Information processing device, control device, and optimization method

Abstract

The system automatically optimizes the values ​​of hyperparameters in the thermal displacement prediction calculation formula. The information processing device includes: a parameter selection method determination unit that selects at least one of the hyperparameters in the thermal displacement prediction calculation formula as a first hyperparameter; a parameter selection unit that sets / changes the value of the first hyperparameter and fixes the values ​​of the remaining hyperparameters as the values ​​of second hyperparameters; a machine learning unit that generates a thermal displacement prediction calculation formula based on pre-acquired teacher data of thermal displacement, according to the value of the first hyperparameter; and a model evaluation unit that stores the error between the thermal displacement estimated by each thermal displacement prediction calculation formula and the measured thermal displacement as an evaluation value, along with the value of the first hyperparameter. The parameter selection method determination unit sets the value of the first hyperparameter that has the lowest evaluation value as the optimal value based on the history of the first hyperparameter value and the evaluation value.

Description

Technical Field

[0001] This invention relates to information processing apparatus, control apparatus, and optimization method. Background Technology

[0002] One of the main causes of machining errors in machine tools is the relative thermal displacement between the tool and the workpiece caused by the thermal expansion of the machine tool's mechanical components. More specifically, these components include, for example, the spindle, spindle assembly, bed, column, worktable, and tool. Heat generated by the spindle's rotation, the spindle drive motor, the coolant supplied to the tool from the coolant supply system, and the lubricating oil supplied to the spindle bearings, especially the thermal deformation of these components centered on the spindle, sometimes results in relative thermal displacement between the tool and the workpiece.

[0003] Regarding this, the following technique is known: The difference between the estimated value of the thermal displacement of a mechanical element calculated using a thermal displacement prediction formula based on the estimated thermal displacement of the mechanical element and the measured value of the thermal displacement of the mechanical element is repeatedly calculated. The coefficients related to the measured data in the thermal displacement prediction formula, except for those related to the time delay factor, are fixed to predetermined values. Machine learning is used to set the processing of the coefficients related to the time delay factor. Based on the difference, the coefficients related to the time delay factor are fixed to predetermined values, and machine learning is used to set the processing of the coefficients related to the measured data. Thus, a high-precision correction formula is derived with less computation, and a high-precision correction is performed based on the correction formula. For example, see Patent Document 1.

[0004] Existing technical documents

[0005] Patent documents

[0006] Patent Document 1: Japanese Patent Application Publication No. 2018-153902 Summary of the Invention

[0007] The problem that the invention aims to solve

[0008] Patent Document 1 discloses the following: using the set of temperature measurement data from temperature sensors and the measured value data of thermal displacement during temperature measurement as teacher data, machine learning is performed on the coefficients related to the time delay factor and the coefficients related to the measurement data of the thermal displacement prediction formula under pre-set conditions (for example, as described below, pre-set condition parameters such as the number of temperature sensors configured on the machine tool and the configuration position of the temperature sensors, the sampling time of the temperature sensor during temperature measurement, and where the time delay is traced back to in the past). thereby calculating the optimal thermal displacement prediction formula (hereinafter also referred to as the "thermal displacement model") under the pre-set conditions.

[0009] Hereinafter, such pre-set conditional parameters (e.g., parameters concerning the number and location of temperature sensors configured on the machine tool, the sampling time for temperature measurement by the temperature sensors, and where the time delay is traced back to in the past) are also referred to as hyperparameters (design variables).

[0010] That is, the optimal formula for predicting thermal displacement (thermal displacement model) can be calculated based on the values ​​of hyperparameters.

[0011] However, the combinations of these hyperparameters (design variables) are numerous (almost infinite). Trying all combinations to calculate the optimal thermal displacement prediction formula (thermal displacement model) for each hyperparameter value is impractical. Furthermore, it is unrealistic to derive a further optimal thermal displacement prediction formula (thermal displacement model) based on the set of optimal thermal displacement prediction formulas (thermal displacement models) for each hyperparameter value. For example, even assuming the user adopts a trial-and-error approach—trying good combinations and then trying the next combination based on a good one—it is difficult to uniquely determine the hyperparameter values. Additionally, generating a practical thermal displacement prediction formula (thermal displacement model) requires a significant amount of time (e.g., more than a week) for trial-and-error of hyperparameter combinations.

[0012] Therefore, it is desirable to automatically optimize the values ​​of the hyperparameters of the thermal displacement prediction formula (thermal displacement model).

[0013] Methods for solving problems

[0014] (1) One aspect of the information processing apparatus of this disclosure optimizes the values ​​of multiple hyperparameters included in the thermal displacement prediction formula when generating the thermal displacement prediction formula through machine learning. The thermal displacement prediction formula estimates the thermal displacement of the mechanical element based on a set of measurement data including temperature data of the mechanical element and its surroundings and / or the operating state data of the mechanical element of a machine tool having a mechanical element with thermal expansion. The information processing apparatus includes: a measurement data acquisition unit that acquires the measurement data set; a thermal displacement acquisition unit that acquires the measured value of the thermal displacement of the mechanical element; a storage unit that stores the measurement data set acquired by the measurement data acquisition unit as input data and the measured value of the thermal displacement of the mechanical element acquired by the thermal displacement acquisition unit as a tag and interconnects them as teacher data; a parameter selection method determination unit that selects at least one of the multiple hyperparameters as the optimization target as a first hyperparameter; and a parameter selection unit that sets / changes the value of the first hyperparameter and selects hyperparameters not selected by the parameter selection method determination unit. The hyperparameter is fixed as a second hyperparameter; the machine learning unit performs machine learning based on the combination of the values ​​of the first hyperparameter and the second hyperparameter, according to the measured data set and the measured value of the thermal displacement of the mechanical element, thereby generating the thermal displacement prediction formula based on the set / modified value of the first hyperparameter with the value of the second hyperparameter as a fixed value; and the model evaluation unit calculates the error between the thermal displacement estimated by inputting the measured data set into the thermal displacement prediction formula for each value of the first hyperparameter and the measured value of the thermal displacement of the mechanical element as an evaluation value, with the value of the second hyperparameter as a fixed value, and stores the calculated evaluation value in the storage unit corresponding to the value of the first hyperparameter; the parameter selection method determination unit uses the value of the second hyperparameter as a fixed value, and selects the value of the first hyperparameter with the smallest evaluation value as the optimal value based on the history of the value of the first hyperparameter and the evaluation value stored in the storage unit.

[0015] (2) One aspect of the control device of the present disclosure has an information processing device of (1).

[0016] (3) One aspect of the optimization method disclosed herein is implemented by a computer. When generating a thermal displacement prediction calculation formula through machine learning, the values ​​of multiple hyperparameters included in the thermal displacement prediction calculation formula are optimized. The thermal displacement prediction calculation formula estimates the thermal displacement of the mechanical element based on a set of measurement data including temperature data of the mechanical element and its surroundings, and / or motion state data of the mechanical element of a machine tool containing mechanical elements with thermal expansion. The optimization method includes: a measurement data acquisition step, acquiring the measurement data set; a thermal displacement acquisition step, acquiring the measured value of the thermal displacement of the mechanical element; a storage step, using the measurement data set acquired in the measurement data acquisition step as input data, and linking the measured value of the thermal displacement of the mechanical element acquired in the thermal displacement acquisition step as a tag and storing it as teacher data in a storage unit; a parameter selection method determination step, selecting at least one of the multiple hyperparameters as the optimization target as a first hyperparameter; a parameter selection step, setting / changing the value of the first hyperparameter, and in the parameter selection method determination step... In the first step, the unselected hyperparameter is fixed as the second hyperparameter; in the machine learning step, machine learning is performed based on the combination of the values ​​of the first and second hyperparameters, according to the measured data set and the measured value of the thermal displacement of the mechanical element, thereby using the value of the second hyperparameter as a fixed value and generating the thermal displacement prediction formula according to the value of the first hyperparameter; and in the model evaluation step, the value of the second hyperparameter is fixed, and an evaluation value representing the error between the thermal displacement estimated by inputting the measured data set into the thermal displacement prediction formula for each value of the first hyperparameter and the measured value of the thermal displacement of the mechanical element is calculated, the value of the second hyperparameter is fixed, and the calculated evaluation value is stored in the storage unit in correspondence with the value of the first hyperparameter; regarding the parameter selection method determination step, the value of the second hyperparameter is fixed, and based on the history of the first hyperparameter value and the evaluation value stored in the storage unit, the value of the first hyperparameter with the smallest evaluation value is taken as the optimal value.

