Engine control unit

The motor control device addresses inaccuracies in torque estimation and abnormality detection by integrating physical and machine learning models to correct for complex non-linear phenomena, enhancing accuracy and precision in motor control.

DE112019007222B4Active Publication Date: 2024-11-21MITSUBISHI ELECTRIC CORP
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
DE112019007222
Authority / Receiving Office
DE · DE
Patent Type
Patents
Current Assignee / Owner
Filing Date
2019-04-16
Publication Date
2024-11-21
Estimated Expiration
2039-04-16

AI Technical Summary

Technical Problem

Existing motor control methods struggle to accurately model and correct for complex non-linear phenomena such as friction and vibration, leading to inaccuracies in torque estimation and difficulty in detecting abnormalities.

Method used

A motor control device that incorporates an engine condition observation unit, a first torque estimation unit using physical model parameters, a physical model storage unit for learning, a feature amount generation unit, and a machine learning model calculation unit with a second torque estimation unit to improve torque estimation accuracy and generalization performance.

Benefits of technology

Enhances torque estimation accuracy and abnormality detection by learning a model that corrects for non-linear friction components, improving the speed and precision of motor control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

Motor control device (100) for controlling a motor to drive a target machine (10), the device comprising the following: an engine condition monitoring unit (1) to monitor the condition of the engine and to output an observation result as an engine condition signal; a first torque estimation unit (21) to output a first estimated torque and physical state variables using an equation of motion based on the motor state signal and physical model parameters of the target machine (10), wherein the first estimated torque is an estimate of a torque of the motor, and the physical state variables are components of the first estimated torque; a storage unit for physical models (22) to learn and store the physical model parameters in such a way as to reduce the difference between an actual torque and the first estimated torque, with the actual torque being included in the motor state signal; a feature amount generation unit (3) to generate and output a feature amount, wherein the feature amount generation unit (3) generates the feature amount based on at least one of the first estimated torque, the physical state variables and the motor state signal; a second torque estimating unit (41) to calculate a machine learning correction amount and to add the calculated machine learning correction amount to the first estimated torque to output a second estimated torque, wherein the second torque estimating unit (41) calculates the machine learning correction amount based on the feature value and a parameter of a machine learning model using machine learning including an autoregressive term in an input-output relationship; and a machine learning model storage unit (42) for learning the parameter of the machine learning model by supervised learning and storing the learned parameter of the machine learning model, wherein the machine learning model storage unit (42) learns the parameter of the machine learning model such that the correction amount of the machine learning is closer to the difference between the actual torque and the first estimated torque.
Need to check novelty before this filing date? Find Prior Art

Description

Area

[0001] The present invention relates to an engine control device that generates a model simulating a control target. General state of the art

[0002] For motor control systems that control target machines such as robots and machine tools, a high-speed, high-accuracy control method and error detection method are required. A model that accurately simulates a control target is essential for improving the performance of these methods. Unfortunately, errors in the physical model parameters, such as mass, barycentric position, and inertia tensor, which are due to the installation of a wire, an end effector, and the like in a control target, cannot be ignored in the practical operation of the motor control system. In addition, many components that are non-linear with motor speed, such as vibration and friction, are also included. Therefore, it has been difficult to provide a model that can reproduce these components with high accuracy in advance.

[0003] In recent years, a control method has been proposed to prevent or reduce errors in force and torque estimation in such a model. For example, Patent Literature 1 discloses a technique in which a controller for an electric motor acquires an M-sequence or sine command and corrects an inertia tensor, a friction coefficient, and a spring constant for each sampling period to minimize a torque estimation error, thereby correcting an error contained in an initial value of a physical model parameter during operation. Patent Literature 2 relates to a machine learning system and a motor control system, and more particularly, relates to a machine learning system and a motor control system having the function of automatically adjusting a parameter.The machine learning system includes: a state observer for observing the speed, torque, current, and voltage values ​​of a motor detected by a motor controller for driving the motor; a motor power calculator for calculating a motor power from the speed, torque, current, and voltage values ​​of the motor observed by the state observer; a reward calculator for calculating a reward based on the motor power; and a learning unit for updating an action value table based on the speed, torque, current, and voltage values ​​of the motor, wherein the learning unit performs an arithmetic operation of a state variable observed by the state observer in a multi-layer structure and updates the action value table in real time. Citation listPatent literature Patent literature 1: JP 2013 - 128 387 A Patent literature 2: DE 10 2016 011 526 B4 Brief descriptionTechnical problem

[0004] However, the technique described in Patent Literature 1 is based on the assumption that the motion of the control device for an electric motor can be modeled using the inertia tensor, friction, and torque from a spring. Therefore, the technique described in Patent Literature 1 poses a problem in that it cannot express a disturbance that is not considered as a model. Furthermore, because a friction model is simplified, the technique described in Patent Literature 1 poses a problem in that the friction model is unable to express complex phenomena with strong nonlinearity, such as the Stribeck effect and hysteresis.

