Learning device for predictive control for controlled object, control device, machine learning program, control program, machine learning method, and control method

Abstract

The present invention improves reliability of predictive control for a controlled object. This learning device generates a pre-learning model (Mr) by a first learning algorithm using a first learning data set (Ds1). Each learning data item included in the first learning data set (Ds1) includes a measurement of a disturbance and a correction value corresponding to the disturbance. The learning device generates a predictive model (Mp) by a second learning algorithm using a second learning data set (Ds2) including an input to the pre-learning model (Mr) and an output of the pre-learning model (Mr) corresponding to the input. The output of the pre-learning model (Mr) includes specific information enabling definition of the reliability of the output.

Description

Learning device, control device, machine learning program, control program, machine learning method, and control method for predictive control of a control object

[0001] The present disclosure relates to a learning device, a control device, a machine learning program, a control program, a machine learning method, and a control method for predictive control of a control object.

[0002] Conventionally, configurations for performing predictive control of a control target are known. For example, Japanese Patent Laid-Open Publication No. 2022-139055 (Patent Document 1) discloses a control system including a controller that controls a control amount of the control target for machining a workpiece by having the control target act on the workpiece via a 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. The predictor is constructed based on a first prediction model that receives a disturbance as an input and outputs a set value for setting a quality characteristic value to a quality characteristic target value.

[0003] Patent Literature 1 discloses a configuration in which a first prediction model is generated by a support vector machine and a predictor has a second prediction model generated by a decision tree model. In machine learning for the second prediction model, a pair of a virtual value of a disturbance and a virtual value of a setting value output by the first prediction model for the virtual value of the disturbance is used as training data. Because the training data generated by the first prediction model is high-quality training data for machine learning, an algorithm that does not require a large amount of computational resources can be used in machine learning for the second prediction model.

[0004] Japanese Patent Application Laid-Open No. 2022-139055

[0005] In controlling a controlled object, it is desirable to perform predictive control only when the output of a predictive model is deemed reliable, from the viewpoints of ensuring product quality and avoiding unexpected behavior of the controlled object, etc. However, Patent Document 1 does not take into consideration the reliability of predictions made by the predictive model.

[0006] The present disclosure has been made to solve the above-mentioned problems, and its purpose is to improve the reliability of predictive control of a control target.

[0007] A learning device according to one aspect of the present disclosure is a learning device for machine learning of a prediction model that predicts a correction value of a manipulated variable for a control target subjected to a disturbance. The learning device includes a first storage unit and a learning unit. A first learning data set is stored in the first storage unit. The learning unit generates a pre-training model by a linear learning algorithm using the first training data set. Each piece of training data included in the first training data set includes a measured value of a disturbance and a correction value corresponding to the disturbance. The learning unit generates a prediction model by a non-linear learning algorithm using a second training data set including an input to the pre-training model and an output of the pre-training model corresponding to the input. The output of the pre-training model includes specific information that can define the reliability of the output.

[0008] According to this disclosure, the output of a predictive model includes specific information that can define the reliability of the output, thereby improving the reliability of predictive control of a controlled object.

[0009] In the above disclosure, the learning unit may identify a first range in which the reliability is greater than a first threshold and a second range in which the reliability is less than a second threshold within the range of input to the prediction model, and output the first range and the second range.

[0010] According to this disclosure, in the operation of predictive control, the reliability of predictive control can be further improved by not performing predictive control outside the first range. Also, in the operation of predictive control, the storage of learning data can be limited to the data storage range, thereby making it possible to efficiently collect learning data for predictive control.

[0011] In the above disclosure, all of the inputs to the pre-trained model included in the second training data set may be included in the first range.

[0012] According to this disclosure, by including all learning data in the first range in which predictive control is performed, it is possible to further improve the reliability of the predictive control range of the trained prediction model.

[0013] In the above disclosure, the linear learning algorithm may include Gaussian process regression, the specific information may include a Gaussian distribution, and the non-linear learning algorithm may include a decision tree.

[0014] According to this disclosure, the linear learning algorithm for the pre-training model includes Gaussian process regression, and the nonlinear learning algorithm for the predictive model includes a decision tree, thereby achieving high speed, reliability, and efficiency of predictive control.

[0015] A control device for controlling a control target using a prediction model generated by the learning device disclosed above may include a control unit and a second storage unit. The control unit may control the control target by correcting an manipulated variable using an output of the prediction model when the reliability is greater than a first threshold, and may control the control target without correcting the manipulated variable when the reliability is less than the first threshold.

[0016] According to this disclosure, predictive control is not performed outside the first range, thereby further improving the reliability of predictive control.

[0017] In the above disclosure, the control unit may store the input and output of the prediction model in the second storage unit when the reliability is less than a second threshold.

[0018] According to this disclosure, the storage of learning data is limited to the data storage range, thereby making it possible to efficiently collect learning data for predictive control.

[0019] A control program for controlling a control object using a predictive model generated by the learning device disclosed above, when executed by a processor, may cause the processor to control the control object by correcting the manipulated variable using the output of the predictive model if the reliability is greater than a first threshold, and to control the control object without correcting the manipulated variable if the reliability is less than the first threshold.

[0020] According to this disclosure, predictive control is not performed outside the first range, thereby further improving the reliability of predictive control.

[0021] A control method for controlling a control object using a predictive model generated by a learning device in the above disclosure may include a step of correcting an operating variable using the output of the predictive model to control the control object when the reliability is greater than a first threshold, and a step of controlling the control object without correcting the operating variable when the reliability is less than the first threshold.

[0022] According to this disclosure, predictive control is not performed outside the first range, thereby further improving the reliability of predictive control.

[0023] A machine learning program according to another aspect of the present disclosure is a machine learning program for machine learning a predictive model that predicts a correction value of a manipulated variable for a controlled object subjected to a disturbance. When executed by a processor, the machine learning program causes the processor to generate a pre-trained model by a linear learning algorithm using a first training data set, where each training data item included in the first training data set includes a measured value of a disturbance and a correction value corresponding to the disturbance. The machine learning program further causes the processor to generate a predictive model by a non-linear learning algorithm using a second training data set including inputs to the pre-trained model and outputs of the pre-trained model corresponding to the inputs. The output of the pre-trained model includes specific information that can define the reliability of the output.

[0024] A machine learning method according to another aspect of the present disclosure is a machine learning method for machine learning a predictive model that predicts a correction value of a manipulated variable for a controlled object subjected to a disturbance. The machine learning method generates a pre-trained model by a linear learning algorithm using a first training data set. Each training data set included in the first training data set includes a measured value of a disturbance and a correction value corresponding to the disturbance. The machine learning method further includes a step of generating a predictive model by a non-linear learning algorithm using a second training data set including an input for the pre-trained model and an output of the pre-trained model corresponding to the input. The output of the pre-trained model includes specific information that can define a reliability of the output.

