JP7634458B2 - Motor control device and motor control method - Google Patents

Description

This disclosure relates to a motor control device and a motor control method that perform trajectory control of a control target.

In motor control devices that drive machines such as machine tools and robots, trajectory control is performed to drive the machine according to commands. When trajectory control is performed, a challenge is how to compensate for trajectory errors that occur due to the effects of frictional forces acting on the machine. A compensation method is known that reduces the effects of frictional forces by calculating a torque that offsets the frictional forces using a friction model that simulates the frictional forces acting on the machine and applying the calculated torque to the torque command.

The higher the accuracy of the friction model, the more accurate trajectory control that reduces the effects of frictional forces becomes possible. The simplest friction model is a linear friction model in which the motor speed and friction are proportional to each other, but this friction model can make it difficult to simulate frictional forces with high accuracy. In an actual machine, the machine and motor are connected via machine components such as a ball screw or a reducer. The effects of the bearings and lubricating oil of the machine components appear as nonlinear components in friction with respect to the motor speed. Conventionally, there is known technology that models nonlinear friction and compensates for friction using the obtained friction model.

Patent Document 1 discloses a motor control device having a speed feedforward control section. The speed feedforward control section has a speed-side acceleration input section that outputs the input upper command acceleration as a speed-side acceleration output, a speed-side speed input section that outputs the input upper command speed as a speed-side speed output, a plurality of speed-side boundary speed input sections that are prepared corresponding to a plurality of boundary speeds that are boundaries of a plurality of speed ranges obtained by dividing a limited speed range and output a speed-side boundary speed output from a position corresponding to the upper command speed, a speed-side first weight learning section that changes a plurality of speed-side first learning weights corresponding to each of the speed-side first outputs that are outputs from each input section according to the speed deviation, and a speed-side output section that adds a plurality of speed-side first multiplication values obtained by multiplying the speed-side first output and the speed-side first learning weight, and outputs the result as a second tentative command current.

JP 2017-127176 A

The technology disclosed in Patent Document 1 uses a neural network learning model to represent a friction model. This technology makes it possible to represent a nonlinear friction model by finely discretizing the input values of the neural network, but the more finely the input values are discretized, the greater the number of weighting coefficients becomes, and so the amount of data and calculation required for learning increases. In other words, this technology has the problem that the calculation load becomes large when attempting to perform highly accurate adjustments.

As described above, the technology disclosed in Patent Document 1 has the problem that it is not possible to provide a motor control device that can adjust friction compensation with high accuracy and with a small calculation load.

The present disclosure has been made in consideration of the above, and aims to provide a motor control device that can adjust friction compensation with high accuracy and low computational load.

In order to solve the above-mentioned problems and achieve the object, a motor control device according to the present disclosure is a motor control device that performs trajectory control to drive a motor based on a position command and control a trajectory of a control target mechanically connected to the motor, and includes a feedback control unit that performs feedback control to calculate a first torque command that is a command for the torque of the motor based on the position command and the position of the motor, and a data storage unit that stores trajectory control data that is data related to the operating state of the motor or the control target. The motor control device according to the present disclosure further includes a learning unit that divides the trajectory control data into a plurality of data sets classified based on the speed and acceleration of the motor , and generates a learning model for each of the plurality of data sets using only data from the corresponding data set of the trajectory control data , a friction compensation torque calculation unit that calculates a friction compensation torque that is a torque of the motor that compensates for friction based on the plurality of learning models, an adder that calculates a second torque command by adding the first torque command and the friction compensation torque, and a current control unit that drives the motor by controlling a current to the motor so as to follow the second torque command.

The motor control device disclosed herein has the advantage of being able to adjust friction compensation with high accuracy and low computational load.

FIG. 1 is a block diagram showing a configuration of a motor control device according to a first embodiment. FIG. 1 is a perspective view showing a schematic example of a motor controlled by a motor control device according to a first embodiment and a target machine that is a drive target; FIG. 1 is a diagram showing an example of setting a threshold value for trajectory control data according to the first embodiment; FIG. 1 is a diagram showing an example of a plurality of learning models generated from divided trajectory control data according to the first embodiment; FIG. 1 is a diagram showing an example of a friction model expressed by Gaussian process regression according to the first embodiment; FIG. 13 is a diagram showing an example of a friction model represented by a decision tree according to the second embodiment; FIG. 11 is a diagram showing a comparison between a single learning model configured with a decision tree according to the second embodiment and multiple learning models constructed for each input space. FIG. 13 is a diagram showing an example of a friction model expressed by a neural network according to the third embodiment; FIG. 11 is a diagram showing a comparison between a single learning model configured with a neural network model according to the third embodiment and multiple learning models constructed for each input space. FIG. 13 is a diagram showing an example of a learning model that employs a Coulomb friction model in a specific input space according to the fourth embodiment. FIG. 13 is a diagram showing division of trajectory control data into a plurality of sets by clustering according to the fifth embodiment. FIG. 1 is a diagram showing a processor in the case where at least some of a model speed calculation unit, a feedback control unit, a friction compensation torque calculation unit, an adder, a current control unit, a data holding unit, and a learning unit included in a motor control device according to a first embodiment are realized by a processor. FIG. 1 is a diagram showing a processing circuit in the case where at least a part of a model speed calculation unit, a feedback control unit, a friction compensation torque calculation unit, an adder, a current control unit, a data holding unit, and a learning unit included in a motor control device according to a first embodiment is realized by the processing circuit.

