JPH0643944A - Vibration prevention control method - Google Patents

Abstract

PURPOSE:To easily and efficiently suppress the vibration of a robot arm. etc., by utilizing a neural network. CONSTITUTION:After the neural network 11 of a forward system which outputs the acceleration estimated value of the tip part of the robot arm learns dynamic characteristics of the robot arm tip part. a robot arm driving motor 15 is brought under driving control while allowing the neural network 12 of a backward system to learn which finds a correction value DELTAVN(t) for a speed command V00(t) to the robot arm driving motor 15 so that an acceleration estimated value alphaa(t) of the arm tip part found by the neural network 11 of the forward system reaches an acceleration command alpha00(t) of the arm tip part.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention mainly relates to an anti-vibration control method for suppressing vibration during operation of an articulated robot arm.

[0002]

2. Description of the Related Art Conventional control for suppressing vibration of an arm tip portion of an industrial robot is a control for compensating the vibration of the arm tip portion by a control theory based on a differential equation expressing the dynamic characteristics of the robot. This is done by configuring the system. However, in order to create a differential equation, a high degree of mathematical knowledge is required, and in the case of a complex controlled object, creating a differential equation is difficult and often impossible. In addition, there is a problem that skill is required for designing the control system based on the control theory.

The present invention has been made in view of such circumstances, and an object thereof is to learn a dynamic characteristic of a controlled object by a neural network so that it is not necessary to create a differential equation and control is performed. An object is to provide a vibration control method that does not require skill in system design.

[0004]

A vibration control method according to the present invention uses a detection value of an acceleration sensor attached to a vibration control target as a teaching signal, and gives a command to the control target and an acceleration estimate already obtained. A new acceleration estimation value is obtained by the first neural network based on the value and the learning process of the first neural network so that the acceleration estimation value matches the teaching signal. After learning, so that the acceleration estimated value obtained by the first neural network matches the acceleration target value, the acceleration estimated value previously obtained by the second neural network and the speed command for the controlled object Based on the above, learning for obtaining the correction value of the speed command for the controlled object is performed, and based on the obtained correction value and the speed command, Characterized in that it comprises a step for production of the control object.

[0005]

According to the present invention, the dynamic characteristics of the controlled object are learned by the first neural network, and the acceleration estimated value obtained by the first neural network by the second neural network becomes the acceleration target value. Learning for obtaining the correction value of the command value for the control target object so as to be the same or a value within the allowable range can surely and quickly bring the acceleration of the control target object to the acceleration target value.

[0006]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A case where the present invention is applied to vibration control of an arm of an industrial robot will be specifically described below.

FIG. 1 is a schematic view of an industrial robot to which a vibration control method according to the present invention is applied. In the figure, 1 is a base part having a driving motor and a control system built therein, 2 is an arm, 3 is a base. It is a joint that connects the arms 2 to each other. The arm 2 is bent and stretched and rotated by the rotating operation of the joint portion 3 by the driving motor, and the manipulator 4 at the tip executes a predetermined work. Reference numeral 5 denotes an acceleration sensor attached to the tip of the arm 2, which is attached to the tip of the arm 2 when learning the neural network 11 of the forward system, which will be described later, and detached from the arm 2 during actual operation.

Next, the learning operation of the neural network 11 of the forward system, which is performed prior to the actual operation, will be described with reference to the block diagram shown in FIG. FIG. 2 is a block diagram showing a configuration at the time of learning of the neural network 11 of the forward system in the vibration control system of the robot arm. The control system is provided with a forward system neural network 11 and an inverse system neural network 12, and when learning the forward system neural network 11, the inverse system neural network 12 is separated from the others and kept inactive.

Here, the neural network is an artificial neural network model, that is, a non-linear / parallel processing industrial system, and its forward system is a system for identifying a controlled object, that is, for giving an input and obtaining an output. And
In the embodiment, the drive voltage for the motor is input and the acceleration of the arm tip is output. The inverse system is a system for identifying the inverse system of the controlled object, that is, for obtaining a certain input from a certain output, and in the embodiment, the acceleration of the arm tip portion is given to obtain the drive voltage for the motor.