[0025] The learning device, control device, machine learning program, control program, machine learning method, and control method according to the present disclosure can improve the reliability of predictive control of a control target.

[0026] 11 is a block diagram illustrating a functional configuration of a control system according to a first embodiment. FIG. 12 is a diagram for explaining the flow of machine learning performed in the learning unit of FIG. 1. FIG. 13 is a diagram illustrating an example of a flow of processes performed by each of the learning unit, the feedback control system, the feedforward control system, and the learning data generation unit of FIG. 1. FIG. 14 is a flowchart illustrating an example of a specific processing flow of the machine learning processing of FIG. 3. FIG. 15 is a flowchart illustrating an example of a specific processing flow of the feedforward compensation value determination processing of FIG. 3. FIG. 16 is a flowchart illustrating an example of a specific processing flow of the learning data storage processing of FIG. 3. FIG. 2(c) is a diagram illustrating a learning dataset used in machine learning for the prediction model according to the first embodiment. FIG. 17 is a flowchart illustrating an example of a specific processing flow of the machine learning processing according to a modification of the first embodiment. FIG. 18 is a schematic diagram illustrating an example of a network configuration of a control system according to a second embodiment. FIG. 19 is a block diagram illustrating an example of a hardware configuration of the control device of FIG. 9. FIG. 19 is a schematic diagram illustrating an example of a hardware configuration of the support device of FIG. 9. FIG. 11 is a flowchart illustrating an example of a flow of a prediction model generation processing performed by a processor that executes the support program of FIG. 11. FIG. 12 is a diagram illustrating an example of a case where a time series data tab is selected on the GUI screen displayed in the processing for setting a range of quality characteristic values ​​of FIG. 12. FIG. 13 is a diagram illustrating an example of a case where a scatter plot matrix tab is selected on the GUI screen displayed in the processing for setting a range of quality characteristic values ​​of FIG. 16 is a diagram showing an example of a GUI screen displayed in the process of setting a range of quality characteristic values ​​in FIG. 12 when a feature table tab is selected. FIG. 17 is a diagram showing an example of a GUI screen displayed in the process of checking a predicted control range in FIG. 12. FIG. 18 is a diagram showing a scatter plot displayed on the GUI screen of FIG. 16 when the dimension of the explanatory variable is 1. FIG. 19 is a diagram showing a scatter plot displayed on the GUI screen of FIG. 16 when the dimension of the explanatory variable is 3. FIG. 19 is a diagram showing an example of a GUI screen displayed in the process of evaluating a prediction model in FIG. 12 when a prediction accuracy evaluation tab is selected. FIG. 20 is a diagram showing an example of a GUI screen displayed in the process of evaluating a prediction model in FIG. 12 when a quality characteristic value evaluation tab is selected.

[0027] Hereinafter, embodiments will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are designated by the same reference numerals, and their description will not be repeated in principle.

[0028] [First Embodiment] <Application Example> Fig. 1 is a block diagram showing the functional configuration of a control system 1 according to a first embodiment. As shown in Fig. 1, the control system 1 includes a control device 100 and a learning device 400. The control device 100 and the learning device 400 may be connected via a network (for example, the Internet or a cloud system) and may be located remotely from each other. Alternatively, the control device 100 and the learning device 400 may be configured as an integrated device.

[0029] The learning device 400 includes a learning unit 410 and a memory unit 420 (first memory unit). The memory unit 420 stores a pre-training model Mr, a prediction model Mp, a training dataset Ds1 (first training dataset), and a training dataset Ds2 (second training dataset).

[0030] The machine learning algorithm (first learning algorithm) used to construct the pre-training model Mr includes a linear learning algorithm (e.g., Gaussian process regression, ensemble learning, or support vector regression). The learning data set Ds1 is used in this machine learning algorithm. In Gaussian process regression, the mean value of the Gaussian distribution is used as the output (predicted value) of the pre-training model Mr, while in ensemble learning, the mean value of the predicted values ​​of multiple learning models (weak learners) is used as the output of the pre-training model Mr.

[0031] The machine learning algorithm (second learning algorithm) used to construct the prediction model Mp includes a nonlinear learning algorithm (e.g., a decision tree or a neural network). The learning dataset Ds2 is used in the machine learning algorithm. The output of the prediction model Mp includes specific information (e.g., a Gaussian distribution or multiple predicted values ​​of multiple learning models used in ensemble learning) that can define the reliability of the output. For example, the smaller the variance of the Gaussian distribution corresponding to the output of the prediction model Mp or the variance of the predicted values ​​of multiple weak learners, the higher the reliability of the output. If the inference time is defined as the time interval from when explanatory variables are input to a model to when a target variable is output, the inference time of the prediction model Mp is shorter than that of the pre-training model Mr. In other words, the prediction model Mp is faster than the pre-training model Mr in terms of predictive control.

[0032] The learning device 400 outputs the trained prediction model Mp and reliability information of the output of the prediction model Mp to the control device 100. The reliability information includes a range of inputs to the prediction model Mp (prediction control range) in which the reliability of the output of the prediction model Mp is greater than a prediction threshold (e.g., 0.5), and a range of inputs to the prediction model Mp (data storage range) in which the reliability of the output of the prediction model Mp is less than a storage threshold (e.g., 0.7).

[0033] The control device 100 includes a feedback control unit 110, a feedforward compensation unit 120 (control unit), a learning data generation unit 130 (control unit), a storage unit 140, a subtractor 150, and an adder 160. The storage unit 140 stores the prediction model Mp, control information Inc, and learning data set Ds1A received from the learning device 400. The control information Inc includes the predicted control range and data storage range received from the learning device 400.

[0034] The control device 100 outputs the sum of a feedback control amount rb and a feedforward compensation value rf to the control object 200 as a control amount r to the control object 200 so that a control amount q, which is an output value of the control object 200 subjected to a disturbance ds, approaches a target value qr. The feedforward compensation value rf corresponds to a correction value of the control amount r. The control device 100 and the control object 200 are connected via a network (for example, the Internet or a cloud system) and may be located remotely from each other.

[0035] The disturbance ds is a quantity that disturbs the state of a control system including the control device 100 and the controlled object 200. The disturbance ds includes a plurality of disturbance elements, such as the amount of light, voltage, current, and temperature that are accidentally or suddenly input to the controlled object. The plurality of disturbance elements can be determined as a plurality of factors that are expected to affect the quality characteristic value (control amount q) of the controlled object 200 based on previously obtained knowledge such as a characteristic diagram of the target device of the controlled object 200.

