JP6814441B2 - Learning control device and learning control method for drive machines - Google Patents

Description

The present invention relates to a learning control device and a learning control method for controlling a driving machine that repeatedly operates.

When a driving machine such as a robot that repeats a certain operation pattern operates, there is a problem that vibration is generated due to the natural vibration of the driving machine, and the operation cannot be controlled with high accuracy. Conventionally, in order to solve such a problem, for example, the position or trajectory error of a target part such as a machine tip point in the previous cycle is obtained, the drive machine is controlled so as to reduce the position or trajectory error in the next cycle, and this is repeated. Therefore, a learning control method for converging the position or orbital error to near 0 has been proposed (see Patent Documents 1 and 2).

In the method disclosed in Patent Document 1, the deviation amount Δθ from the original position of the machine tip point is obtained based on the detected acceleration by the acceleration sensor attached to the machine tip point, and the position feedback P1 is subtracted from the position command Pc. The first position deviation ε1 is obtained, the deviation amount Δθ is added to the first position deviation ε1 to obtain the second position deviation ε2, the correction amount is obtained from the second position deviation ε2, and this is the first position. The speed command Vc is obtained by adding to the deviation ε1.

The method disclosed in Patent Document 2 calculates the position of the control target part from the detection result of the sensor provided in the control target part of the articulated robot, calculates the learning correction amount for correcting the position error, and learns. Based on the correction amount, the position deviation data calculated from the position command data related to the target position of the control target part is corrected, and the articulated robot is operated at a predetermined operation speed based on the corrected position deviation data. is there. In this method, in the process of calculating the learning correction amount, the learning correction amount is calculated while increasing the operation speed of the articulated robot up to the maximum operation speed.

Japanese Unexamined Patent Publication No. 2006-172149 Japanese Unexamined Patent Publication No. 2012-240142

In a multi-input / output system such as an articulated robot, one input not only affects one output, but also interferes with other outputs. Therefore, it is not possible to accurately control this output by using only one input for controlling one output. However, the above-mentioned Patent Documents 1 and 2 do not consider the interference problem in such a multi-input / output system, and accurate control that suppresses the interference cannot be performed.

The present invention has been made in view of such circumstances, and a main object thereof is to provide a learning control device and a learning control method for a drive machine capable of solving the above problems.

In order to solve the above-mentioned problems, the learning control device for a driving machine according to one aspect of the present invention is for learning a driving machine that controls a driving machine having a plurality of movable parts to execute a certain repetitive operation a plurality of times. An observation physical quantity acquisition means for acquiring an observation physical quantity which is a physical quantity related to a position observed at each of a plurality of target parts which are the movable parts of the drive machine, and the plurality of the control devices in the Nth repetitive operation. Based on each of the target physical quantities which are the target physical quantities for each target part and each of the observed physical quantities for each of the plurality of target parts acquired in the Nth repetitive operation by the observed physical quantity acquisition means. A control signal for causing the driving machine to perform the N + 1th repetitive operation based on the learning control means that generates a learning output signal for each of a plurality of target parts and the learning output signal generated by the learning control means. Is provided with a control signal generation means for generating each of the plurality of target parts.

In this embodiment, the learning control means projects the target physical quantity and the observed physical quantity in the Nth repetitive operation in the time space of the driving machine onto a projection space of a predetermined base signal, and the target physical quantity projection component and the observed physical quantity. Based on the projection means for calculating the projection component, the target physical quantity projection component in the Nth repetitive operation calculated by the projection means, and the observed physical quantity projection component, in the projection space in the N + 1th repetitive operation. Based on the learning output projection component calculating means for calculating the learning output signal projection component and the learning output signal projection component calculated by the learning output projection component calculating means, the said in the time space in the N + 1th repetitive operation. It may have a learning output signal calculating means for calculating a learning output signal.

Further, in the above aspect, the learning output projection component calculation means is based on a projection space model obtained by converting a mathematical model showing the relationship between the observed physical quantity in the time space and the learning output signal into a mathematical model in the projection space. The controller may be configured to calculate the learning output signal projection component.

Further, in the above aspect, the learning output projective component calculation means may be configured as a controller that stabilizes the projective space model with a steady error of zero.

Further, in the above aspect, the base signal may be a signal relating to the tracking error of the target portion.

