DE112021006961T5 - Numerical control device - Google Patents

Abstract

Numerical control device (100) includes: an engine condition observation unit (1) that outputs engine condition signals indicating results of observation of a condition of the engine; a cutting condition observation unit (5) that outputs a cutting condition signal indicating whether the machine tool is cutting; a first torque estimation unit (21) that outputs a first estimated torque and physical state variables based on the engine state signals and physical model parameters of the machine tool, the first estimated torque being an estimated value of the torque of the engine, the physical state variables being components of the first estimated torque ; a feature value generation unit (3) which generates feature values from the cutting state signal and the first estimated torque, the physical state variables and/or the engine state signals; and a machine learning model calculation unit (4) which calculates a machine learning correction value from the feature values and the machine learning model parameters, which learns machine learning model parameters such that the calculated machine learning correction value approximates a difference between an actual torque included in the engine state signals and the first estimated torque, and the learned machine learning model parameters.

Description

Area

The present disclosure relates to a numerical control device that generates a model that simulates an object to be controlled.

background

There is a need for a rapid and highly accurate control method and error detection method for a numerical control device used to control a machine tool. A model that accurately simulates an object to be controlled is essential to improving such methods. However, in the actual operation of the numerical control device, errors in physical model parameters such as a mass, a barycentric position, an inertia tensor and friction against a cover may occur due to installation of a wire, a spindle, a tool, a workpiece or the like in one to be controlled Object should not be neglected. It also includes components that are non-linear with respect to engine rotation, such as vibration and friction. Therefore, it has been difficult to provide a model that pre-reproduces these components with high accuracy.

In recent years, control methods have been proposed to prevent or reduce errors in the estimation of forces and torques in such models. For example, Patent Literature 1 discloses a technique in which an electric motor control device corrects an error contained in initial values of physical model parameters during operation by obtaining an M-sequence or a sine command and an inertia tensor, a coefficient of friction, and a Spring constant for each sampling period can be corrected to minimize torque estimation error.

Citation list

Patent literature

Patent Literature 1: disclosed Japanese Patent Application No. 2013-128387

overview

Technical problem

However, the technique described in Patent Literature 1 is based on the assumption that the movement of the electric motor control device can be modeled using the inertia tensor, friction and torque of a spring. Consequently, the technique described in Patent Literature 1 has a problem in that a disorder not considered in a model cannot be expressed. In addition, since a friction model is also simplified, the technique described in Patent Literature 1 has a problem in that the friction model is unable to express complex phenomena with strong nonlinearity such as Stribeck effect and hysteresis.

The present disclosure has been made in view of the above, and it is an object of the present disclosure to obtain a numerical control device that enables high-accuracy torque estimation for a phenomenon with strong nonlinearity.

the solution of the problem

In order to solve the above problem and achieve the goal, a numerical control device according to the present disclosure controls a motor of a machine tool that cuts a workpiece. The numerical control device includes: an engine condition monitoring unit for outputting engine condition signals indicating observation results of a condition of the engine; a cutting condition observation unit for outputting a cutting condition signal indicating whether the machine tool is performing cutting; a first torque estimation unit for outputting a first estimated torque and physical state variables based on the motor state signals and physical model parameters of the machine tool, the first estimated torque being an estimate of the torque of the motor, the physical state variables being components of the first estimated torque are; a feature value generation unit for generating feature values from the cutting state signal and from the first estimated torque, the physical state variables and/or the engine state signals and outputting the generated feature values; and a machine learning model calculation unit for calculating a machine learning correction value from the feature values and machine learning model parameters, learning the machine learning model parameters such that the calculated machine learning correction value approximates a difference between an actual torque included in the engine state signals and the first estimated torque, and the learned Store machine learning model parameters.

Advantageous effects of the invention

The numerical control device according to the present disclosure has an effect of enabling highly accurate torque estimation for a phenomenon with strong nonlinearity.

Brief description of the drawings

• 1 is a block diagram showing a configuration example of a numerical control device according to a first embodiment.
• 2 is a schematic diagram showing a target machine to be controlled by the numerical control device according to the first embodiment.
• 3 is a diagram showing an example of a nonlinear friction model, which is a physical model parameter to be estimated by a physical model storage unit according to the first embodiment.
• 4 is a block diagram showing a configuration example of a feature value generation unit and a machine learning model calculation unit according to the first embodiment.
• 5 is a diagram showing an autoregressive term in the configuration of the feature value generation unit and the machine learning model calculation unit according to the first embodiment.
• 6 Fig. 10 is a flowchart showing a process of learning physical model parameters and machine learning model parameters in the numerical control apparatus according to the first embodiment.
• 7 is a diagram representing an engine speed representing the operation of the target machine to be controlled by the numerical control device according to the first embodiment, a first estimated torque estimated by the numerical control device, and a second estimated torque estimated by the numerical control device is, shows.
• 8th is a diagram showing irregularity detection to be performed by the numerical control device of the first embodiment.
• 9 is a diagram for describing the cutting load detection to be performed by the numerical control device of a second embodiment.
• 10 is a diagram for describing the detection of wear or loss of a tool to be performed by the numerical control device of a third embodiment.
• 11 is a diagram showing an example in which processing circuits included in each of the numerical control devices according to Embodiments 1 to 3 include a processor and a memory.
• 12 is a diagram showing an example in which the processing circuits included in the numerical control devices according to Embodiments 1 to 3 include dedicated hardware.

Description of embodiments

Below, numerical control devices according to embodiments will be described in detail with reference to the drawings.

First embodiment.

1 is a block diagram showing a configuration example of a numerical control device 100 according to a first embodiment. The numerical control device 100 controls a motor of a target machine 10, which is a machine tool. The numerical control device 100 includes an engine state observation unit 1, a physical model calculation unit 2, a feature value generation unit 3, a machine learning model calculation unit 4, and a cutting state observation unit 5. The physical model calculation unit 2 includes a first torque estimation unit 21 and a physical model storage unit 22. The machine learning model calculation unit 4 includes a second torque estimation unit 41 and a machine learning model storage unit 42.

Now, the machine tool to be controlled by the numerical control device 100 will be described. 2 is a schematic diagram showing the target machine 10 to be controlled by the numerical control device 100 according to the first embodiment. The target machine 10 is a machine tool that performs a cutting operation on a workpiece 15 by moving a tool 14 relative to the workpiece 15. Here, a machining center is described as an example of the target machine 10.

The target machine 10 includes columns 11, a spindle head 12, a table 13 and the tool 14. The tool 14 rotates by receiving a driving force generated by a spindle motor. The columns 11 and the spindle head 12 or the table 13 are driven in a straight line by receiving a driving force generated by a servo motor. The spindle motor and servo motor are not shown. The numerical control device 100 controls the position and orientation of the tool 14 with respect to the workpiece 15.

In the first embodiment, a three-moving axis machining center is described as an example of the machine tool 10. Of course, the machine tool that is the target machine 10 is not limited to a three-axis machining center. The numerical control device 100 can be used for various machine tools driven by motors. Examples of the machine tool as the target machine 10 further include machine tools such as a five-axis machining center, a milling center, a lathe and a combination lathe, and an assembly machine.

The engine condition observation unit 1 observes the condition of the engine of the target machine 10 and outputs engine condition signals indicating the observation results. In particular, the motor state observation unit outputs at least the position, the speed and/or the acceleration of each actuator corresponding to a first axis, a second axis and a third axis, which are moving axes, and the spindle, and an actual torque, which is the torque of the Spindle motor and servo motor are output as the motor status signals.