The motor control device and motor control method according to the embodiment are described in detail below with reference to the drawings.

Embodiment 1.
Fig. 1 is a block diagram showing the configuration of a motor control device 10 according to a first embodiment. The motor control device 10 is a device that performs trajectory control by driving a motor 20 based on a position command and controlling a trajectory along which a control target mechanically connected to the motor 20 moves. The motor control device 10 is connected to a motor 20 that drives a target machine 30, which is a machine to be controlled, and controls the motor 20. The motor 20 and the target machine 30 are also shown in Fig. 1. The motor 20 is mechanically connected to the target machine 30. The target machine 30 operates as a result of the motor control device 10 controlling the motor 20.

Figure 2 is a perspective view showing a schematic example of a motor 20 controlled by the motor control device 10 according to the first embodiment and a target machine 30 that is the object to be driven. Figure 2 shows a case where the target machine 30 is a ball screw device. The ball screw device has a ball screw 31, a table 32, and a coupling 33.

The motor 20 is connected to a ball screw 31 via a coupling 33. The ball screw 31 converts the rotational motion of the motor 20 into linear motion. A table 32 is attached to the ball screw 31, and the table 32 performs linear motion. The motor 20 can switch its rotation direction between a first direction and a second direction opposite to the first direction. The motor 20 has a position detector 21. The position detector 21 detects the position of the motor 20 and outputs the position of the motor 20 to the motor control device 10. Hereinafter, the position of the motor 20 will be referred to as the motor position.

In the first embodiment, the target machine 30 is a ball screw device, but the target machine 30 is not limited to a ball screw device. The motor control device 10 is widely applied to the target machine 30 driven by the motor 20, in addition to machine components such as guide mechanisms. Examples of the target machine 30 include a vertical articulated robot, an NC (Numerical Control) machine tool, or a mounting machine, in addition to a ball screw device.

Returning to FIG. 1, the motor control device 10 has a model speed calculation unit 1, a feedback control unit 2, a friction compensation torque calculation unit 3, an adder 4, a current control unit 5, a data retention unit 6, and a learning unit 7.

The model speed calculation unit 1 receives a position command as input and outputs a model speed. The position command is input to the model speed calculation unit 1 from a command value generating device such as a controller or an NC device. The model speed is obtained by time-differentiating the position command and applying a filter as a value indicating the target speed of the motor 20. Differential processing related to time differentiation is replaced by difference processing in a discrete time system, and the value obtained by dividing the difference between the current motor command and the motor command one sample before by the control processing cycle is used as an approximate differential value. The position command is a target value for the rotational position of the motor 20, but the target value for the translational position of the table 32 may be used as the position command by performing unit conversion.

The feedback control unit 2 executes feedback control to calculate a first torque command, which is a command for the torque of the motor 20, based on the position command and the motor position. That is, the feedback control unit 2 calculates a first torque command, which is a command for the torque of the motor 20, by feedback control based on the position command and the motor position. The feedback control unit 2 receives the position command and the motor position, and outputs a first torque command. The motor position is input to the feedback control unit 2 from a position detector 21 of the motor 20. The feedback control unit 2 performs a position control process of proportional control or integral control so as to reduce a position deviation, which is a deviation between the position command and the motor position, and outputs a first torque command that reduces the position deviation. The motor position may be a value obtained by reading the position of the table 32 by a sensor such as a linear scale. The feedback control unit 2 may be any type of control as long as it causes the motor position to follow the position command. For example, the feedback control unit 2 may perform two-degree-of-freedom control.

The friction compensation torque calculation unit 3 receives the model speed from the model speed calculation unit 1 and outputs the friction compensation torque. The friction compensation torque calculation unit 3 calculates the friction compensation torque based on a friction model configured using multiple learning models. As described below, each learning model is learned based on different trajectory control data. The friction model outputs the friction compensation torque based on the model speed and the model acceleration, which is the differential value of the model speed. The friction compensation torque, which is the output of the friction model, is added to the torque command of the motor 20 so as to reduce the influence of friction disturbance that causes trajectory error. In the above description, the friction model takes the model speed and the model acceleration as inputs, but any data related to the position, speed, or acceleration of the motor 20, including the speed of the motor 20 obtained from the motor position, can be used. Hereinafter, the speed of the motor 20 is described as the motor speed.

The adder 4 calculates the second torque command by adding the first torque command and the friction compensation torque. In other words, the adder 4 calculates the sum of the first torque command and the friction compensation torque, and outputs the calculation result as the second torque command.

The current control unit 5 drives the motor 20 by controlling the current to the motor 20 so that the current follows the second torque command. In other words, the current control unit 5 supplies a current to the motor 20 so that the torque of the motor 20 follows the second torque command.

