JPH07261853A - Vibration reduction device for robot - Google Patents

Abstract

PURPOSE:To securely reduce vibration generated during the operation of a robot at low cost by removing from a control signal a frequency component corresponding to the natural vibration frequency of a robot shaft outputted from a neural network. CONSTITUTION:During the robot operation, a neural network arithmetic part 5 inputs target positions of respective robot shafts J1-J6 outputted from an inverse transformation arithmetic part 2 and sequentially calculates and outputs natural vibration frequencies fn1-fn6 of the respective robot shafts J1-J6. Then a variable notch filter arithmetic part 4 removes a frequency component in a specific range having its center frequency at a notch frequency fN from an inputted signal of a manipulated variable (v) while denoting the natural vibration frequency fn of the controlled system (robot shaft) as the notch frequency fN, and applies the result to a robot shaft driving source in sequence. Consequently, even if the robot attitude changes abruptly, that can be followed up and the vibration that the robot generates is securely removed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for reducing vibration generated during operation of an industrial robot.

[0002]

2. Description of the Related Art Generally, in order to generate a large torque, a motor 7 is provided in a joint portion of a robot as shown in FIG.
The reduction gear 9 is incorporated between the arm 10 and the arm 10. Since the speed reducer 9 has low rigidity, vibration is likely to occur particularly during high-speed operation. The frequency of this vibration is the natural frequency of the entire system including the control system and the mechanical system.

This natural frequency greatly varies depending on the posture of the robot. Therefore, the frequency of vibration generated during the high speed operation also changes according to the posture of the robot.

Therefore, in order to reduce the vibration generated in such a robot, a notch filter is provided in the control loop to suppress the vibration by removing the frequency component corresponding to the natural frequency of the controlled object from the control signal. Attempts are being made. In this case, it is necessary to change the notch frequency of the notch filter according to the posture taken by the robot.
As a conventional technique of this kind, for example, JP-A-62-126
No. 402, JP-A-62-226317, and JP-A-1-304511.

In the above-mentioned Japanese Patent Laid-Open No. 62-126402, first, the natural frequencies at a plurality of set positions of the XY stage are measured in advance, and the measured natural frequencies are stored in the ROM. Then, during the operation of the XY stage, the natural frequency is read from the ROM according to the position information of the XY stage, and the frequency of the notch filter is switched to the read natural frequency by the analog switch.

Further, in the technique described in Japanese Patent Laid-Open No. 62-226317, the frequency of the machine is detected during operation of the machine by using the phase-locking circuit, and the frequency of the variable notch filter is detected by the machine. The frequency is changed.

Further, the technique described in Japanese Patent Laid-Open No. 1-304511 has a means for detecting or identifying the moment of inertia of a mechanical system by using a linear motion equation, and the natural frequency changes according to the change of the moment of inertia. On the other hand, the frequency of the notch filter is changed accordingly.

[0008]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention
The method of Japanese Patent No. 6402 is such that the notch frequency of the notch filter is switched to a natural frequency prepared in advance, and the switching is performed discretely. Therefore, in the case of a robot in which the posture suddenly changes and the natural frequency also suddenly changes, it is difficult to accurately change the notch frequency of the notch filter to follow the change in posture. Also,
In order to change the notch frequency more accurately,
Many data must be prepared in ROM beforehand,
There is also a cost problem. Further, this technique also requires an analog circuit, which also increases the cost.

Further, in the method disclosed in Japanese Patent Laid-Open No. 62-226317, it is necessary to detect the vibration frequency. Therefore, the vibration frequency cannot be detected unless the vibration is maintained to some extent. Therefore, there is a problem that the vibration cannot be suppressed for several cycles when the vibration starts to occur. Further, since a circuit for detecting the signal to be controlled is required, the cost is increased.

In the method disclosed in Japanese Patent Laid-Open No. 1-304511, the natural frequency is calculated from the moment of inertia using a linear motion equation. However, in an actual robot, the speed reducer has lost motion (low torque). In the area,
There is a region with low rigidity, and this is called lost motion.) Therefore, even if the inertia is the same, the spring constant changes depending on the posture, and the natural frequency also changes significantly, which is a non-linear characteristic. Cannot be applied as is. Further, although the arm is assumed to be a concentrated mass system, since it is actually a distributed mass system, there is a problem that the natural frequency cannot be accurately obtained.