[0036] Hereinafter, the configuration including the feedback control unit 110 and the subtractor 150 will also be referred to as a feedback control system, and the configuration including the feedforward compensation unit 120 and the adder 160 will also be referred to as a feedforward control system.

[0037] The subtractor 150 outputs the error eq (=qr-q) between the target value qr and the controlled variable q to the feedback control unit 110. The feedback control unit 110 determines a feedback manipulated variable rb based on the error eq and outputs it to the adder 160. When the measured value do at each sampling timing is within the predicted control range, the feedforward compensation unit 120 predicts a feedforward compensation value rf from the measured value do of the disturbance ds using the prediction model Mp and outputs it to the learning data generation unit 130. When the measured value do at each sampling timing is within the predicted control range, the feedforward compensation unit 120 outputs the feedforward compensation value rf to the adder 160. When the measured value do at each sampling timing is not within the predicted control range, the feedforward compensation unit 120 sets the feedforward compensation value rf to 0 and outputs it to the adder 160. The adder 160 outputs the sum of the feedback manipulated variable rb and the feedforward compensation value rf to the controlled object 200 as the manipulated variable r.

[0038] The training data generation unit 130 generates a training data set Ds1A (second training data set) to be used in additional training of the prediction model Mp (updating the prediction model Mp). When the measurement value do at each sampling timing is included in the data storage range, the training data generation unit 130 acquires the measurement value do, the manipulated variable r, the feedforward compensation value rf, and the controlled variable q. The training data generation unit 130 stores training data including a combination of the measurement value do (explanatory variable) and the feedforward compensation value rf (objective variable) in the storage unit 140 as elements of the training data set Ds1A. The training data generation unit 130 may transmit training data to the training device 400 each time it generates it, or may transmit the training data set Ds1A collectively to the training device 400. The training data set Ds1A is used by the training device 400 in additional training of the prediction model Mp (updating the prediction model Mp). The structure of each piece of training data included in the training data set Ds1 of the training device 400 is the same as the structure of each piece of training data included in the training data set Ds1A.

[0039] FIG. 2 is a diagram illustrating the flow of machine learning performed by the learning unit 410 in FIG. 1 . In FIG. 2 , graphs (a), (b), (c), and (d) showing the correspondence between explanatory variables (input) and response variables (output) are also shown. In FIG. 2 , the curve Vt shows the true correspondence between the explanatory variables and response variables. In FIG. 2 , in order to clearly illustrate the features of the present disclosure, a case where the explanatory variables are one-dimensional is shown. The number of dimensions of the explanatory variables may be two or more. The explanatory variables x1, x2, x3, x4, x5, and x6 are in descending order of magnitude. The expected range of explanatory variables is from x1 to x6.

[0040] 2(a) is a graph showing a prediction model Mp9 according to a comparative example. The prediction model Mp9 has been trained using a decision tree that uses the training dataset Ds1. As shown in FIG. 2(a), in the range of explanatory variables in which the training data included in the training dataset Ds1 are distributed, the error between the output of the prediction model Mp9 and the true value is relatively small. However, in the range of explanatory variables in which the training data included in the training dataset Ds1 are not distributed, the error is relatively large.

[0041] Therefore, in order to distribute the training data across the entire range of expected explanatory variables, the training unit 410 of the training device 400 according to embodiment 1 first generates the trained pre-trained model Mr shown in FIG. 2(b) by Gaussian process regression using the training dataset Ds1. A linear learning algorithm such as Gaussian process regression makes it easier to reduce the error between the objective variable and the true value in the range of explanatory variables where the training data is not distributed compared to the case of using a nonlinear learning algorithm such as a decision tree. For example, as shown in FIG. 2(b), in the range of explanatory variables where the training data included in the training dataset Ds1 is not distributed, the error between the output of the pre-trained model Mr and the true value for the explanatory variables x3 and x6 is smaller than that shown in FIG. 2(a).

[0042] Next, as shown in FIG. 2( c), the learning unit 410 inputs multiple explanatory variables distributed throughout the entire range of the expected explanatory variables into the pre-training model Mr and obtains multiple objective variables output from the pre-training model Mr. The learning unit 410 generates multiple combinations (predicted surface data) in which the multiple explanatory variables correspond to the multiple objective variables, as a training dataset Ds2 (second training dataset). The multiple explanatory variables input to the pre-training model Mr are preferably mesh data that are evenly distributed within the range of the expected explanatory variables to prevent bias in the distribution of the multiple explanatory variables. FIG. 2( c) illustrates a case in which the multiple explanatory variables distributed throughout the entire range of the expected explanatory variables are mesh data.

[0043] When the linear learning algorithm for generating the pre-trained model Mr is Gaussian process regression, the output of the pre-trained model Mr includes a Gaussian distribution. The learning unit 410 defines the reliability of the output of the pre-trained model Mr based on the variance of the Gaussian distribution. The smaller the variance of the Gaussian distribution, the greater the reliability of the output of the pre-trained model Mr.

[0044] The lower the reliability of the output of the pre-training model Mr, the lower the reliability of the prediction accuracy and the greater the need for additional training. Therefore, in order to limit the cases in which predictive control is performed to a range in which the reliability of the explanatory variables is relatively high, the learning unit 410 generates the training data set Ds2 and identifies a range (first range) of explanatory variables having reliability greater than a prediction threshold (first threshold) as the predictive control range. In FIG. 2( c), the range of explanatory variables from x1 to x5 is the predictive control range. The prediction threshold can be determined appropriately based on, for example, actual machine experiments, simulations, product specifications, or manufacturing process control values.

[0045] The learning unit 410 specifies a range (second range) of explanatory variables with reliability lower than a storage threshold (second threshold) as the data storage range in order to limit the range in which learning data for additional learning is saved to a range in which the reliability of the explanatory variables is relatively low. In FIG. 2( c), the range of explanatory variables from x2 to x4 and from x5 to x6 are the data storage ranges. The storage threshold can be determined appropriately based on, for example, actual machine experiments, simulations, product specifications, or manufacturing process control values. Note that the magnitude relationship between the prediction threshold and the storage threshold is not particularly limited; the prediction threshold may be greater than the storage threshold, the prediction threshold may be smaller than the storage threshold, or the prediction threshold and the storage threshold may be the same.

[0046] The learning unit 410 generates a prediction model Mp by a decision tree using the learning data set Ds2. As shown in Fig. 2(d), the error between the output of the prediction model Mp and the true value for the explanatory variables x3 and x6 is smaller than that shown in Fig. 2(a).

[0047] According to the learning device 400, the predictive model Mp used in the operation of predictive control is trained using a nonlinear learning algorithm, thereby enabling high-speed predictive control. Furthermore, in the operation of predictive control, predictive control is not performed outside the predictive control range, thereby further improving the reliability of predictive control. Furthermore, in the operation of predictive control, learning data is saved only within the data storage range, thereby enabling more efficient collection of learning data for predictive control. In other words, according to the learning device 400, high-speed, reliable, and efficient predictive control can be achieved.