[0017] Invention Effects

[0018] According to one method, the values ​​of hyperparameters in the thermal displacement prediction calculation formula (thermal displacement model) can be automatically optimized. Attached Figure Description

[0019] Figure 1 This is a functional block diagram illustrating an example of the functional structure of an information processing system according to one embodiment.

[0020] Figure 2This is a flowchart illustrating the optimization of the values ​​of the first and second hyperparameters of the thermal displacement prediction calculation formula (thermal displacement model) in the information processing device.

[0021] Figure 3 Yes Figure 2 The flowchart describes the detailed processing steps S14 and S16.

[0022] Figure 4 This is a diagram illustrating an example of the optimization process in step S14.

[0023] Figure 5 This is a diagram illustrating an example of the optimization process in step S16.

[0024] Figure 6 Yes Figure 2 The flowchart describes the detailed processing steps of the optimization process in step S15.

[0025] Figure 7 This is a diagram illustrating an example of the optimization process in step S15.

[0026] Figure 8 This is a diagram illustrating an example of the order in which four hyperparameters are optimized. Detailed Implementation

[0027] <One Implementation Method>

[0028] First, a general overview of this embodiment will be provided. In this embodiment, the information processing device selects at least one of the multiple hyperparameters to be optimized as the first hyperparameter when calculating the thermal displacement prediction formula (thermal displacement model) for the estimated mechanical elements of the machine tool. The information processing device sets / changes the value of the first hyperparameter, uses other unselected hyperparameters as second hyperparameters, and sets the value of the second parameter as a fixed value. Based on the set / changed value of the first hyperparameter, for example as described above, the set of temperature measurement data from the temperature sensor and the measured value data of thermal displacement at the time of temperature measurement are used as teacher data. Machine learning is performed on the coefficients related to the time delay element and the coefficients related to the measurement data of the thermal displacement prediction formula (thermal displacement model), thereby calculating the optimal thermal displacement prediction formula (thermal displacement model).

[0029] Furthermore, the information processing device stores the evaluation value calculated by the model evaluation unit (described later) in relation to the value of the first hyperparameter, based on the optimal thermal displacement prediction formula (thermal displacement model) calculated according to the value of the first hyperparameter. Thus, the information processing device can, based on the stored history of the first hyperparameter value and the evaluation value of the optimal thermal displacement prediction formula (thermal displacement model) calculated according to the first hyperparameter value, designate other unselected hyperparameters as second hyperparameters, set the value of the second parameter to a fixed value, and, when the value of the first hyperparameter is changed, set the value of the first hyperparameter when the evaluation value of the optimal thermal displacement prediction formula (thermal displacement model) calculated according to the changed first hyperparameter value as the optimal value of the first hyperparameter when other unselected hyperparameters are used as second hyperparameters and the value of the second parameter is set to a fixed value (referred to as the "first optimal value of the first hyperparameter").

[0030] Next, based on setting the value of the first parameter to the optimal value of the first hyperparameter for the first time, the value of the second hyperparameter is set / changed. According to the set value of the changed second hyperparameter, for example, as mentioned above, the set of temperature data measured by the temperature sensor and the measured value data of thermal displacement during temperature measurement are used as teacher data. The coefficients related to the time delay element and the coefficients related to the measurement data of the thermal displacement prediction calculation formula (thermal displacement model) are machine-learned. Thus, the optimal thermal displacement prediction calculation formula (thermal displacement model) is calculated.

[0031] Furthermore, similar to the aforementioned process, the information processing device stores the evaluation value calculated by the model evaluation unit (described later) in relation to the value of the second hyperparameter for the optimal thermal displacement prediction calculation formula (thermal displacement model) calculated according to the value of the second hyperparameter, based on the stored value of the second hyperparameter and the history of the evaluation value of the optimal thermal displacement prediction calculation formula (thermal displacement model) calculated according to the value of the second hyperparameter. When the value of the second hyperparameter is changed, the second hyperparameter value at which the evaluation value of the optimal thermal displacement prediction calculation formula (thermal displacement model) calculated according to the changed second hyperparameter value is minimized is set as the optimal second hyperparameter value when the value of the first parameter is set to the first optimal first hyperparameter value (referred to as the "first optimal second hyperparameter value").

[0032] The aforementioned process can also be repeated. Specifically, the value of the second parameter is set to the optimal value of the second hyperparameter in the first step, and the value of the first hyperparameter is set / changed again. For example, as described above, the set of temperature data measured by the temperature sensor and the measured value of thermal displacement during temperature measurement is used as teacher data. Machine learning is performed on the coefficients related to the time delay factor and the coefficients related to the measurement data in the thermal displacement prediction formula (thermal displacement model). Thus, the optimal thermal displacement prediction formula (thermal displacement model) is calculated.

[0033] Furthermore, the information processing device stores the evaluation value calculated by the model evaluation unit (described later) in relation to the value of the first hyperparameter, based on the optimal thermal displacement prediction formula (thermal displacement model) calculated according to the value of the first hyperparameter which has been changed. Thus, the information processing device can set the value of the second hyperparameter calculated in the aforementioned process as the optimal value of the second hyperparameter based on the history of the stored values ​​of the first hyperparameter and the evaluation values ​​of the optimal thermal displacement prediction formula (thermal displacement model) calculated according to the value of the first hyperparameter. When the value of the first hyperparameter has been changed, the value of the first hyperparameter at which the evaluation value of the optimal thermal displacement prediction formula (thermal displacement model) calculated according to the changed value of the first hyperparameter is minimized is taken as the optimal value of the first hyperparameter when the value of the second parameter is set to a fixed value (referred to as the "second optimal value of the first hyperparameter").

[0034] By repeating this process, under any pre-set conditions (e.g., a pre-set number of processing iterations, an evaluation value below a pre-set threshold, etc.), the optimal hyperparameter value obtained at that point in time can be set as the optimal value. Therefore, according to this embodiment, the problem of "automatically optimizing the value of hyperparameters" can be solved.

[0035] The above is a summary of this implementation method.

[0036] Next, the structure of this embodiment will be described in detail with reference to the accompanying drawings.

[0037] Figure 1 This is a functional block diagram illustrating an example of the functional structure of an information processing system according to one embodiment. For example... Figure 1 As shown, the information processing system 1 includes: a machine tool 10, a control device 20, and an information processing device 30.

[0038] The machine tool 10, control device 20, and information processing device 30 can also be directly connected to each other via a connection interface not shown. Alternatively, the machine tool 10, control device 20, and information processing device 30 can also be connected to each other via a network such as a LAN (Local Area Network). In this case, the machine tool 10, control device 20, and information processing device 30 may also have a communication unit (not shown) for communicating with each other through such a connection.

[0039] Furthermore, as described later, the information processing device 30 may also be included in the control device 20. Additionally, as described later, the control device 20 may also be included in the machine tool 10.

[0040] Machine tool 10 is a machine tool known to those skilled in the art. For example, machine tool 10 performs cutting operations on a workpiece disposed on machine tool 10 using tools such as cutting tools mounted on the spindle included in machine tool 10, according to operation commands from control device 20.

