JP2022139055A - Control system, prediction model generation device, and computer program - Google Patents

Abstract

To automatically suppress deviation from a target state of processing results.SOLUTION: A control system (1) comprises a controller (13) controlling a controlled variable of a control target (mechanism unit 14) for processing a workpiece, and a predictor (15) constructed on the basis of a first prediction model which outputs a setting value for using disturbance (X) as input and a quality characteristic value (R) as a quality characteristic target value (Ro).SELECTED DRAWING: Figure 1

Description

The present disclosure relates to a control system, a prediction model generation device, and a computer program.

In the field of Factory Automation (FA), a control system that automatically controls a control target for processing a work (material) is widely used. In such a control system, feedback control is performed such that the controlled variable of the controlled object becomes a predetermined set value.

JP 2019-141869 A

However, a disturbance outside the control of the above control system may act on the workpiece or the like and affect the work as a result of machining the workpiece. To give a specific example, for example, disturbances such as temperature fluctuations in the factory act on tools such as tools and dies that are operated by the workpiece or the control object to process the workpiece, and the result of machining differs from the target state. may deviate.

Conventionally, when the result of processing deviates from the target state, a skilled worker corrects the above set values based on experience to adjust the workmanship. However, such an adjustment method depends on the individual, and it is desired to be able to automatically and appropriately adjust the set value of the control amount of the controlled object.

The present disclosure has been made in view of the above problems, and an object thereof is to realize a control system that can automatically suppress the deviation from the target state of the machining result of the workpiece due to disturbance. .

The present disclosure employs the following configurations in order to solve the above-described problems. A control system according to one aspect of the present disclosure includes a controller that controls a control amount of the controlled object for machining the workpiece by causing the controlled object to act on the workpiece through a tool according to a set value; or a predictor that corrects the set value according to a disturbance acting on at least one of the tools, wherein the predictor corrects the set value, the disturbance, and the workpiece as a result of the machining. Whether the quality characteristic value belongs to a first class larger than the quality characteristic target value plus a predetermined value or belongs to a second class smaller than the quality characteristic target value minus the predetermined value and the set value for setting the quality characteristic value as the quality characteristic target value by performing the first machine learning with the set of the classification result as the first teacher data. It is built on the first prediction model, which outputs

According to the above configuration, it is possible to realize a control system capable of automatically suppressing deviation from the target state of the machining result of the workpiece due to disturbance.

In the control system according to the aspect described above, the predetermined value may be 0 or a positive value. According to the above configuration, the above classification can be performed appropriately. In particular, by setting the predetermined value to a positive value, it becomes possible to easily determine the boundary of the classification.

In the control system according to the above aspect, the first prediction model may be a prediction model generated by a support vector machine. According to the above configuration, it is possible to appropriately generate the first prediction model.

In the control system according to the above aspect, the predictor may have the first prediction model. According to the above configuration, the configuration of the predictor is more specific, and it is possible to realize a control system that can automatically suppress deviation from the target state of the machining result of the workpiece due to disturbance. .

In the control system according to the above aspect, the predictor inputs the virtual value of the disturbance input to the first prediction model and the set value output by the first prediction model for the virtual value of the disturbance. By performing second machine learning using a pair of a virtual value and a virtual value as second teacher data, the disturbance is input and the set value for making the quality characteristic value the quality characteristic target value is output. , may have a configuration with a second prediction model. According to the above configuration, it is possible to construct the predictor with small resources.

In the control system according to the one aspect described above, the second prediction model may be a prediction model generated by a decision tree model. According to the above configuration, it is possible to appropriately generate the second prediction model.

A predictive model generation device according to one aspect of the present disclosure includes a controller that controls a control amount of the controlled object for machining the workpiece by causing the controlled object to act on the workpiece through a tool according to a set value; A predictive model generating device for generating a predictive model applicable to the predictor of a control system, comprising: a predictor that corrects the set value in accordance with a disturbance acting on at least one of the workpiece and the tool wherein the set value, the disturbance, and the quality characteristic value of the workpiece as a result of the machining belong to a first class larger than a value obtained by adding a predetermined value to the quality characteristic target value, or the quality characteristic The first machine learning is performed using a combination of a result of classification as to whether it belongs to a second class smaller than the value obtained by subtracting the predetermined value from the target value and a set of the first teacher data, thereby reducing the disturbance. A configuration is provided for generating a first prediction model that receives as input and outputs the set value for making the quality characteristic value the quality characteristic target value.