The data storage unit 6 stores trajectory control data, which is data related to the operating state of the motor 20 or the controlled object. The data storage unit 6 stores trajectory control data, which is data related to the operating state of the motor 20 or the controlled object during execution of trajectory control using friction compensation torque.

The data storage unit 6 receives the motor position detected by the position detector 21 of the motor 20 and a current value, which is the value of the current supplied to the motor 20, and outputs trajectory control data. The trajectory control data is a data set including the motor speed obtained by differentiating the motor position, the motor acceleration obtained by differentiating the motor speed, and a current value, which is torque information. The data storage unit 6 samples and stores the trajectory control data for the motor position and current value acquired for each control period. The data storage unit 6 samples the trajectory control data so that the trajectory control data is included before and after the positive/negative sign of the motor speed is switched. In other words, the data storage unit 6 samples and stores the trajectory control data before and after the rotation direction is switched between the first direction and the second direction. For example, a part of the data storage unit 6 is realized by a semiconductor memory.

The learning unit 7 divides the trajectory control data into a plurality of data sets, and performs learning based on the data set for each of the plurality of data sets and generates a learning model corresponding to the data set. That is, the learning unit 7 divides the trajectory control data into a plurality of data sets, and generates a learning model corresponding to each of the plurality of data sets by learning based on each of the plurality of data sets. For example, the learning unit 7 learns a learning model in which the number of parameters is variable. The friction compensation torque calculation unit 3 calculates the friction compensation torque, which is the torque of the motor 20 that compensates for friction, based on a plurality of learning models. The learning unit 7 divides the input trajectory control data into a plurality of groups. The learning unit 7 divides the trajectory control data into a plurality of groups based on the motor speed value and the motor acceleration value contained in the trajectory control data. The learning unit 7 learns each learning model corresponding to the trajectory control data of each group. The learning unit 7 has a data boundary that defines a boundary for dividing the trajectory control data set in advance, and the learning unit 7 divides the trajectory control data based on the data boundary.

The learning unit 7 learns the relationship between the motor speed and the friction compensation torque corresponding to the motor acceleration by learning the learning model. The friction compensation torque is calculated by subtracting the inertia torque expressed as the product of the moment of inertia of the load connected to the motor 20 and the motor acceleration from the actual torque obtained by converting the current value of the motor 20 into units based on the trajectory control data. Each learning model is learned by supervised learning using the corresponding trajectory control data.

When learning the learning model, the learning unit 7 learns the relationship between the motor speed and motor acceleration and the friction compensation torque corresponding to the motor speed and motor acceleration, and when in use, the learning model receives the model speed and model acceleration as inputs and outputs the friction compensation torque.

The processing of the learning unit 7 will be described below. The processing of the learning unit 7 will be described using an example in which the trajectory control data is divided into eight groups. It is assumed that thresholds related to the trajectory control data are set. First, three thresholds V1, V2, and V3 are set as thresholds for the motor speed. The first threshold V1 is a negative value, the second threshold V2 is a value of 0, and the third threshold V3 is a positive value. In other words, the learning unit 7 divides the trajectory control data into four cases corresponding to the motor speed: when the motor speed is less than the threshold V1, when the motor speed is equal to or greater than the threshold V1 and less than the threshold V2, when the motor speed is equal to or greater than the threshold V2 and less than the threshold V3, and when the motor speed is equal to or greater than the threshold V3. A1 is set as a threshold for the motor acceleration, and the value of the threshold A1 is set to 0. In other words, the learning unit 7 divides the trajectory control data into two in accordance with the sign of the motor acceleration.

The trajectory control data is divided into eight groups, from group 1 to group 8, by the above-mentioned threshold value. The areas of the trajectory control data corresponding to the eight groups are as shown in FIG. 3. FIG. 3 is a diagram showing an example of threshold setting for the trajectory control data according to the first embodiment. For the sake of explanation, the i-th group of trajectory control data divided into eight groups will be called the i-th group of trajectory control data. i is an integer from 1 to 8. The learning model corresponding to the i-th group of trajectory control data and learned using the i-th group of trajectory control data will be called the i-th learning model. FIG. 4 is a diagram showing an example of multiple learning models generated from the divided trajectory control data according to the first embodiment. FIG. 4 shows the relationship between speed and friction when each learning model from the first learning model to the fourth learning model is generated in a region where the acceleration is positive.

Only the i-th set of trajectory control data is used to train the i-th learning model. Therefore, compared to training one learning model using all trajectory control data, the amount of data in the dataset used to train one learning model is small.

The learning unit 7 constructs a friction model using the multiple learning models described above. Specifically, the learning unit 7 constructs a friction model that outputs the sum of the outputs of the eight learning models as the friction compensation torque, as shown in the following formula (1).

In equation (1), T is a friction compensation torque, F _i is a friction torque that is the output of the i-th learning model, and w _i is a coefficient multiplied to the output of the i-th learning model.