[0041] The control device 20 is a numerical control device known to those skilled in the art. It controls the machine tool 10 to perform predetermined machining operations, such as cutting, by sending control signals to the machine tool 10. The control device 20 stores multiple machining programs determined according to the machining content of the workpiece. Furthermore, the control device 20 reads and interprets the machining program for the predetermined machining operation, extracts the predetermined machining conditions (e.g., the frequency of spindle acceleration / deceleration, rotational speed, cutting load, and cutting time), and inputs position command data, etc., to the thermal displacement correction device 201. Based on the thermal displacement amount prediction calculation formula (thermal displacement model) generated by the information processing device 30 (described later), the control device 20 generates and sends control signals to the machine tool 10 from the thermal displacement correction device 201, thereby causing the machine tool 10 to execute the predetermined machining operation.

[0042] Furthermore, the thermal displacement correction device 201 can be, for example, a thermal displacement correction device such as Patent Document 1, which will be omitted in detail.

[0043] Additionally, the control device 20 can also extract machining conditions (e.g., the frequency of spindle acceleration / deceleration, rotational speed, cutting load, and cutting time) and output them to the information processing device 30, which will be described later. Furthermore, regarding rotational speed, cutting time, etc., the control device 20 can also output information obtained in real time from the spindle motor (not shown) included in the machine tool 10, etc., to the information processing device 30.

[0044] Furthermore, in the control device 20, multiple terminals for connecting to sensors mounted on the machine tool 10 are provided to acquire measurement data (e.g., temperature, displacement). Sensors connected to the control device 20 and installed on the machine tool 10 can be added or removed by plugging and unplugging the sensor cables relative to these terminals. Additionally, the configuration of the sensors installed on the machine tool 10 can be changed. Moreover, changing the sensor configuration can be done by removing the sensor from its original location on the machine tool 10 and adding it to the new location. Therefore, the number of temperature sensors and their configuration positions, which are considered hyperparameters, can be changed.

[0045] Furthermore, the control device 20 can also output measurement data from the sensor to the information processing device 30.

[0046] <Information Processing Device 30>

[0047] The information processing device 30 is a computer device known to those skilled in the art, such as... Figure 1 As shown, the system includes a control unit 310 and a storage unit 320. The control unit 310 further includes: a measurement data acquisition unit 311, a thermal displacement acquisition unit 312, a parameter selection method determination unit 313, a parameter selection unit 314, a machine learning unit 315, and a model evaluation unit 316. The storage unit 320 stores historical data 321 and measurement data 322.

[0048] <Storage Unit 320>

[0049] The storage unit 320 can be ROM (Read Only Memory), HDD (Hard Disk Drive), etc., and can also store history data 321 and measurement data 322 together with various control programs.

[0050] In the optimization process, for example as described later, the history data 321 is input into the set of measurement data obtained from the control device 20 into a thermal displacement prediction calculation formula (thermal displacement model) generated according to the value set for the hyperparameter as the optimization object. The verification error calculated based on the thermal displacement estimated by the thermal displacement prediction calculation formula (thermal displacement model) and the measured value of the thermal displacement obtained from the control device 20 is used as the evaluation value of the thermal displacement prediction calculation formula (thermal displacement model) and stored in correspondence with the value set for the hyperparameter.

[0051] Regarding the measurement data 322 measured according to the hyperparameter setting value as the optimization object, the measurement data set obtained by the measurement data acquisition unit 311 from the control device 20 is used as input data, and the measured value of the thermal displacement of the mechanical element obtained by the thermal displacement acquisition unit 312 from the control device 20 is used as a label. They are correlated with each other and stored as teacher data for machine learning of the optimal thermal displacement prediction calculation formula (thermal displacement model) for the hyperparameter setting value.

[0052] <Control Unit 310>

[0053] The control unit 310 includes a CPU (Central Processing Unit), ROM, RAM (Random Access Memory), CMOS (Complementary Metal-Oxide-Semiconductor) memory, etc., which are configured to communicate with each other via a bus, as is known to those skilled in the art.

[0054] The CPU is the processor that controls the overall information processing device 30. The CPU reads the system program and application program stored in the ROM via the bus, and controls the information processing device 30 as a whole according to the system program and application program. Thus, as... Figure 1 As shown, the control unit 310 is configured to perform the functions of the measurement data acquisition unit 311, the thermal displacement acquisition unit 312, the parameter selection method determination unit 313, the parameter selection unit 314, the machine learning unit 315, and the model evaluation unit 316. Various data, such as temporary calculation data and display data, are stored in RAM. Furthermore, the CMOS memory is configured as a non-volatile memory that is backed up by a battery (not shown) and maintains its stored state even when the power supply to the information processing device 30 is disconnected.

[0055] The measurement data acquisition unit 311 acquires a set of measurement data from the control device 20. Here, the measurement data may also include temperature data of the mechanical elements of the machine tool 10 and their surroundings as measured by temperature sensors. Furthermore, the measurement data may also include operational status data of the mechanical elements of the machine tool 10, specifically, physical property values ​​of parts where temperature sensors are not attached, such as the spindle rotation speed, coolant flow rate to the spindle, and lubricating oil quantity to the spindle bearings.

[0056] The thermal displacement acquisition unit 312 acquires, for example, the measured value of the thermal displacement of the mechanical elements of the machine tool 10 detected by the probe.

[0057] The parameter selection method determination unit 313 selects at least one hyperparameter of the optimization target from among the multiple hyperparameters included in the thermal displacement prediction calculation formula (thermal displacement model). Here, the thermal displacement prediction calculation formula (thermal displacement model) of this embodiment is illustrated by incorporating the time offset element of the measurement data.

[0058] More specifically, since there are multiple independent variables within the measurement data set, for example, an example is given of a thermal displacement prediction formula (thermal displacement model) obtained by incorporating the time offset factor of the measurement data into a polynomial set according to a multiple regression analysis based on a generalized linear model.

[0059] Furthermore, the thermal displacement prediction formula (thermal displacement model) is not limited to the examples shown. This invention can be applied to any thermal displacement prediction formula (thermal displacement model).

[0060] Let the estimated value of the thermal displacement at time t be Y(t), and let the temperature sensor at time t be X k The measured value is set as X. k In the case of (t), the formula for predicting the thermal displacement using the time offset element of the measured data as the time delay element (thermal displacement model) is as follows: Equation 1. Here, Δt k It is temperature sensor X k The sampling time of the measured value, coefficient T k This sets the delay coefficient for tracing back to a specific point in the past. Furthermore, when assuming a total of N temperature sensors pre-configured on machine tool 10, n is the number of temperature sensors used (1 ≤ n ≤ N). Note that n here represents the number of placement locations containing the temperature sensors. That is, the total number of combinations of selecting n sensors from a set of N different temperature sensors. N C n Each of the different configuration positions corresponds to each n.

[0061] Thus, the sampling time Δt k coefficient T k The number of temperature sensors (including their placement) *n* is a hyperparameter of the thermal displacement prediction calculation formula (thermal displacement model) in Formula 1. According to the present invention, the optimal hyperparameter values ​​can be automatically calculated in the thermal displacement prediction calculation formula (thermal displacement model) of Formula 1.

[0062] [Mathematical Expression 1]

[0063]

[0064] Here, a0, a1…a n The coefficients, b, are determined through multiple regression analysis. k0 b k1…b kTk It is a coefficient corresponding to the delay.

[0065] As mentioned above, the coefficient T corresponding to the delay amount k and temperature sensor X k The sampling time Δt of the measured value k These are hyperparameters (hereinafter also referred to as "time delay hyperparameters") related to the time delay element of the measured data. That is, based on the two time delay hyperparameters, the thermal displacement prediction calculation formula (thermal displacement model) in Formula 1 refers to the reference interval and the number of references for the time delay measured data. In other words, it means to trace back to where several data points were observed regarding the time delay.