A value obtained by a sensor based on a direct measurement on the engine can be used as the actual torque. Alternatively, a value obtained by converting a current value obtained from the actuator may be used as the actual torque. The motor status signals are not limited to only signals of the position, speed and acceleration of each actuator and the actual torque. The motor state signals may include values of, for example, sensor data from an inertial sensor or an external sensor provided on each axis of the target machine 10 or provided at a predetermined position on the target machine 10, and may include feature value data obtained by image processing performed on the sensor data become. Examples of the sensor data include temperature data at the actuator of each axis, distance data obtained from a sensor such as an infrared sensor or ultrasonic sensor, and motion image data obtained from an image sensor.

The first torque estimation unit 21 calculates a first estimated torque, which is an estimated value of the actual torque, using the equation of motion with the engine state signals and physical model parameters as input. The engine condition signals are output from the engine condition observation unit 1. The physical model parameters described below are output from the physical model storage unit 22. Accordingly, the first torque estimation unit 21 estimates the first estimated torque, which is the estimated Value of the motor's torque is by using the equation of motion based on the motor state signals and the physical model parameters of the target machine 10. The first torque estimation unit 21 outputs the first estimated torque and the physical state variables that are components of the first estimated torque.

The physical state variables may include not only an inertia term, a centrifugal force term, a Coriolis force term and a gravity term but also a Coulombic friction term, a viscous friction term, an arm vibration term and an elastic deflection term. In addition, the physical state variables may also include a term of a function, such as a nonlinear friction model described below, which is introduced based on a simulation of some physical phenomena in calculating the first estimated torque. The physical state variables include an estimate of friction. As a result, the numerical control device 100 enables the machine learning model calculation unit 4 to correct the uncertainty of a friction model and an error of the friction estimate with respect to a friction component of the actual torque.

The motor state signals to be used in the equation of motion correspond to the position, velocity and acceleration of each actuator. The physical model parameters to be used in the equation of motion correspond to the mass, barycentric position and inertia tensor of each member. Accordingly, the physical model parameters include the mass of a member of the target machine 10, the barycentric position of the member of the target machine 10, and the inertia tensor of the member of the target machine 10. Thus, the numerical controller 100 can determine the accuracy of the estimate of the first estimated torque with respect to the actual torque by learning the mass, the barycentric position and the inertia tensor, which are coefficients, as the physical model parameters in inverse dynamics using information related to a known mechanism of the target machine 10. Further, a correction value in the machine learning model calculation unit 4 in the numerical control device 100 becomes relatively small. Accordingly, the numerical control device 100 can improve the generalization ability of a second estimated torque. Here, for example, a Newton-Euler method or the Lagrangian equations of motion can be used as the equation of motion.

The first torque estimation unit 21 uses the equation of motion including a steady and smooth friction model. The friction model takes as input the speed of the engine obtained from the engine state signal and outputs an estimated value of a friction component of the torque. As will be described below, the machine learning model calculation unit 4 uses a neural network. Now that the neural network can calculate a function that is continuous with respect to the inputs, it is desirable that feature values and inputs to the feature value generation unit 3 be continuous. Therefore, the numerical control device 100 defines a friction model that is continuous even near zero speed. Therefore, by using the neural network, the numerical control device 100 can also learn and correct a friction component that cannot be learned in the physical model storage unit 22.

The physical model storage unit 22 uses, as training data, engine state data output when the target machine 10 is started up in a non-cutting state and operated for a certain period of time, and learns physical model parameters such that the difference between the actual torque , which is included in the engine status signals, and the first estimated torque is reduced. The physical model storage unit 22 stores the learned physical model parameters. As a result of learning the physical model parameters, the numerical control device 100 can correct changes in physical properties due to the installation of a sensor, a wire, and the like and machining errors, assembly errors, and the like, and improve the accuracy of estimating the actual torque. Further, in general, approximate values of physical model parameters may be calculated from a specification table, computer-aided design (CAD) data, and the like. Therefore, it is possible to improve the speed, convergence and the like for learning by assigning these values initial values at the time of learning. However, if these values are unknown or difficult to calculate, the physical model parameters can be initialized by any given constants.

A linear equation of an overdetermined system is used for learning the physical model parameters in the physical model storage unit 22. The one described above The equation of motion is modified and only the physical model parameters were separated from it to obtain the linear equation. For a position, velocity and acceleration, if nonlinear terms are not considered, Newton's equation of motion, Euler's equation of motion and Lagrange's equation of motion can each be expressed by a linear combination of products of constant coefficients and variables, the constant coefficients include a mass, a barycentric Position and an inertia tensor. The variables include a position, a speed and an acceleration. That is, if Δf denotes an error in the dynamic calculation with respect to the actual torque, then the equation of motion of each axis can be described in a variable-separated form as described in Equation (1) by using physical model parameters w, which are the mass, the barycentric position and the inertia tensor, and a coefficient matrix A(v,a) containing a velocity v and an acceleration a can be used.

$$\Delta \text{f}=\text{A}\left(\text{v},\text{a}\right)\text{w}$$

Therefore, the physical model storage unit 22 can calculate the physical model parameters for each axis by solving the equation (1) for w for each axis by the least square method. In addition, the physical model storage unit 22 can obtain a more probable solution for the physical model calculation unit 2 by inputting the equation (1) with respect to w into a single linear equation for each axis and solving the single linear equation for a prime mover at the same time where interactions can occur between the axes, such as in a machine tool with five axes. As a method for identifying the physical model parameters w, the physical model storage unit 22 may use a Bayesian estimation, a genetic algorithm, a neural network and the like in addition to the least squares method, as well as the regularized least squares method and the Kalman method. Filter.

It is noted that when the influence of friction in the equation of motion is taken into account, the physical model storage unit 22 can simultaneously also estimate a coefficient of friction as a physical model parameter as in equation (1) by expressing the friction by an equation , which is represented by a linear combination of a velocity function and a coefficient of friction, from which the coefficient of friction can be separated. Examples of the equation include a linear friction model to be shown below, which is a friction model that includes Coulombic friction and viscous friction. The coefficient of friction to be estimated by the physical model storage unit 22 is expressed by equation (2). It should be noted that in equation (2) shown as a picture and in equation (3) to be described below, a point above the q; is placed, which, however, cannot be expressed in the following description. For this reason, in the following description, a qi over which a dot is placed is simply referred to as q; shown.
Formula 1: $${\text{F}}_{\text{i}}\left({\dot{\text{q}}}_{\text{i}}\right)={\text{r}}_{\text{FC}\text{,i}}\text{sgn}\left({\dot{\text{q}}}_{\text{i}}\right)+{\text{r}}_{\text{FV}\text{,i}}{\dot{\text{q}}}_{\text{i}}$$