[0005] The present invention has been made in view of the above, and an object of the present invention is to provide an engine control device capable of learning a model with improved torque estimation accuracy and generalization performance and improving the speed and accuracy of abnormality detection. Solution to the problem

[0006] To solve the above problems and achieve the object, the present invention is a motor control device for controlling a motor to drive a target machine. The device includes: a motor state observation unit for observing a state of the motor and outputting an observation result as a motor state signal; a first torque estimation unit for outputting a first estimated torque and physical state variables using an equation of motion based on the motor state signal and physical model parameters of the target machine, the first estimated torque being an estimated value of a torque of the motor, the physical state variables being components of the first estimated torque;a physical model storage unit for learning and storing the physical model parameters in such a manner as to reduce a difference between an actual torque and the first estimated torque, the actual torque being included in the engine state signal;and a feature amount generation unit for generating and outputting a feature amount, wherein the feature amount generation unit generates the feature amount based on at least one of the first estimated torque, the physical state variable, and the engine state signal. The device further comprises: a second torque estimation unit for calculating a machine learning correction amount and adding the calculated machine learning correction amount to the first estimated torque to output a second estimated torque, wherein the second torque estimation unit calculates the machine learning correction amount based on the feature amount and a parameter of a machine learning model using machine learning including an autoregressive term in an input-output relationship;and a machine learning model storage unit for learning the parameter of the machine learning model through supervised learning and storing the learned parameter of the machine learning model, wherein the machine learning model storage unit learns the parameter of the machine learning model such that the machine learning correction amount is closer to the difference between the actual torque and the first estimated torque.; Advantageous effects of the invention

[0007] According to the present invention, an engine control device achieves the effect of learning a model with improved torque estimation accuracy and generalization performance and improves the speed and accuracy of abnormality detection. Brief description of the drawings Fig. 1 is a block diagram illustrating a configuration example of a motor control device. Fig. 2 is a schematic diagram illustrating a target machine to be controlled by the engine control device. Fig. Figure 3 is a diagram illustrating an example of a nonlinear friction model, which is a physical model parameter estimated by a physical model storage unit. Fig. 4 is a block diagram illustrating a configuration example of a feature amount generation unit and a machine learning model calculation unit. Fig. Figure 5 is a diagram illustrating an autoregressive term in the configuration of the feature magnitude generation unit and the machine learning model calculation unit. Fig. 6 is a flowchart illustrating the process of learning physical model parameters and machine learning model parameters in the engine control device. Fig. 7 is a diagram illustrating an engine speed representing the operation of the target machine to be controlled by the engine controller, a first estimated torque estimated by the engine controller, and a second estimated torque estimated by the engine controller. Fig. 8 is a diagram illustrating an example in which the processing circuit included in the engine control device includes a processor and a memory. Fig. 9 is a diagram illustrating an example in which the processing circuit included in the engine control device includes dedicated hardware. Description of the embodiment

[0008] An engine control device according to an embodiment of the present invention will be described in detail below with reference to the drawings. Note that the present invention is not limited to the embodiment. Embodiment

[0009] Fig. 1 is a block diagram illustrating a configuration example of a motor control device 100 according to an embodiment of the present invention. The motor control device 100 controls a motor that drives a target machine 10. The motor control device 100 includes an engine state observation unit 1, a physical model calculation unit 2, a feature amount generation unit 3, and a machine learning model calculation unit 4. 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.

[0010] The target machine 10 to be controlled by the engine control device 100 is described. Fig. 2 is a schematic diagram illustrating the target machine 10 to be controlled by the motor control device 100 according to the present embodiment. A horizontal articulated robot will be described as an example of the target machine 10. A horizontal articulated robot, which is the target machine 10, includes a first axis 11j, a second axis 12j, a third axis 13j, a fourth axis 14j, a first link 11l, a second link 12l, a third link 13l, a fourth link 14l, and an arm tip portion 15e. The motor control device 100 drives the second member 12l with an actuator installed on the first axis 11j, drives the third member 13l with an actuator installed on the second axis 12j, and drives the fourth member 14l with actuators installed on the third axis 13j and the fourth axis 14j.Thus, the motor controller 100 can control the position and orientation of the arm tip section 15e of the horizontal articulated robot. The third axis 13j corresponds to a rotational movement of the fourth link 14l about the axis extending in a direction along the arm tip section 15e, and the fourth axis 14j corresponds to a translational movement of the fourth link 14l in a direction along the arm tip section 15e.

[0011] Although the present embodiment specifies the horizontal articulated robot as an example of the target machine 10, the target machine 10 is not limited thereto, and the motor control device 100 can be widely applied to the target machine 10 to be driven by a motor. Further, examples of the target machine 10 may include a vertical articulated robot, a numerical control (NC) machine tool, and an assembly machine.

[0012] The motor state monitoring unit 1 monitors the state of the motor driving the target machine 10 and outputs an observation result as a motor state signal. Specifically, at least one of a position, a speed, and an acceleration of the actuator corresponding to each of the first to fourth axes 11j to 14j, as well as a current torque, are output as motor state signals from the motor state monitoring unit 1. The actual torque is a torque of the motor described above. The actual torque may be a value directly measured by a sensor or may be a value obtained by converting an input current value obtained from the actuator. Furthermore, the motor state signals are not limited to signals about the position, speed, and acceleration of each actuator and the current torque.The motor state signals may include values ​​from, for example, sensor data from an internal sensor or an external sensor provided on each axis or at a predetermined position of the robot. Examples of sensor data include temperature data of the actuator of each axis, data obtained from a force sensor of an arm tip, distance data obtained from a sensor such as an infrared sensor or an ultrasonic sensor, and moving image data obtained from an image sensor, as well as feature amount data obtained by image processing performed on the moving image data.