The friction model constructed by the learning unit 7 is calculated using the model speed and model acceleration as inputs. Regarding the method of determining w _i , w _i may be determined so that only w _i corresponding to the i-th set of trajectory control data including the input model speed and model acceleration is increased. For example, when the model speed and model acceleration within the range of the first set of trajectory control data are input, w _i may be determined so that w ₁ is 1 and w ₂ to w ₈ are 0.

Next, learning of the ith learning model based on the ith group trajectory control data will be described. Supervised learning is performed. Since the input of the ith learning model is the motor speed and motor acceleration included in the ith group trajectory control data, and the output of the ith learning model is friction torque, in supervised learning, learning is performed using a data set including the motor speed, motor acceleration, and friction torque. With regard to the data set, data included in the ith group trajectory control data is used for the motor speed and motor acceleration. The friction torque is calculated by subtracting the inertia torque, which is expressed as the product of the moment of inertia of the load connected to the motor 20 and the motor acceleration, from the actual torque obtained by converting the current value of the motor 20 into units based on the ith group trajectory control data.

An example of a learning model is a learning model expressed by Gaussian process regression. Since a learning model expressed by Gaussian process regression has the same number of parameters as the number of data, the number of data is the number of parameters. A learning model expressed by Gaussian process regression outputs a friction torque corresponding to the distance between each data and the input motor speed.

At this time, the predicted distribution of the friction torque is also calculated at the same time as the variance of the Gaussian distribution. However, the distance used when calculating the predicted distribution of the friction torque is calculated by a kernel function k, and specifically, the distance between two points x _n and x _n' is generally expressed by using the Gaussian kernel shown in the following formula (2). Note that x _n and x _n' are vectors. n is an integer equal to or greater than 1.

In formula (2), θ ₁ and θ ₂ are hyperparameters set in the learning unit 7. However, the kernel function is not limited to the Gaussian kernel, and may be a linear kernel, an exponential kernel, a periodic kernel, or a Matern kernel, or may be a kernel function formed by a combination of the sum or product of these. At the same time, it is considered that the observed current value and the calculated friction torque contain observation noise, and the observation noise follows a Gaussian distribution shown in formula (3) below, in which the average of the observation noise is 0 and the variance is σ ^2. As shown in formula (4) below, a kernel function k with a noise term added thereto may be used.

In formula (4), the delta function δ(x _n , x _n′ ) is a function that returns 1 when n=n′ in distance calculation between the same trajectory control data, and returns 0 when n=n′. Fig. 5 is a diagram showing an example of a friction model expressed by Gaussian process regression according to the first embodiment. Fig. 5 shows a learning model expressed by Gaussian process regression.

The learning unit 7 estimates hyperparameters in the kernel function of the Gaussian process regression, and outputs a learning model after identifying the hyperparameters. The hyperparameters are estimated so that a likelihood function p(y|X, θ) of friction torque y for X consisting of trajectory control data and θ=(θ ₁ , θ ₂ , σ ² ) summarizing the hyperparameters is maximized. As a method for identifying the hyperparameter θ that maximizes the likelihood function, a search method using a grid search in the θ space or an MCMC (Markov Chain Monte Carlo) method, or a gradient method such as a steepest descent method, a conjugate gradient method, or a quasi-Newton method can be used. It should be noted that θ summarizing X, y, and the hyperparameters is a vector.

By using a learning model expressed by Gaussian process regression, when the learning model outputs a friction torque, the learning unit 7 can simultaneously calculate the variance of the friction torque. Therefore, the accuracy of the friction torque output by the learning model can be evaluated. For example, if the accuracy is low, the learning unit 7 can increase the trajectory control data or reacquire the trajectory control data to improve the accuracy.

However, when the number of data used for learning in Gaussian process regression is N, the amount of calculation required is proportional to the cube of N, as shown in the following formula (5). N is a natural number.

In other words, the increase in the amount of calculations as the number of data items increases is very large, which makes efficient adjustment difficult. In the first embodiment, the trajectory control data is divided into eight pieces. For simplicity of explanation, if the data is divided equally, the amount of calculations is as shown in the following formula (6).

In other words, compared to the amount of calculation before the data was divided, the amount of calculation can be reduced to 1/64, enabling efficient adjustment. In the above example, the trajectory control data is divided equally, but the division does not necessarily have to be equal.

In the first embodiment, Gaussian process regression is used as the learning model, but other learning methods may be used as long as a learning model can be constructed from the trajectory control data. In other learning methods, efficient adjustment can also be achieved by dividing the trajectory control data and learning multiple learning models corresponding to each of the multiple data groups obtained by the division.

As described above, according to the first embodiment, it is possible to provide a motor control device 10 that can adjust friction compensation with high accuracy and low computational load.

Embodiment 2.
Next, a motor control device according to embodiment 2 will be described. In embodiment 1, the motor control device 10 is described in which the learning model in the learning unit 7 is expressed by Gaussian process regression and feedforward correction of friction disturbance is performed. The motor control device according to embodiment 2 expresses the friction model by a decision tree. Since the components other than the friction compensation torque calculation unit 3 and the learning unit 7 are the same as those in embodiment 1, the description will be omitted.