[0066] On the other hand, temperature sensor X k Imagine multiple possibilities, but since (1) if all temperature sensors X are used k Therefore, the cost increases, and (2) it is explained that calculation formulas with many variables are prone to overlearning. Thus, not all temperature sensors X are included. k The value is taken into the thermal displacement prediction calculation formula (thermal displacement model) of mathematical formula 1. Therefore, the parameter selection unit 314, described later, selects the value from multiple temperature sensors X. k The combination selected is the one used in the thermal displacement prediction calculation formula (thermal displacement model) of mathematical formula 1. The temperature sensor X used in this combination... k The number of temperature sensors is n. Hereinafter, the hyperparameter of the number of temperature sensors n is also called the "hyperparameter of the temperature sensor combination".

[0067] Furthermore, the parameter selection method determination unit 313 may, for example, use the hyperparameter of time delay as the first hyperparameter, or the hyperparameter of temperature sensor combination as the second hyperparameter, and alternately select the first hyperparameter and the second hyperparameter as the optimization object.

[0068] The parameter selection unit 314 sets / changes the value of the hyperparameter of one party selected by the parameter selection method determination unit 313, and fixes the value of the hyperparameter of the other party not selected by the parameter selection method determination unit 313.

[0069] Specifically, the parameter selection unit 314, for example, when selecting a time delay hyperparameter as the first hyperparameter, sets / changes the temperature sensor X included in the time delay hyperparameter within a predetermined search range. k The sampling time Δt of the measured value k and the coefficient T corresponding to the delay amount kOn the other hand, the value of the hyperparameter of the temperature sensor combination is fixed as the second hyperparameter. Specifically, the parameter selection unit 314 may be fixed, for example, as the value of the hyperparameter of the temperature sensor combination set as the second hyperparameter, or as a combination predetermined by using the temperature sensor combination as the initial value (e.g., all temperature sensors), or as the combination where the evaluation value is minimized according to the history data 321.

[0070] Therefore, regarding the temperature sensor X as the first hyperparameter k The sampling time Δt of the measured value k For example, you can search for a range of 1 to 10 seconds, regarding the coefficient T corresponding to the delay. k The search can be performed on coefficients ranging from 1 to 10. However, the search range is not limited to this and can be appropriately determined based on factors such as the prediction accuracy of the thermal displacement prediction formula (thermal displacement model). For example, regarding temperature sensor X... k The sampling time Δt of the measured value k For example, you can also search within a range of 1 minute to 10 minutes, or 1 hour to 10 hours.

[0071] Furthermore, the parameter selection unit 314 can also set / change the value of the hyperparameter of the time delay, which is the first hyperparameter, within the search range using a full-search (grid search) method. Alternatively, the parameter selection unit 314 can also use random search, Bayesian optimization, etc., to set / change the value of the hyperparameter of the time delay, which is the first hyperparameter, within the search range. By using random search, Bayesian optimization, etc., the information processing device 30 can reduce computational costs. In this case, it is preferable to perform convergence determination (convergence of model performance) in all combination implementation decisions. However, if performance is not satisfied, the search range can be expanded and the implementation repeated.

[0072] On the other hand, when the parameter selection unit 314 selects a hyperparameter for the temperature sensor combination as the second hyperparameter, it sets / changes the value of the hyperparameter of the temperature sensor combination within the search range. In this case, the parameter selection unit 314 fixes the value of the hyperparameter that is the time delay, which is the first hyperparameter. Specifically, for example, the value of the hyperparameter that is the time delay, which is set as the first hyperparameter, can also be fixed to the value at which the evaluation value is minimized based on the history data 321.

[0073] Here, regarding the hyperparameters of the temperature sensor combination as the second hyperparameter, it is also possible to search for 1 to N temperature sensors X for each temperature sensor. k The range of combinations. Furthermore, the search range is not limited to this and can be appropriately determined based on factors such as the prediction accuracy of the thermal displacement prediction formula (thermal displacement model).

[0074] Furthermore, the parameter selection unit 314 can set / change the values ​​of the hyperparameters of the temperature sensor combination as the second hyperparameter within the search range using a full-search (grid search) method, or it can use random search, Bayesian optimization, etc., to set / change the values ​​of the hyperparameters of the temperature sensor combination as the second hyperparameter within the search range. By using random search, Bayesian optimization, etc., the information processing device 30 can reduce computational costs. In this case, it is preferable to perform convergence determination (convergence of model performance) during the full-combination implementation decision. However, if the performance is not satisfied, the temperature sensor X can be changed. k The pasting location, etc., and the process is repeated from the data collection point.

[0075] The machine learning unit 315 performs machine learning based on the measured values ​​of the measured data set and the measured thermal displacement of the mechanical elements, using a combination of the values ​​of the hyperparameters representing the time delay (first hyperparameter) and the hyperparameters representing the temperature sensor combination (second hyperparameter). The machine learning unit 315 generates a thermal displacement prediction calculation formula (thermal displacement model) based on the value of the hyperparameter representing the time delay of the optimized object after setting / changing the parameters (first hyperparameter) or the value of the hyperparameters representing the temperature sensor combination (second hyperparameter).

[0076] Specifically, the machine learning unit 315 uses the value of the hyperparameter of time delay (first hyperparameter) set by the parameter selection unit 314 and the value of the hyperparameter of the temperature sensor combination (second hyperparameter) to set the thermal displacement prediction calculation formula (thermal displacement model) of mathematical formula 1. The machine learning unit 315 sets the thermal displacement prediction calculation formula (thermal displacement model) based on the difference between the estimated value of the thermal displacement of the mechanical element calculated by substituting the measurement data set within a predetermined period stored as teacher data in the measurement data 322 and the measured value of the thermal displacement of the mechanical element within a predetermined period stored as a tag in the measurement data 322, for example, by using the least squares method to minimize the difference.

[0077] That is, the machine learning unit 315 can infer and set a coefficient that minimizes the square error between the estimated value of the thermal displacement calculated by the thermal displacement prediction formula (thermal displacement model) of the time offset element of the measurement data using the least squares method and the measured value of the thermal displacement.

[0078] More specifically, when the measurement data is set as X k Set the label to Y L At that time, the Machine Learning Department targeted 315

[0079] [Mathematical Expression 2]

[0080]

[0081] Find the coefficient 'a' that minimizes the sum of multiple teacher data. k and b kτ The group.

[0082] That is, the machine learning unit 315 can also be similar to the case in Patent Document 1, for the coefficient a of mathematical formula 2 k and b kτ The parameter a, which is not related to time, is repeatedly fixed. k The time-dependent parameter b in the state. k0 b k1 …b kTk The machine learning-based setup and the fixed time-related parameter b k0 b k1 …b kTk The parameter a in the state that is not related to time. k Based on machine learning settings, this determines a k and b kτ .

[0083] The model evaluation unit 316 calculates the error between the estimated thermal displacement and the measured thermal displacement of the mechanical element, obtained by inputting the measured data set into a thermal displacement prediction formula (thermal displacement model) generated based on the value of the hyperparameter of the time delay of the optimized object (first hyperparameter) or the value of the hyperparameter of the temperature sensor combination (second hyperparameter), as an evaluation value. The model evaluation unit 316 stores the calculated evaluation value in the history data 321, corresponding to the set values ​​of the hyperparameter of the time delay of the optimized object or the hyperparameter of the temperature sensor combination. Furthermore, the parameter selection method determination unit 313 sets the value (or combination) of the hyperparameter of the time delay of the optimized object or the hyperparameter of the temperature sensor combination when the evaluation value is minimized according to the history data 321.

[0084] <Optimization Processing of Information Processing Device 30>

[0085] Next, while referring to Figure 2 The process of optimizing the values ​​of the first and second hyperparameters of the thermal displacement prediction calculation formula (thermal displacement model) in the information processing device 30 is explained.