[0013] With the input motor state signals and physical model parameters, the first torque estimation unit 21 calculates a first estimated torque through the equation of motion. The first estimated torque is an estimated value of the actual torque. The actual torque is the torque of the motor described above. The physical model parameters described below are output from the physical model storage unit 22. That is, the first torque estimation unit 21 estimates the first estimated torque, which is the estimated value of the torque of the motor, 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 physical state variables. The physical state variables 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 Coulomb friction term, a fluid friction term, an arm vibration term, and an elastic deformation term. Furthermore, the physical state variables may also include a term of a function, such as a nonlinear friction model described below, which is introduced to simulate some physical phenomena in calculating the first estimated torque. The physical state variables include a friction estimate. Due to this, the engine controller 100 enables the machine learning model calculation unit 4 to correct the uncertainty of a friction model and an error in the friction estimate with respect to a friction component of the actual torque.The machine learning model calculation unit 4 is described below.

[0014] The motor state signals to be used in the equation of motion correspond to the position, velocity, and acceleration of the individual actuators. The physical model parameters to be used in the equation of motion correspond to the mass, barycentric position, and inertia tensor of each link. That is, the physical model parameters include the mass of the target machine 10 in a linkage, the barycentric position of the target machine 10 in the linkage, and the inertia tensor of the target machine 10 in the linkage.As a result, the motor controller 100 can improve the accuracy of estimating the first estimated torque relative to the actual torque by learning the mass, barycentric position, and inertia tensor, which are coefficients, as the physical model parameters in inverse dynamics using information about a known mechanism of the target machine 10. Furthermore, in the motor controller 100, a correction value in the machine learning model calculation unit 4 described below becomes relatively small. Accordingly, the motor controller 100 can improve the generalization performance of a second estimated torque. The equation of motion used herein is, for example, the Newton-Euler method or the Lagrange equation of motion.

[0015] The first torque estimation unit 21 uses the equation of motion that includes a model with continuous and smooth friction. The friction model takes the engine speed obtained from the engine state signal as input and outputs an estimated value for a friction component of the torque. As described below, the machine learning model calculation unit 4 uses a neural network. Since the neural network can calculate a function that is continuous with respect to the input, it is desirable that a feature amount and the input to the feature amount generation unit 3 be continuous. For this reason, the engine controller 100 defines a friction model that is continuous even near a zero speed.Thus, by using the neural network, the engine control device 100 can also learn and correct a friction component that cannot be learned in the physical model storage unit 22.

[0016] The physical model storage unit 22 uses motor state signals output during operation of the target machine 10 over a certain period of time as teaching data and learns physical model parameters in such a way as to reduce the difference between the actual torque contained in the motor state signals and the first estimated torque. The physical model storage unit 22 stores the learned physical model parameters. As a result of learning the physical model parameters, the motor control device 100 can correct changes in physical characteristics due to the installation of a sensor, a wire, and the like, machining errors, assembly errors, and the like, and improve the accuracy of estimating the actual torque.In general, the approximate values ​​of the physical model parameters can be calculated using a specification table, computer-aided design (CAD) data, and the like. Using these values ​​as initial values ​​at the time of learning, it is possible to improve, for example, learning speed and learning convergence. It should be noted that if these values ​​are unknown or difficult to calculate, the physical model parameters can be initialized with arbitrary given constants.

[0017] A linear equation of an overdetermined system is used for learning the physical model parameters in the physical model storage unit 22. The above-described equation of motion is deformed, and only the physical model parameters are separated from it to obtain the linear equation of the overdetermined system. If nonlinear terms for position, velocity, and acceleration are not considered, the Newtonian equation of motion, the Euler equation of motion, and the Lagrange equation of motion can 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 velocity, and an acceleration.That is, if Δf indicates an error in the dynamics calculation with respect to the actual torque, the equation of motion of each axis can be described in a variable-separated form as in equation (1) by using physical model parameters w, including the mass, the barycentric position and the inertia tensor, and a coefficient matrix A(v,a) containing a velocity v and an acceleration a. Δf=A(v,a)w

[0018] The physical model storage unit 22 obtains w for each axis from equation (1) by the least squares method, thereby calculating the physical model parameters for each axis. In the case of a driving machine that may cause interference between the axes, such as an industrial robot with series-connected drive axes, the physical model storage unit 22 summarizes the w-related equation (1) for each axis into a single linear equation and solves the single linear equation simultaneously, so that the entire physical model calculation unit 2 can obtain a more probable solution. When identifying the physical model parameters w, the physical model storage unit 22 may use Bayesian estimation, a genetic algorithm, a neural network, and the like, in addition to the least squares method, the regularized least squares method, and the Kalman filter.

[0019] It should be noted that when the influence of friction is considered in the equation of motion, the physical model storage unit 22 expresses the friction with an equation whose friction coefficients are separable and represented in a linear combination with a velocity function, so that the physical model storage unit 22 can simultaneously estimate the friction coefficients as physical model parameters as in equation (1). Examples of an equation in which the friction coefficients are separable include a linear friction model, which is a friction model with Coulomb friction and fluid friction, as shown below. The friction coefficient to be estimated by the physical model storage unit 22 is expressed by equation (2). It should be noted that a point mark over q iin equation (2) and equation (3), as described below, which is an image acquisition, does not appear in the following description. For this reason, q i , above which a point marker is placed in the equations, is simply referred to as q in the following description i designated. [Formula 1] Fi(q˙i)=rFC,isgn(q˙i)+rFV,iq˙i