Figure 6 shows an example of a friction model represented by a decision tree according to embodiment 2. In learning the decision tree, node splitting is performed to maximize the objective function IG shown in the following equation (7).

In formula (7), D _p is a parent dataset or a parent node, f is the number of features, D _left is a left child node from the parent node, D _right is a right child node from the parent node, N _p is the number of data included in the parent node D _p , N _left is the number of data included in the left child node D _left , N _right is the number of data included in the right child node D _right , and I is an impurity index. Regarding the impurity index I, since a regression problem for friction torque is dealt with in the second embodiment, the mean square error shown in the following formula (8) is used.

In formula (8), y _i is the value of each sample at node D _t , and y _t is the average value of the values of each sample. _{N t} is the number of samples at node D _t . That is, y _i corresponds to the output of the trajectory control data, and the branch is generated so that the objective function IG obtained by subtracting the sum of the variances of the two child nodes from the variance of the parent node is maximized. Regarding the method of selecting branch candidates at node D _t , the inputs x _i of the multiple trajectory control data corresponding to the output y _i are arranged in ascending order with attention paid to the axis of the motor speed or motor acceleration, and the midpoint of the adjacent trajectory control data is set as the branch candidate point, and formula (6) is evaluated for each of the multiple branch candidates.

As in the first embodiment, the reduction in the amount of calculations when using multiple learning models in a decision tree will be explained using an example in which trajectory control data is divided into eight groups. The method of setting the thresholds V1, V2, V3, and A1 for the trajectory control data, and the correspondence between the i-th group of trajectory control data and the i-th learning model are the same as in the first embodiment. When the number of trajectory control data used for learning is N, and a friction model of the same size as the single learning model shown in FIG. 6 is learned, it is necessary to evaluate formula (6) a maximum of (N-1) times in each layer. When the number of layers is sufficiently smaller than N, a calculation amount proportional to the cube of N is required, as shown in the following formula (9).

In other words, the amount of calculation increases significantly as the amount of data increases, making efficient adjustments difficult.

In the second embodiment, the trajectory control data and learning model are divided into eight. Since thresholds V1, V2, V3, and A1 are set in advance as known values, there is no need to calculate equations (7) and (8). It is only necessary to determine which learning model each trajectory control data corresponds to, and the amount of calculation is proportional to N, as shown in equation (10) below.

In other words, compared to the amount of calculation before the data was divided, the amount of calculation can be reduced to one squared N, enabling efficient adjustment. Figure 7 is a diagram showing a comparison between a single learning model configured with a decision tree according to the second embodiment and multiple learning models constructed for each input space. Figure 7 shows the learning models after division into each group.

Decision trees have a tendency to become overlearned when the data given is biased or when a single deep tree is constructed. For this reason, it is known that the generalization performance of a learning model can be improved by setting a threshold value based on known physical properties of the target, dividing the target into multiple shallow trees, and training each of the multiple shallow trees separately.

Therefore, according to the second embodiment, it is possible to provide a motor control device that can adjust friction compensation with high accuracy with a small calculation load.

The friction model may use an ensemble method such as a random forest or gradient boosting decision tree that includes multiple decision trees as described above, and may apply one or both of pre-pruning, which stops tree generation early during the construction process, and post-pruning, which removes nodes with little information after the tree is constructed.

Embodiment 3.
Next, a motor control device according to a third embodiment will be described. In the first embodiment, the learning model in the learning unit 7 is expressed by Gaussian process regression, and in the second embodiment, the learning model in the learning unit 7 is expressed by a decision tree, and a motor control device that performs feedforward correction of friction disturbance has been described. In the motor control device according to the third embodiment, the friction model is expressed by a neural network. As the components other than the friction compensation torque calculation unit 3 and the learning unit 7 are the same as those in the first and second embodiments, the description will be omitted.

8 is a diagram showing an example of a friction model expressed by a neural network according to embodiment 3. If each layer of the neural network is represented as X _i , the input and output of layer X _i are expressed by the following recurrence formula (11).

In formula (11), W _i is a matrix, X _i and B _i are column vectors, W is called a weight matrix, and B is called a bias vector. Therefore, the computational complexity of a neural network is determined by the number of nodes C per layer, which is equal to the number of rows of W and B, and the number of layers L included in the entire neural network. By increasing the number of layers and nodes of a neural network, the learning model becomes a more complex learning model of a higher order, so there is generally a trade-off between the estimation accuracy and the computational complexity of a neural network.

As in the case of the first and second embodiments, the third embodiment also shows an example in which the trajectory control data is divided into eight sets. As shown in Fig. 8, when learning is performed using a single learning model having a number of layers _Ls and a number of nodes _Cs , the number of layers _Ls and the number of nodes _Cs need to be relatively large values in order to simultaneously express a complex curve near zero speed and a smooth curve above a certain speed, and the amount of calculation in this case is expressed as _Ls × _Cs as shown in the following formula (12).

In other words, the increase in computational complexity associated with using only a single learning model is significant, which makes efficient adjustment difficult.