[0086] Figure 2 This is a flowchart illustrating the optimization process related to the values ​​of the first and second hyperparameters of the thermal displacement prediction calculation formula (thermal displacement model) in the information processing device 30.

[0087] Furthermore, in the process shown here, the value of the hyperparameter of time delay (first hyperparameter) is initially optimized, then the value of the hyperparameter of temperature sensor combination (second hyperparameter) is optimized, and then the value of the hyperparameter of time delay (first hyperparameter) is optimized again. However, the optimization process of the information processing device 30 is not limited to this. For example, the information processing device 30 may repeat the optimization process of the hyperparameter of time delay (first hyperparameter) and the optimization process of the hyperparameter of temperature sensor combination (second hyperparameter) more than twice.

[0088] In step S11, the measurement data acquisition unit 311 acquires a set of measurement data from the control device 20. More specifically, the measurement data acquisition unit 311 acquires temperature data and / or operating status data of the mechanical elements of the machine tool 10 and their surroundings. Operating status data may include, for example, the spindle rotation speed, coolant flow rate, and lubricating oil flow rate.

[0089] Furthermore, for example, as measurement data, it is possible to obtain data on the amount of temperature change instead of the temperature itself. Moreover, as data on the amount of temperature change, it is possible to obtain data on the temperature change from the initial temperature, or data on the temperature change from the last measured temperature to the current measured temperature.

[0090] In addition, as operational status data, the heat absorbed by the coolant and the heat absorbed by the lubricating oil can also be included.

[0091] In step S12, the thermal displacement acquisition unit 312 acquires, for example, the measured value of the thermal displacement of the mechanical elements of the machine tool 10 detected by the probe. Specifically, for example, the X, Y, and Z axis components of the thermal displacement can also be measured, and a set of these measured values ​​can be used as the actual measured value.

[0092] In step S13, the storage unit 320 uses the measurement data group obtained in step S11 as input data, and uses the measured value of the thermal displacement of the mechanical element obtained in step S12 as a tag, sets them as an interrelated group, and stores them as teacher data in the measurement data 322.

[0093] Furthermore, the information processing device 30 can perform the machine learning described later based on continuously acquiring the teacher data online. Alternatively, it can use all the teacher data acquired in advance as batch data to perform the machine learning described later. Alternatively, it can use the batch data divided into multiple small groups as mini-batch data to perform the machine learning described later.

[0094] In step S14, the information processing device 30 selects the hyperparameter of the time delay (the first hyperparameter) as the hyperparameter of the optimization object, and uses the teacher data obtained in step S13 to optimize the value of the hyperparameter of the time delay (the first hyperparameter). Details of the processing in step S14 will be described later.

[0095] In step S15, the information processing device 30 selects the hyperparameter (second hyperparameter) of the temperature sensor combination as the hyperparameter to be optimized, and optimizes the value of the hyperparameter (second hyperparameter) of the temperature sensor combination based on the teacher data obtained in step S13 and the processing result of step S14. Details of the processing in step S15 will be described later.

[0096] In step S16, the information processing device 30 selects the hyperparameter of time delay (the first hyperparameter) as the hyperparameter to be optimized, and optimizes the value of the hyperparameter of time delay (the first hyperparameter) based on the teacher data obtained in step S13 and the processing result of step S15. The processing of step S16 is the same as that of step S14, and details will be described later.

[0097] In step S17, the information processing device 30 uses the optimized time delay hyperparameter (first hyperparameter) and the temperature sensor combination hyperparameter (second hyperparameter) to send the thermal displacement prediction calculation formula (thermal displacement model) set by machine learning to the thermal displacement correction device 201 of the control device 20.

[0098] Figure 3 Yes Figure 2 The flowcharts provide a detailed explanation of the optimization processes in steps S14 and S16.

[0099] Figure 4 This is a diagram illustrating an example of the optimization process in step S14. Figure 5 This is a diagram illustrating an example of the optimization process in step S16.

[0100] In the optimization process of step S14, such as Figure 4As shown, the value of the hyperparameter (second hyperparameter) of the temperature sensor combination is fixed, for example, using a combination of all N temperature sensors. While sequentially changing the value of the hyperparameter of time delay within the detection range, a thermal displacement prediction calculation formula (thermal displacement model) of Mathematical Formula 1 is generated according to the value of the hyperparameter of time delay (first hyperparameter). Furthermore, in the optimization process of step S14, the error between the estimated thermal displacement amount estimated by inputting the measurement data set into the thermal displacement prediction calculation formula (thermal displacement model) generated according to the value of the hyperparameter of time delay (first hyperparameter) and the measured value of the thermal displacement amount of the mechanical element is calculated as an evaluation value (verification error). The value of the hyperparameter of time delay that minimizes the evaluation value (verification error) is selected as the optimal value.

[0101] On the other hand, in the optimization process of step S16, such as Figure 5 As shown, the value of the hyperparameter (second hyperparameter) of the temperature sensor combination is fixed to the temperature sensor combination optimized in step S15 (for example, a combination of m temperature sensors is used). While the value of the hyperparameter of time delay (first hyperparameter) is changed sequentially within the detection range, a thermal displacement prediction calculation formula (thermal displacement model) of Mathematical Formula 1 is generated according to the value of the hyperparameter of time delay (first hyperparameter). Furthermore, m is an integer of 1 ≤ m ≤ N. In the optimization process of step S16, as in step S14, the error between the estimated thermal displacement amount estimated by inputting the measurement data set into the thermal displacement prediction calculation formula (thermal displacement model) generated according to the value of the hyperparameter of time delay (first hyperparameter) and the measured value of the thermal displacement amount of the mechanical element is calculated as an evaluation value (verification error). The value of the hyperparameter of time delay (first hyperparameter) when the evaluation value (verification error) is minimized is selected as the optimal value.

[0102] In step S21, the parameter selection method determination unit 313 selects the hyperparameter of time delay (first hyperparameter) as the hyperparameter of the optimization object.

[0103] In step S22, the parameter selection unit 314 uses random search, Bayesian optimization, etc., to set / change the hyperparameter (first hyperparameter) of the time delay selected in step S21 for the temperature sensor X within each detection range. k The sampling time Δt of the measured value k and the coefficient T corresponding to the delay amount k In addition, the parameter selection unit 314 fixes the value of the temperature sensor combination (second hyperparameter) to a predetermined combination.

[0104] Furthermore, the predetermined combination, such as in the case of the optimization process in step S14, uses all N temperature sensors X. k The combination, in the case of the optimization process in step S16, is the combination selected in the process in step S15.

[0105] In step S23, the machine learning unit 315 performs machine learning based on the measurement data set, the measured value of the thermal displacement of the mechanical element, the value of the hyperparameter of the time delay (first hyperparameter), and the value of the hyperparameter of the temperature sensor combination (second hyperparameter), thereby generating the thermal displacement prediction calculation formula (thermal displacement model) of mathematical formula 1.

[0106] In step S24, the model evaluation unit 316 calculates the error between the estimated thermal displacement amount, derived from inputting the measurement data set into the thermal displacement prediction formula (thermal displacement model) generated in step S23, and the measured value of the thermal displacement amount of the mechanical element, as an evaluation value. The model evaluation unit 316 then compares the calculated evaluation value with the time delay hyperparameter (first hyperparameter) of the temperature sensor X. k The sampling time Δt of the measured value k and the coefficient T corresponding to the delay. k The set values ​​are stored in the corresponding resume data 321.

[0107] In step S25, the parameter selection unit 314 determines, based on the history data 321, whether the temperature sensor X is in operation. k The sampling time Δt of the measured value k The value and the coefficient T corresponding to the delay amount k The process is applied to all combinations within the search range of values. If all combinations have been applied, the process proceeds to step S25. Otherwise, if not all combinations have been applied, the process returns to step S22.