Further, in the above aspect, the base signal may be a combination of sinusoidal signals having a frequency component that is an integral multiple of the frequency of the repetitive operation.

Further, the learning control method for a driving machine according to another aspect of the present invention is a learning control method for a driving machine that controls a driving machine having a plurality of movable parts to execute a certain repetitive operation a plurality of times. The step of acquiring the observed physical quantity, which is a physical quantity related to the position observed in each of the plurality of target parts, which are the movable parts of the drive machine, and the target physical quantity for each of the plurality of target parts in the Nth repetitive operation. A step of generating a learning output signal for each of the plurality of target parts based on each of the target physical quantities and each of the observed physical quantities for each of the plurality of target parts acquired in the Nth repetitive operation. Based on the learned output signal, the step includes a step of generating a control signal for causing the driving machine to execute the N + 1th repetitive operation for each of the plurality of target parts.

According to the learning control device and the learning control method of the drive machine according to the present invention, it is possible to suppress the interference of the input with respect to the output in the multi-input / output system.

The schematic diagram which shows the structure of the automatic welding system which concerns on embodiment. The block diagram which shows the structure of the learning control apparatus which concerns on embodiment. The functional block diagram for demonstrating the principle of control by the learning control apparatus which concerns on embodiment. The functional block diagram which shows the structure of the learning control part. The schematic diagram which shows the structure of the 2 link robot. A functional block diagram for explaining the construction of the learning output projection component calculation unit. The functional block diagram which shows the construction example of the learning output projection component calculation part. The flowchart which shows the procedure of operation of the learning control device which concerns on embodiment. The flowchart which shows the procedure of the learning output signal generation processing. A functional block diagram showing the configuration of the conventional method. The graph which shows the time change of the angle command signal and the joint angle of the 1st axis in the conventional method. The graph which shows the angle command signal of the 2nd axis and the time change of a joint angle in the conventional method. The graph which shows the lateral movement and the vertical movement of the tip part of a 2-link robot during a weaving operation in the conventional method. The graph which shows the movement of the tip part of a 2-link robot during a weaving operation in a conventional method. The graph which shows the time change of the angle command signal and the joint angle of the 1st axis in the early stage of learning of this method. The graph which shows the time change of the angle command signal and the joint angle of the 2nd axis in the early stage of learning of this method. The graph which shows the lateral movement and the vertical movement of the tip part of the 2-link robot during the weaving operation in the early stage of learning of this method. The graph which shows the movement of the tip part of a 2-link robot during the weaving operation in the early stage of learning of this method. The graph which shows the time change of the angle command signal and the joint angle of the 1st axis in the learning period of this method. The graph which shows the time change of the angle command signal of the 2nd axis and the joint angle in the latter part of learning of this method. The graph which shows the lateral movement and up-and-down movement of the tip part of a 2-link robot during the weaving operation in the latter part of learning of this method. The graph which shows the movement of the tip part of the 2-link robot during the weaving operation in the latter part of learning of this method.

Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings. It should be noted that each of the embodiments shown below exemplifies a method and an apparatus for embodying the technical idea of the present invention, and the technical idea of the present invention is not limited to the following. Absent. The technical idea of the present invention can be modified in various ways within the technical scope described in the claims. Further, in each of the following embodiments, a welding robot of an articulated manipulator will be described as an example, but the application target of the present invention is not limited to these, and if it is an articulated input / output system, the articulated manipulator It is also possible to apply other drive machines.

<Structure of automatic welding system>
FIG. 1 is a schematic view showing a configuration of an automatic welding system according to the present embodiment. The automatic welding system 10 includes a welding robot 20, a learning control device 30, and a power supply device 40.

The welding robot 20 is composed of a vertical articulated manipulator, and has a welding torch 21 at its tip. The welding robot 20 according to the present embodiment performs welding electrode type arc welding such as MIG (Metal Inert Gas) welding or MAG (Metal Active Gas) welding. The welding robot 20 is connected to each of the learning control device 30 and the power supply device 40.

The welding wire 24 is sent from the wire feeding device 23 to the welding torch 21, and is sent out from the tip of the welding torch 21. The power supply device 40 is a constant voltage power supply device and supplies electric power to the welding wire 24. As a result, a welding voltage is applied between the welding wire 24 and the work (material to be welded) 50, and an arc is generated. Further, an acceleration sensor 25 is provided at the tip portion of the welding robot 20, that is, the tip portion of the arm that supports the welding torch 21. The acceleration sensor 25 is connected to the learning control device 30. The power supply device 40 includes a current sensor (not shown) that detects a welding current generated during welding.