[0048] 3 is a flowchart showing an example of the flow of processing performed by each of the learning unit 410, the feedback control system, the feedforward control system, and the learning data generating unit 130 in FIG. 1. The routines corresponding to the flowcharts of the feedback control system and the feedforward control system are executed, for example, at each sampling time. The routines corresponding to the flowcharts of the learning data generating unit 130 are executed, for example, in response to the first execution of the routines corresponding to the flowcharts of the feedforward control system. Hereinafter, steps will simply be abbreviated as S.

[0049] 3, in S400, the learning unit 410 performs machine learning processing to generate a trained prediction model Mp, and then the process proceeds to S410. In S410, the learning unit 410 transmits the trained prediction model Mp, the predicted control range, and the data storage range to the control device 100, and then the process ends.

[0050] In S111, the subtractor 150 of the feedback control system calculates the error eq between the target value qr and the controlled variable q, and outputs the error eq to the feedback control unit 110. In S312, the feedback control unit 110 determines the feedback manipulated variable rb based on the error eq, outputs the determined feedback manipulated variable rb to the adder 160, and ends the process.

[0051] In S121, the feedforward compensation unit 120 of the feedforward control system determines a feedforward compensation value rf from the measurement value do and outputs it to the learning data generation unit 130 and the adder 160. In S122, the adder 160 outputs the sum of the feedback manipulated variable rb and the feedforward compensation value rf as the manipulated variable r to the control object 200, and the process ends.

[0052] The training data generation unit 130 saves the training data in S130 and ends the process. FIG. 4 is a flowchart showing an example of a specific processing flow of the machine learning process S400 of FIG. 3 . As shown in FIG. 4 , the training unit 410 generates a pre-training model Mr by Gaussian process regression using the training data set Ds1 in S401, and proceeds to S402. In S402, the training unit 410 generates mesh data consisting of multiple explanatory variables in an expected range for the explanatory variables input to the prediction model Mp so that the explanatory variables are evenly distributed throughout the entire expected range, and proceeds to S403. In S403, the training unit 410 identifies a predicted control range and a data storage range based on a Gaussian distribution included in the output of the pre-training model Mr when mesh data in the expected range is input to the pre-training model Mr, and proceeds to S404. In S404, the learning unit 410 inputs mesh data of an expected range into the pre-trained model Mr, generates a training data set Ds2 including a combination of the mesh data and the output of the pre-trained model, and proceeds to S405. In S405, the learning unit 410 generates a prediction model Mp by a decision tree using the training data set Ds2, and then ends the process.

[0053] 5 is a flowchart showing an example of a specific processing flow of the feedforward compensation value determination processing of FIG. 3. As shown in FIG. 5, in S1211, the feedforward compensation unit 120 predicts a feedforward compensation value rf from the measured value d o of the disturbance ds using the prediction model Mp, and proceeds to S1212. In S1212, the feedforward compensation unit 120 outputs the feedforward compensation value rf to the learning data generation unit 130, and proceeds to S1213. In S1213, the feedforward compensation unit 120 determines whether the measured value d o of the disturbance ds is within the predicted control range. If the measured value d o of the disturbance ds is within the predicted control range (YES in S1213), the feedforward compensation unit 120 proceeds to S1215. If the measured value do of the disturbance ds is not included in the predictive control range (NO in S1213), the feedforward compensation unit 120 sets the feedforward compensation value rf to 0 in S1214 and proceeds to S1215. In S1215, the feedforward compensation unit 120 outputs the feedforward compensation value rf to the adder 160 and ends the process.

[0054] 6 is a flowchart showing an example of a specific processing flow of the learning data storage process S130 in FIG. 3. As shown in FIG. 6, the learning data generation unit 130 determines in S131 whether the measured value do of the disturbance ds is included in the data storage range. If the measured value do is included in the data storage range (YES in S131), the learning data generation unit 130 stores the feedforward compensation value rf and the measured value do in the storage unit 140 in S132 and ends the processing. If the measured value do is not included in the data storage range (NO in S131), the learning data generation unit 130 ends the processing.

[0055] [Variation of First Embodiment] In the description of the first embodiment, a learning dataset used in machine learning for a prediction model Mp is generated using mesh data generated so that explanatory variables input to the prediction model Mp are evenly distributed throughout the entire expected range. Predictive control is not performed for explanatory variables that are not included in the predictive control range. Therefore, in machine learning for the prediction model Mp, the need for learning data corresponding to explanatory variables included in the predictive control range is relatively high, and the need for learning data corresponding to explanatory variables that are not included in the predictive control range is relatively low. Therefore, in a variation of the first embodiment, a learning dataset used in machine learning for the prediction model Mp is generated using mesh data generated so that explanatory variables are evenly distributed throughout the predictive control range. By using this mesh data in machine learning for the prediction model Mp, the reliability of the predictive control range of the trained prediction model Mp can be further improved.

[0056] 7 is a diagram illustrating FIG. 2(c) together with a learning dataset Ds2A used in machine learning for the prediction model Mp according to embodiment 1. In FIG. 7, explanatory variables x1, x2, and x4 to x6 are the same as explanatory variables x1, x2, and x4 to x6, respectively, in FIG. 2. That is, similar to FIG. 2(c), the range of explanatory variables from x1 to x5 is the predictive control range.

[0057] As shown in FIG. 7 , the learning unit 410 generates mesh data consisting of multiple explanatory variables evenly distributed throughout the entire predictive control range. The learning unit 410 inputs the mesh data to a pre-training model Mr and acquires multiple objective variables output from the pre-training model Mr. The learning unit 410 generates multiple combinations (prediction surface data) in which the multiple explanatory variables correspond to the multiple objective variables, as a training dataset Ds2A (second training dataset). Comparing the training datasets Ds2 and Ds2A, the training dataset Ds2A is more densely distributed in the predictive control range than the training dataset Ds2. By performing machine learning on the predictive model Mp using the training dataset Ds2A, the reliability of the trained predictive model Mp in the predictive control range can be further improved compared to when the training dataset Ds2 is used.

[0058] FIG. 8 is a flowchart showing an example of a specific processing flow of the machine learning process S400 according to the modification of the first embodiment. The flowchart shown in FIG. 8 is a flowchart in which S414 is added to the flowchart of FIG. 4 and S404 and S405 of FIG. 4 are replaced with S415 and S416, respectively. As shown in FIG. 8, the learning unit 410 executes S401 and S402 as in the first embodiment, and then in S414 generates mesh data composed of multiple explanatory variables included in the predicted control range so as to be distributed throughout the entire predicted control range, and the process proceeds to S415. In S415, the learning unit 410 inputs the mesh data of the predicted control range into the pre-training model Mr to generate a training data set Ds2A including a combination of the mesh data and the output of the pre-training model Mr, and the process proceeds to S416. In S416, the learning unit 410 generates a prediction model Mp using a decision tree that uses the training data set Ds2A, and then ends the process.