[0108] Furthermore, the parameter selection unit 314 determines, based on the history data 321, whether the temperature sensor X is in operation. k The sampling time Δt of the measured value k The value and the coefficient T corresponding to the delay amount k The search range of values ​​is applied to all combinations, but it is not limited to this. For example, the parameter selection unit 314 may also determine whether the evaluation value (verification error) calculated by the model evaluation unit 316 in step S24 is below a preset threshold, or whether the number of repetitions is a preset predetermined number.

[0109] In step S26, the parameter selection method determination unit 313 selects the temperature sensor X from the hyperparameters (first hyperparameters) of time delay with the minimum evaluation value based on the history data 321. k The sampling time Δt of the measured value k The value and the coefficient T corresponding to the delay amount k The value is taken as the optimal value.

[0110] Figure 6 Yes Figure 2 The flowchart describes the detailed processing steps of the optimization process in step S15.

[0111] Figure 7 This is a diagram illustrating an example of the optimization process in step S15. (See diagram for example.) Figure 7 As shown, in the optimization process of step S15, the value of the hyperparameter of time delay (first hyperparameter) is fixed to the value selected in step S14. While the value of the hyperparameter of the temperature sensor combination (second hyperparameter) is changed sequentially within the detection range, the thermal displacement prediction calculation formula (thermal displacement model) of Mathematical Formula 1 is generated according to the value of the hyperparameter of the temperature sensor combination (second hyperparameter). Furthermore, in the optimization process of step S15, the error between the estimated thermal displacement amount estimated by inputting the measurement data set into the thermal displacement prediction calculation formula (thermal displacement model) generated according to the value of the hyperparameter of the temperature sensor combination (second hyperparameter) and the measured value of the thermal displacement amount of the mechanical element is calculated as an evaluation value (verification error). The combination of hyperparameters (second hyperparameter) of the temperature sensor combination that minimizes the evaluation value (verification error) is selected as the optimal combination.

[0112] Furthermore, in the following description, in temperature sensor X k The hyperparameter values ​​of the temperature sensor combination are optimized within the detection range of 1 to N combinations.

[0113] In step S31, the parameter selection method determination unit 313 selects the hyperparameter (second hyperparameter) of the temperature sensor combination as the hyperparameter of the optimization object.

[0114] In step S32, the parameter selection unit 314 uses random search, Bayesian optimization, etc., to set / change the value of the hyperparameter (second hyperparameter) of the temperature sensor combination selected in step S31 within the exploration range. Additionally, the parameter selection unit 314 selects the temperature sensor X with the time-delayed hyperparameter (first hyperparameter). k The sampling time Δt of the measured value k and the coefficient T corresponding to the delay amount k It is fixed to the value selected in step S14.

[0115] In step S33, the machine learning unit 315 performs a process of interaction with... Figure 3 The same process as step 23 is used. Machine learning is performed based on the measured data set, the measured value of the thermal displacement of the mechanical elements, the value of the hyperparameter of the time delay (first hyperparameter), and the value of the hyperparameter of the temperature sensor combination (second hyperparameter). As a result, the thermal displacement prediction calculation formula (thermal displacement model) of mathematical formula 1 is generated.

[0116] In step S34, the model evaluation unit 316 calculates the error between the estimated thermal displacement amount, which is input into the thermal displacement prediction formula (thermal displacement model) generated in step S33, and the measured value of the thermal displacement amount of the mechanical element, as an evaluation value. The model evaluation unit 316 stores the calculated evaluation value in the history data 321, corresponding it to the combination of hyperparameters (second hyperparameters) set for the temperature sensor combination.

[0117] In step S35, the parameter selection unit 314 determines, based on the history data 321, whether all combinations of temperature sensor combinations have been implemented within the search range. If all combinations have been implemented, the process proceeds to step S35. Otherwise, if not all combinations have been implemented, the process returns to step S32.

[0118] Furthermore, the parameter selection unit 314 determines, based on the history data 321, whether all combinations of temperature sensor combinations have been implemented within the search range, but is not limited to this. For example, the parameter selection unit 314 may also determine whether the evaluation value (verification error) calculated by the model evaluation unit 316 in step S34 is below a preset threshold, or whether the number of repetitions is a preset predetermined number.

[0119] In step S36, the parameter selection method determination unit 313 selects the value of the hyperparameter (second hyperparameter) of the temperature sensor combination with the minimum evaluation value as the optimal value based on the history data 321.

[0120] Based on the above, the information processing device 30 of one embodiment selects either the hyperparameter of time delay (first hyperparameter) or the hyperparameter of temperature sensor combination (second hyperparameter) included in the thermal displacement prediction calculation formula (thermal displacement model). The information processing device 30 repeatedly optimizes the value of the selected hyperparameter of time delay (first hyperparameter) when the value of the hyperparameter of temperature sensor combination (second hyperparameter) is fixed while the hyperparameter of time delay (first hyperparameter) is selected, and optimizes the value of the selected hyperparameter of temperature sensor combination (second hyperparameter) when the value of the hyperparameter of time delay (first hyperparameter) is fixed while the hyperparameter of temperature sensor combination (second hyperparameter) is selected, thereby determining the values ​​of the hyperparameter of time delay (first hyperparameter) and the hyperparameter of temperature sensor combination (second hyperparameter).

[0121] Therefore, the information processing device 30 can automatically optimize the values ​​of the hyperparameters of the thermal displacement prediction calculation formula (thermal displacement model).

[0122] Furthermore, in optimizing the values ​​of the selected time delay hyperparameter (first hyperparameter) or the hyperparameter of the temperature sensor combination (second hyperparameter), the information processing device 30 uses random search, Bayesian optimization, etc., to set / change the values ​​of the selected time delay hyperparameter (first hyperparameter) or the hyperparameter of the temperature sensor combination (second hyperparameter) within the search range. As a result, the information processing device 30 can reduce computational costs.

[0123] The above describes one embodiment, but the information processing device 30 is not limited to the above embodiment and includes variations and improvements within the scope of achieving the purpose.

[0124] <Variation Example 1>

[0125] In the above embodiments, the information processing device 30 is shown to be a device different from the machine tool 10 and the control device 20, but the machine tool 10 or the control device 20 may also have some or all of the functions of the information processing device 30.

[0126] Alternatively, for example, the server may also include some or all of the measurement data acquisition unit 311, thermal displacement acquisition unit 312, parameter selection method determination unit 313, parameter selection unit 314, machine learning unit 315, and model evaluation unit 316 of the information processing device 30. Furthermore, the various functions of the information processing device 30 can also be implemented in the cloud using virtual server functions, etc.

[0127] Furthermore, the information processing device 30 can also be a distributed processing system in which the functions of the information processing device 30 are appropriately distributed to multiple servers.

[0128] <Variation Example 2>

[0129] Furthermore, for example, in the above-described embodiment, the information processing device 30 utilizes the thermal displacement prediction calculation formula (thermal displacement model) of Formula 1, which uses the time offset element of the measurement data, based on multiple regression analysis of a generalized linear model, but it is not limited to this. For example, the information processing device 30 may also use the thermal displacement prediction calculation formula (thermal displacement model) of the first delay element of the measurement data.

[0130] [Mathematical Expression 3]

[0131]

[0132] Alternatively, the information processing device 30 can also use a device based on multiple regression analysis using a nonlinear model.

[0133] <Variation Example 3>

[0134] Furthermore, for example, in the above embodiment, the information processing device 30 optimizes the value of the hyperparameter of the time delay in step S14, then optimizes the value of the hyperparameter of the temperature sensor combination in step S15, and then optimizes the value of the hyperparameter of the time delay again in step S16, but is not limited to this. For example, the information processing device 30 may repeat the optimization processing of the hyperparameter of the temperature sensor combination in step S15 and the optimization processing of the hyperparameter of the time delay in step S16 two or more predetermined times.