The power supply device 40 includes a CPU and a memory, and controls welding power by executing a computer program for power supply control by the CPU. Further, the power supply device 40 is connected to the wire feeding device 23, and the CPU controls the wire feeding speed. The power supply device 40 performs data communication with the learning control device 30.

Next, the configuration of the learning control device 30 will be described. The learning control device 30 controls the operation of the welding robot 20. FIG. 2 is a block diagram showing the configuration of the learning control device 30. The learning control device 30 includes a CPU 301, a memory 302, an operation panel 303 including a plurality of switches, a teaching pendant 304, an input / output unit 305, and a communication unit 306.

A learning control program 310, which is a computer program for learning control of the welding robot 20, is stored in the memory 302. When the CPU 301 executes the learning control program 310, the welding robot 20 performs learning control of the welding operation. ..

The operation panel 303 and the teaching pendant 304 are used for inputting instructions to the learning control device 30. The operator can input the teaching program to the teaching pendant 304. The learning control device 30 controls the welding robot 20 according to the teaching program input from the teaching pendant 304. Further, this teaching program can also be created by a computer (not shown). In this case, the teaching program can be given to the learning control device 30 by being delivered by a portable recording medium or transmitted by data communication.

The above-mentioned acceleration sensor 25, a current sensor provided in the power supply device 40, a drive circuit of a motor which is an actuator of the welding robot 20, an angle sensor, and an angular velocity sensor (not shown) are connected to the input / output unit 305. ing. The acceleration detected by the acceleration sensor 25, the current value of the welding current detected by the current sensor, the rotation angle and the angular velocity of the motor of each axis of the welding robot 20 are input to the input / output unit 305 and given to the CPU 301. Further, the CPU 301 performs learning control of the welding robot 20 as described later by the learning control program 310, and outputs a control signal to each of the drive circuits of the motor of the welding robot 20.

The communication unit 306 has a wired or wireless communication function. The communication unit 306 performs data communication with the power supply device 40 using a predetermined communication protocol.

The learning control device 30 having the above configuration controls the motor of each axis of the welding robot 20 to control the position and speed of the welding torch 21. The learning control device 30 causes the welding robot 20 to perform a weaving operation. The weaving operation is a repetitive operation in which the welding torch 21 is alternately swung in a direction intersecting the welding direction. The learning control device 30 controls the welding robot 20 so as to perform the weaving operation according to the set weaving cycle, amplitude, and welding speed.

<Principle of learning control>
FIG. 3 is a functional block diagram for explaining the principle of control by the learning control device according to the present embodiment. The learning control device 30 calculates an angle command for each motor in order to cause the welding robot 20 to perform a required operation, and based on each angle command and the measured values of the angle and the angular velocity of each motor, as per the angle command. Apply current to the motor so that it operates. The learning control device 300 has each functional block of the first-axis motor control unit 311, 2nd-axis motor control unit 321, ... That controls the motor for each axis, and the motor control units 311, 321, ... Each of them has an angle / angular velocity control unit 312, 322, ... And a current control unit 313, 323, .... The angle command signal θ _{REF, 1} (t) to the motor of the first axis and the measured values of the angle and the angular velocity of the motor of the first axis are input to the angle / angular velocity control unit 312 of the first axis. The angle / angular velocity control unit 312 calculates a torque command to the motor of the first axis from the input angle command signal θ _{REF, 1} (t) and the measured values of the angle and the angular velocity. The second-axis motor control unit 321 also calculates the torque command to the second-axis motor in the same manner. The same applies to the motor control unit for the third and subsequent axes.

The detected acceleration of the acceleration sensor 25 provided at the tip portion of the welding robot 20 is converted into position data (robot tip position signal) by the data conversion unit 314 and input to the coordinate conversion unit 315. The coordinate conversion unit 315 calculates a joint angle position signal indicating the joint angle of each axis from the robot tip position signal (position in the Cartesian coordinate system) using the kinematic model of the welding robot 20. Hereinafter, the converted joint angle position signal is referred to as a “target site position signal”.