[0020] For a friction model F i an axis i moving with a speed q i is operated, r indicates FC,i a Coulomb friction coefficient and denotes r FV,ia fluid friction coefficient. It is often observed that fluid friction during the operation of a multi-axis mechanism, such as an industrial robot, does not exhibit a completely linear relationship with velocity, and especially during high-speed actuator operations, it is smaller than the value given by assuming a linear relationship between fluid friction and velocity. In such a case, it is desirable to estimate friction using a friction model with a higher degree of freedom that accounts for nonlinearity with respect to velocity, remove a friction component from the difference between the actual torque and the first estimated torque, and then estimate other physical model parameters.Examples of a nonlinear friction model, i.e. such a friction model, include the following function represented by equation (3). [Formula 2] Fi(q˙i)=Fc(q˙i)+Fv(q˙i)Fc(q˙i)={rFC,isin(π2ωlsq˙i)|q˙i|<ωlsrFC,isgn(q˙i)|q˙i|≥ωlsFv( q˙i)={rFV,iq˙i|q˙i|<ωhssgn(q˙i){rFV,iωhs+rST,i(1−e(−rFV,irST,i(|q˙i|−ωhs)))}|q˙i|≥ωhs

[0021] ω ls indicates a limit speed for low speed operation, ω hs a limit speed indicates operation at high speed and r ST,i denotes the saturation friction coefficient. A diagram of the nonlinear friction model is shown in Fig. 3 illustrates. Fig. 3 is a diagram illustrating 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 present embodiment. In Fig. 3, the horizontal axis represents the engine speed and the vertical axis represents the friction. The non-linear friction model has, in addition to the Coulomb friction coefficient r FC,i and the fluid friction coefficient r FV,i has a total of four parameters and is a non-linear function with respect to the speed q i. 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 is dependent on the initial value. Thus, the Coulomb friction coefficient r FC,i and the fluid friction coefficient r FV,iis added once to the physical model parameter w, assuming that the friction is linear, and each friction coefficient is calculated by the above equation (1) and then set as an initial value, so that estimation failure due to dependence on the initial value can be prevented. That is, the physical model storage unit 22 can use the least squares method, the regularized least squares method, the nonlinear least squares method, the Kalman filter, or the like when learning the physical model parameters. By using the least squares method, the nonlinear least squares method, or the Kalman filter, the motor controller 100 can estimate unknown coefficients in the equation of motion and the equation of state for the target machine 10 from the motion data.The movement data, for example, is data contained in the engine status signals.

[0022] The physical model storage unit 22 stores the thus obtained nonlinear friction model coefficients as physical model parameters and adds the output of the nonlinear friction model to the physical state variables each time. The engine control device 100 can detect fluid friction due to aging deterioration, abnormalities, or the like, as well as a change in Coulomb friction, by storing friction model coefficients as physical model parameters in the physical model storage unit 22. 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 engine control device 100 can design the friction model more complicated and detailed by alternately estimating the friction model coefficients and other physical model parameters.In addition, since the engine controller 100 can also estimate other physical model parameters while substantially eliminating the influence of friction, it is possible to learn the physical model parameters with higher accuracy.

[0023] The nonlinear friction model can be a multivariable function that depends not only on a velocity function but also on a position, acceleration, motor temperature, and the like. Each coefficient of the nonlinear friction model can be identified by the Gauss-Newton method, the steepest descent method, an extended Kalman filter, an unscented Kalman filter, or the like. For the nonlinear friction model, a generation method based on a machine learning algorithm can be used, such as a support vector regression model, a regression tree, a random forest, a multivariate adaptive regression spline model, or a neural network.However, the non-linear friction model, considered only as depending on an explicitly given speed, is sufficient, since the non-linear friction model only needs to eliminate the macro influence focused on the static properties of friction, provided that the first estimated torque is corrected by a machine learning unit 411 containing a neural network in the machine learning model calculation unit 4, as described below.

[0024] The feature amount generation unit 3, to which the first estimated torque, the physical state variables, and the motor state signals are input, outputs feature amounts to be input to the machine learning unit 411. The values ​​of these inputs to the feature amount generation unit 3 and the value of the actual torque to be estimated change based on a time series. It is necessary to consider a correlation therebetween in the time axis direction. Thus, the machine learning model storage unit 42 also provides the feature amounts as vectors that also include values ​​obtained in a past specific period. Furthermore, the influence of mutual interference between the axes is not small in a multi-axis mechanism such as the horizontal articulated robot described in the present embodiment.Therefore, it is desirable for the feature magnitude to be a vector that also includes an input value corresponding to each axis. However, a torque component in the feature magnitude 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.

[0025] Fig. 4 is a block diagram illustrating a configuration example of the feature amount generation unit 3 and the machine learning model calculation unit 4 according to the present embodiment. Fig. 4, z is a unit delay operator and N d is an integer of 0 or more, which specifies the number of unit times of the delay. It can also be said that the Fig. 4 are storage devices provided in the feature amount generation unit 3. That is, the feature amount generation unit 3 may include storage devices capable of storing the first estimated torque, the physical state variables, and the engine state signals, and may output, as feature amounts, those extracted from the first estimated torque, the physical state variables, and the engine state signals stored in the storage devices. Since the engine control device 100 includes time-series change information about the physical state variables and the engine state signals in the feature amounts input to the machine learning model calculation unit 4, the machine learning model calculation unit 4 can estimate the operation of the target machine 10 in more detail.Thus, it is possible to improve the accuracy of estimating the second estimated torque with respect to the actual torque.

[0026] The feature amount generation unit 3 can generate and output feature amounts using only some of the first estimated torque, the physical state variables, and the engine state signals, without using all of the first estimated torque, the physical state variables, and the engine state signals. That is, the feature amount generation unit 3 can generate a feature amount based on at least one of the first estimated torque, the physical state variables, and the engine state signals and output the generated feature amount.