[0135] Therefore, the information processing device 30 can set more optimized hyperparameters for time delay and temperature sensor combination.

[0136] <Variation Example 4>

[0137] Furthermore, for example, in the above-described embodiment, the thermal displacement prediction calculation formula (thermal displacement model) has two types: a time delay hyperparameter (first hyperparameter) and a temperature sensor combination hyperparameter (second hyperparameter), but it is not limited to these. For example, the information processing device 30 may also use a thermal displacement prediction calculation formula (thermal displacement model) with three or more hyperparameters.

[0138] For example, when the thermal displacement prediction calculation formula (thermal displacement model) has four hyperparameters, the information processing device 30 optimizes the value of the first hyperparameter while keeping the values ​​of the other three hyperparameters fixed. Next, the information processing device 30 optimizes the value of the second hyperparameter while keeping the values ​​of the other three hyperparameters fixed. Then, the information processing device 30 optimizes the value of the third hyperparameter while keeping the values ​​of the other three hyperparameters fixed. Finally, the information processing device 30 optimizes the value of the fourth hyperparameter while keeping the values ​​of the other three hyperparameters fixed. Furthermore, the information processing device 30 can repeat these four optimization processes.

[0139] Furthermore, for example, in the optimization process of the first type of hyperparameter value, if the search result does not converge, or if the search result converges but the effect is lower than that of the optimization of other types of hyperparameter values ​​(e.g., the evaluation value (verification error) is large), the information processing device 30 can flexibly change the order of optimization processes according to the processing results.

[0140] Figure 8 This is a diagram illustrating an example of the order in which the values ​​of the four hyperparameters are optimized. Figure 8 The upper part indicates the case where the optimization process for the first hyperparameter value is repeated sequentially up to the optimization process for the fourth hyperparameter value. On the other hand, Figure 8The following paragraph, for example, indicates the order of optimization processes when the priority of optimizing the first type of hyperparameter value is reduced because the initial optimization process for the first type of hyperparameter value did not converge. That is, in Figure 8 In the next section, because the priority of the initial optimization processing of the first type of hyperparameter value is reduced, the information processing device 30 starts the second initial optimization processing from the optimization processing of the second type of hyperparameter value.

[0141] Therefore, the information processing device 30 can optimize the adjustment schedule of the order of optimization processing of various hyperparameter values, thereby reducing the processing time and improving the processing results (i.e., obtaining a better thermal displacement prediction calculation formula (thermal displacement model)).

[0142] <Variation Example 5>

[0143] Furthermore, for example, in the above-described embodiment, the information processing device 30 performs machine learning based on the measurement data set, the measured value of the thermal displacement of the mechanical element, the value of the hyperparameter of the time delay (first hyperparameter), and the value of the hyperparameter of the temperature sensor combination (second hyperparameter), thereby generating the thermal displacement prediction calculation formula (thermal displacement model) of Formula 1, but it is not limited to this. For example, the information processing device 30 may also use the measurement data set and the measured value of the thermal displacement of the mechanical element to generate the thermal displacement prediction calculation formula (thermal displacement model) of Formula 1 through a known cross-validation method.

[0144] <Variation Example 6>

[0145] Furthermore, for example, in the above embodiment, the machine tool 10 is configured as a cutting machine, but it is not limited to this. The machine tool 10 may also be, for example, a wire electrical discharge machining (EDM) machine or a laser processing machine.

[0146] Furthermore, the functions included in the information processing apparatus 30 in one embodiment can be implemented separately by hardware, software, or a combination thereof. Here, implementation by software means implementation by loading and executing a program into a computer.

[0147] Programs can be stored and provided to a computer using various types of non-transitory computer-readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (e.g., floppy disks, magnetic tapes, hard disks), optical-magnetic recording media (e.g., optical discs), CD-ROMs (Read Only Memory), CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash memory ROMs, and RAM). Alternatively, programs can also be provided to a computer using various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transient computer-readable media can provide programs to a computer via wired communication paths such as wires and optical fibers, or via wireless communication paths.

[0148] Furthermore, the steps describing a program recorded in a recording medium naturally include processing performed sequentially in that order, as well as processing that may not be performed sequentially, and processing that is performed in parallel or individually.

[0149] In other words, the information processing apparatus, control apparatus, and optimization method disclosed herein can be implemented in a variety of ways having the following structure.

[0150] (1) One aspect of the information processing apparatus 30 disclosed herein is an information processing apparatus that optimizes the values ​​of multiple hyperparameters included in a thermal displacement prediction calculation formula when generating such a formula through machine learning. The thermal displacement prediction calculation formula estimates the thermal displacement of a mechanical element based on a set of measurement data including temperature data of the mechanical element of a machine tool 10 with thermal expansion and its surroundings, and / or motion state data of the mechanical element. The information processing apparatus includes: a measurement data acquisition unit 311 that acquires a set of measurement data; a thermal displacement acquisition unit 312 that acquires measured values ​​of the thermal displacement of the mechanical element; a storage unit 320 that stores the set of measurement data acquired by the measurement data acquisition unit 311 as input data and the measured values ​​of the thermal displacement of the mechanical element acquired by the thermal displacement acquisition unit 312 as tags, interconnecting them as teacher data; a parameter selection method determination unit 313 that selects at least one of the multiple hyperparameters as the optimization target as a first hyperparameter; and a parameter selection unit 314 that sets / changes the first hyperparameter. The parameter selection method determination unit 313 fixes the value of the second hyperparameter as the second hyperparameter, which is not selected by the parameter selection method determination unit 313; the machine learning unit 315 performs machine learning based on the combination of the values ​​of the first hyperparameter and the second hyperparameter, according to the measured values ​​of the measurement data set and the measured values ​​of the thermal displacement of the mechanical element, thereby fixing the value of the second hyperparameter as the fixed value and generating a thermal displacement prediction calculation formula according to the set / changed value of the first hyperparameter; and the model evaluation unit 316 fixes the value of the second hyperparameter, calculates the error between the thermal displacement estimated by inputting the thermal displacement prediction calculation formula of each value of the first hyperparameter into the measurement data set and the measured value of the thermal displacement of the mechanical element as the evaluation value, fixes the value of the second hyperparameter as the fixed value, and stores the calculated evaluation value and the value of the first hyperparameter in the storage unit 320, while the parameter selection method determination unit 313 fixes the value of the second hyperparameter, and, according to the history of the first hyperparameter value and the evaluation value stored in the storage unit 320, takes the value of the first hyperparameter with the smallest evaluation value as the optimal value.

[0151] According to the information processing device 30, the values ​​of the hyperparameters of the thermal displacement prediction calculation formula (thermal displacement model) can be automatically optimized.

[0152] (2) In the information processing apparatus 30 described in (1), the parameter selection method determination unit 313 may select at least one other hyperparameter from a plurality of hyperparameters as the first hyperparameter after optimizing the value of the first hyperparameter by using the value of the second hyperparameter as a fixed value.

[0153] Therefore, the information processing device 30 is able to optimize the values ​​of all hyperparameters.

[0154] (3) In the information processing device 30 described in (1) or (2), the parameter selection unit 314 may set / change the value of the first hyperparameter according to a predetermined detection range.

[0155] Therefore, the information processing device 30 can quickly obtain the optimal hyperparameter values.

[0156] (4) In any of the information processing apparatus 30 described in (1) to (3), the parameter selection unit 314 may use random search or Bayesian optimization to set / change the value of the first hyperparameter.

[0157] As a result, the information processing device 30 is able to obtain the optimal hyperparameter values ​​more efficiently.

[0158] (5) The control device 20 of this disclosure has an information processing device 30 as described in any one of (1) to (4).