[0059] As described above, the learning device, control device, machine learning method, and control method according to the first embodiment and the modifications can improve the reliability of predictive control of a control target.

[0060] Second Embodiment In a second embodiment, as an example of the control device according to the first embodiment, a configuration in which the control device includes a PLC (Programmable Logic Controller) will be described.

[0061] <Network Configuration Example of Control System> Fig. 9 is a schematic diagram showing a network configuration example of a control system 2 according to embodiment 2. As shown in Fig. 9, the control system 2 includes a device group in which a plurality of devices are configured to be able to communicate with each other. Typically, the devices may include a control device 100 that is a processing entity that executes a control program, and peripheral devices connected to the control device 100. The control device 100 in Fig. 9 has the same functional configuration as the control device 100 shown in Fig. 1.

[0062] The control device 100 corresponds to an industrial controller that controls control targets such as various facilities or devices. The control device 100 is a type of computer that executes control calculations and typically includes a PLC (Programmable Logic Controller). The control device 100 is connected to field devices 200 via a field network 20. The control device 100 exchanges data with at least one field device 200 via the field network 20.

[0063] The control calculations executed by the control device 100 include a process of collecting data collected or generated in the field device 200, a process of generating data such as command values ​​(operation variables) for the field device 200, and a process of transmitting the generated output data to the target field device 200. The data collected or generated in the field device 200 includes data related to disturbances input to the field device 200 and control variables resulting from the actual operation of the field device 200 in accordance with the command values. The command value for the field device 200 is determined by adding a feedforward compensation value predicted from the disturbance using a prediction model to a control variable provisionally calculated based on the error between a control target value (target value) calculated based on a control program executed by the control device 100 and the actual control variable.

[0064] The field network 20 preferably employs a bus or network that performs periodic communication. Known examples of such buses or networks that perform periodic communication include EtherCAT (registered trademark), EtherNet / IP (registered trademark), DeviceNet (registered trademark), and CompoNet (registered trademark). EtherCAT (registered trademark) is preferred because it guarantees data arrival time.

[0065] Any field device 200 can be connected to the field network 20. The field device 200 includes an actuator that exerts some physical action on a robot or conveyor in the field, and an input / output device that exchanges information with the field.

[0066] In the control system 2, a field device 200 includes a plurality of servo drivers 220_1 and 220_2 and a plurality of servo motors 222_1 and 222_2 connected to the plurality of servo drivers 220_1 and 220_2, respectively. The field device 200 is an example of the control target 200 in Fig. 1. The control target may include, for example, a sensor or an arm robot.

[0067] The servo drivers 220_1 and 220_2 drive the corresponding one of the servo motors 222_1 and 222_2 in accordance with a command value (for example, a position command value or a speed command value) from the control device 100. In this manner, the control device 100 can control the field device 200.

[0068] The control device 100 is also connected to other devices via a higher-level network 32. The higher-level network 32 is connected to the Internet 900, which is an external network, via a gateway 700. The higher-level network 32 may employ Ethernet (registered trademark) or EtherNet / IP (registered trademark), which are common network protocols. More specifically, at least one server device 600 and at least one display device 500 may be connected to the higher-level network 32.

[0069] The server device 600 may be a database system or a manufacturing execution system (MES). The manufacturing execution system acquires information from controlled manufacturing equipment or facilities to monitor and manage the entire production, and may also handle order information, quality information, shipping information, and the like. In addition to these, a device providing information services may be connected to the upper network 32. Possible information services include processing that acquires information from controlled manufacturing equipment or facilities and performs macro or micro analysis. For example, possible information services include data mining that extracts some characteristic trend contained in information from controlled manufacturing equipment or facilities, or machine learning tools that perform machine learning based on information from controlled facilities or machines.

[0070] The display device 500 receives operations from the user, outputs commands to the control device 100 in response to the user operations, and graphically displays the results of calculations performed by the control device 100, etc.

[0071] A support device 400 can be connected to the control device 100. The support device 400 may be connected to the control device 100 via the higher-level network 32 or the Internet 900. The support device 400 is a device that assists the control device 100 in making the preparations necessary for controlling the control target. Specifically, the support device 400 provides a development environment (such as a program creation and editing tool, a parser, and a compiler) for programs executed by the control device 100, a setting environment for setting configuration information (configuration) for the control device 100 and various devices connected to the control device 100, a function for outputting generated programs to the control device 100, a function for online correction and modification of programs executed on the control device 100, and a machine learning function for the prediction model Mp. The support device 400 has a functional configuration similar to that of the learning device 400 shown in FIG. 1.

[0072] In the control system 2, the control device 100, the support device 400, and the display device 500 are each configured as separate entities, but a configuration may be adopted in which all or part of these functions are integrated into a single device.

[0073] The control device 100 is not limited to being used only at one production site, but may also be used at other production sites, and may also be used on multiple different lines within one production site.

[0074] <Example of Hardware Configuration of Control Device> Fig. 10 is a block diagram showing an example of the hardware configuration of the control device 100 of Fig. 9. As shown in Fig. 14, the control device 100 includes a processor 302, a main memory 304, a storage 140, a memory card interface 312, a host network controller 306, a field network controller 308, a local bus controller 316, and a USB (Universal Serial Bus) controller 370 that provides a USB interface. These components are connected via a processor bus 318.

[0075] 10 , the processor 302 corresponds to an arithmetic processing unit that executes control calculations, and is configured with a CPU (Central Processing Unit) and / or a GPU (Graphics Processing Unit), etc. Specifically, the processor 302 reads out a program stored in the storage 140, expands it in the main memory 304, and executes it to realize control calculations for a control target.

[0076] The main memory 304 is configured with a volatile storage device such as a dynamic random access memory (DRAM) and / or a static random access memory (SRAM). The storage 140 is configured with a non-volatile storage device such as a solid state drive (SSD) and / or a hard disk drive (HDD).

[0077] The storage 140 stores a control program Pc, a learning data set Ds1A, a prediction model Mp, and control information Inc. The storage 140 corresponds to the memory unit 140 in FIG. 1. The control program Pc includes a program for comprehensively controlling the control device 100 and realizing each function of the control device 100. That is, the processor 302 that executes the control program Pc realizes the feedback control system (feedback control unit 110 and subtractor 150), the feedforward control system (feedforward compensation unit 120 and adder 160), and the learning data generation unit 130 in FIG. 1.