[0027] The second torque estimation unit 41 includes the machine learning unit 411. The second torque estimation unit 41 receives the feature amounts generated by the feature amount 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. The second torque estimation unit 41 calculates a machine learning correction amount based on the machine learning performed by the machine learning unit 411 using the feature amounts and the machine learning model parameters. The second torque estimation unit 41 outputs, as the second estimated torque, a value obtained by adding the machine learning correction amount to the first estimated torque.

[0028] The machine learning unit 411 is provided by a neural network. The neural network of the machine learning unit 411, which inputs the feature amounts generated by the feature amount generation unit 3 and the machine learning model parameters output from the machine learning model storage unit 42, outputs the machine learning correction amount in such a way as to minimize an error of the second estimated torque with respect to the actual torque.That is, an ideal value of the machine learning correction amount to be output by the machine learning unit 411 corresponds to the difference between the actual torque and the first estimated torque, and the feature amount to be input to the machine learning unit 411 includes a lagged difference between a past actual torque and the first estimated torque. Thus, the machine learning unit 411 has an autoregressive term in an input-output relationship, as shown in FIG. Fig. 5 illustrates.

[0029] Fig. 5 is a diagram illustrating an autoregressive term in the configuration of the feature amount generation unit 3 and the machine learning model calculation unit 4 according to the present embodiment. Fig. Figure 5 illustrates a part related to the autoregressive term used in Fig. 4. The parameters of the machine learning model specify a weight matrix and a bias vector in the neural network. Furthermore, using a neural network called a recurrent neural network, such as a recurrent neural network, a long-short-term memory (LSTM), or a recurrent gate unit (GRU), with a recursive structure in the input-output relationship of an intermediate layer is also an effective means of improving the estimation accuracy of such time series data. Furthermore, to reduce the learning load and the amount of calculation, 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 mutual interference between the axes can be ignored, the machine learning unit 411 may be applied separately for each axis.

[0030] Universal function approximation models to be used in the machine learning unit 411, such as a neural network, are generally not suitable for approximating a discontinuous function. When the torque corresponding to friction is approximated and corrected in the first estimated torque by using the linear friction model of equation (2), the difference between the actual torque and the first estimated torque may cause a discontinuous jump, and the machine learning unit 411 includes its input / output including these values ​​and forms the autoregressive term, which may lead to a significant deterioration in the results of learning the machine learning model parameters and the results of calculating the machine learning correction amount.Furthermore, it is very difficult to explicitly apply these models to an actual prime mover. This is because the actual prime mover is affected by, for example, lubrication-induced friction, as represented by the Stribeck curve, or by a stick-slip phenomenon caused by the alternating occurrence of adhesion and sliding, and the friction exhibits strong nonlinearity at a speed close to zero and during low-speed operation. In view of this, a continuous and smooth friction model, such as the nonlinear friction model shown in Equation (3), is used near a speed close to zero, thereby providing continuous values ​​for the initial estimated torque and feature magnitudes.Thus, even a friction component that is difficult to reproduce with the first estimated torque and the nonlinear friction model described above can be estimated with high accuracy by adding the machine learning correction amount. The feature amounts are vectors that simultaneously include, for example, speeds and outputs of the nonlinear friction model obtained in a certain past period. Thus, 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 make a correction in the form of a machine learning correction amount.

[0031] As described above, the second torque estimation unit 41 calculates a machine learning correction amount based on the feature amounts and the parameters of the machine learning model by using machine learning including the autoregressive term in the input-output relationship. The second torque estimation unit 41 adds the calculated machine learning correction amount to the first estimated torque to output the second estimated torque. Furthermore, the second torque estimation unit 41 may use machine learning using a neural network as the machine learning including an autoregressive term in an input-output relationship.By using a neural network as a machine learning method, the motor controller 100 can learn and store a torque component that is difficult to express in the equation of motion as a nonlinear model with multiple degrees of freedom. Furthermore, the second torque estimation unit 41 can use the difference between the actual torque and the first estimated torque as an autoregressive term in machine learning, including the autoregressive term in an input-output relationship used to calculate the machine learning correction amount. The difference between the actual torque and the first estimated torque is a target value for a correction amount with respect to the first estimated torque. The target value of the correction amount changes with the operation of the motor of the target machine 10 based on a time series.In this regard, the engine controller 100 provides an autoregressive term that includes a past target value of a correction amount as an input. Due to this, the engine controller 100 can estimate an optimal correction amount for the next time with higher accuracy.

[0032] After the physical model storage unit 22 learns the physical model parameters, the machine learning model storage unit 42 uses the difference between the actual torque and the first estimated torque and the feature amounts during operation of the target machine 10 for a certain period of time as training data for input / output, and learns the machine learning model parameters in such a way as to reduce an error in the machine learning correction amount with respect to the difference between the actual torque and the first estimated torque. That is, the machine learning model storage unit 42 learns the machine learning model parameters through supervised learning, so that the machine learning correction amount is closer to the difference between the actual torque and the first estimated torque.The machine learning model storage unit 42 stores the learned parameters of the machine learning model. Although the machine learning model parameters can be initialized with arbitrary given constants at the beginning of learning, 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 parameters of the machine learning model. Thus, in general, 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 a probability distribution derived from a normal distribution include the truncated normal distribution, the He normal distribution, and the Glorot normal distribution.