In FIG. 2, the rotation angle θ ₀₀ (t) of the motor, which is a command value, is input to the linear controller 13 and is the output of a rotary encoder (not shown) attached to the motor 15 at the addition point. The current rotation angle of the motor 15 is subtracted, and a value obtained by multiplying the deviation by the position gain K _p is output to the servo amplifier 14 and the neural network 11 of the forward system as a speed command V ₀₀ (t) for the motor. The servo amplifier 14 sends a control signal corresponding to the input value to the motor.
It outputs to 15 and drives the motor 15 to rotate the joint part 3 to bend / extend and rotate the arm 2.

The acceleration sensor 5 mounted on the tip of the arm 2 detects the acceleration of the tip of the arm 2 and stores it in the memory 16 as an actual acceleration value α (t) of the tip of the arm. The memory 16 outputs the stored measured acceleration value α (t-1) of the arm tip portion one control cycle before to the other input terminal in the neural network 11 of the forward system. Forward system neural network 11 is a linear controller
Based on the motor speed command V ₀₀ (t) input from 13 and the measured acceleration value α (t-1) of the arm tip part one control cycle before,
The acceleration α _a (t) at the tip of the arm is estimated, and the estimated acceleration value α _a (t) is output to the comparison and collation unit 17.

The comparison and collation unit 17 uses the estimated acceleration value α _a (t) of the arm tip portion inputted from the neural network 11 of the forward system and the measured acceleration value α (t) of the arm tip portion inputted from the acceleration sensor 5. Then, learning for correcting the coupling coefficient of the neural network 11 is performed so that the estimated acceleration value α _a (t) of the arm tip portion matches the measured acceleration value α (t) of the arm tip portion which is a teaching signal.

Next, the configurations of the forward system and inverse system neural networks will be described. Forward system,
The configuration of the inverse system neural network is substantially the same, and the forward system neural network will be specifically described below. FIG. 3 is an explanatory diagram showing the principle of the neural network of the forward system. The neural network consists of three layers: input layer I, middle layer M, and output layer O.
The input layer I has a hierarchical structure, and the input layer I is composed of two neurons I ₁ and I ₂ (FIG. 3 shows a case of having n neurons), and the intermediate layer M is composed of j neurons. From M ₁ , M ₂ , ... M _j , the output layer O further comprises one neuron O ₁ , and each neuron I ₁ , I _{2 of the} input layer I and each neuron M ₁ -M _j of the intermediate layer M The neurons M _{1 to} M _j of the intermediate layer M and the neuron O ₁ of the output layer O are coupled with different coupling coefficients.

Next, a learning method based on the back propagation rule will be described. Learning of the back propagation rule is divided into forward calculation and backward calculation. In the forward calculation, first, input data x _i (collective name of input data, which is described as being assumed that the number of neurons in the input layer I is _{n from} I ₁ to In) is input from the input layer I,
From the equations (1) and (2) below, each neuron M ₁ to
The output value z _j of M _j is obtained. Then, from this output value z _j , the neuron O _{1 of the} output layer O is calculated according to the following equations (1) and (2).
Output value y _k of

[0015]

[Equation 1]

[0016]

[Equation 2]

That is, in the input layer I, the motor speed command value V ₀ (t) for each control cycle is used for the neuron I ₁ , and the measured acceleration value α (t-1) of the arm tip portion one control cycle before is used for the neuron. Input to I ₂ . Other neurons I _{3 to} I of the input layer I
_{The n} may or may not be input as necessary.

Each neuron M ₁ , M ₂ , of the intermediate layer M,
M _3, ... M _j is the output value z according to the sigmoid function based on the sum of the input from the input layer I which is obtained by the equation (1)
_j (general term for z ₁ , z ₂ , z ₃ ... z _j ) is determined. That is, the linear addition value net of the input value a _i weighted by the coupling coefficient w ₁ as shown in the equation (1) is obtained, and the value is obtained by substituting it into the sigmoid function as shown in the equation (2). The output value b obtained is determined as the output value of each neuron M ₁ , M _{2 to} M _j .