According to the above configuration, it is possible to realize a control system capable of automatically suppressing deviation from the target state of the machining result of the workpiece due to disturbance.

One aspect of the present disclosure is a computer program for causing a computer to function as the predictive model generation device according to the above aspect. According to the above configuration, it is possible to realize a control system capable of automatically suppressing deviation from the target state of the machining result of the workpiece due to disturbance.

According to one aspect of the present disclosure, it is possible to realize a control system capable of automatically suppressing deviation from a target state of a result of machining a workpiece due to disturbance.

1 is a block diagram showing a control system according to Embodiment 1; FIG. 1 is a block diagram showing a control system based on the prior art for comparison; FIG. 4 is a flow chart showing a method of constructing a predictor in the control system according to Embodiment 1; FIG. 4 is a diagram schematically showing an X (disturbance)-S (set value)-R (quality characteristic value) space; It is a figure for demonstrating classification. FIG. 4 is a diagram schematically showing an X (disturbance)-S (set value) space representing a decision boundary for class classification; FIG. 10 is a diagram showing data of processing results in which the quality characteristic target value Ro was the required value, plotted together in the X (disturbance)-S (set value) space representing the decision boundary of class classification. 9 is a flowchart for explaining a predictor generation method in the control system according to the second embodiment; FIG. 10 is a diagram for explaining second teacher data for generating a predictor in the control system according to Embodiment 2; FIG. 10 is a diagram showing an example of quality characteristic value results from a control system based on the prior art for comparison; FIG. 11 is a diagram showing an example of quality characteristic values obtained by the control system of the control system according to the second embodiment;

§1 Application Example First, an example of a scene to which the present disclosure is applied will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration example of a control system 1 related to an application example. The control system 1 includes a control system for machining the work WK. The control system 1 can be applied to the FA field.

The control system 1 includes a controller 13 and a mechanism section 14 that is controlled by the controller 13 . The controller 13 controls the amount of control of the mechanism part 14 (controlled object) for machining the workpiece WK by the mechanism part 14 (controlled object) acting on the workpiece WK through the tool TL according to the set value S. do. In addition, in FIG. 1, the work WK and/or the tool TL are expressed as a functional block "tool/work 20".

The control system 1 further includes a predictor 15 and a learning device 30 (prediction model generation device). The predictor 15 receives disturbance X that acts on at least one of the workpiece WK and the tool TL, and outputs a correction value D that corrects the set value S that is input to the controller 13 . The predictor 15 is constructed based on the first prediction model generated by the learner 30 .

The learning device 30 determines that the set value S input to the controller 13, the disturbance X, and the quality characteristic value R of the work WK as a result of the above-described machining are obtained by adding a predetermined value ε to the quality characteristic target value Ro. The result of classifying whether it belongs to a first class that is larger than the quality characteristic target value Ro or belongs to a second class that is smaller than the value obtained by subtracting a predetermined value ε from the quality characteristic target value Ro. Perform the first machine learning. Then, the learning device 30 generates a first prediction model that receives the disturbance X as an input and outputs a set value S for setting the quality characteristic value R as a result of machining the workpiece WK to the quality characteristic target value Ro. Generate.

Therefore, according to this application example, a control system capable of automatically suppressing the deviation from the target state of the machining result of the workpiece due to the disturbance, that is, the divergence of the quality characteristic value R from the quality characteristic target value Ro. can be realized.

§2 Configuration example [Embodiment 1]
<Overview of control system>
An embodiment of the present disclosure will be described in detail below. FIG. 1 is a block diagram showing a configuration example of a control system 1 according to the first embodiment. In the following, as a specific example to facilitate understanding of the configuration of the control system 1, a case in which the control system 1 includes a servo press that presses a workpiece using a die will be described as appropriate. show.

The control system 1 includes a control system 10 provided with a set value generator 11 , an adder 12 , a controller 13 , a mechanism section 14 and a predictor 15 . The mechanism section 14 is a control target of the controller 13 . The controller 13 controls the control amount of the mechanism section 14 for machining the work WK by the mechanism section 14 acting on the work WK through the tool TL. The controller 13 feedback-controls the mechanism section 14 to be controlled so that the control amount of the mechanism section 14 becomes a value corresponding to the set value S that is input.