[0033] Furthermore, the machine learning model storage unit 42 uses the mini-batch gradient descent method and the error backpropagation method as the learning method, and updates the machine learning model parameters represented as the weight matrix and the bias vector batch by batch. That is, the machine learning model storage unit 42 can use the error backpropagation method as a supervised learning method when learning the machine learning model parameters. For example, in a case where a sufficient amount of training data cannot be prepared, the machine learning model storage unit 42 can use the general gradient descent method, stochastic gradient descent, or the like instead of the mini-batch gradient descent method.Furthermore, using a method such as dropout or early stopping is also an effective approach that can be used by the machine learning model storage unit 42 to improve the generalization performance of the neural network in the machine learning unit 411. Dropout is a method in which neurons are randomly removed at the time of learning. Early stopping is a method that stops learning early, monitoring an error for each epoch. It is necessary to apply a dropout probability to the weight matrix at the time of deriving the machine learning correction amount when dropout is applied to learning. Therefore, the machine learning model storage unit 42 also adds and stores such a hyperparameter as a parameter of the machine learning model.

[0034] The previous process of learning the physical model parameters and the machine learning model parameters is described with reference to a flowchart. Fig. 6 is a flowchart illustrating the process of learning physical model parameters and machine learning model parameters in the engine control device 100 according to the present embodiment. The physical model storage unit 22 sets initial values ​​of a mass, a barycentric position, and an inertia tensor, which are physical model parameters of each link, based on values ​​in the specification table (step S1). The engine condition monitoring unit 1 outputs, as training data, time series data on actual torque, position, speed, and acceleration, which are engine condition signals of the target engine 10 (step S2).

[0035] The first torque estimation unit 21 calculates a first estimated torque through dynamic calculation with respect to position, velocity, and 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 fluid 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 using the dynamic equation converted into a linear equation and the least squares method (step S6).The physical model storage unit 22 sets the Coulomb friction coefficient and the fluid friction coefficient as initial values ​​in the identified physical model parameters, 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, in turn, 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 initial estimated torque and the nonlinear friction model from the actual torque, converting the dynamic equation into a linear equation by eliminating the friction model term (step S8).The first torque estimation unit 21 adds a term of the nonlinear friction model to the dynamic equation and in turn calculates a first estimated torque using the learned physical model parameters (step S9).

[0036] The feature magnitude generation unit 3 receives the first estimated torque, physical state variables, and the engine state signals as inputs, records these values ​​in vectors that also include values ​​obtained in a certain past period, and outputs the vectors as feature magnitudes (step S10). The machine learning model storage unit 42 initializes a weight matrix and a bias, which are parameters of the machine learning model, so 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, such as early stopping and dropout, as conditions for learning and adds their values ​​to the parameters of the machine learning model (step S12).The second torque estimation unit 41, which receives the input of the feature amounts and outputs a machine learning correction amount whose ideal value is the difference between the actual torque and the first estimated torque, learns the weight matrix and the bias of the machine learning model parameters by the error backpropagation method according to the conditions for learning the machine learning model parameters (step S13).

[0037] As described above, the physical model storage unit 22 learns the physical model parameters in advance, and the machine learning model storage unit 42 learns the machine learning model parameters with the machine learning correction amount for the first estimated torque, which is used as an autoregressive term. This significantly contributes to improving the generalization performance of the second estimated torque.This is because the motor control device 100 distinguishes the physical model parameters from the machine learning model parameters, thereby separately learning a torque estimation error component resulting from a known physical phenomenon that can be easily modeled in the equation of motion and a torque estimation error component that is difficult to model in the equation of motion, thus reducing the absolute value of the machine learning correction amount 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 amount, the motor control device 100 can also reduce the number of parameters included in the machine learning model parameters.Due to this, it is possible to prevent overlearning, increase the learning speed, and reduce the training data used for learning in the machine learning model storage unit 42.

[0038] With regard to the autoregressive term related to the machine learning correction amount, the absolute value of the machine learning correction amount is similarly reduced, and the range of the torque estimation error in the neural network or the like is reduced, compared with a configuration in which the machine learning model calculation unit 4 learns a difference between the actual torque and the calculation result provided by the equation of motion using initial values ​​for the physical model parameters, and uses the difference as the machine learning correction amount, or a configuration in which calculation results provided by the equation of motion using initial values ​​for the physical model parameters,are included in feature amounts, and the machine learning unit 411 learns the parameters of the machine learning model and outputs the second estimated torque directly. Furthermore, the advantages of using the autoregressive term include, for example, the ease of learning of estimation error factors that change periodically along the time axis and are difficult to capture in the equation of motion. Such error factors include, for example, the cogging torque dependent on the motor characteristics, a torque ripple, and the influence of backlash in the torque transmission between gears. In practice, Fig. 7 the first estimated torque obtained after learning the physical model parameters and the second estimated torque obtained by the machine learning model parameters learned from the difference between the actual torque and the first estimated torque. Fig. 7 is a diagram illustrating an engine speed representing the operation of the target engine 10 to be controlled by the engine control device 100, the first estimated torque estimated by the engine control device 100, and the second estimated torque estimated by the engine control device 100 according to the present embodiment. For each item, the horizontal axis represents time. As shown in Fig. As illustrated in Figure 7, the second estimated torque is closer to the actual torque than the first estimated torque.

[0039] Furthermore, in a case where the neural network is used for the machine learning unit 411, it is unclear which parameter included in the feature amounts influenced the machine learning correction amount and to what extent the parameter influenced the machine learning correction amount. Therefore, it is very difficult to estimate a change in the physical characteristics of the target machine 10 from a change in the machine learning model parameters. However, in the present embodiment, in which the individual values ​​of the physical characteristics can be directly verified when learning the physical model parameters, the engine control device 100 distinguishes between the physical model parameters and the machine learning model parameters when learning these parameters. Therefore, it can be said that the configuration of the present embodiment is suitable for anomaly detection.