[0078] The memory card interface 312 accepts a memory card 314, which is an example of a removable storage medium. The memory card interface 312 is capable of reading and writing any data from and to the memory card 314.

[0079] The host network controller 306 exchanges data with any information processing device connected to the host network 32 via the host network 32 (for example, a local area network).

[0080] The field network controller 308 exchanges data with any device such as the servo motors 222_1 and 222_2 via the field network 20.

[0081] The local bus controller 316 exchanges data with any functional unit 380 constituting the control device 100 via the local bus 122. The functional unit 380 may include, for example, an analog I / O unit responsible for inputting and / or outputting analog signals, a digital I / O unit responsible for inputting and / or outputting digital signals, and a counter unit that receives pulses from an encoder or the like.

[0082] The USB controller 370 exchanges data with any information processing device via a USB connection. The USB controller 370 is connected to, for example, a support device 400.

[0083] Fig. 11 is a schematic diagram showing an example of the hardware configuration of the support device 400 shown in Fig. 9. As an example, the support device 400 is realized by a computer that follows a general-purpose architecture executing a program.

[0084] 11, the support device 400 includes a processor 430, a volatile memory 404, a non-volatile memory 420, a display 450, a keyboard 411, a mouse 412, a memory card interface 414, and an external device interface 418. These components are connected to each other via a processor bus 408 so as to be able to communicate with each other.

[0085] The processor 430 is composed of a CPU, a GPU, etc., and performs various processes by reading a program (e.g., a support program Psp) stored in the non-volatile memory 420, expanding it in the volatile memory 404, and executing it. The support program Psp includes a machine learning program for the prediction model Mp. In other words, the processor 430 that executes the support program Psp realizes the learning unit 410 in FIG. 1. The non-volatile memory 420 corresponds to the storage unit 420 in FIG. 1.

[0086] In the support device 400, the processor 430 executes a predetermined program to provide the functions required for the support device 400. Some or all of the functions may be implemented using a dedicated hardware circuit (for example, an ASIC or an FPGA).

[0087] The volatile memory 404 is configured by various volatile storage devices such as DRAM, SRAM, etc. The non-volatile memory 420 is configured by non-volatile storage devices such as SSD and / or HDD, for example.

[0088] The non-volatile memory 420 stores an operating system (OS) for realizing basic functions, as well as a support program Psp, training data sets Ds1 and Ds2, a pre-training model Mr, and a prediction model Mp for providing the functions of the support device 400. The support program Psp defines instructions for causing a computer to function as the support device 400.

[0089] The display 450 outputs processing results and the like from the processor 430. The keyboard 411 and mouse 412 accept user operations when creating various programs such as user programs or when the user inputs setting values ​​and the like according to the object to be controlled.

[0090] The memory card interface 414 is configured so that the memory card 115 can be attached and detached, and various data (such as user programs) can be written to the memory card 115 and various data can be read from the memory card 115 .

[0091] The external device interface 418 exchanges data with any external device such as the control device 100 via a network.

[0092] Although not shown in the figure, the support device 400 may be equipped with an optical drive, and a program (such as a support program) stored in an optical recording medium such as a DVD (Digital Versatile Disc) that non-transiently stores computer-readable programs is read and installed in the non-volatile memory 420.

[0093] The support program Psp and the like executed in the support device 400 may be installed via a computer-readable DVD, or may be installed by downloading from a server device 600 or the like on a network. In addition, the functions provided by the support device 400 may be realized by using some of the modules provided by the OS.

[0094] FIG. 12 is a flowchart illustrating an example of the flow of the predictive model generation process performed by the processor 430 executing the support program Psp of FIG. 11 . As shown in FIG. 12 , in S101, the processor 430 loads the training dataset Ds1. For example, the processor 430 loads feature data from a pre-specified path (target folder). If the target folder contains multiple comma separated values ​​(CSV) files, the processor 430 merges and loads the CSV files present in the first three levels below the target folder in chronological order. If the sampling function, the number of feature data files present in the target folder, or the file size is large, the processor 430 samples the data to a readable size. The processor 430 loads the roles of each variable (quality characteristic value, objective variable, and explanatory variable) assigned by the event definition file.

[0095] In S102, the processor 430 cleanses the training dataset Ds1. For example, if there is missing data in the data read in S101 (read data), the processor 430 skips the corresponding record (row) and displays the number of missing data on the history screen. If there is a type conversion error in the read data, the processor 430 skips the corresponding record (row) and displays the number of type conversion errors on the history screen. A type conversion error can occur, for example, when a record contains data other than numeric values.

[0096] In S103, the processor 430 displays a GUI (Graphical User Interface) screen on which the range of quality characteristic values ​​of the training data used in machine learning can be set. After S103, the processor 430 displays a GUI screen on which the range of explanatory variables (prediction control range) using the prediction model Mp can be confirmed in S104. After S104, the processor 430 displays a GUI screen on which the evaluation results of the trained prediction model can be confirmed in S105. After S105, the processor 430 ends the prediction model generation process.

[0097] FIG. 13 is a diagram illustrating an example of a GUI screen 810 displayed in the quality characteristic value range setting process S103 of FIG. 12 when the time-series data tab is selected. As shown in FIG. 13 , the GUI screen 810 displays time-series data for a quality characteristic value (product thickness) and two explanatory variables (material thickness and mold temperature). The user can toggle the display of the time-series data for the quality characteristic value and the explanatory variables by entering values ​​into checkboxes on the left side of the screen 810. The user can set a training range 811 for the quality characteristic value of the training data used for machine learning in the training dataset Ds1 by entering values ​​into a numerical value setting control 812.

[0098] The user can determine a range of quality characteristic values ​​appropriate for machine learning for the predictive model Mp based on the number of training data (learning data), the estimated Cpk (process capability index) when the predictive model Mp is applied to predictive control (after application of the predictive model), and the training data coverage. When the training range 811 is narrowed, the estimated Cpk improves, but the number of training data decreases. As a result, the training data coverage of the trained predictive model Mp decreases. When the training range 811 is widened, the estimated Cpk decreases, but the number of training data increases. As a result, the training data coverage of the trained predictive model Mp increases. If the user's setting of the training range 811 results in the number of training data being below the minimum number required to generate the predictive model Mp, a warning is displayed on the GUI screen 810.

[0099] FIG. 14 is a diagram showing an example in which the scatter plot matrix tab is selected on the GUI screen 810 displayed in the quality characteristic value range setting process S103 of FIG. 12 . FIG. 14 illustrates the scatter plot matrix tab when the dimension of the explanatory variables is two. As shown in FIG. 14 , the scatter plot matrix tab displays a scatter plot 821 of training data indicated by circles and non-training data indicated by crosses. The scatter plot 821 is shown as a diagram showing the correspondence between two explanatory variables and is divided into multiple rectangles. A range 822 within the scatter plot 821 is the range (upper and lower limits) of the explanatory variables set in the event definition file. The training data coverage rate is the ratio of the number of rectangles in which training data is plotted to the number of rectangles included in the range 822.