The quality characteristic value R is an index for evaluating the quality of the machining result of the work WK. The quality characteristic value R is generally a multidimensional quantity, ie can be represented by a vector. Note that the quality characteristic value R may be a scalar quantity. When the quality characteristic target value Ro, which is the target value of the quality characteristic value R, is input to the setting value generation unit 11, the setting value generation unit 11 outputs the basic setting value So according to the quality characteristic target value Ro. The set value S input to the controller 13 is equal to the basic set value So output by the set value generator 11 when the correction value D is absent. When the set value S input to the controller 13 is corrected, the correction value D is added to the basic set value So by the adder 12 and input to the controller 13 as the set value S.

As shown in FIG. 1, it is assumed that at least one of the work WK and the tool TL is subjected to a disturbance X outside the feedback control by the controller 13 and the mechanism section 14 . The predictor 15 calculates a correction value D according to the disturbance X. The predictor 15 is constructed based on a prediction model (first prediction model) generated by the learning device 30 provided in the control system 1 . Details of the learner 30 and the predictor 15 will be described later.

In the example of a servo press, the mechanism section 14 is a mechanism composed of a servo motor, a slider, etc. in the servo press, and the control amount is the position of the slider, and the like. The controller 13 may be a servomotor controller, and the manipulated variable is a quantity relating to the motion that the servomotor controller instructs the servomotor.

The work WK is a material to be pressed. A tool TL is a die or the like attached to a slider. The quality characteristic value R can be, for example, a dimension such as a thickness of a predetermined portion of the workpiece WK after press working. In this case, the quality characteristic target value Ro may be the target value of the thickness.

As the disturbance X, the temperature of the mold can be exemplified. When the control system 10 repeatedly performs press working, the controller 13 controls the mechanism unit 14 according to the basic set value So output by the set value generation unit 11 according to the target value of thickness (target quality characteristic value Ro). It is assumed that In this case, the thickness of a predetermined portion of the workpiece WK may fluctuate as a result of press working due to the temperature of the mold.

It can be seen from the above example that the quality characteristic value R can thus be influenced by the disturbance X. However, in the control system 1 according to the first embodiment, the predictor 15 outputs the correction value D according to the disturbance X, thereby adjusting the set value S input to the controller 13, and the quality characteristic value due to the disturbance X Fluctuations in R can be suppressed.

<Comparative example, effect>
FIG. 2 is a diagram showing a control system 10P of a comparative example to which the prior art is applied, for comparison with the control system 10 of the control system 1 according to the first embodiment. For convenience of description, members having the same functions as those of the members described in the above embodiments are denoted by the same reference numerals, and description thereof will not be repeated. As illustrated, the control system 10P of the comparative example differs from the control system 10 in that the predictor 15 is not provided.

In the control system 10 of the comparative example, similarly to the above description, when the quality characteristic value R deviates from the quality characteristic target value Ro due to the influence of the disturbance X, the skilled worker W controls the quality characteristic value R, that is, the workpiece A correction value D is determined based on feeling and experience by judging the workmanship of WK processing. Then, the set value S input to the controller 13 is corrected to bring the quality characteristic value R closer to the quality characteristic target value Ro.

However, such a method is individualized and inefficient, and it has been desired to automate the determination of the correction value D. Moreover, it is not guaranteed that the correction value D is appropriate, and it is desired that the correction value D is properly calculated according to the disturbance X. The control system 1 according to the first embodiment solves these problems by providing the above configuration.

That is, according to the control system 1 according to the first embodiment, the deviation of the quality characteristic value R from the quality characteristic target value Ro due to the influence of the disturbance X is effectively suppressed. Since the correction value D is automatically adjusted by the predictor 15, the influence of the disturbance X is corrected immediately and efficiently. Furthermore, since the construction of the predictor 15 is performed using machine learning as will be described later, in the control system 1, the correction value D can be adjusted accurately.

<Construction of predictor>
Along the flow chart of FIG. 3, the construction method and principle of the predictor 15 of the control system 1 according to the first embodiment will be described in detail below with reference to FIGS. 4 to 7 as well. FIG. 3 is a flow chart showing how the predictor 15 is constructed.

First, the learning device 30 acquires from the control system 10 the quality characteristic target value Ro, the set value S, and the disturbance X for a large number of machining by the control system 10 . Below, the index indicating each processing is represented by the symbol i. In an initial state in which the predictor 15 is not in operation, the basic set value So may be acquired instead of the set value S (step S11).

Furthermore, the learning device 30 acquires the quality characteristic value R, which is the performance of machining of the workpiece WK for many machining operations by the control system 10 . Since the quality characteristic value R is evaluated by an inspection machine, a measuring device, etc., the learning device 30 directly or indirectly acquires the quality characteristic value R from these devices (step S12).