The target site position signals θ _{L, 1} (t) of the first axis, the target site position signals θ _{L, 2} (t), ... Of the second axis are acquired in each cycle of the weaving operation and given to the learning control unit 316. Be done. Further, the learning control unit 316 is also given angle command signals θ _{REF, 1} (t), θ _{REF, 2} (t), ... For each axis. The learning control unit 316 sets each target site position signal θ _{L, i} (t) (that is, observed physical quantity) and each angle command signal θ _{REF, i} (t) (that is, target physical quantity) in the Nth weaving operation. based on, and outputs the N + 1 th learning output signal z _i for weaving operation _(t) axis (i.e., target site) for each. Note that i is an index indicating the joint (axis) of the welding robot 20.

Learning output signal z _i of each axis is added to the torque command for the N + 1 th weaving operation of the corresponding axis, the torque command current control unit 313 and 323, are input to .... In other words, learning the output signal z ₁ of the first axis is added to the torque command output from the first axis angular-velocity control section 312, the torque command is given to the first axis of the current control unit 313. Biaxial subsequent learning output signal _z 2, _{z 3,} ... is the same. The current control units 313, 323, ... Calculate the current value for the corresponding motor to output the given torque command, and output this as a control signal. The current of the current value calculated in this way is supplied to each motor, and each motor is driven.

Here, the learning control unit 316 will be described in more detail. FIG. 4 is a functional block diagram showing the configuration of the learning control unit 316. The learning control unit 316 has a signal storage unit 331, a signal projection unit 332, a learning output projection component calculation unit 333, and a learning output signal generation unit 334. The angle command signals θ _{REF, i} (t) and the target site position signals θ _{L, i} (t) input to the learning control unit 316 are stored in the signal storage unit 331. The signal storage unit 331 includes an angle command signal for one cycle of the Nth weaving operation and a target site position signal θ _{REF, i} (t), θ _{L, i} (t) (t ∈ [(N-1) T,). NT], T: The period of one weaving operation) is stored.

Angle command signal and target site position signal for one cycle of the Nth weaving operation stored by the signal storage unit 331 θ _{REF, i} (t), θ _{L, i} (t) (t ∈ [(N-1) T) , NT]) is given to the signal projection unit 332. The signal projection unit 332 applies the angle command signal and the target site position signal θ _{REF, i} (t), θ _{L, i} (t) (t ∈ [(N-1) T, NT]) to the following equation (1). Projects onto the basis F defined in, and calculates the angle command projection component and the target site position projection component.
Here, it is assumed that the projection of the signals θ _{REF, i} (t) on the basis F is expressed by the following equation (2).
In a drive machine of a multi-input / output system having a plurality of actuators such as the welding robot 20 according to the present embodiment, the angle command projection component of each joint is arranged so as to express the entire angle command projection component. The same applies to the target site position projection component.

Here, the base signal F can be a combination of sinusoidal signals having a frequency component that is an integral multiple of the frequency 1 / T of the repetitive operation, as represented by the following equation (3).

The base signal is preferably selected so as to include the frequency component of the tracking error (= angle command signal-target site position signal) of the target site. For example, when a tracking error occurs due to the natural vibration of the robot, a base signal including a sinusoidal signal having a natural vibration frequency component is selected. This makes it possible to accurately extract the tracking error even if the target site position signal contains noise. Therefore, by selecting such a base signal, it is not necessary to provide a noise removal circuit such as a low-pass filter, and the necessary information may be lost or a phase delay may occur due to the provision of the low-pass filter or the like. Absent.

Each of the angle command projection component and the target site position projection component generated by the signal projection unit 332 is input to the learning output projection component calculation unit 333. The learning output projection component calculation unit 333 calculates the learning output projection component for the N + 1th weaving operation from the angle command projection component and the target site position projection component for one cycle of the given Nth weaving operation. The learning output projective component calculation unit 333 is configured based on a projective space model described later.

Derivation of the projective space model will be described. First, consider a model in which the learning output signal is input and the target site position signal is output in time and space. The model P of the welding robot 20 is represented by the following state space.
For the sake of simplicity, the model of the current control unit is assumed to be 1.