Next, the neuron O ₁ of the output layer O is
Based on the total value of the inputs from, the estimated acceleration value α _a (t) of the arm tip in each control cycle, which is the output value y _k, is determined. This decision is also made according to the sigmoid function as in the above-mentioned decision.

On the other hand, in the backward calculation, the error E between the teaching data t _k expressed by the following equation (3) and the output value y _k of the output layer O:
The coupling coefficient w _ji between the intermediate layer M and the input layer I and the coupling coefficient w _kj between the output layer O and the intermediate layer M are changed so as to decrease gradually.

[0021]

[Equation 3]

When the error E is partially differentiated by the coupling coefficient w _kj between the output layer O and the intermediate layer M, the following equation (4) is obtained.

[0023]

[Equation 4]

[0024] The correction amount Δw _kj of the coupling coefficient w _kj, the correction amount Δ
_The learning coefficient η, which is a parameter that controls the size of _kj , is used to express the equation (5).

[0025]

[Equation 5]

Similarly, when the error E is partially differentiated by the coupling coefficient w _ji between the intermediate layer M and the input layer I, the following equation (6) is obtained.

[0027]

[Equation 6]

[0028] The correction amount Δw _ji of the coupling coefficient w _ji, the correction amount Δ
The learning coefficient η, which is a parameter for controlling the size of w _ji , is used to represent the following equation (7).

[0029]

[Equation 7]

In this way, the correction amount Δ of the coupling coefficient w _kj
The correction amount Δw _ji of w _kj and the coupling coefficient w _ji is calculated to correct the coupling coefficient w _kj and the coupling coefficient w _ji . However, actually, in consideration of the convergence of learning, the correction amount Δw _kj (t-
1) and Δw _ji (t-1) are stored, and the correction amounts Δw _kj (t) and Δw _ji (t) are obtained using the following equations (8) and (9).

[0031]

[Equation 8]

[0032]

[Equation 9]

In the back-propagation rule, the forward calculation and the backward calculation as described above are repeated for all learning patterns, and learning is performed until the error E becomes sufficiently small, in other words, until the error E is reduced to within an allowable range. I do.

Next, the actual operation is carried out, but in this actual operation process, the teaching signal to the neural network 11 of the forward system is unnecessary, and the acceleration sensor 5 is removed.

FIG. 4 is a block diagram showing the control system during the actual operation of the industrial robot, and FIG. 5 is a flow chart showing the control process during the actual operation of the industrial robot. In this actual operation process, one input terminal and one output terminal of the inverse system neural network 12 are connected to the output terminal side of the linear controller 13, and the other input terminal is connected to the memory 16 together with the above-mentioned forward system neural network 11. To do.

The rotation angle θ ₀₀ (t) of the motor, which is a position command, is input to the linear controller 13 and the second-order differentiator 17. In this second-order differentiator 17, each rotation θ ₀₀ (t) of the motor is second-order differentiated to obtain an acceleration target value α ₀₀ (t) at the tip of the arm. In the linear controller 13, a value obtained by negatively feeding back the detection value of the rotary encoder provided with the motor is added at the addition point, and the difference is multiplied by the position gain K _p to output the motor speed command (motor voltage value) V ₀₀ ( t) and outputs it to one input of the inverse system neural network 12. To the other input end of the neural network 12 of the inverse system, the arm tip acceleration estimated value α _a (t-1) one control cycle before is input from the memory 16 (step S1).
.