Here, for a friction model F _i of an axis i operating at a speed q _i , r _FC,i denotes a Coulombic friction coefficient, and r _FV,i denotes a viscous friction coefficient. In the operation of a multi-axis mechanism, such as a machine tool, it is now often observed that the viscous friction does not have a completely linear relationship with the speed and is less than a value to be obtained in a case when it is assumed that the The relationship between viscous friction and speed is assumed to be linear, especially when actuators operate at high speed. In such a case, it is desirable to estimate friction using a higher degree of freedom friction model that takes into account nonlinearity with respect to speed, as well as a friction component from a difference between the actual torque and the first estimated torque and then estimate other physical model parameters. Examples of a nonlinear friction model, which is such a friction model, include the following function expressed by equation (3).
Formula 2: $$\begin{array}{l}{\text{F}}_{\text{i}}\left({\dot{\text{q}}}_{\text{i}}\right)={\text{F}}_{\text{c}}\left({\dot{\text{q}}}_{\text{i}}\right)+{\text{F}}_{\text{v}}\left({\dot{\text{q}}}_{\text{i}}\right)\\ {\text{F}}_{\text{c}}\left({\dot{\text{q}}}_{\text{i}}\right)=\{\begin{array}{l}{\text{r}}_{\text{FC}\text{,i}}\text{sin}\left(\frac{\mathrm{<figure-callout\; id="2"\; label="\pi "\; filenames="DE112021006961T5\_0009.png,DE112021006961T5\_0010.png"\; state="\{\{state\}\}">\pi </figure-callout>}}{2{\mathrm{\omega }}_{\text{Is}}}{\dot{\text{q}}}_{\text{i}}\right)\left|{\dot{\text{q}}}_{\text{i}}\right|<{\mathrm{\omega }}_{\text{Is}}\\ {\text{r}}_{\text{FC}\text{,i}}\text{sgn}\left({\dot{\text{q}}}_{\text{i}}\right)\left|{\dot{\text{q}}}_{\text{i}}\right|\ge {\mathrm{\omega }}_{\text{Is}}\end{array}\\ {\text{F}}_{\text{v}}\left({\dot{\text{q}}}_{\text{i}}\right)=\{\begin{array}{l}{\text{r}}_{\text{FV}\text{,}}{}_{\text{i}}{\dot{\text{q}}}_{\text{i}}\left|{\dot{\text{q}}}_{\text{i}}\right|<{\mathrm{\omega }}_{\text{hs}}\\ \text{sgn}\left({\dot{\text{q}}}_{\text{i}}\right)\left\{{\text{r}}_{\text{FV}\text{,i}}{\mathrm{\omega }}_{\text{hs}}+{\text{r}}_{\text{ST}\text{,i}}\left(1-{\text{e}}^{\left(-\frac{{\text{r}}_{\text{FV}\text{,i}}}{{\text{r}}_{\text{ST}\text{,i}}}\left(\left|{\dot{\text{q}}}_{\text{i}}\right|<{\mathrm{\omega }}_{\text{hs}}\right)\right)}\right)\right\}\left|{\dot{\text{q}}}_{\text{i}}\right|\ge {\mathrm{\omega }}_{\text{hs}}\end{array}\end{array}$$

Here, ω _1s denotes a limit speed for low-speed operation, ω _hs denotes a limit speed for high-speed operation and r _ST,i denotes the saturation coefficient of friction. A graph of the nonlinear friction model is in 3 shown. 3 is a diagram showing an example of a nonlinear friction model, which is a physical model parameter to be estimated by the physical model storage unit 22 according to the first embodiment. In 3 the horizontal axis represents the engine speed, and the vertical axis represents the friction.

The nonlinear friction model has a total of four parameters in addition to the Coulombic friction coefficient r _FC,i and the viscous friction coefficient r _FV,i , and it is a nonlinear function depending on the speed qi. Therefore, these parameters are obtained from the difference between the actual torque and the first estimated torque using the Levenberg-Marquardt method, which is a nonlinear least squares method. The Levenberg-Marquardt method has an initial value dependence. Thus, the Coulombic friction coefficient r _FC,i and the viscous friction coefficient r _FV,i are added once to the physical model parameters w assuming the linearity of friction, and each friction coefficient is calculated by the above equation (1) and then as an initial value is set so that the failure of the estimation due to the initial value dependency can be prevented. The physical model storage unit 22 may thus use the least squares method, the regularized least squares method, the nonlinear least squares method, the Kalman filter, or the like in learning the physical model parameters. The numerical control device 100 can estimate unknown coefficients in the equation of motion and the equation of the state of the target machine 10 from motion data by using the least squares method, the nonlinear least squares method, the nonlinear least squares method, or the Kalman filter. The movement data is, for example, data contained in the engine status signals.

The physical model storage unit 22 stores coefficients of the nonlinear friction model obtained here as physical model parameters, and adds the output of the nonlinear friction model to the physical state variables each time. The numerical control device 100 can detect viscous friction due to aging, irregularities or the like and a change in Coulombic friction by storing coefficients of the friction model in the physical model storage unit 22 as physical model parameters. For example, when learning the physical model parameters, the physical model storage unit 22 alternately estimates the coefficients of the nonlinear friction model and other physical model parameters. The numerical control device 100 can design the friction model in more complexity and detail by alternately estimating the coefficients of the friction model and the other physical model parameters. In addition, since the numerical control device 100 can also estimate the other physical model parameters in a state where the influence of friction has been substantially eliminated, it is possible to estimate the physical model parameters with greater accuracy.

The nonlinear friction model can be a function of several variables, which depends not only on a velocity function but also on a position, a speed, a motor temperature and the like, and each coefficient of the nonlinear friction model can be determined by the Gauss-Newton method, the method of the steepest Descent, an extended Kalman filter, an unscented Kalman filter or the like can be identified. In addition, as the nonlinear friction model, a model generation method based on a machine learning algorithm such as a support Vector regression model, a regression tree, a random forest, a multivariate adaptive regression spline model or a neural network. However, only the macro influence that focuses on static properties of friction needs to be removed from the nonlinear friction model, so it is only necessary to consider the nonlinear friction model as a function of an explicitly stated velocity. This is because the first estimated torque is corrected by a machine learning unit 411 that includes a neural network in the machine learning model calculation unit 4 to be described below.

The cutting state observation unit 5 observes the state of cutting performed by the target machine 10 and outputs a cutting state signal indicating an observation result. The cutting status signal is a signal indicating whether the target machine 10 is cutting. The cutting state observation unit 5 obtains a machining program from the target machine 10 and observes the state of cutting based on the contents of the machining program.

Engine state signals output from the engine state observation unit 1, a first estimated torque and physical state variables output from the first torque estimating unit 21, and a state signal output from the cutting state observation unit 5 are input to the feature value generation unit 3. The feature value generation unit 3 uses the first estimated torque, the physical state variables, the engine state signals and the cutting state signal as inputs, and outputs feature values to be input to the machine learning unit 411.

For example, in a case where a cutting load is estimated, change information, machining information and application determination information are additionally input to the feature value generation unit 3. The tool information is information about the type and shape of the tool 14. The tool information includes information about a tool diameter, a tool length, a tool shape and the number of blades of the tool 14 or information about the length of a projection of the tool 14. If the tool 14 for Turning is used, the tool information may further include information about the shape of a tip of the tool 14 or information about a tool clearance angle of the tool 14. The machining information is information about a machining condition of the machining to be performed with the target machine 10. The machining information includes information about a machining condition, such as a feed speed or a feed width.

The application specification relates to a determination of a cutting irregularity determination, a cutting load determination and a tool wear or loss determination 14. The application specification information indicates which of the provisions for the irregularity determination, the cutting load determination and the wear or tear . Loss determination must be carried out. That is, the application determination information is information indicating which of the determinations is to be made for a plurality of events that may occur during processing.