Here, a model of a two-link robot having only two joints will be described as an example. FIG. 5 is a schematic view showing the configuration of a two-link robot. The two-link robot 200 has two motors 211 and 212 and two arms 221,222. The angle of rotation of the motor 211 of the first axis is θ _{M, 1} , and the angle of rotation of the motor 212 of the second axis is θ _{M, 2.} The arm 221 of the first axis (arm driven by the motor 211 of the first axis) ) Is represented by θ _{L, 1} , and the rotation angle of the second axis arm 221 (arm driven by the second axis motor 212) is represented by θ _{L, 2} . At this time, the model of the 2-link robot 200 is expressed as follows.

In the case of the above-mentioned two-link robot 200, the state space expression is as follows.

Further, the model C on the controller side has the following state expression.

By combining the above models P and C, a mathematical model G showing the relationship between the target site position signal θ _L (t) and the learning output signal z (t) can be obtained as follows.
However, in the above equation, the angle command signal θ _REF is omitted because it has no relation.

Here, a projective space model is derived by converting the above-mentioned mathematical model G on time space into a representation on the projective space of the basis signal. The state x ^N at the start of the Nth repeated operation is defined as in the following equation (5).
At this time, the projective space model showing the relationship between the learning output signal projection component and the target site position projection component is expressed as follows.
The above-mentioned projective space model is a discrete-time dynamic system in which one trial of repeated operation is used as one sample.

Next, the learning output projection component calculation unit 333 is constructed using the projective space model derived as described above. FIG. 6 is a functional block diagram for explaining the construction of the learning output projection component calculation unit 333. The projective space model 350 is treated as a discrete time dynamic system to be trained, and a controller 351 for matching the target site position projection component, which is the output of the projective space model 350, with the angle command projection component is designed, and this controller 351 is used. Is a learning output projective component calculation unit 333. That is, the learning output projection component calculation unit 333 is a controller configured based on the projection space model.

Here, in the case of repeated operation, the angle command projection component becomes a constant value. By considering in projective space, the tracking control problem for the angle command signal θ _REF (t) that changes with time can be set as the convergence problem to a constant value. By designing a controller (learning output projection component calculation unit) that stabilizes this discrete-time dynamic system with zero steady-state error, it is possible to match the target site position projection component with the angle-command projection component. Therefore, the target site position signal θ _L (t) = angle command signal θ _REF (t).

FIG. 7 is a functional block diagram showing a construction example of the learning output projection component calculation unit 333. As shown in FIG. 7, the controller 351 (learning output projection component calculation unit 333) that stabilizes with zero stationary error is designed to include a state estimation unit 341, a state feedback unit 342, and an integration servo unit 343. be able to. Here, the state estimation unit 341 estimates the state of the projective space model 350 from the target site position projection component and the learning output projection component by using a state estimation method such as a state estimation observer or a Kalman filter. The state feedback unit 342 outputs the state estimated by the state estimation unit 341 by multiplying the state feedback gain. Integrating servo section 343 integrates the error between the angle command projection component and the target portion position projection component, it outputs the result integral gain K _I.

Here, when the operation is stopped after the end of the Nth repeated operation, that is, when the vibration is settled, the state x ^N = 0, and the projective space model 350 becomes a static system. In the present embodiment, since the learning output projection component calculation unit 333, which is a controller, is configured based on a dynamic system considering x ^N ≠ 0, the N + 1th time is performed without waiting after the Nth time of the repeated operation is completed. The learning control can be performed without any problem even when the repetitive operation is started and the influence of the vibration generated in the Nth repetitive operation exists in the N + 1th repetitive operation. Further, the above dynamic system is derived from the model of the drive machine of the multi-input / output system, and since the learning output projection component calculation unit 333 is configured based on this dynamic system, the drive machine of the multi-input / output system It is also possible to perform control that suppresses input interference with the output.

See FIG. 4 again. The learning output projection component output by the learning output projection component calculation unit 333 as described above is input to the learning output signal generation unit 334. Learning output signal generation unit 334, and the input learning output projection component, based on the baseband signal F, the learning output signal z _{i (t)} as in the following equation (6) (t∈ [(N -1) T, NT]) is generated.

<Operation of learning control device>
Hereinafter, the operation of the learning control device 30 according to the present embodiment will be described. FIG. 8 is a flowchart showing the operation procedure of the learning control device according to the present embodiment. The CPU 301 of the learning control device 30 calculates the angle command of the motor of each axis so that the welding robot 20 executes the Nth weaving operation, and generates each angle command signal (step S101). Further, the learning control device 30 is given the measured values of the angle and the angular velocity output from the angle sensor and the angular velocity sensor provided in each joint of the welding robot 20. The CPU 301 calculates a torque command (hereinafter, referred to as “pre-correction torque command”) to the motor of each axis based on the generated angle command signal and the input measured values of the angle and the angular velocity (step S102). ). The process of step S102 is executed by the angle / angular velocity control units 312, 322, ... (See FIG. 3).