The quality characteristic value R varies depending on the set value S input to the controller 13, and is also affected by the disturbance X as described above. ). Here, the set value S is not limited to a one-dimensional quantity, and may be a multi-dimensional quantity, that is, a vector. Moreover, the disturbance X is not limited to a one-dimensional quantity, and may be a multi-dimensional quantity, that is, a vector.

FIG. 4 is a diagram schematically showing the XSR space, where for the sake of illustration, the quality characteristic value R, the set value S, and the disturbance X are each represented as one-dimensional quantities. . Note that the numerical values of each parameter are in arbitrary units. FIG. 4 schematically shows a surface representing the function R=R(S,X) in the XSR space. FIG. 4 also schematically shows a plane showing the condition R=Ro where the quality characteristic value R matches the quality characteristic target value Ro. The area where these two planes intersect shows the value of the setpoint S to keep the quality characteristic value R in line with the quality characteristic target value Ro when the disturbance X is applied.

Next, the learning device 30 classifies each processing i based on the acquired data for each processing i. The learning device 30 uses a predetermined value ε to determine that processing that satisfies the conditional expression R−Ro>ε belongs to class C1 (first class). That is, it is determined that the machining in which the quality characteristic value R exceeds the quality characteristic target value Ro belongs to class C1. Machining that satisfies the conditional expression R-Ro<-ε is determined to belong to class C2 (second class). That is, it is determined that the processing in which the quality characteristic value R is lower than the quality characteristic target value Ro belongs to class C2 (step S13).

FIG. 5 is a diagram showing the result of classification performed by the learning device 30 in the XS space. Also in FIG. 5, the set value S and the disturbance X are represented as one-dimensional quantities as described above, so the XS space is two-dimensional here. In FIG. 5, processing data belonging to class C1 and class C2 are plotted with white circles (open circles) and squares, respectively.

Further, in FIG. 5, machining data satisfying the conditional expression |R−Ro|≦ε that does not belong to either class C1 or class C2 is plotted with black dots. Machining that satisfies the conditional expression |R−Ro|≦ε corresponds to machining in which the quality characteristic value R matches the quality characteristic target value Ro within a range in which the error is the predetermined value ε. That is, it corresponds to the processing in which the workmanship was good.

Next, the learning device 30 uses machine learning (first machine learning) to obtain a prediction model (first prediction model) for determining the decision boundary B of two class classifications of class C1 and class C2 in the XS space. to generate. At that time, the learning device 30 converts a set of data of the set value S, the disturbance X, and the class (class C1 or class C2) for each processing i as plotted in FIG. training data).

As an algorithm for such machine learning (first machine learning), it is desirable to apply a support vector machine (SVM). This is because the support vector machine is known as machine learning with extremely high performance for generating two-class pattern classifiers (step S14).

FIG. 6 is a diagram representing the determination boundary B of class classification by the prediction model (first prediction model) thus generated by the learning device 30 on the XS space shown in FIG. A decision boundary B can correspond in XS space to a function f defined as S=f(X). 4 and 6, the procedure for obtaining the decision boundary B as described above consists of a surface representing the function R=R(S, X) in the XSR space of FIG. It can be understood that this corresponds to finding an area that intersects with the plane that satisfies the condition R=Ro.

Therefore, the generated prediction model (first prediction model) is a learning model that predicts the set value S for matching the quality characteristic value R with the quality characteristic target value Ro when the disturbance X is input. is. Also, the generated prediction model (first prediction model) corresponds to a computing unit that executes the function f as a result.

FIG. 7 is a diagram obtained by plotting processing data that satisfies the conditional expression |R−Ro|≦ε on the graph of FIG. 6 . Arrows in the figure are symbols for emphasizing the data. Whether or not the obtained decision boundary B is appropriate is judged by confirming whether or not the data of the machining with good performance matches the calculated decision boundary B, using a graph such as that shown in FIG. be able to. Further, it is understood from FIG. 7 that even if it is difficult to prepare a large number of processing data with good quality, the decision boundary B can be determined well by using the above procedure.

Subsequently, the learning device 30 sets a predictor for calculating the correction value D in the control system 10 using the generated prediction model (first prediction model). Here, it corresponds to the correction value D=f(X)-So (step S15).