[0100] FIG. 15 is a diagram showing an example of a GUI screen 810 displayed in the quality characteristic value range setting process S103 of FIG. 12 when the feature table tab is selected. As shown in FIG. 15 , the training data set Ds1 is shown as table data. Each training data item included in the training data set Ds1 is shown as a row (record) of the table data. Each row is classified into training data, non-training data, and deletion data by different colors. The user can change the classification of the training data corresponding to the selected row to training data, non-training data, or deletion data.

[0101] FIG. 16 is a diagram showing an example of a GUI screen 840 displayed in the predicted control range confirmation process S104 of FIG. 12 . Below, FIGS. 14 and 13 will also be referenced in addition to FIG. 16 as needed. As shown in FIG. 16 , the GUI screen 840 displays a scatter plot 841, the coverage rate before data expansion, and the coverage rate after data expansion. The scatter plot 841 is a diagram showing the correspondence between two explanatory variables and is divided into multiple rectangles. Training data indicated by circles, as in FIG. 14 , is plotted on the scatter plot 841. A range 842 within the scatter plot 841 is the range of the explanatory variables set in the event definition file. A range 843 within the range 842 corresponds to the predicted control range. A range 844 within the range 842 corresponds to the data storage range. The coverage rate before data expansion is the ratio of the number of rectangles in which training data is plotted to the number of rectangles included in the range 842, and corresponds to the training data coverage rate in FIG. 14 . The coverage rate after data expansion is the training data coverage rate corresponding to the learning dataset Ds2. That is, the coverage rate after data expansion is the ratio of the number of rectangles included in ranges 843 and 844 to the number of rectangles included in range 842. Depending on the coverage rate after data expansion, the user can select whether to output the prediction model Mp or return to screen 810 in FIG. 13 and reset the quality characteristic ranges.

[0102] Fig. 17 is a diagram showing a scatter plot displayed on the GUI screen 840 of Fig. 16 when the dimension of the explanatory variable is 1. Fig. 18 is a diagram showing a scatter plot displayed on the GUI screen 840 of Fig. 16 when the dimension of the explanatory variable is 3. Note that even when the dimension of the explanatory variable is 1 or 3, scatter plots corresponding to Figs. 17 and 18, respectively, are displayed on the GUI screen 820 of Fig. 8.

[0103] 19 is a diagram showing an example of a case where the prediction accuracy evaluation tab is selected on the GUI screen 850 displayed in the evaluation process S105 of the prediction model Mp in FIG. 12. As shown in FIG. 19, the prediction model Mp is evaluated by the root mean squared error (RMSE) between the predicted corrected value (predicted value) and the actually measured corrected value (actual value), and the coefficient of determination R2. The smaller the RMSE, the higher the prediction accuracy of the prediction model Mp. Furthermore, the closer R2 is to 1, the higher the prediction accuracy of the prediction model Mp.

[0104] 20 is a diagram showing an example of a case where the quality characteristic value evaluation tab is selected on the GUI screen 850 displayed in the evaluation process S105 of the prediction model Mp in FIG. 12. As shown in FIG. 20, the Cpk before and after application of the prediction model are shown. The user can compare the two and select whether to start operation of predictive control using the trained prediction model Mp or to reconstruct the prediction model Mp.

[0105] As described above, the learning device, control device, machine learning program, control program, machine learning method, and control method according to the second embodiment can improve the reliability of predictive control of a control target.

[0106] <Additional Notes> The present embodiment as described above includes the following technical ideas.

[0107] [Configuration 1] A learning device (400) for machine learning of a prediction model (Mp) that predicts a correction value (rf) of a manipulated variable (r) for a control target (200) subjected to a disturbance (ds), the learning device (400) comprising: a first memory unit (420) in which a first learning dataset (Ds1) is stored; and a learning unit (410) that generates a pre-trained model (Mr) by a linear learning algorithm using the first learning dataset (Ds1), wherein each piece of learning data included in the first learning dataset (Ds1) includes a measurement value of the disturbance (ds) and the correction value (rf) corresponding to the disturbance (ds), the learning unit (410) generates the prediction model (Mp) by a non-linear learning algorithm that uses a second learning dataset (Ds2) that includes an input to the pre-trained model (Mr) and an output of the pre-trained model (Mr) corresponding to the input, and the output of the pre-trained model (Mr) includes specific information that can define the reliability of the output.

[0108] [Configuration 2] The learning device (400) according to Configuration 1, wherein the learning unit (410) identifies a first range in which the reliability is greater than a first threshold and a second range in which the reliability is less than a second threshold within an input range of the prediction model (Mp), and outputs the first range and the second range.

[0109] [Configuration 3] The learning device (400) according to Configuration 2, wherein all of the inputs to the pre-trained model (Mr) included in the second training data set (Ds2) are included in the first range.

[0110] [Configuration 4] The learning device (400) according to any one of configurations 1 to 3, wherein the linear learning algorithm includes Gaussian process regression, the specific information includes a Gaussian distribution, and the non-linear learning algorithm includes a decision tree.

[0111] [Configuration 5] A control device (100) comprising: a control unit (120, 130) that controls the control object (200) using the prediction model (Mp) generated by the learning device (400) according to configuration 2 or 3; and a second storage unit (140), wherein the control unit (120, 130): when the reliability is greater than the first threshold, corrects the manipulated variable (r) using the output of the prediction model (Mp) to control the control object (200); when the reliability is less than the first threshold, controls the control object (200) without correcting the manipulated variable (r); and when the reliability is less than the second threshold, stores the input and output of the prediction model (Mp) in the second storage unit (140).

[0112] [Configuration 6] The control device according to Configuration 5, wherein the control unit (120, 130) stores the input and output of the prediction model (Mp) in the second storage unit (140) when the reliability is smaller than the second threshold.

[0113] [Configuration 7] A control program (Pc) for controlling the controlled object (200) using the prediction model (Mp) generated by the learning device (400) described in configuration 2 or 3, wherein the control program (Pc), when executed by a processor (302), causes the processor (302) to: control the controlled object (200) by correcting the manipulated variable (r) using the output of the prediction model (Mp) when the reliability is greater than the first threshold; and control the controlled object (200) without correcting the manipulated variable (r) when the reliability is less than the first threshold.

[0114] [Configuration 8] A control method for controlling the controlled object (200) using the prediction model (Mp) generated by the learning device (400) described in configuration 2 or 3, comprising: a step of correcting the manipulated variable (r) using an output of the prediction model (Mp) to control the controlled object (200) when the reliability is greater than the first threshold; and a step of controlling the controlled object (200) without correcting the manipulated variable (r) when the reliability is less than the first threshold.