The application determination information is input into the feature value generation unit 3 by a user of the target machine 10. The user selects an object for which the application determination is to be performed, such as “irregularity detection,” “cutting load detection,” and “tool wear and loss detection,” and inputs the application determination information indicating the selected object into the feature value generation unit 3. The object for which the application determination is to be carried out also includes a function in which “irregularity detection”, “cutting load detection” or “tool wear or loss detection” is used for internal processing. Examples of such a function include a function in which the user directly specifies a cutting load value via parameters such as cutting load control, and a function in which a user controls a load value via machining conditions without specifying the load value, such as friction stir welding. For each of these functions, “cutting load detection” is performed through internal processing. Note that when the numerical control device 100 performs only one detection among irregularity detection, cutting load detection, and wear or loss detection, the application determination information may be a fixed value. Alternatively, the application determination information does not need to be input into the feature value generation unit 3.

The feature value generation unit 3 determines which application determination to perform for “irregularity detection,” “cutting load detection,” and “tool wear or loss detection.” is from the application determination information entered. Based on the result of the determination, the feature value generation unit 3 determines whether to use machine learning model parameters based on the learning performed during cutting or machine learning model parameters based on the parameters learned during non-cutting for estimating a second estimated torque. In addition, the feature value generation unit 3 determines, based on the result of the determination, whether an actual torque obtained in cutting or an actual torque obtained in non-cutting in calculating subtracting the second estimated torque from the actual torque is to be used. The feature value generation unit 3 outputs feature values that contain information about the specific content.

These values input to the feature value generation unit 3 and the value of the actual torque to be estimated change in a time series manner. It is also necessary to consider a correlation between these in the timeline direction. Therefore, the machine learning model storage unit 42 also provides the feature values as vectors that also contain values obtained in the past. In addition, the influence of interactions between axes is not small in a multi-axis mechanism such as a machine tool for which the first embodiment is intended. Therefore, it is desirable that the feature value be a vector that also contains an input value corresponding to each axis. However, a torque component in the feature value is specified as a difference between the actual torque and the first estimated torque, and only a past value is used for the actual torque to be estimated.

4 is a block diagram showing a configuration example of the feature value generation unit 3 and the machine learning model calculation unit 4 according to the first embodiment. In 4 z is a unit delay operator, and N _d is an integer greater than or equal to zero indicating the number of unit times of delay. You can also say that the in 4 The unit delay operators shown are storage devices provided in the feature value generation unit 3. This means that the feature value generation unit 3 includes storage devices that store the first estimated torque, the physical state variables, the motor state signals, the cutting state signal, the tool information, the machining information and the application determination information.

The feature value generation unit 3 extracts the information to be used as a feature value from the information stored in the storage devices. Since the numerical control device 100 includes as input time series change information on the physical state variables and the engine state signals in the feature values, the machine learning model calculation unit 4 can predict the operation of the target machine 10 more accurately. As a result, the numerical control device 100 can improve the accuracy of estimating the second estimated torque with respect to the actual torque.

When generating the feature values, the feature value generation unit 3 does not necessarily need to use all of the information from the first estimated torque, the physical state variables, the engine state signals, the cutting state signal, the tool information, the machining information, and the application determination information. The feature value generation unit 3 can generate feature values using only some of the first estimated torque, the physical state variables, the engine state signals, the cutting state signal, the tool information, the machining information, and the application determination information. The feature value generation unit 3 can generate feature values from the cutting state signal and at least one of the first estimated torque, the physical state variables and the engine state signals, and output the generated feature values.

The second torque estimation unit 41 includes the machine learning unit 411. The second torque estimation unit 41 uses the feature values generated by the feature value generation unit 3, the first estimated torque output from the first torque estimation unit 21, and the machine learning model parameters output from the machine learning model storage unit 42 , as inputs, and calculates a machine learning correction value based on machine learning performed by the machine learning unit 411 using the feature values and the machine learning model parameters. The second torque estimation unit 41 outputs a value obtained by adding the machine learning correction value and the first estimated torque as the second estimated torque.

The machine learning unit 411 includes a neural network. The neural network uses the feature values generated by the feature value generation unit 3 and the machine learning model parameters output from the machine learning model storage unit 42 as inputs, and outputs the machine learning correction value so that an error in the second estimated torque with respect to the actual torque is minimized. This means that an ideal value of the machine learning correction value to be output corresponds to the difference between the actual torque and the first estimated torque and includes a difference between a delayed past actual torque and the first estimated torque in the input feature values. Thus, the machine learning unit 411 in one in 5 The input-output relationship shown has an autoregressive term.

5 is a diagram showing an autoregressive term in the configuration of the feature value generation unit 3 and the machine learning model calculation unit 4 according to the first embodiment. 5 shows part of the in 4 configuration shown, which is related to the autoregressive term. Here, the machine learning model parameters refer to a weighting matrix and a bias vector in the neural network. Further, the use of a neural network called a recurrent type neural network is also an effective means of improving the accuracy of prediction of such time series data. The recurrent type neural network has a recursive structure in the input-output relationship of an intermediate layer. Examples of the recurrent type neural network include a recurrent neural network, a long-short-term memory (LSTM), and a gated recurrent unit (GRU).

To reduce the learning load and the amount of calculations, a polynomial model, a multiple regression model, a support vector regression model, a regression tree, a random forest, a multivariate adaptive regression spline model, or the like can be used instead of the neural network. Alternatively, in a case where, for example, the interaction between the axes is neglected, the machine learning unit 411 may be applied to each axis separately.

However, general-purpose function approximation models for use in the machine learning unit 411, such as a neural network, are generally not suitable for approximating discontinuous functions. Therefore, when a torque corresponding to friction is approximated at the first estimated torque by using the linear function model of equation (2), there is a possibility that the difference between the actual torque and the first estimated torque produces a discontinuous jump , and this in turn may cause a significant deterioration in the results of learning the machine learning model parameters and calculating the machine learning correction value performed by the machine learning unit 411, then these values are included as input values and output values and the autoregressive term is formed. Furthermore, it is very difficult to explicitly apply these models to an actual machine tool. This is because the friction that needs to be generated when the speed is close to zero and the operation is performed at low speed has a strong nonlinearity due to the influence of a stick-slip phenomenon occurring alternately adhesion and slippage, friction related to lubrication as represented by the Stribeck curve, and the like.

Therefore, using a friction model with smooth continuity near zero velocity, such as the nonlinear friction model shown in Equation (3), makes each value of the first estimated torque and feature values continuous. As a result, it is possible to estimate even a friction component that is difficult to reproduce with high accuracy with the first estimated torque and the above-described linear friction model by adding a machine learning correction value. In addition, the feature values are vectors that also contain, for example, velocities and outputs of the nonlinear friction model that were obtained in a certain past time. Therefore, the machine learning unit 411 can also learn an estimation error factor that changes depending on an immediately preceding condition, such as the hysteresis characteristics of friction, and can make a correction in the form of a machine learning correction value.

As described above, the second torque estimation unit 41 calculates a machine learning correction value from the feature values and the machine learning model parameters by using machine learning with the autoregressive term in the input-output relationship. The second torque estimation unit 41 adds the calculated machine learning correction value to the first estimated one Torque and outputs the second estimated torque. In addition, the second torque estimation unit 41 may use machine learning using a neural network as machine learning that includes an autoregressive term in the input-output relationship. The numerical control device 100 can learn and store a torque component that is difficult to express in the equation of motion as a nonlinear model with many degrees of freedom by using a neural network as a machine learning method.

The second torque estimation unit 41 may use the difference between the actual torque and the first estimated torque in machine learning to calculate the machine learning correction value as an autoregressive term to be included in the input-output relationship. The difference between the actual torque and the first estimated torque is a target value of a correction quantity with respect to the first estimated torque. The target value of the correction quantity changes as the engine of the target machine 10 operates on a time series basis. Therefore, the numerical control device 100 provides an autoregressive term having a past target value of a correction quantity as input. As a result, the numerical control device 100 can estimate an optimal correction value for the next time with greater accuracy.