Next, the CPU 301 adds the learning output signal for the Nth weaving operation (where N ≧ 2 in this case) to the pre-correction torque command, and calculates the torque command for the Nth weaving operation for each axis. (Step S103). This learning output signal is calculated from the angle command signal for the N-1th weaving operation and the position data of the tip portion of the welding robot 20 detected in the N-1th weaving operation by the learning control unit 316 described above. It is generated based on the target site position signal.

The CPU 301 calculates the current value for outputting the torque indicated by the torque command by the corresponding motor for each axis (step S104). A control signal indicating the calculated current value is output (step S105). As a result, an electric current is supplied to each motor, and the welding robot 20 executes the Nth weaving operation. The process of step S104 is executed by the current control units 313, 323, ... (See FIG. 3).

The acceleration of the tip portion of the welding robot 20 in the Nth weaving operation is detected by the acceleration sensor 25. The detected acceleration by the acceleration sensor 25 is given to the learning control device 30, and the CPU 301 calculates the position of the tip portion of the welding robot 20 from the detected acceleration of the acceleration sensor 25, and from this position, the target portion position indicating the joint angle of each axis. The signal is calculated (step S106). The process of step S106 is executed by the data conversion unit 314 and the coordinate conversion unit 315 (see FIG. 3).

The CPU 301 executes a learning output signal generation process for generating a learning output signal for the N + 1th weaving operation (step S107). The process of step S107 is executed by the learning control unit 316 (see FIG. 3).

FIG. 9 is a flowchart showing the procedure of the learning output signal generation processing. In the learning output signal generation process, the CPU 301 first stores the target site position signal of each axis in the Nth weaving operation and the angle command signal for the Nth weaving operation in a signal storage unit which is an area in the memory 302. It is stored in 331 (step S201). The signal storage unit 331 stores the target site position signal and the angle command signal of each axis for one cycle of the Nth weaving operation.

Next, the CPU 301 projects the angle command signal and the target site position signal stored in the signal storage unit 331 onto the base signal, and calculates the angle command projection component and the target site position projection component (step S202). The process of step S202 is executed by the signal projection unit 332 (see FIG. 4).

The CPU 301 calculates the learning output projection component for the N + 1th weaving operation from the angle command projection component and the target site position projection component for one cycle of the Nth weaving operation (step S203). The process of step S203 is executed by the learning output projection component calculation unit 333 (see FIG. 4). Next, the CPU 301 generates a learning output signal for the N + 1th weaving operation based on the learning output projection component and the base signal (step S204). The process of step S204 is executed by the learning output signal generation unit 334 (see FIG. 4). This completes the learning output signal generation process.

When the learning output signal for the N + 1th weaving operation is generated, the CPU 301 waits until the timing to start the calculation of the angle command signal for the weaving operation in the next cycle (N + 1th time) (step S108). When the calculation timing of the angle command signal for the weaving operation in the next cycle is reached, the CPU 301 returns to step S101 to generate the angle command signal for the N + 1th weaving operation (step S101), and the N + 1th weaving operation. The pre-correction torque command for is calculated (step S102). Then, the CPU 301 adds the learning output signal for the N + 1th weaving operation to the pre-correction torque command for the N + 1th weaving operation, and calculates the torque command for the N + 1th weaving operation for each axis (step S103). ). After that, the CPU 301 executes the processes of steps S104 to S108. As described above, by providing the timing adjustment process of step S108, the learning output signal generated based on the angle command signal and the target site position signal related to the weaving operation of the previous cycle can be used for the weaving operation of the next cycle. It can be added to the pre-correction torque command.

By repeatedly executing the processes of steps S101 to S108 as described above, the learning control of the weaving operation is repeatedly performed.

With the above configuration, according to the learning control device 30 according to the present embodiment, each of the angle command signals for each axis in the Nth weaving operation of the welding robot 20 which is a multi-input / output system and the Nth time. In order to generate the N + 1th learning output signal for each axis based on each of the target site position signals for each axis acquired in the weaving operation of, the angle command signal (input) for one axis of the welding robot 20 is generated. Interfering with the angle (output) of another axis, that is, the interference of the input with respect to the output can be suppressed.