As described above, in the control system 1 according to the first embodiment, when constructing a learning model for calculating a function f corresponding to a so-called regression curve, the quality characteristic value R is above or below the quality characteristic target value Ro. A procedure for classifying into two classes, ie, whether it is swinging or not, is used. Therefore, the decision boundary B can be determined satisfactorily even for a data set that has been processed well and has relatively little processing. As a result, even in a data set with relatively little processing that was good in performance, the learning model that predicts the set value S for matching the quality characteristic value R with the quality characteristic target value Ro is improved. can be built to

Furthermore, in the control system 1 according to the first embodiment, the quality characteristic value R skillfully utilizes a procedure for classifying into two classes, whether the quality characteristic value R is above or below the quality characteristic target value Ro, A support vector machine, which is a two-class pattern classifier, can be applied.

Therefore, according to the control system 1, in constructing the predictor 15 that obtains the correction value D from the disturbance X, a support vector machine known as machine learning with extremely high performance can be used. Therefore, it is possible to construct the predictor 15 that correctly predicts the correction value D from complicated teacher data.

Although the predetermined value ε may be 0, it may also be a positive value. By setting the predetermined value ε to an appropriate positive value, the class classification of each data in the teacher data (first teacher data) becomes clear, and machine learning (first machine learning) for generating a prediction model for determining the decision boundary B learning) can be performed more appropriately.

[Embodiment 2]
Other embodiments of the disclosure are described below. For convenience of description, members having the same functions as those of the members described in the above embodiments are denoted by the same reference numerals, and description thereof will not be repeated.

<Construction of predictor>
The control system 1 according to the second embodiment is the same as the first embodiment except that the prediction model (second prediction model) generated by the learner 30 and applied to the predictor 15 is different from that of the first embodiment. is. Along the flow chart of FIG. 8, the construction method and principle of the predictor 15 of the control system 1 according to the second embodiment will be described in detail below while also referring to FIG.

FIG. 8 is a flow chart showing a method of constructing the predictor 15 in the control system 1 according to the second embodiment. Steps S21 to S24 are the same as steps S11 to S14 in the first embodiment.

Subsequently, the learning device 30 uses the generated prediction model (first prediction model) to generate virtual data of a set of the disturbance X and the set value S as second teacher data. At that time, the learning device 30 selects the virtual values of the disturbance so that they are uniformly distributed within the possible range of the disturbance X. The virtual value of the setting value is the output value obtained by inputting the virtual value of the disturbance to the first prediction model (step S25).

FIG. 9 is a graph showing the second teacher data obtained in this manner on the XS space similar to FIG. In the drawing, as in FIG. 7, arrows indicate the positions where machining data that satisfies the conditional expression |R−Ro|≦ε, ie, machining data with good performance exists. As schematically shown in FIG. 9, the second teacher data are uniformly distributed with respect to the disturbance X. In FIG. Therefore, it is good quality as teacher data for machine learning. In addition, even if a large amount of processing data with good performance cannot be prepared, it is understood that good teacher data for machine learning can be prepared.

Next, the learning device 30 generates a second prediction model that receives the disturbance X and outputs the set value S by second machine learning using the second teacher data. Since the second teacher data is used, the set value S output by the second prediction model corresponds to the set value S for setting the quality characteristic value R to the quality characteristic target value Ro.

As the second machine learning algorithm, it is desirable to apply a decision tree model. This is because the second prediction model based on the decision tree model can be generated even if the amount of computation and resources for operations such as memory are relatively small. A Deep Binary Tree algorithm is an example of such a decision tree model (step S26).

Furthermore, the learning device 30 uses the generated second prediction model to set the predictor 15 for calculating the correction value D in the control system 10 (step S27).

As described above, in the control system 1 according to the first embodiment, the second machine learning is performed using the second teacher data generated by the first prediction model. Generate a second prediction model of Since the second teacher data is good quality teacher data for machine learning, the second machine learning can use an algorithm that does not require a large amount of computing resources. Therefore, for example, the second machine learning algorithm can be implemented in a programmable logic controller (PLC) or the like, and the control system 10 can be easily constructed at the FA site.

<Verification example>
10 and 11 are graphs showing the results of variation in the product quality value R when the control system 10P of the comparative example shown in FIG. 2 and the control system 10 according to the second embodiment are operated, respectively. Each figure shows the results when the control system is applied to press working by a servo press machine.

In each figure, the product thickness after press working of the workpiece is employed as the quality characteristic value R and the quality characteristic target value Ro. The correction amount D is a correction amount for the slider position, which is a control amount. In each figure, the horizontal axis of the graph is the working index i, which corresponds to the press working shot.