The inverse system neural network 12 uses the speed command V ₀₀ (t) for the motor and the estimated acceleration value α _a (t-1) of the arm tip portion one control cycle before, and the neural network 12 of the forward system. The velocity command V ₀₀ necessary to match the estimated value α _a (t) of the arm tip acceleration output from 11 with the acceleration target value α ₀₀ (t) of the arm tip obtained by the second-order differentiator 17. Correction value for (t) ΔV _N (t)
Then, the motor speed command V ₀₀ (t), which is the output of the linear controller 13, is added to this, and the final motor speed command V _amp (t) is obtained (step S2).
It is input to one input terminal of the neural network 11 of 14 and the forward system (step S3). The acceleration estimation value α _a (t-1) of the arm tip one control cycle before is input from the memory 16 to the other input terminal of the forward system neural network 11. Based on these, the forward system neural network 11 is input. 11 outputs the estimated acceleration value α _a (t) of the arm tip in the next control cycle.

Next, the teaching signal V _Nteach (t) to the neural network 12 of the inverse system one control cycle before is given below.
It is manufactured according to the equation (10) (step S5). V _Nteach (t) = V _Nteach (t-1) + δ {α _a (t) -α ₀₀ (t)} (10) where δ: coefficient (determined by experiment)

In the inverse system neural network 12, the speed command value V ₀₀ (t) for the motor 15 and the estimated acceleration value α _a (t-1) of the arm tip portion one control cycle before are input, and the above-mentioned V _Nteach is input. Learning is performed using (t) as a teaching signal (step S6). This learning is performed only once in each control cycle. The newly obtained teaching signal V _Nteach (t) estimated acceleration value α _a (t) of the arm tip is stored in the memory 16 or the like so as to be used in the next control cycle.

Neural network of this inverse system
The specific contents of the learning of 12 are substantially the same as the first changing process of the coupling coefficient performed in learning of the neural network 11 of the forward system. The final motor speed command V _amp (t) is input to the servo amplifier 14 and the motor 15
Is controlled to operate the arm 2 in a state in which its vibration is suppressed. It returns to step S1 again and repeats the above-mentioned process.

In the above-described embodiment, when the next teaching signal is obtained, the measured acceleration value α (t-1) of the arm tip portion one control cycle before, and the estimated acceleration value α _a (t of the arm tip portion α
-1), the previous teaching signal V _Nteach (t-1) or the like was used, but the present invention is not limited to this, and a value of any number of cycles before may be used. Further, in the above-described embodiment, the case where the invention is applied to the vibration isolation of the arm of the industrial robot has been described, but the present invention is not limited to this and can be applied to the vibration isolation of various similar devices.

[0042]

As described above, in the method of the present invention, the first
Since the dynamic characteristics of the robot arm are learned by the neural network of No. 2, it is not necessary to create a mathematical model such as a differential equation, and the second neural network learns so as to obtain a desired output. The present invention has excellent effects such that the design is easy and the applicable range is wide because it can be applied to a non-linear controlled object.

[Brief description of drawings]

FIG. 1 is a perspective view of an industrial robot to which the method of the present invention is applied.

FIG. 2 is a block diagram showing a control system during learning of a neural network of a forward system implemented in the method of the present invention.

FIG. 3 is an explanatory diagram of a neural network.

FIG. 4 is a block diagram showing a control system during actual operation of the method of the present invention.

FIG. 5 is a flowchart showing a control process during actual operation of the method of the present invention.

[Explanation of symbols]

 1 base part 2 arm 3 joint part 4 manipulator 5 acceleration sensor 11 forward system neural network 12 reverse system neural network 13 linear controller 14 servo amplifier 15 motor

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification number Office reference number FI technical display location G05B 19/407 F 9064-3H // F16F 15/02 A 9138-3J

Claims (1)

[Claims] 1. A new acceleration is applied by a first neural network based on a command to a control object and an estimated acceleration value already obtained, using a detection value of an acceleration sensor attached to a vibration control object as a teaching signal. An estimated value is obtained, and the first estimated value is set so that the estimated acceleration value matches the teaching signal.
The process of learning the neural network, and the first
After the learning of the neural network is completed, the acceleration estimated value previously obtained by the second neural network and the control target are controlled so that the acceleration estimated value obtained by the first neural network matches the acceleration target value. Based on a speed command for an object, learning for obtaining a correction value of the speed command for the controlled object, and actually operating the controlled object based on the calculated correction value and the speed command are included. Anti-vibration control method.