In the third embodiment, the trajectory control data and the learning model are divided into eight parts. However, the thresholds V1, V2, V3, and A1 are known. For a curve between the thresholds V1 and V3 near zero speed that changes in a complex manner, if L of the learning model corresponding to this region is set as _Lm and C of the learning model is set as _Cm , the amount of calculation for each model is expressed as _Lm x _Cm . Since _Lm x _Cm can be obtained without the need to learn friction changes in all speed ranges, it is a small value compared to _Ls x _Cs , as shown in the following formula (13). m is an integer of 1 or more.

Furthermore, if L of the learning model corresponding to the region below threshold V1 and above threshold V3 is denoted as _Lh and C of the learning model corresponding to the region is denoted as _Ch , the amount of calculation is expressed as _Lh × _Ch . Since it is known that the curve in the region has a monotonous shape without abrupt changes compared to the curve from threshold V1 to threshold V3 near zero speed, _Lh × _Ch is a smaller value than _Ls × _Cs , as shown in the following formula (14).

In other words, the amount of calculations after dividing the data can be reduced compared to the amount of calculations before dividing the data, enabling efficient adjustment. Figure 9 is a diagram showing a comparison between a single learning model configured with a neural network model according to embodiment 3 and multiple learning models constructed for each input space. Figure 9 shows the learning models after division into each group.

As with the decision tree described in the second embodiment, neural networks have a tendency to become overtrained when the data provided is biased or large in size. Therefore, if it is known that the output characteristics of the trajectory control data change significantly in a specific region of the input space, by dividing the neural network model for each region and training them individually, each learning model can be changed to an appropriate size according to the complexity of the output, thereby improving generalization performance.

Therefore, according to the third embodiment, it is possible to provide a motor control device that can adjust friction compensation with relatively high accuracy with a relatively small computational load.

As for the neural network, a recurrent type neural network such as an RNN (Recurrent Neural Network), LSTM (Long Short Term Memory) or GRU (Gated Recurrent Unit) that has a recursive structure in the input/output relationship of the intermediate layer, or a convolution type neural network such as a CNN (Convolutional Neural Network), FCN (Fully Convolutional Network) or TCN (Temporal Convolutional Network) may be used.

Embodiment 4.
Next, a motor control device according to embodiment 4 will be described. In embodiments 1 to 3, the learning model in the learning unit 7 is constructed for each input region using the same machine learning algorithm. The motor control device according to embodiment 4 expresses a friction model by using a different machine learning algorithm or function form for each region. As the components other than the friction compensation torque calculation unit 3 and the learning unit 7 are the same as those in embodiments 1 and 2, description thereof will be omitted.

In the fourth embodiment, as in the first to third embodiments, an example is used in which the trajectory control data is divided into eight sets. As described in the third embodiment, it is known that in the regions below threshold V1 and above threshold V3, the curve has a monotonous shape without any abrupt changes compared to the curve from threshold V1 to threshold V3 near zero speed. Therefore, the friction acting on a ball screw device, which is an example of the target machine 30, is generally accurately represented by Coulomb friction that balances with the driving force below a certain threshold at the motor speed of zero speed shown in the following formula (15), and viscous friction that is expressed in a proportional relationship to the motor speed.

In equation (15), f is the friction force, v is the viscous friction coefficient, w is the motor speed, and c is the Coulomb friction term.

It is known that nonlinear friction that depends on load conditions other than the motor speed, such as the stick-slip phenomenon, occurs before and after the motor speed switches from positive to negative, that is, in the region between thresholds V1 and V3. This friction is a cause of trajectory errors that appear as quadrant projections when the direction of motion of the axis of the target machine 30 is reversed.

In the fourth embodiment, a learning model is constructed using the machine learning algorithm described in the first to third embodiments only in the region from threshold V1 to threshold V3 near zero speed, and inference is performed using the Coulomb friction model shown in equation (15) in the region below threshold V1 and above threshold V3. This allows the motor control device according to the fourth embodiment to reduce the amount of calculations without degrading the accuracy of the friction model. FIG. 10 is a diagram showing an example of a learning model that employs the Coulomb friction model in a specific input space according to the fourth embodiment. FIG. 10 shows an example of a learning model that employs the Coulomb friction model in the region below threshold V1 and above threshold V3.

For the friction model used in the region below threshold V1 and above threshold V3, a friction model such as the LuGre Model, the Seven parameter model, or the State variable model may be used instead of the Coulomb friction model.

The machine learning algorithm and learning model given in the first to third embodiments may be switched between the region below threshold V1 and above threshold V3 and the region between threshold V1 and threshold V3. By using Gaussian process regression with a small number of data N set in the region below threshold V1 and above threshold V3, and using a neural network in the region between threshold V1 and threshold V3, a learning model with a smaller amount of calculations can be configured compared to the case where a single friction model is used. The machine learning algorithm used is not limited to the methods given in the first to third embodiments, and any algorithm capable of learning a regression model may be used.

Therefore, according to the fourth embodiment, it is possible to provide a motor control device that can adjust friction compensation with high accuracy and low computational load.