Each figure also shows the results of a series of tests in which press working shots were continuously performed multiple times. In each series of tests, successive press shots are performed to increase the temperature of the mold. The results shown in each figure were obtained by adjusting the correction value D to suppress the phenomenon in which the product quality value R tends to drift as the temperature of the mold rises.

In the case of FIG. 11, in which the predictor 15 adjusts the correction value D, adjustment of the correction value D is timely and appropriate compared to the case of FIG. 10, in which the correction value D is manually adjusted by the skilled worker W. It is clear that the variation of the quality characteristic value R is suppressed as a result.

[Example of realization by software]
The functional blocks of the control system 1 (particularly, the set value generator 11, the adder 12, the controller 13, the predictor 15, and the learner 30) are logic circuits (hardware) formed in an integrated circuit (IC chip) or the like. may be realized by , or may be realized by software.

In the latter case, the control system 1 comprises at least one computer that executes instructions of a program (computer program) that is software that implements each function. Each computer has, for example, one or more processors and a computer-readable recording medium storing the program. In the computer, the processor reads the program from the recording medium and executes it, thereby achieving the object of the present disclosure.

As the processor, for example, a CPU (Central Processing Unit) can be used. As the recording medium, a "non-temporary tangible medium" such as a ROM (Read Only Memory), a magnetic disk, a card, a semiconductor memory, a programmable logic circuit, or the like can be used. In addition, a RAM (Random Access Memory) for developing the above program may be further provided.

Also, the program may be supplied to the computer via any transmission medium (communication network, broadcast wave, etc.) capable of transmitting the program. Note that one aspect of the present invention can also be implemented in the form of a data signal embedded in a carrier wave in which the program is embodied by electronic transmission.

The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims, and embodiments obtained by appropriately combining technical means disclosed in different embodiments is also included in the technical scope of the present disclosure.

Reference Signs List 1 control system 10 control system 11 set value generator 12 adder 13 controller 14 mechanism (controlled object)
15 predictor 20 tool/work 30 learner (prediction model generation device)
D Correction value So Basic setting value S Setting value Ro Quality characteristic target value R Quality characteristic value X Disturbance

Claims (8)

a controller that controls the amount of control of the controlled object for machining the workpiece by causing the controlled object to act on the workpiece through the tool according to a set value;
a predictor that corrects the set value according to a disturbance acting on at least one of the workpiece or the tool;
The predictor is
the set value and the disturbance;
The quality characteristic value of the workpiece as a result of the machining belongs to a first class greater than a value obtained by adding a predetermined value to the quality characteristic target value, or is greater than the quality characteristic target value minus the predetermined value. By performing the first machine learning using the combination of the classification result as to whether it belongs to the second class, which is smaller, as the first teacher data,
A control system built based on a first predictive model that receives the disturbance as an input and outputs the set value for making the quality characteristic value the quality characteristic target value. 2. The control system of claim 1, wherein said predetermined value is 0 or a positive value. 2. The control system of claim 1, wherein the first predictive model is a support vector machine generated predictive model. The predictor is
4. A control system according to any preceding claim, comprising said first predictive model. The predictor is
A set of a virtual value of the disturbance input to the first prediction model and a virtual value of the set value output by the first prediction model for the virtual value of the disturbance is set as second teacher data, By performing the second machine learning,
4. The control system according to any one of claims 1 to 3, comprising a second predictive model that receives said disturbance and outputs said set value for making said quality characteristic value said quality characteristic target value. . 6. The control system of claim 5, wherein the second predictive model is a predictive model generated by a decision tree model. a controller that controls the amount of control of the controlled object for machining the workpiece by causing the controlled object to act on the workpiece through the tool according to a set value;
and a predictor that corrects the set value in response to a disturbance acting on at least one of the workpiece or the tool,
A prediction model generation device that generates a prediction model applicable to the predictor,
the set value and the disturbance;
The quality characteristic value of the workpiece as a result of the machining belongs to a first class that is greater than the quality characteristic target value plus a predetermined value, or is greater than the quality characteristic target value minus the predetermined value. By performing the first machine learning using the combination of the classification result as to whether it belongs to the second class, which is smaller, as the first teacher data,
A prediction model generation device for generating a first prediction model that receives the disturbance as an input and outputs the set value for making the quality characteristic value the quality characteristic target value. A computer program for causing a computer to function as the prediction model generation device according to claim 7.

Images