[0159] According to the control device 20, the same effect as (1) to (4) can be obtained.

[0160] (6) The optimization method disclosed herein is a computer-implemented optimization method that optimizes the values ​​of multiple hyperparameters included in the thermal displacement prediction calculation formula when generating the thermal displacement prediction calculation formula through machine learning. The thermal displacement prediction calculation formula estimates the thermal displacement of the mechanical element based on a set of measurement data including temperature data and / or motion state data of the mechanical element of the machine tool 10 containing mechanical elements with thermal expansion. The optimization method includes: a measurement data acquisition step, acquiring a set of measurement data; a thermal displacement acquisition step, acquiring the measured value of the thermal displacement of the mechanical element; a storage step, using the set of measurement data acquired in the measurement data acquisition step as input data, and using the measured value of the thermal displacement of the mechanical element acquired in the thermal displacement acquisition step as a tag to associate them with each other and store them as teacher data in the storage unit 320; a parameter selection method determination step, selecting at least one of the multiple hyperparameters as the optimization object as the first hyperparameter; and a parameter selection step, setting / changing the first hyperparameter. The value of the number is used as the second hyperparameter, which is not selected in the parameter selection method determination step, and the value of the second hyperparameter is fixed. In the machine learning step, machine learning is performed based on the combination of the values ​​of the first hyperparameter and the second hyperparameter, according to the measured values ​​of the measurement data set and the measured values ​​of the thermal displacement of the mechanical element. Thus, the value of the second hyperparameter is fixed, and a thermal displacement prediction calculation formula is generated according to the value of the first hyperparameter. In the model evaluation step, the value of the second hyperparameter is fixed, and an evaluation value is calculated, which represents the error between the thermal displacement estimated by inputting the measurement data set into the thermal displacement prediction calculation formula for each value of the first hyperparameter and the measured value of the thermal displacement of the mechanical element. The value of the second hyperparameter is fixed, and the calculated evaluation value is stored in the storage unit 320 in correspondence with the value of the first hyperparameter. In the parameter selection method determination step, the value of the second hyperparameter is fixed, and the value of the first hyperparameter with the smallest evaluation value is taken as the optimal value based on the history of the value of the first hyperparameter and the evaluation value stored in the storage unit 320.

[0161] According to this optimization method, the same effect as (1) can be obtained.

[0162] Symbol Explanation

[0163] 1. Information processing system

[0164] 10 Machine tools

[0165] 20. Control devices

[0166] 30 Information processing devices

[0167] 310 Control Department

[0168] 311 Measurement Data Acquisition Department

[0169] 312 Thermal displacement acquisition section

[0170] 313 Parameter Selection Method Determination Section

[0171] 314 Parameter Selection Section

[0172] 315 Machine Learning Department

[0173] 316 Model Evaluation Department

[0174] 320 Storage Department

[0175] 321 Resume Data

[0176] 322 Measurement data.

Claims

1. An information processing apparatus, which optimizes the values ​​of multiple hyperparameters included in a thermal displacement prediction formula when generating such a formula through machine learning, wherein the thermal displacement prediction formula estimates the thermal displacement of a mechanical element based on a set of measurement data of temperature data and / or motion state data of the mechanical element and its surroundings of a machine tool containing a mechanical element with thermal expansion, characterized in that, The information processing device has: The measurement data acquisition unit acquires the measurement data set; The thermal displacement acquisition unit acquires the measured value of the thermal displacement of the mechanical element. The storage unit uses the measurement data set obtained by the measurement data acquisition unit as input data and stores the measured values ​​of the thermal displacement of the mechanical elements obtained by the thermal displacement acquisition unit as tags, and links them together as teacher data. The parameter selection method determination unit selects at least one of the plurality of hyperparameters as the hyperparameter to be optimized as the first hyperparameter; The parameter selection unit sets / changes the value of the first hyperparameter, and fixes the value of the second hyperparameter as a second hyperparameter by using a hyperparameter that was not selected by the parameter selection method determination unit. The machine learning department performs machine learning based on the combination of the values ​​of the first hyperparameter and the second hyperparameter, according to the measured values ​​of the measurement data set and the measured values ​​of the thermal displacement of the mechanical elements. Thus, the value of the second hyperparameter is used as a fixed value, and the thermal displacement prediction calculation formula is generated according to the set / changed value of the first hyperparameter. as well as The model evaluation unit uses the value of the second hyperparameter as a fixed value, calculates the error between the estimated thermal displacement value (derived from the thermal displacement prediction formula input to the measurement data set for each value of the first hyperparameter) and the measured value of the thermal displacement of the mechanical element, and uses this error as an evaluation value. It also uses the value of the second hyperparameter as a fixed value and stores the calculated evaluation value in the storage unit, corresponding to the value of the first hyperparameter. The parameter selection method determination unit takes the value of the second hyperparameter as a fixed value, and takes the value of the first hyperparameter with the smallest evaluation value as the optimal value based on the value of the first hyperparameter stored in the storage unit and the history of the evaluation value.

2. The information processing device according to claim 1, characterized in that, The parameter selection method determination unit takes the value of the second hyperparameter as a fixed value, optimizes the value of the first hyperparameter, and selects at least one other hyperparameter from the plurality of hyperparameters as the first hyperparameter.

3. The information processing apparatus according to claim 1 or 2, characterized in that, The parameter selection unit sets / changes the value of the first hyperparameter according to a predetermined exploration range.

4. The information processing apparatus according to claim 1 or 2, characterized in that, The parameter selection unit uses random search or Bayesian optimization to set / change the value of the first hyperparameter.

5. A control device, characterized in that, have: The information processing apparatus according to any one of claims 1 to 4.

6. An optimization method, implemented by a computer, optimizes the values ​​of multiple hyperparameters included in the thermal displacement prediction formula when generating the thermal displacement prediction formula through machine learning. The thermal displacement prediction formula estimates the thermal displacement of the mechanical element based on a set of measurement data including temperature data of the mechanical element and its surroundings, and / or motion state data of the mechanical element of a machine tool containing mechanical elements with thermal expansion. The method is characterized in that... The optimization method has the following characteristics: The measurement data acquisition step involves acquiring the measurement data set. The thermal displacement acquisition step involves obtaining the measured value of the thermal displacement of the mechanical element. In the storage step, the measurement data set obtained in the measurement data acquisition step is used as input data, and the measured values ​​of the thermal displacement of the mechanical element obtained in the thermal displacement acquisition step are used as tags and linked together as teacher data and stored in the storage unit. The parameter selection method determines the steps, selecting at least one of the multiple hyperparameters as the hyperparameter to be optimized as the first hyperparameter; In the parameter selection step, the value of the first hyperparameter is set / changed, and the value of the second hyperparameter is fixed by taking the hyperparameter that was not selected in the parameter selection method determination step as the second hyperparameter. The machine learning step involves performing machine learning based on the combination of the values ​​of the first hyperparameter and the second hyperparameter, according to the measured values ​​of the measurement data set and the measured values ​​of the thermal displacement of the mechanical element. Thus, the value of the second hyperparameter is used as a fixed value, and the thermal displacement prediction calculation formula is generated according to the value of the first hyperparameter. as well as The model evaluation step involves using the value of the second hyperparameter as a fixed value, calculating an evaluation value representing the error between the estimated thermal displacement amount, derived from inputting the measured data set into the thermal displacement prediction formula for each value of the first hyperparameter, and the measured value of the thermal displacement amount of the mechanical element. The second hyperparameter is then used as a fixed value, and the calculated evaluation value is stored in the storage unit in correspondence with the value of the first hyperparameter. In the parameter selection method determination step, the value of the second hyperparameter is taken as a fixed value, and the value of the first hyperparameter with the smallest evaluation value is taken as the optimal value based on the value of the first hyperparameter stored in the storage unit and the history of the evaluation value.

Images