After the physical model parameters in the physical model storage unit 22 are learned, the machine learning model storage unit 42 learns the machine learning model parameters so that an error in the machine learning correction amount with respect to the difference between the actual torque and the first estimated torque is reduced by obtaining feature values when the target machine is operated for a certain period of time, are used as training data, and by using the difference between the actual torque and the first estimated torque as training data. Accordingly, the machine learning model storage unit 42 learns the machine learning model parameters through supervised learning such that the machine learning correction value approaches the difference between the actual torque and the first estimated torque. The machine learning model storage unit 42 stores the learned machine learning model parameters.

When learning begins, the machine learning model parameters can be initialized with any given constants. Specifically, the machine learning unit 411 expressed by a neural network has an initial value dependence on a weight matrix, a bias and the like, which are machine learning model parameters. Therefore, the machine learning unit 411 initializes the weight matrix with a probability distribution derived from a normal distribution, in addition to a uniform distribution and a general normal distribution. Examples of the probability distribution derived from a normal distribution include a truncated normal distribution, the He normal distribution, and the Glorot normal distribution.

In addition, the machine learning model storage unit 42 uses the mini-batch gradient descent method and the error back propagation method as the learning method and updates the machine learning model parameters represented as the weight matrix and the bias vector for each batch. That is, the machine learning model storage unit 42 can use the error back propagation method as a supervised learning method at the time of learning the machine learning model parameters. However, in a case where a sufficient amount of training data cannot be prepared, for example, the machine learning model storage unit 42 may use the general gradient descent method, the stochastic gradient descent or the like instead of the mini-batch gradient descent method.

In addition, using a method such as dropout or early stopping is also an effective approach that can be addressed by the machine learning model storage unit 42 to improve the generalization ability of the neural network in the machine learning unit 411. Dropout is a procedure in which neurons are randomly removed at the time of learning. Early stopping is a technique that allows learning to stop soon based on monitoring an error for each epoch. It is necessary to apply the dropout probability to the weight matrix at the same time as inferring the machine learning correction amount when using dropout in learning. Therefore, the machine learning model storage unit 42 also adds such a hyperparameter as a machine learning model parameter and stores this point

The process of learning the physical model parameters and the machine learning model parameters is described below with reference to a flowchart. 6 is a flowchart that shows a flow of learning physical model parameters and machine learning model parameters in the numerical control device 100 according to the first embodiment.

The physical model storage unit 22 sets initial values of a mass, a barycentric intermediate position and an inertia tensor, which are physical model parameters for each link, based on values in the specification table (step S1). The engine condition observation unit 1 outputs time series data on an actual torque, a position, a speed and an acceleration, which are engine condition signals of the target machine 10, as training data (step S2).

The first torque estimation unit 21 calculates a first estimated torque by dynamic calculation with respect to the position, the speed and the acceleration (step S3). The first torque estimation unit 21 adds a linear friction model term to a dynamic equation (step S4). The physical model storage unit 22 adds the Coulomb friction coefficient and the viscous friction coefficient to the physical model parameters (step S5). The physical model storage unit 22 identifies the physical model parameters based on the difference between the actual torque and the first estimated torque by using the dynamic equation transformed into a linear equation and the least square method (step S6).

The physical model storage unit 22 sets the Coulomb friction coefficient and the viscous friction coefficient in the identified physical model parameters as initial values, identifies a nonlinear friction model by the Levenberg-Marquardt method, and adds each of the obtained coefficients to the physical model parameters (step S7 ). The physical model storage unit 22 again identifies physical model parameters other than the friction coefficients by applying the dynamic equation and the least squares method to a result obtained by subtracting the values of the first estimated torque and the nonlinear friction model from the actual one Torque was obtained, wherein the dynamic equation is transformed into a linear equation by eliminating the term of the friction model (step S8). The first torque estimation unit 21 adds a term of the nonlinear friction model to the dynamic equation and recalculates the first estimated torque using the learned physical model parameters (step S9).

The feature value generation unit 3 takes the first estimated torque, physical state variables, the engine state signals, cutting state information, machining information, tool information and application determination information as inputs and assembles these values as vectors that also include values obtained in a certain period of time in the past. Then, the feature value generation unit 3 outputs the vectors as feature values (step S10). The machine learning model storage unit 42 initializes a weight matrix and a bias, which are machine learning model parameters, such that the weight matrix is initialized with a normal distribution and the bias is initialized with a constant (step S11).

The machine learning model storage unit 42 sets hyperparameters for early stopping, dropout, or the like as conditions for learning and adds their values to the machine learning model parameters (step S12). The second torque estimation unit 41 uses the feature values as inputs and learns the weight matrix and the bias, which are machine learning model parameters, by the error back propagation method according to the conditions for learning in the machine learning model parameters so that the ideal value of a machine learning correction value is output, which is the difference between the actual torque and the first estimated torque (step S13).

Thus, as described above, the physical model parameters are pre-learned in the physical model storage unit 22, and the machine learning model parameters are learned in the machine learning model storage unit 42 by using the machine learning correction value for the first estimated torque as the autoregressive term. This greatly contributes to improving the generalization ability of the second estimated torque. This is because the numerical control device 100 can separately learn an error component of torque estimation that comes from a known physical phenomenon that can be easily modeled in the equation of motion and an error component of torque estimation that is difficult to model in the equation of motion. by distinguishing the physical model parameters and the machine learning model parameters, resulting in a reduction in the absolute value of the machine learning correction value which is output by the machine learning unit 411. In other words, since no explicitly specified torque estimation error component is included in the machine learning correction value, the numerical control device 100 can also reduce the number of parameters included in the machine learning model parameters. As a result, it is possible to prevent over-learning, increase the speed of learning, and reduce the training data to be used for learning in the machine learning model storage unit 42.

Similarly, regarding the autoregressive term with respect to the machine learning correction value, it is possible to reduce the range of the torque estimation error in the neural network or the like. This is because the absolute value of the machine learning correction value is reduced compared to a configuration in which the machine learning model calculation unit 4 learns a difference between the actual torque and the result of calculation based on the equation of motion using initial values for the physical model parameters and the difference is used as the machine learning correction value, or a configuration in which the result of the calculation based on the equation of motion using initial values for the physical model parameters is included in the feature values and the machine learning model parameters are learned so that the second estimated torque is obtained from the machine learning unit 411 directly is issued. Advantages of using the autoregressive term also include, for example, the ease of learning estimation error factors that change periodically over time and are difficult to provide for in the equation of motion, such as cogging torque that depends on motor characteristics, torque ripple, and the influence of kickback in the Torque transmission between gearboxes.

7 shows the first estimated torque as currently obtained after learning the physical model parameters and the second estimated torque as currently obtained by learning the machine learning model parameters based on the difference between the actual torque and the first estimated torque. 7 is a diagram showing an engine speed representing the operation of the target machine 10 to be controlled by the numerical control device 100 according to the first embodiment, the first estimated torque estimated by the numerical control device 100, and the second estimated torque determined by the numerical control device 100 is estimated. For each element, the horizontal axis represents time. As in 7 shown, the second estimated torque is closer to the actual torque than the first estimated torque.