Embodiment 5.
Next, a motor control device according to embodiment 5 will be described. In the first to fourth embodiments, the motor control device has been described in which the trajectory control data is divided into a plurality of groups by the operation of the learning unit 7, and each learning model corresponding to the trajectory control data of each group is learned. In the first to fourth embodiments, the trajectory control data is divided based on a preset threshold value.

In the fifth embodiment, a threshold value for dividing the trajectory control data is not set in advance. In the fifth embodiment, a motor control device having a learning unit that divides the trajectory control data into a plurality of groups using a clustering technique and learns each learning model corresponding to the trajectory control data of each group is described. The learning unit in the fifth embodiment is referred to as learning unit 7e. The components other than learning unit 7e are the same as those in the first embodiment, so the description is omitted.

As described above, the motor control device according to the fifth embodiment has a learning unit 7e. The learning unit 7e divides the input trajectory control data into a plurality of groups. The learning unit 7e divides the trajectory control data by clustering based on the trajectory control data. The learning unit 7e divides the trajectory control data into a plurality of groups by k-means based on the motor speed value and acceleration value contained in the trajectory control data. The learning unit 7e learns each learning model corresponding to each group of trajectory control data. The input of the learning model is the motor speed and motor acceleration, and the output of the learning model is the friction torque. Each learning model is learned by supervised learning using the corresponding trajectory control data.

The process of dividing the trajectory control data into multiple sets performed by the learning unit 7e will be described below. Assuming an example in which the learning unit 7e divides the trajectory control data into eight sets, the k-means method process used will be described.

The trajectory control data used for learning is represented by n points, and data including the i-th motor speed value and the i-th acceleration value among them is collectively represented as x _i , where i is an integer from 1 to n, and n is an integer equal to or greater than 1. To divide the trajectory control data into eight sets, the learning unit 7e first randomly allocates the n points of data to the eight sets.

Next, for each of the eight pairs, the learning unit 7e calculates the center _Mj of the pair, where j is an integer from 1 to 8. The learning unit 7e calculates the center _Mj of the pair so that it is the arithmetic mean of the data included in the pair.

Next, the learning unit 7e calculates the distance between the i-th data x _i and each of the centers M _j of the eight sets, and assigns the i-th data to the set with the smallest distance among the eight distances. The learning unit 7e performs this operation on all n data points.

As a result, the data belonging to each group is replaced, and the center of the group moves. Therefore, the learning unit 7e performs calculations again so that the center _Mj of the group becomes the arithmetic mean of the data of the group.

The learning unit 7e calculates the center of each group, calculates the distance between each data item and the center of each group, and allocates each data item to each group until the allocation of each data item converges or a predetermined number of repetitions are completed. FIG. 11 is a diagram showing the division of trajectory control data into a plurality of groups by clustering according to the fifth embodiment. FIG. 11 shows the correspondence between each data item and the center of each group.

By the above procedure, eight groups are formed by clustering based on the trajectory control data without determining a threshold value in advance. In the above explanation, an example was given in which the k-means method was used as the clustering method, but the clustering method in embodiment 5 is not limited to the k-means method, and other clustering methods such as the single-link method or group average method may be used.

The learning unit 7e divides the trajectory control data into eight groups, group 1 to group 8. The i-th group of the trajectory control data divided into the eight groups is referred to as the i-th group trajectory control data. The learning model that corresponds to the i-th group trajectory control data and is learned using the i-th group trajectory control data is referred to as the i-th learning model.

Only the i-th set of trajectory control data is used to train the i-th learning model. Therefore, the amount of data set used for training is smaller than when all trajectory control data is used.

The learning unit 7e constructs a friction model using multiple learning models. The method for constructing the friction model is the same as the method described in embodiment 1.

According to the motor control device of embodiment 5, the trajectory control data is divided into multiple groups by clustering based on the trajectory control data executed by the learning unit 7e. Therefore, it is not necessary to set a threshold value for dividing the trajectory control data in advance. Data with similar magnitudes of the speed and acceleration values, which are inputs to the learning model, are used as learning data for the same learning model, making it possible to perform efficient, highly accurate learning using a relatively small amount of learning data. By using multiple highly accurate learning models, it becomes possible to construct a highly accurate friction model.

By using a highly accurate friction model, it is possible to achieve friction compensation to improve the accuracy of trajectory control.

According to embodiment 5, it is possible to provide a motor control device that can adjust friction compensation with high accuracy and low computational load.

FIG. 12 is a diagram showing a processor 91 in the case where at least some of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data retention unit 6, and learning unit 7 of the motor control device 10 according to the first embodiment are realized by the processor 91. In other words, at least some of the functions of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data retention unit 6, and learning unit 7 may be realized by the processor 91 that executes a program stored in the memory 92. The processor 91 is a CPU (Central Processing Unit), a processing system, an arithmetic system, a microprocessor, or a DSP (Digital Signal Processor). The memory 92 is also shown in FIG. 12.

When at least some of the functions of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data storage unit 6, and learning unit 7 are realized by the processor 91, the functions are realized by the processor 91 and software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 92. The processor 91 realizes at least some of the functions of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data storage unit 6, and learning unit 7 by reading and executing the program stored in the memory 92.