<Evaluation test>
The inventor has implemented feedback control without learning control (hereinafter referred to as "conventional method") and learning control method according to the present embodiment (hereinafter referred to as "this method") for the two-link robot. The performance of the method was evaluated. FIG. 10 is a functional block diagram showing the configuration of the conventional method. As shown in FIG. 10, a control device having an angle / angular velocity control unit 312, 322, ..., a current control unit 313, 323, ..., And not having a data conversion unit 314, a coordinate conversion unit 315, and a learning control unit 316. The conventional method was carried out by. Further, in this evaluation test, the two-link robot 200 was made to execute the weaving operation under the conditions of the weaving frequency of 2 Hz and the total amplitude of 4 mm by the conventional method and the present method.

11A to 11D show the behavior of the 2-link robot 200 when the conventional method is carried out. FIG. 11A is a graph showing the time change of the angle command signal and the joint angle (target site position signal) of the first axis, and FIG. 11B is the angle command signal and the joint angle (target site position signal) of the second axis. 11C is a graph showing a time change, FIG. 11C is a graph showing lateral movement and vertical movement of the tip portion of the 2-link robot 200 during the weaving operation (horizontal crossing direction with respect to the welding direction), and FIG. 11D is a weaving operation. It is a graph which shows the movement of the tip part of the 2-link robot 200 inside. In FIGS. 11A and 11B, the vertical axis represents the joint angle and the horizontal axis represents time. In FIG. 11C, the vertical axis represents distance and the horizontal axis represents time. Further, in FIG. 11D, the vertical axis represents the vertical movement distance of the tip portion of the two-link robot 200, and the horizontal axis represents the horizontal movement distance thereof.

From FIGS. 11A and 11B, it can be seen that the vibration causes a deviation between the angle command signal of each joint and the joint angle. Further, from FIGS. 11C and 11D, it can be seen that the lateral movement of the weaving operation does not show a sinusoidal waveform, and that the vertical movement occurs within a range of ± about 0.7 mm. These behaviors are considered to be the influence of the vibration generated in the previous weaving cycle, which remains in the next weaving operation.

12A to 12D show the behavior of the two-link robot 200 at the initial stage of learning when this method is performed, and FIGS. 13A to 13D show the behavior of the latter stage of learning. 12A and 13A are graphs showing the time change of the angle command signal and the joint angle of the first axis, and FIGS. 12B and 13B are graphs showing the time change of the angle command signal and the joint angle of the second axis. 12C and 13C are graphs showing the time change of lateral movement and vertical movement of the tip portion of the 2-link robot 200 during the weaving operation, and FIGS. 12D and 13D show the 2-link robot 200 during the weaving operation. It is a graph which shows the movement of the tip part of. In FIGS. 12A, 12B, 13A, and 13B, the vertical axis represents an angle and the horizontal axis represents time. In FIGS. 12C and 13C, the vertical axis represents distance and the horizontal axis represents time. Further, in FIGS. 12D and 13D, the vertical axis represents the vertical movement distance of the tip portion of the 2-link robot 200, and the horizontal axis represents the horizontal movement distance thereof.

From FIGS. 12A and 12B, it can be seen that the deviation between the angle command signal of each joint and the joint angle is significantly reduced as compared with the conventional method. Further, from FIGS. 12C and 12D, it can be seen that the lateral movement of the weaving operation is closer to the sinusoidal waveform and the range of the vertical movement is reduced as compared with the conventional method. As described above, in the initial stage of learning, the behavior of the weaving motion is improved as compared with the conventional method. Further, from FIGS. 13A and 13B, there is almost no deviation between the angle command signal of each joint and the joint angle, and from FIGS. 13C and 13D, the lateral movement of the weaving operation is sinusoidal. It shows a waveform, and it can be seen that almost no vertical movement occurs. In this way, when this method is applied to a series of continuous operations (that is, there is no waiting time between repeated operations) in a multi-input / output system, the vibration gradually decreases as the iterative learning control progresses. You can see that there is.