[0040] A method for detecting abnormalities using the physical model calculation unit 2 and the machine learning model calculation unit 4 that have performed learning mainly involves, for example, monitoring a difference between the actual torque and the second estimated torque. The engine control device 100 monitors the difference between the actual torque and the second estimated torque to provide observation values ​​roughly concentrated in a range around 0, so that the absolute value itself is regarded as the degree of the actual torque abnormality. A specific abnormality detection method may be a simple method that detects an abnormality depending on whether the absolute value of the difference between the actual torque and the second estimated torque exceeds a preset threshold.Another anomaly detection method may be a method that distinguishes between two classes of normality and anomaly in a two-class classification problem by using logistic regression, a support vector machine, a decision tree, a neural network, or the like, taking the feature magnitudes and the difference between the actual torque and the second estimated torque as inputs and preparing labeled teaching data.

[0041] When it is difficult to collect training data that is motion data corresponding to an abnormal condition, the motor controller 100 may perform oversampling on motion data corresponding to an abnormal condition acquired using the Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic (ADASYN), or the like. Alternatively, the motor controller 100 may classify the motion data into two clusters of normality and anomaly through unsupervised learning, such as the K-Means method and a self-organizing map. Furthermore, the number of classes and the number of clusters in the motor controller 100 are not limited to two of normality and anomaly, but can be set gradually more finely according to the degree of anomaly.In addition, the degree of anomaly can be freely set to a value other than the absolute value of the difference between the actual torque and the second estimated torque.

[0042] It is assumed that the times at which the physical model parameters are learned in the physical model storage unit 22 and the machine learning model parameters are learned in the machine learning model storage unit 42 are Fig. 1 by dashed lines, for which fault detection is intended. In this case, it is desirable to perform learning at the time of, for example, test operation in the delivery phase in an environment where the motor control device 100 is in an ideal state as a target for comparison with respect to the degree of deterioration. Many industrial robots, including horizontal articulated robots, now have a force sensor, an end effector, and the like installed at the arm tip. Even in such a case, physical properties such as the mass, barycentric position, and inertia tensor of each link change. Therefore, both the physical model parameters and the machine learning model parameters must be relearned in each state.

[0043] When installing a sensor, an end effector, or the like, it is generally time-consuming and labor-consuming to determine the influence of the installation on the physical model parameters using a robot controller or the like. Meanwhile, in the present embodiment, the physical model storage unit 22 learns the influence from motor state signals provided when the target machine 10 is operated for a certain period of time, so the determination process can also be automated. Furthermore, since the values ​​of the physical model parameters are correctly set, there is also the advantage of reducing the cycle time, trajectory deviation, pressure force at the time of collision, and the like.This is a technique that can also be applied to increase the speed and accuracy of force control and impedance control, in addition to position control, speed control, etc. The application of the method described above is therefore not limited to anomaly detection.

[0044] Next, a hardware configuration of the engine control device 100 will be described. In the engine control device 100, the engine state observation unit 1, the physical model calculation unit 2, the feature amount generation unit 3, and the machine learning model calculation unit 4 are implemented by a processing circuit. The processing circuit may be a memory and a processor that executes a program stored in the memory, or it may be dedicated hardware.

[0045] Fig. 8 is a diagram illustrating an example in which the processing circuit included in the motor control device 100 according to the present embodiment includes a processor and a memory. In a case where the processing circuit includes a processor 91 and a memory 92, each function of the processing circuit of the motor 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 the memory 92. The processor 91 reads the program stored in the memory 92 and executes it to implement each function of the processing circuit. That is, the processing circuit includes the memory 92 for storing programs. As a result of executing the programs, the motor control device 100 is caused to perform processing.In addition, it can also be said that these programs cause a computer to execute a procedure and method for the engine control device 100.

[0046] Here, 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. Furthermore, for example, 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), or an electrically erasable programmable ROM (EEPROM) (registered trademark), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a digital versatile disk (DVD) is applicable as the memory 92.

[0047] Fig. 9 is a diagram illustrating an example in which the processing circuit included in the motor control device 100 according to the present embodiment includes dedicated hardware. In a case where the processing circuit includes dedicated hardware, for example, a single circuit, a compound circuit, a programmed processor, a parallel programmed processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a combination thereof is used as the processing circuit shown in Fig. 9 is applicable. The functions of the engine control device 100 can be implemented separately by the processing circuit 93 or can be implemented collectively by the processing circuit 93.

[0048] It should be noted that some of the functions of the engine control device 100 may be implemented by dedicated hardware, and some of its other functions may be implemented by software or firmware. Thus, the processing circuitry may implement any of the functions described above using dedicated hardware, software, firmware, or a combination thereof.

[0049] As described above, in the present embodiment, the model of the motor 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 the information about a physical phenomenon explicitly given in the equation of motion. The machine learning model calculation unit 4 estimates a torque component with strong nonlinearity that is difficult to express in the equation of motion, including a correction amount for the torque estimated by the physical model calculation unit 2 as an autoregressive term. The motor control device 100 learns these elements sequentially.As a result, the engine control device 100 can learn the model of the target machine 10 from a small amount of high-speed motion data, and the model of the target machine 10 has high torque estimation accuracy and high generalization performance with respect to the actual torque. Furthermore, the engine control device 100 can also detect a minor anomaly by comparing the actual torque with a torque estimate, that is, by comparing the actual machine with the learned model.