Further, in a case where the neural network is used for the machine learning unit 411, it is not clear which parameter included in the feature values influenced the machine learning correction value and to what extent the parameter influenced the machine learning correction value. It is therefore very difficult to estimate a change in the physical properties of the target machine 10 and a change in the machine learning model parameters. However, in the first embodiment, in which the respective values of the physical properties can be directly checked when learning the physical model parameters, the numerical control device 100 distinguishes the physical model parameters and the machine learning model parameters when learning these parameters. Therefore, it can be said that the configuration of the first embodiment is suitable for irregularity detection.

A method that substantially monitors a difference between the actual torque and the second estimated torque may be mentioned as an example of the irregularity detection using the physical model calculation unit 2 and the machine learning model calculation unit 4 that have performed learning. This is because the numerical control device 100 monitors the difference between the actual torque and the second estimated torque as a result, the observed values are substantially close to zero, so the absolute value is considered as the measure of irregularity for the actual torque can be. Particular irregularity detection methods that may be used include a simple method of detecting an irregularity based on whether the absolute value of the difference between the actual torque and the second estimated torque exceeds a predetermined threshold. In addition, a determination can be made using a method for distinguishing between two classes of normality and irregularity in a two-class classification problem using logistic regression, a support vector machine, a decision tree, a neural network or the like , where the feature values and the difference between the actual torque and the second estimated torque are used as inputs and labeled training data is prepared.

When it is difficult to collect motion data corresponding to an irregular state as training data, the numerical control device 100 may perform oversampling with motion data corresponding to an irregular state using a Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic (ADASYN ) or the like, or it can classify motion data into two clusters of normality and irregularity by using unsupervised learning such as the K-Means method or self-organizing map. Further, in the numerical control device 100, the number of classes and the number of clusters are not limited to two (normality and irregularity), and they can be gradually finer adjusted in accordance with the degree of irregularity. In addition, the degree of irregularity can be freely set to a value different from the absolute value of the difference between the actual torque and the second estimated torque.

Learning the physical model parameters in the physical model storage unit 22 and learning the machine learning model parameters in the machine learning model storage unit 42 are in 1 indicated by dashed lines. Assuming that the learning of these parameters is for failure detection, it is desirable to perform the learning of these parameters in an environment in which the numerical control device 100 is in an ideal state for comparison with respect to the degree of deterioration. For example, it is desirable to perform the learning of these parameters at the time of test operation upon delivery. However, the spindle and the tool 14 are attached to the tip, and then the workpiece 15 arranged on the table 13 is cut with the machine tool. This changes physical properties of each term, such as in this case the mass, the barycentric position and the inertia tensor. Therefore, for example, the numerical control device 100 needs to relearn both the physical model parameters and the machine learning model parameters in each state to take the mass of the workpiece 15 into account.

When the workpiece 15 is placed, it generally takes time and effort to adjust the influence on these physical model parameters using the numerical control device 100 or the like. In the first embodiment, the physical model storage unit 22 learns engine state signals that are output when the target machine 10 is operated for a certain period of time. As a result, a hiring process can also be automated. Furthermore, since the values of the physical model parameters are set correctly, there is the advantage that the cycle time, trajectory deviation, pressure forces in collisions and the like are reduced. This is a technique that can also be used to improve the speed and accuracy of force control and, in addition to position control, velocity control and the like, impedance control. The application of the method described above is therefore not limited to irregularity detection.

The machine learning model storage unit 42 separately performs supervised learning for a period during which the target machine 10 is cutting and a period during which the target machine 10 is not cutting. The machine learning model storage unit 42 stores machine learning model parameters learned based on various values input and output during cutting and machine learning parameters input and output during non-cutting. Feature values including a cutting state signal and application determination information are input to the machine learning model calculation unit 4. As a result, the machine learning model calculation unit 4 can perform machine learning by switching between learning for the cutting period and learning for the non-cutting period.

8th is a diagram for describing the irregularity detection to be performed by the numerical control device 100 in the first embodiment. 8th shows the actual torque and the second estimated torque to be estimated by the numerical control device 100. The horizontal axis represents time. The processing described here is executed when the application determination information indicates the determination for detecting an irregularity.

In the first embodiment, when the application determination information indicating the irregularity detection is inputted as a feature value, the second torque estimation unit 41 of the machine learning model calculation unit 4 monitors a difference between the actual torque and a second estimated torque calculated by the machine learning model calculation unit 4. The second torque estimation unit 41 calculates the second estimated torque the machine learning model parameters learned based on values input and output while not cutting, and monitors a difference between the actual torque and the calculated second estimated torque. The second torque estimation unit 41 detects the occurrence of the irregularity when the difference reaches or exceeds a certain value. The irregularity concerns an irregularity in the torque of the engine.

When the cutting state signal indicates that no cutting is taking place, the numerical control device 100 may detect an irregularity in torque by causing the machine learning model calculation unit 4 to perform the calculation using a learned model intended for use in a non-cutting state. When an irregularity is detected, the numerical control device 100 adjusts a parameter of a machining condition so that the difference between the actual torque and the second estimated torque is equal to or smaller than the predetermined value described above. For example, the numerical control device 100 makes an adjustment to reduce the feed rate. When an irregularity is detected, the numerical control device 100 may issue an alarm indicating the detected irregularity.

Second embodiment.

In a second embodiment, a case where a cutting load is detected by the numerical control device 100 will be described. In the second embodiment, the same constituent elements as in the above first embodiment are denoted by the same reference numerals, and description overlapping with the first embodiment is omitted. 9 is a diagram for describing the cutting load detection to be performed by the numerical control device 100 in the second embodiment. 9 shows the actual torque and the second estimated torque to be estimated by the numerical control device 100. The horizontal axis represents time.

The processing in the second embodiment is executed when the application determination information indicates the determination for detecting a cutting load. When a cutting state signal indicates that cutting is to be performed, the first torque estimation unit 21 adds a nonlinear friction model term to the dynamic equation and calculates a first estimated torque using the learned physical model parameters. The feature value generation unit 3 uses the first estimated torque, physical state variables, engine state signals, a cutting state signal, tool information, machining information and application determination information as inputs, and outputs feature values to be input to the machine learning unit 411.

When application determination information indicating cutting load detection and a cutting state signal indicating not to cut are input as feature values, a weight matrix that is a machine learning model parameter is stored in the machine learning model storage unit 42. The second torque estimation unit 41 uses the feature values as inputs and learns the weight matrix and a bias, which are machine learning model parameters, by the error back propagation method in accordance with conditions for learning in the machine learning model parameters so that the ideal value of a machine learning correction value, which is a Difference between the actual torque and the first estimated torque is output.

In a learning model based on learning based on values input and output when a cutting state signal indicates not to cut, a second estimated torque calculated by the second torque estimation unit 41 approaches an actual torque to operate the target machine 10. However, when the cutting state signal indicates that cutting is to be performed, a torque equal to the sum of a torque for operating the target machine 10 and a torque for cutting the workpiece 15 is measured as an actual torque to be generated when the Target machine 10 is currently cutting the workpiece 15. Therefore, the numerical control device 100 can obtain a cutting load torque, which is a torque required when the target machine 10 cuts the workpiece 15, by subtracting the second estimated torque from the measured actual torque. Cutting load detection concerns obtaining the cutting load torque.

In the second embodiment, when the application determination information indicating the determination for detecting a cutting load is inputted as a feature value, the second torque estimation unit 41 of the machine learning model calculation unit 4 obtains the cutting load torque, which is a difference between the actual torque and the second estimated torque obtained by the machine learning model calculation unit 4 is calculated. The second torque estimation unit 41 calculates the second estimated torque from the machine learning model parameters learned based on values input and output while not cutting, and obtains the cutting load torque, which is the difference between the actual torque and the calculated second estimated torque is. When the cutting state signal indicates that no cutting is taking place, the numerical control device 100 may detect a cutting load by causing the machine learning model calculation unit 4 to perform the calculation using a learning model intended for use in a non-cutting state.