(Other embodiments)
In the above-described embodiment, the configuration in which the learning output signal is generated for each cycle of the weaving operation, which is a repetitive operation, has been described, but the present invention is not limited to this. The angle command signal and the target site position signal for the continuous operation of a plurality of cycles may be collectively processed to generate the learning output signal for the plurality of cycles.

Further, in the above-described embodiment, the configuration in which all the processes of the learning control program 310 are executed by the single learning control device 30, but the present invention is not limited to this, and the learning control program is not limited to this. It is also possible to make a distributed system in which the same processing as 310 is distributed and executed by a plurality of devices (computers).

The learning control device and learning control method for a drive machine of the present invention are useful as a learning control device and a learning control method for controlling a drive machine that performs repetitive operations.

10 Automatic welding system 20 Welding robot 25 Accelerometer 30 Learning control device 301 CPU
302 Memory 310 Learning control program 316 Learning control unit 331 Signal storage unit 332 Signal projection unit 333 Learning output projection component calculation unit 334 Learning output signal generation unit 2002 Link robot

Claims (6)

A learning control device for a drive machine that controls a drive machine having a plurality of movable parts to execute a certain repetitive operation a plurality of times.
An observation physical quantity acquisition means for acquiring an observation physical quantity which is a physical quantity related to a position observed at each of a plurality of target parts which are the movable parts of the drive machine.
The target physical quantity, which is the target physical quantity for each of the plurality of target parts in the Nth repetitive motion, and the plurality of target parts acquired in the Nth repetitive motion by the observation physical quantity acquisition means. A learning control means that generates a learning output signal for each of the plurality of target parts based on each of the observed physical quantities, and
A control signal generating means for generating a control signal for causing the driving machine to execute the N + 1th repetitive operation based on the learning output signal generated by the learning control means for each of the plurality of target parts. equipped with a door,
The learning control means
A projection means for calculating the target physical quantity projection component and the observed physical quantity projection component by projecting the target physical quantity and the observed physical quantity in the Nth repetitive operation in the time space of the driving machine into the projective space of a predetermined base signal.
Learning to calculate the learning output signal projection component in the projection space in the N + 1th repetitive operation based on the target physical quantity projection component and the observed physical quantity projection component in the Nth repetitive operation calculated by the projection means. Output projection component calculation means and
A learning output signal calculating means for calculating the learning output signal in the time space in the N + 1th repetitive operation based on the learning output signal projection component calculated by the learning output projection component calculating means.
Have ,
Learning control device for drive machines. The learning output projective component calculation means is provided by a controller configured based on a projective space model obtained by converting a mathematical model showing the relationship between the observed physical quantity in the time space and the learning output signal into a mathematical model in the projective space. , It is configured to calculate the learning output signal projection component,
The learning control device for a drive machine according to claim 1. The learning output projective component calculation means is configured as a controller that stabilizes the projective space model with zero steady-state error.
The learning control device for a drive machine according to claim 2. The basal signal is a signal relating to the tracking error of the target portion.
The learning control device for a drive machine according to any one of claims 1 to 3. The basis signal is a combination of sinusoidal signals having a frequency component that is an integral multiple of the frequency of the repetitive operation.
The learning control device for a drive machine according to claim 4. It is a learning control method of a drive machine that controls a drive machine having a plurality of movable parts to execute a certain repetitive operation a plurality of times.
A step of acquiring an observed physical quantity which is a physical quantity related to a position observed in each of a plurality of target parts which are the movable parts of the driving machine.
The target physical quantity, which is the target physical quantity for each of the plurality of target parts in the Nth repetitive motion, and the observed physical quantity for each of the plurality of target parts acquired in the Nth repetitive motion, respectively. Based on the step of generating the learning output signal for each of the plurality of target parts,
Generated on the basis of the learning output signal, a control signal for executing the N + 1 th of the repetitive operation in the drive machine, possess and generating to said plurality of sites,
In the step of generating the learning output signal,
The target physical quantity and the observed physical quantity in the Nth repetitive operation in the time space of the driving machine are projected onto the projective space of a predetermined base signal, and the target physical quantity projection component and the observed physical quantity projection component are calculated.
Based on the calculated target physical quantity projection component and the observed physical quantity projection component in the Nth repetitive operation, the learning output signal projection component in the projection space in the N + 1th repetitive operation is calculated.
Based on the calculated learning output signal projection component, the learning output signal in the time space in the N + 1th repetitive operation is calculated .
Learning control method for drive machines.