[0050] That is, the motor controller 100 learns the physical model parameters and the machine learning model parameters sequentially. As a result, the motor controller 100 can separately learn and estimate a torque component resulting from the mechanical and physical properties of the target machine 10, a physical phenomenon explicitly given in the equation of motion, and a torque component with strong nonlinearity that is difficult to express in the equation of motion. As a result, the motor controller 100 can shorten the time required for the entire learning process and reduce the absolute value of the machine learning correction amount by learning the physical model parameters, so that the generalization performance of the second estimated torque with respect to the actual torque can be improved.Furthermore, the engine control device 100 can improve the accuracy of estimating the second estimated torque because the second torque estimation unit 41 efficiently learns the correlation between the feature amounts in the time axis direction through the autoregressive term of the input-output relationship.

[0051] Note that although the horizontal articulated robot is the object to be controlled in the present embodiment, the motor control device 100 can also be applied to a vertical articulated robot, an NC machine tool, an assembly machine, and the like. This is because the motor control device 100 can be applied regardless of the number of axes of a machine and the mechanism of linear motion and rotation of a joint.

[0052] The configurations set forth in the above embodiment show examples of the subject matter of the present invention. List of reference symbols

[0053] 1 Motor state observation unit; 2 Physical model calculation unit; 3 Feature amount generation unit; 4 Machine learning model calculation unit; 10 Target machine; 11l First link; 11j First axis; 12l Second link; 12j Second axis; 13l Third link; 13j Third axis; 14l Fourth link; 14j Fourth axis; 15e Arm tip section; 21 First torque estimation unit; 22 Physical model storage unit; 41 Second torque estimation unit; 42 Machine learning model storage unit; 100 Motor control device; 411 Machine learning unit.

Claims

[1] Motor control device (100) for controlling a motor to drive a target machine (10), the device comprising: an engine condition observation unit (1) for observing a condition of the engine and outputting an observation result as an engine condition signal; a first torque estimation unit (21) for outputting a first estimated torque and physical state variables using an equation of motion based on the motor state signal and physical model parameters of the target machine (10), wherein the first estimated torque is an estimated value of a torque of the motor, the physical state variables being components of the first estimated torque; a physical model storage unit (22) for learning and storing the physical model parameters in such a manner as to reduce a difference between an actual torque and the first estimated torque, the actual torque being included in the engine state signal; a feature amount generation unit (3) for generating and outputting a feature amount, wherein the feature amount generation unit (3) generates the feature amount based on at least one of the first estimated torque, the physical state variable, and the engine state signal; a second torque estimation unit (41) for calculating a machine learning correction amount and adding the calculated machine learning correction amount to the first estimated torque to output a second estimated torque, wherein the second torque estimation unit (41) calculates the machine learning correction amount based on the feature amount and a parameter of a machine learning model using machine learning including an autoregressive term in an input-output relationship; and a machine learning model storage unit (42) for learning the parameter of the machine learning model through supervised learning and storing the learned parameter of the machine learning model, wherein the machine learning model storage unit (42) learns the parameter of the machine learning model such that the machine learning correction amount is closer to the difference between the actual torque and the first estimated torque. [2] Engine control device (100) according to claim 1, wherein the second torque estimation unit (41) uses machine learning using a neural network as the machine learning including the autoregressive term in the input-output relationship, and the machine learning model storage unit (42) uses an error backpropagation method as a supervised learning method when learning the parameter of the machine learning model. [3] The engine control device (100) according to claim 2, wherein the second torque estimation unit (41) uses the difference between the actual torque and the first estimated torque as the autoregressive term in the machine learning, including the autoregressive term in the input-output relationship used to calculate the correction amount of the machine learning. [4] The engine control device (100) according to any one of claims 1 to 3, wherein the feature amount generation unit (3) includes a storage device capable of storing the first estimated torque, the physical state variables, and the engine state signal, and outputs as a feature amount what is extracted from the first estimated torque, the physical state variables, and the engine state signal stored in the storage device. [5] The engine control device (100) according to any one of claims 1 to 4, wherein the physical model parameters include a mass of the target machine (10) in a linkage, a barycentric position of the target machine (10) in the linkage, and an inertia tensor of the target machine (10) in the linkage. [6] The engine control device (100) according to any one of claims 1 to 5, wherein the first torque estimation unit (21) uses an equation of motion including a continuous and smooth friction model, the friction model using a rotational speed of the engine as input and outputting an estimated value of a friction component of the torque, the rotational speed of the engine being obtained from the engine state signal. [7] The engine control device (100) according to claim 6, wherein the physical model storage unit (22) learns and stores the coefficients of the friction model as the physical model parameters. [8] The engine control device (100) of claim 7, wherein the physical state variables include the friction model. [9] The engine control device (100) according to claim 8, wherein the physical model storage unit (22) alternately estimates the coefficients of the friction model and other physical model parameters when learning the physical model parameters. [10] The engine control device (100) according to claim 9, wherein the physical model storage unit (22) uses a least squares method, a regularized least squares method, a non-linear least squares method, or a Kalman filter in learning the physical model parameters.

Citation Information

Patent Citations

  • Machine learning system and motor control system with function of automatically adjusting a parameter

    DE102016011526B4

  • Controller of motor having function of estimating inertia, coefficient of friction and spring constant simultaneously

    JP2013128387A

  • JP002013128387A