Third embodiment.

In a third embodiment, a case in which the numerical control device 100 detects wear or loss of the tool 14 will be described. In the third embodiment, constituent elements that are the same as those of the first or second embodiment are denoted by the same reference numerals, and description overlapping with the first or second embodiment is omitted. 10 is a diagram for describing the detection of wear or loss of the tool 14 to be performed by the numerical control device 100 in the third embodiment. 10 shows the actual torque and the second estimated torque to be estimated by the numerical control device 100. The horizontal axis represents time.

The processing in the third embodiment is executed when the application determination information indicates the determination for detecting wear or loss of the tool 14. When a cutting state signal indicates to be cut, the first torque estimation unit 21 adds a nonlinear friction model term to the dynamic equation and calculates a first estimated torque using the learned physical model parameters. The feature value generation unit 3 uses the first estimated torque, physical state variables, engine state signals, a cutting state signal, tool information, machining information and application determination information as inputs, and outputs feature values to be input to the machine learning unit 411.

When the application determination information indicating the determination for detecting wear or loss of the tool 14 and a cutting state signal indicating that cutting is to be performed are inputted as feature values, a weight matrix that is a machine learning model parameter is stored in the machine learning model storage unit 42. The second torque estimation unit 41 uses the feature values as inputs and learns the weight matrix and a bias, which are machine learning model parameters, by the error back propagation method in accordance with the conditions for learning in the machine learning model parameters so that the ideal value of a machine learning correction value is output , which is a difference between the actual torque and the first estimated torque.

In a learning model based on learning based on values input and output when the cutting state signal indicates that is being cut, a second estimated torque calculated by the second torque estimation unit 41 may be estimated to be is an actual torque, which is a sum of an actual torque for operating the target machine 10 and a cutting load torque. In a case where the application determination information indicates the determination for detecting wear or loss of the tool 14, the numerical control device 100 may obtain a cutting load torque increased by wear of the tool 14 by taking a measured actual torque from the second estimated torque deducts. Therefore, the numerical control device 100 can detect the degree of wear of the tool 14 from the cutting load torque, which is a difference between the second estimated torque and the measured actual torque. In a case where the application determination information indicates the determination for detecting the wear or loss of the tool 14, the numerical control device 100 may further detect the loss of the tool when an actual torque for operating the target machine 10 is only the second estimated torque to be measured is.

In the third embodiment, when the application determination information indicating the determination for detecting wear or loss of the tool 14 is inputted as a feature value, the second torque estimation unit 41 of the machine learning model calculation unit 4 monitors a difference between the actual torque and the second estimated torque determined by the machine learning model calculation unit 4 is calculated. The second torque estimation unit 41 calculates the second estimated torque from the machine learning model parameters learned based on values input and output while cutting, and monitors a difference between the actual torque and the calculated second estimated torque.

When the measured actual torque is larger than the second estimated torque, the second torque estimation unit 41 estimates an increase in the difference between the second estimated torque and the measured actual torque as an increase in the cutting load torque due to wear of the tool 14. The second torque estimation unit 41 detects the Degree of wear of the tool 14 based on the increase in cutting load torque. In addition, when the measured actual torque is smaller than the second estimated torque or when the measured actual torque is extremely smaller than the second estimated torque, the second torque estimation unit 41 detects that the loss of the tool 14 has occurred.

When the cutting state signal indicates that it is being cut, the numerical control device 100 can detect the wear or loss of the tool 14 by causing the machine learning model calculation unit 4 to perform the calculation using a learning model intended for use in a cutting state. The numerical control device 100 delays and stops the target machine 10 when the wear or loss of the workpiece 14 is detected. If the wear or loss of the tool 14 is detected, the numerical control device 100 may issue an alarm indicating that the wear or loss of the tool 14 has been detected.

Now, a hardware configuration of the numerical control device 100 according to each of Embodiments 1 to 3 will be described. In the numerical control device 100, the engine state observation unit 1, the physical model calculation unit 2, the feature value generation unit 3, the machine learning model calculation unit 4, and the cutting state observation unit 5 are implemented by processing circuits. The processing circuitry may be a memory and a processor that executes programs stored in the memory, or they may be dedicated hardware.

11 is a diagram showing an example in which processing circuits included in each of the numerical control devices 100 according to Embodiments 1 to 3 include a processor 91 and a memory 92. In a case where the processing circuits include the processor 91 and the memory 92, each function of the processing circuit of the numerical control device 100 is implemented by software, firmware, or a combination of software and firmware. The software or firmware is described as a program and stored in memory 92. The processor 91 reads the program stored in the memory 92 and executes it to implement each function of the processing circuits. Thus, the processing circuits include memory 92 for storing programs. As a result of executing the programs, the numerical control device 100 performs processing. In addition, these programs can also be said to cause the computer to execute a procedure and procedure for the numerical control device 100.

The processor 91 may be a central processing unit (CPU), a processing device, an arithmetic device, a microprocessor, a microcomputer, a digital signal processor (DSP), or the like. Further, examples of the memory 92 include a non-volatile or volatile semiconductor memory, such as a random access memory (RAM), a read-only memory (ROM), a flash memory, an erasable programmable ROM (EPROM), and an electrical EPROM (EEPROM (Registered Trade Name). )), a magnetic disk, a flexible disk, an optical disk, a compact disc, a mini disc and a DVD.

12 is a diagram showing an example in which processing circuits included in each of the numerical control devices 100 according to Embodiments 1 to 3 include dedicated hardware. In a case where the processing circuitry includes dedicated hardware Examples of the in 12 Processing circuit 93 shown includes a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), and a combination thereof. The functions of the numerical control device 100 may be implemented separately by the processing circuit 93, or they may be implemented collectively by the processing circuit 93.

It should be noted that some of the functions of the numerical control device 100 may be implemented by dedicated hardware and some of the other functions thereof may be implemented by software or firmware. The processing circuitry may implement any of the functions described above using dedicated hardware, software, firmware, or combinations thereof.

As described above, in Embodiments 1 to 3, the model of the numerical control device 100 is expressed by a combination of the physical model calculation unit 2 and the machine learning model calculation unit 4. The physical model calculation unit 2 estimates the torque using the mechanical physical properties of the target machine 10 and information about a physical phenomenon explicitly stated in the equation of motion. The machine learning model calculation unit 4 estimates a torque component that has a strong nonlinearity and is difficult to express in the equation of motion, wherein a correction value for the torque estimated by the physical model calculation unit 2 is included as an autoregressive term. The numerical control device 100 learns these elements sequentially. As a result, the numerical control device 100 can learn the model of the target machine 10, which has a large torque estimation accuracy and a large generalization ability with respect to the actual torque, with a small amount of motion data and at a high speed. Furthermore, the numerical control device 100 can also detect a small irregularity by comparing the actual torque with a torque estimate, i.e. H. compares the actual machine with the learned model.

