 and presents the results on a user interface such as that shown in FIG. 13. As an example, the response rate, polarity agreement rate, model change degree, etc. of the initial learning model A and the corrected learning model C can be presented. Furthermore, by listing similar evaluation indices for a conventional learning model that does not use the method of the present invention, it is also possible to compare the conventional model with the present invention. The user interface can be provided in any suitable form, for example, as a screen interface on a display provided on a computer, or any other suitable form.

[0048] <Fourth Embodiment> FIG. 14 shows an example of an objective function used when a learning device generated by the model generation computer 11 in any of the first to third embodiments performs machine learning. The learning device can be configured, for example, by a neural network. In this example, machine learning is performed to minimize the objective function. The calculation formula is just an example, and an appropriate objective function may be configured depending on the model that constitutes the learning device.

[0049] In the calculation formula in Figure 14, the coefficient λ1 term is obtained by the method explained in Figure 3, and the coefficient λ2 term is obtained by the method explained in Figure 4. By adjusting the coefficients λ1 and λ2, the weights of these two methods can be adjusted. If one of the coefficients is set to 0, only the other method will be used.

[0050] <Modifications of the present invention> The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and are not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

[0051] The procedure for sampling shape deviations has been explained in Figure 9. This procedure involves sampling data points from within a hypersphere centered on a reference point within the data space to which the shape deviations belong. On the other hand, inputs to the learning machine other than shape deviations can be sampled, for example, from a uniform distribution within a predetermined data space range.

[0052] In the above embodiment, the learning device is configured using a differential model B as shown in FIG. 2, but the present invention can also be implemented without using a differential model B. For example, the method described in FIG. 3 can be implemented by excluding training data near the modified data, and the method described in FIG. 4 can be implemented by sampling from an area somewhat distant from the modified data. To that extent, the initial learning model A itself may be updated with the modified data without using a differential model B. When the method described in FIG. 5 is implemented without using a differential model B, the output from the learning device when data near the modified data is input may be overwritten by the model before relearning. [Explanation of symbols]

[0053] A: Early learning model B: Differential model C: Modified learning model 1: Shape control system 11: Model generation calculator 12: Control computer 14: Shape control device

Claims

1. A shape control system for controlling a shape of an object, a shape control device including an actuator for controlling the shape; a control computer that instructs the shape control device on the amount of operation of the actuator; a model generation computer that generates, by machine learning, a controller model used by the control computer to calculate the manipulated variable; Equipped with the model generation computer is configured to modify the controller model with modification data used to modify the controller model; The model generation computer corrects the controller model so that a difference between an output from the controller model before the correction and an output from the controller model after the correction approaches zero when data sampled in a region that is at least a first threshold distance away from the correction data in a data space to which the correction data belongs is input to the controller model. A shape control system characterized by:

2. the model generation computer is configured to perform machine learning of the controller model using training data used to perform the machine learning; the model generation computer creates removed training data by removing a portion of the training data whose distance from the modified data in the data space is within the first threshold distance; The model generation computer modifies the controller model using data in the data space that is within a second threshold distance from the removed training data, so that the difference tends toward zero. The shape control system according to claim 1 .

3. The model generation computer corrects the controller model using data in a region in the data space where the distance from the corrected data is equal to or greater than the first threshold distance and is within a third threshold distance greater than the first threshold distance, so that the difference approaches zero. The shape control system according to claim 1 .

4. the model generation computer is configured to modify the controller model by adding a base model and a differential model; The model generation computer modifies the controller model by changing the parameters of the differential model without changing the parameters of the base model, so that the difference approaches zero. The shape control system according to claim 1 .

5. the model generation computer is configured to modify the controller model by adding a base model and a differential model; the control computer is configured to calculate the manipulated variable based on the shape; When calculating the manipulated variable, the control computer sets the output from the differential model to 0 if the distance from the modified data of the shape in the data space is equal to or greater than the first threshold distance. The shape control system according to claim 1 .

6. the model generation computer is configured to perform machine learning of the controller model using training data used to perform the machine learning; The model generation computer samples the vicinity of the correction data or the vicinity of the training data in the data space, and modifies the controller model using the sampled data. The shape control system according to claim 1 .

7. the controller model is configured to receive as input a deviation between a target shape and a current shape of the object; When sampling the deviation from within the data space, the model generation computer samples within a hypersphere centered on the correction data or within a hypersphere centered on the training data. The shape control system according to claim 6.

8. The shape control system further includes a terminal that provides a user interface used to input parameters that configure the controller model. The shape control system according to claim 1 .

9. The shape control system further includes a terminal that provides a user interface that displays the degree of change of the controller model before and after the correction. The shape control system according to claim 1 .

10. The shape control system further comprises a terminal providing a user interface that presents performance of the controller model; The user interface includes: a response rate indicating whether the actuator has responded to correct data input to the controller model, and whether the estimated value of the estimation result output by the controller model matches the correct value of the estimation result; a polarity agreement rate representing the degree to which the sign of the manipulated variable of the actuator with respect to the correct answer data input to the controller model matches between the estimated value of the estimation result output by the controller model and the correct answer value of the estimation result; Present at least one of the following: The shape control system according to claim 1 .

Images