[0115] [Configuration 9] A machine learning program (Psp) for machine learning of a prediction model (Mp) that predicts a correction value (rf) of a manipulated variable (r) for a control target (200) that is subjected to a disturbance (ds), wherein the machine learning program (Psp), when executed by a processor (430), causes the processor (430) to: generate a pre-trained model (Mr) by a linear learning algorithm using a first training data set (Ds1); each training data included in the first training data set (Ds1) includes a measurement value of the disturbance (ds) and the correction value (rf) corresponding to the disturbance (ds); and the machine learning program (Psp) further causes the processor (302) to: generate the prediction model (Mp) by a non-linear learning algorithm using a second training data set (Ds2) including an input to the pre-trained model (Mr) and an output of the pre-trained model (Mr) corresponding to the input; A machine learning program (Psp) in which the output of the pre-trained model (Mr) includes specific information that can define the reliability of the output.

[0116] [Configuration 10] A machine learning method for machine learning of a prediction model (Mp) that predicts a correction value (rf) of a manipulated variable (r) for a control object (200) subjected to a disturbance (ds), comprising: a step of generating a pre-trained model (Mr) by a linear learning algorithm using a first training data set (Ds1); each training data included in the first training data set (Ds1) includes a measurement value of the disturbance (ds) and the correction value (rf) corresponding to the disturbance (ds); the machine learning method further comprises a step of generating the prediction model (Mp) by a non-linear learning algorithm using a second training data set (Ds2) that includes an input to the pre-trained model (Mr) and an output of the pre-trained model (Mr) corresponding to the input; and the output of the pre-trained model (Mr) includes specific information that can define the reliability of the output.

[0117] The embodiments disclosed herein are intended to be implemented in appropriate combinations within the scope of compatibility. The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present invention is defined by the claims, not the above description, and is intended to include all modifications within the meaning and scope of the claims.

[0118] 1, 2 Control system, 20 Field network, 32 Upper network, 100 Control device, 110 Feedback control unit, 115, 314 Memory card, 120 Feedforward compensation unit, 122 Local bus, 130 Learning data generation unit, 140 Storage (storage unit), 150 Subtractor, 160 Adder, 200 Field device (controlled object), 222 Servo motor, 302, 430 Processor, 304 Main memory, 306 Upper network controller, 308 Field network controller, 312, 414 Memory card interface, 316 Local bus controller, 318, 408 Processor bus, 370 Controller, 380 Functional unit, 400 Support device (learning device), 404 Volatile memory, 410 Learning unit, 411 Keyboard, 412 Mouse, 418 External device interface, 420 Non-volatile memory (storage unit), 450 Display, 500 display device, 600 server device, 700 gateway, 810, 820, 840, 850 GUI screen, 811 training range, 812 numerical setting control, 821, 841 scatter plot, 822, 842 to 844 range, 900 Internet, Ds1, Ds1A, Ds2, Ds2A learning data set, Inc control information, Mp, Mp9 prediction model, Mr pre-learning model, Pc control program, Psp support program, Vt curve, do measured value, ds disturbance, eq error, q controlled variable, qr target value, r manipulated variable, rb feedback manipulated variable, rf feedforward compensation value.

Claims

1. A learning device for machine learning of a predictive model that predicts a correction value of a manipulated variable for a controlled object subjected to a disturbance, comprising: a first memory unit in which a first learning dataset is stored; and a learning unit that generates a pre-learning model by a first learning algorithm using the first learning dataset, wherein each piece of learning data included in the first learning dataset includes a measured value of the disturbance and the correction value corresponding to the disturbance, and the learning unit generates the predictive model by a second learning algorithm that uses a second learning dataset that includes an input to the pre-learning model and an output of the pre-learning model corresponding to the input, and the output of the pre-learning model includes specific information that can define the reliability of the output.

2. The learning device described in claim 1, wherein the learning unit identifies a first range in which the reliability is greater than a first threshold and a second range in which the reliability is less than a second threshold within the input range of the predictive model, and outputs the first range and the second range.

3. The learning device according to claim 2, wherein all of the inputs to the pre-trained model included in the second training data set are included in the first range.

4. A learning device according to any one of claims 1 to 3, wherein the first learning algorithm includes Gaussian process regression, the specific information includes a Gaussian distribution, and the second learning algorithm includes a decision tree.

5. A control device for controlling the controlled object using the predictive model generated by the learning device described in claim 2 or 3, comprising: a control unit; and a second memory unit, wherein the control unit controls the controlled object by correcting the manipulated variable using the output of the predictive model when the reliability is greater than the first threshold, and controls the controlled object without correcting the manipulated variable when the reliability is less than the first threshold.

6. The control device according to claim 5, wherein the control unit stores the input and output of the prediction model in the second storage unit when the reliability is less than the second threshold.

7. A control program for controlling the controlled object using the predictive model generated by the learning device described in claim 2 or 3, wherein when the control program is executed by a processor, the control program causes the processor to: if the reliability is greater than the first threshold, correct the manipulated variable using the output of the predictive model to control the controlled object; and if the reliability is less than the first threshold, control the controlled object without correcting the manipulated variable.

8. A control method for controlling the controlled object using the predictive model generated by the learning device described in claim 2 or 3, comprising: a step of correcting the manipulated variable using the output of the predictive model to control the controlled object when the reliability is greater than the first threshold; and a step of controlling the controlled object without correcting the manipulated variable when the reliability is less than the first threshold.

9. A machine learning program for machine learning of a predictive model that predicts a correction value of a manipulated variable for a controlled object subjected to a disturbance, wherein the machine learning program, when executed by a processor, causes the processor to: generate a pre-training model by a first learning algorithm using a first learning dataset, each piece of training data included in the first training dataset includes a measured value of the disturbance and the correction value corresponding to the disturbance, the machine learning program further causes the processor to: generate the predictive model by a second learning algorithm using a second training dataset including an input to the pre-training model and an output of the pre-training model corresponding to the input, and the output of the pre-training model includes specific information that can define the reliability of the output.

10. A machine learning method for machine learning a predictive model that predicts a correction value of a manipulated variable for a controlled object subjected to a disturbance, comprising: a step of generating a pre-trained model by a first learning algorithm using a first training data set, wherein each training data included in the first training data set includes a measured value of the disturbance and the correction value corresponding to the disturbance, the machine learning method further comprising: a step of generating the predictive model by a second learning algorithm using a second training data set including an input to the pre-trained model and an output of the pre-trained model corresponding to the input, the output of the pre-trained model including specific information that can define the reliability of the output.

Images