The numerical control device 100 thus learns the physical model parameters and the machine learning model parameters sequentially. As a result, the numerical control device 100 can control the mechanical physical properties of the target machine 10, a torque component that arises from a physical phenomenon that is explicitly specified in the equation of motion, and a torque component that has strong nonlinearity and is difficult to express in the equation of motion , learn and appreciate separately. As a result, the numerical control device 100 can shorten the time necessary for the entire learning and reduce the absolute value of the machine learning correction value by learning the physical model parameters, so that the generalization ability of the second estimated torque with respect to the actual torque can be improved. Further, the numerical control device 100 can improve the accuracy of estimation of the second estimated torque by causing the second torque estimation unit 41 to efficiently learn the correlation between the feature values in the time direction based on the autoregressive term of the input-output relationship. As described above, the numerical control device 100 enables highly accurate torque estimation even for complicated phenomena with strong nonlinearity.

Note that Embodiments 1 to 3 are based on the assumption that the target machine 10, which is a machine tool, is controlled by the numerical control device 100, however, the numerical control device 100 can also be used for any other machine regardless Number of axes of the machine and mechanism for linear movement or rotation of a joint. The numerical control device 100 can also be used to control cutting performed by a robot or the like, a assembler or the like.

The configurations set forth in each of the above embodiments show examples of the subject matter of the present disclosure. The configuration of each embodiment may be combined with other known techniques. The respective configurations of the embodiments can be suitably combined with each other. It is possible to partially omit or change parts of the configuration of each embodiment without departing from the scope of the present disclosure.

Reference symbol list

1

engine condition observation unit;

2

model calculation unit;

3

feature size generation unit;

4

machine learning model calculation unit;

5

cutting condition observation unit;

10

target machine;

11

Pillar;

12

spindle head;

13

Table;

14

Tool;

15

Workpiece;

21

first torque estimation unit;

22

Physical model storage unit;

41

second torque estimation unit;

42

machine learning model storage unit;

41

Processor;

92

Storage;

93

processing circuit;

100

numerical control device;

411

Machine learning unit.

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• JP 2013128387 [0004]

Claims (19)

A numerical control device for controlling a motor of a machine tool that cuts a workpiece, the numerical control device comprising: an engine condition observation unit for outputting engine condition signals indicating results of observation of a condition of the engine; a cutting condition observation unit for outputting a cutting condition signal indicating whether the machine tool is cutting; a first torque estimation unit for outputting a first estimated torque and physical state variables based on the engine state signals and physical model parameters of the machine tool, the first estimated torque being an estimated value of the torque of the engine, the physical state variables being components of the first estimated torque; a feature value generation unit for generating feature values from the cutting state signal and the first estimated torque, and/or the physical state variables and/or the engine state signals, and outputting the generated feature values; and a machine learning model calculation unit for calculating a machine learning correction value from the feature values and the machine learning model parameters, learning the machine learning model parameters such that the calculated machine learning correction value approximates a difference between an actual torque included in the engine state signals and the first estimated torque, and the learned Store machine learning model parameters. Numerical control device according to Claim 1 , wherein the feature value generation unit includes a storage device and extracts information to be used as the feature values from the information stored in the storage device, the storage device storing the first estimated torque, the physical state variables, the engine state signals, the cutting state signal, tool information and machining information, wherein the tool information is information about a tool to be used by the machine tool, wherein the machining information is information about a machining condition for the machining to be performed by the machine tool. Numerical control device according to Claim 2 , wherein the machine learning model calculation unit stores the machine learning model parameters learned based on values input and output while cutting, and the machine learning model parameters learned based on values input and output while not cutting. Numerical control device according to Claim 3 , wherein the machine learning model calculation unit calculates a second estimated torque, monitors a difference between the actual torque and the calculated second estimated torque, and detects an irregularity based on the difference, wherein the machine learning model calculation unit calculates the second estimated torque by adding the machine learning correction value to the first estimated torque added, which was calculated from the machine learning model parameters learned based on the values input and output while not cutting. Numerical control device according to Claim 3 , wherein the machine learning model calculation unit calculates a second estimated torque and obtains a cutting load torque that is a difference between the actual torque and the calculated second estimated torque, wherein the machine learning model calculation unit calculates the second estimated torque by adding to the first estimated torque the machine learning correction value, which was calculated from the machine learning model parameters based on the values input and output while not cutting. Numerical control device according to Claim 3 , wherein the machine learning model calculation unit calculates a second estimated torque, monitors a difference between the actual torque and the calculated second estimated torque, and detects wear or loss of the tool based on the difference, the machine learning model calculation unit calculates the second estimated torque by adding to the first estimated torque the machine learning correction value calculated from the machine learning model parameters learned based on values input and output while cutting. Numerical control device according to one of Claims 2 until 6 , wherein the storage device further stores determination information indicating which determination to make for detecting a plurality of events in editing. Numerical control device according to one of Claims 2 until 7 , wherein the tool information includes information about a number of blades of the tool or information about a length of a projection of the tool. Numerical control device according to one of Claims 2 until 7 , wherein the tool is a tool to be used for turning, and wherein the tool information includes information about a shape of a tip of the tool or information about a clearance angle of the tool. Numerical control device according to one of Claims 1 until 9 , wherein the machine learning model calculation unit calculates the machine learning correction value using a neural network as the machine learning that includes an autoregressive term in an input-output relationship, and uses an error back propagation method when learning the machine learning model parameters. Numerical control device according to Claim 10 , where the machine learning model calculation unit uses the difference between the actual torque and the first estimated torque as the autoregressive term. Numerical control device according to one of Claims 1 until 11 , wherein the first torque estimation unit outputs the first estimated torque and the physical state variables using an equation of motion based on the engine state signals and the physical model parameters. Numerical control device according to Claim 12 , wherein the equation of motion includes a steady and smooth friction model, the friction model using as an input a speed of the engine obtained from the engine state signals, the friction model outputting an estimated value of a friction component of the torque. Numerical control device according to one of Claims 1 until 11 , further comprising: a physical model calculation unit for learning the physical model parameters so that the difference between the actual torque and the first learned torque is reduced, and for storing the learned physical model parameters. Numerical control device according to Claim 14 , wherein the first torque estimation unit outputs the first estimated torque and the physical state variables by using an equation of motion based on the engine state signals and the physical model parameters, the equation of motion using a steady and smooth friction model, the friction model a speed of the engine obtained from the engine state signals used as an input, wherein the friction model outputs an estimated value of a friction component of the torque, and wherein the physical model calculation unit learns coefficients of the friction model as the physical model parameters and stores the learned coefficients. Numerical control device according to Claim 15 , where the physical state variables comprise the friction model. Numerical control device according to Claim 16 , wherein the physical model calculation unit alternately estimates the coefficients of the friction model and other physical model parameters when learning the physical model parameters. Numerical control device according to Claim 16 or 17 , wherein the physical model calculation unit uses a least squares method, a regularized least squares method, a nonlinear least squares method or a Kalman filter in learning the physical model parameters. A numerical control device for controlling a motor of a machine tool that cuts a workpiece, the numerical control device comprising: an engine condition observation unit for outputting engine condition signals indicating results of observation of a condition of the engine; a cutting condition observation unit for outputting a cutting condition signal indicating whether the machine tool is cutting; a first torque estimation unit to output a first estimated torque and physical state variables based on the motor state signals and physical model parameters of the machine tool, where the first estimated torque is an estimate of the torque of the engine, wherein the physical state variables are components of the first estimated torque; a feature value generation unit for generating feature values from the cutting state signal and the first estimated torque, the physical state variables and/or the engine state signals and outputting the generated feature values; and a second torque estimation unit for outputting a second estimated torque by adding to the first estimated torque a machine learning correction value calculated from the feature values and the machine learning model parameters such that the calculated machine learning correction value is a difference between an actual torque included in the engine state signals , and approximates the first estimated torque.

Images