When at least some of the functions of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data storage unit 6, and learning unit 7 are realized by the processor 91, the motor control device 10 has a memory 92 for storing a program that results in at least some of the steps executed by the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data storage unit 6, and learning unit 7. It can also be said that the program stored in the memory 92 causes a computer to execute at least some of the procedures or methods executed by the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data storage unit 6, and learning unit 7.

The memory 92 may be, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), or an EEPROM (registered trademark) (Electrically Erasable Programmable Read-Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disk).

FIG. 13 is a diagram showing a processing circuit 93 in a case where at least a part of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data holding unit 6, and learning unit 7 of the motor control device 10 according to embodiment 1 is realized by the processing circuit 93. In other words, at least a part of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data holding unit 6, and learning unit 7 may be realized by the processing circuit 93.

The processing circuitry 93 is dedicated hardware. For example, the processing circuitry 93 is a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination of these.

The model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data retention unit 6, and part of the learning unit 7 may be realized by dedicated hardware separate from the remaining units.

Regarding the multiple functions of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data retention unit 6, and learning unit 7, some of the multiple functions may be realized by software or firmware, and the remaining multiple functions may be realized by dedicated hardware. In this way, the multiple functions of the model speed calculation unit 1, feedback control unit 2, friction compensation torque calculation unit 3, adder 4, current control unit 5, data retention unit 6, and learning unit 7 can be realized by hardware, software, firmware, or a combination of these.

Each of the friction compensation torque calculation units 3 in the second to fifth embodiments may be realized by a processor that executes a program stored in a memory. The memory is a memory similar to the memory 92. The processor is a processor similar to the processor 91. Each of the friction compensation torque calculation units 3 in the second to fifth embodiments may be realized by a processing circuit. The processing circuit is a processing circuit similar to the processing circuit 93.

Each of the learning units 7 in the second to fourth embodiments and the learning unit 7e in the fifth embodiment may be realized by a processor that executes a program stored in a memory. The memory is a memory similar to the memory 92. The processor is a processor similar to the processor 91. Each of the learning units 7 in the second to fourth embodiments and the learning unit 7e in the fifth embodiment may be realized by a processing circuit. The processing circuit is a processing circuit similar to the processing circuit 93.

The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or the embodiments may be combined with each other. In addition, parts of the configurations may be omitted or modified without departing from the spirit of the invention.

1 Model speed calculation unit, 2 Feedback control unit, 3 Friction compensation torque calculation unit, 4 Adder, 5 Current control unit, 6 Data storage unit, 7, 7e Learning unit, 10 Motor control device, 20 Motor, 21 Position detector, 30 Target machine, 31 Ball screw, 32 Table, 33 Coupling, 91 Processor, 92 Memory, 93 Processing circuit.

Claims (6)

A motor control device that performs trajectory control to drive a motor based on a position command and control a trajectory of a control target mechanically connected to the motor,
a feedback control unit that executes feedback control to calculate a first torque command, which is a command for a torque of the motor, based on the position command and a position of the motor;
a data storage unit that stores trajectory control data that is data related to an operation state of the motor or the controlled object;
a learning unit that divides the trajectory control data into a plurality of data sets classified based on a speed of the motor and an acceleration of the motor , and generates a learning model for each of the plurality of data sets using only data of a corresponding data set among the trajectory control data ;
a friction compensation torque calculation unit that calculates a friction compensation torque, which is a torque of the motor that compensates for friction, based on a plurality of the learning models;
an adder that calculates a second torque command by adding the first torque command and the friction compensation torque;
a current control unit that drives the motor by controlling a current to the motor so as to follow the second torque command. The motor control device according to claim 1 , wherein the learning unit learns a learning model in which the number of parameters is variable. The motor control device according to claim 1 , wherein the data storage unit stores trajectory control data that is data related to an operational state of the motor or the controlled object during execution of the trajectory control using the friction compensation torque. The motor control device according to claim 1 , wherein the learning unit is configured to have a data boundary that defines a boundary for dividing the trajectory control data, and the learning unit divides the trajectory control data based on the data boundary. The motor control device according to claim 1 , wherein the learning unit divides the trajectory control data by clustering based on the trajectory control data. A motor control method for performing trajectory control in which a motor is driven based on a position command and a trajectory of a control target mechanically connected to the motor is controlled, the method comprising the steps of:
calculating a first torque command, which is a command for torque of the motor, by feedback control based on the position command and a position of the motor;
storing trajectory control data relating to a state of an operation of the motor or the controlled object during execution of the trajectory control;
Dividing the trajectory control data into a plurality of data sets classified based on a speed of the motor and an acceleration of the motor , and generating a learning model for each of the plurality of data sets using only data of a corresponding data set among the trajectory control data ;
calculating a friction compensation torque, which is a torque of the motor that compensates for friction, based on a plurality of the learning models;
calculating a second torque command by adding the first torque command and the friction compensation torque;
and controlling a current to the motor so as to follow the second torque command, thereby driving the motor.

Images