The present invention has been made in order to solve the above-mentioned problems of the prior art, and provides a vibration reducing device capable of surely reducing the vibration generated during the operation of a robot without increasing the cost. It is intended to be provided.

[0012]

Therefore, in the main invention of the present invention, a robot axis drive source which is provided for each axis of the robot and drives the robot axis according to an input control signal, and the control signal from Signal processing means for removing frequency components corresponding to the natural frequency of the robot axis,
In a vibration reduction apparatus for a robot that reduces a vibration generated in the robot axis by applying a control signal processed by the signal processing means to the robot axis drive source, inputting the current position of each axis of the robot. According to the above, a neural network for calculating and outputting the natural frequency of each axis of the robot and adding it to the signal processing means is provided, and the frequency component corresponding to the natural frequency of the robot axis output from the neural network is controlled by the control. I try to remove it from the signal.

[0013]

According to this structure, the current position of each axis of the robot is sequentially input to the neural network during the operation of the robot, and the natural frequency of each axis of the robot is sequentially calculated and output. Then, the corresponding frequency component is sequentially removed from the control signal by the natural frequency of the robot axis which is sequentially calculated and output from the neural network, and is sequentially added to the robot axis drive source. As a result, even if the robot posture changes suddenly, it can be followed, and the vibration generated in the robot can be reliably reduced.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT An embodiment of a vibration reducing apparatus for a robot according to the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the apparatus of the embodiment. In the embodiment, 6 of J1 to J6 shown in FIG.
It is assumed that the vibration generated in the vertical articulated robot 13 composed of axes is reduced.

The target trajectory calculation unit 1 calculates a target trajectory rd of the arm tip of the robot 13, and the calculated tip target trajectory rd is input to the inverse transformation calculation unit 2.
In the inverse conversion calculation unit 2, the input position on the tip target trajectory rd is the target position θd of each axis J1 to J6 of the robot 13.
It is converted into (θd1 to θd6) and is output to the servo control calculation unit 3 provided for each robot axis. In FIG. 1, the part surrounded by the broken line, that is, the servo control calculation unit 3, the variable notch filter calculation unit 4, the servo amplifier 6, the motor 7, the position detector 8, the speed reducer 9, and the arm 10 are
It is provided for each of the robot axes J1, J2 .... In the servo control calculation unit 3, servo control calculation is performed based on the deviation between the target position θd and the current position θ of the robot axis detected by the position detector 8 to calculate the operation amount v, which is the variable notch filter calculation. Added to Part 4.

In the variable notch filter calculation unit 4, the natural frequency fn of the controlled object (robot axis) is used as the notch frequency fN, and the notch frequency fN is used as the center frequency in a predetermined range from the input signal of the operation amount v. Processing for removing frequency components is performed. The manipulated variable v * signal-processed by the variable notch filter calculation unit 4 is output to the servo amplifier 6.

The servo amplifier 6 receives the input operation amount v *
The motor 7 is applied with a drive signal according to the above, and the motor 7 is driven. Therefore, the driving force of the motor 7 is transmitted to the speed reducer 9 and the arm 10 is driven.

The neural network operation unit 5 receives the target positions θd1 to θd6 of the robot axes J1 to J6 output from the inverse transformation operation unit 2 as input, and performs the operation described later to uniquely identify the robot axes J1 to J6. The frequencies fn1 to fn6 are output.

The natural frequencies fn1 to f of the robot axes J1 to J6 calculated and output by the neural network calculation unit 5
n6 is input to the variable notch filter calculation units 4 of the corresponding robot axes.

As described above, the variable notch filter calculation unit 4 changes the center frequency fN according to the input natural frequency fn.

FIG. 2 shows the configuration of the neural network operation unit 5 shown in FIG.

As shown in FIG. 2, the neural network operation unit 5 is composed of three layers of an input layer, an intermediate layer, and an output layer. In the case of the 6-axis robot 13 assumed in the embodiment, The target positions θd1 to θd6 of the axes J1 to J6 are input, and the natural frequencies fn1 to fn6 of the axes J1 to J6 are calculated and output.

However, in practice, the fluctuation of the natural frequency does not pose a problem for all the axes, and in the case of the robot 13 shown in FIG. 6, the first axis is the J1 axis and the second axis is the J2 axis. Since the fluctuation of the natural frequency of
It is sufficient if the natural frequency can be obtained using a neural network only for the axes.

FIG. 3 shows the configuration of the neural network operation unit 5 for calculating the natural frequencies fn1 and fn2 of only the J1 axis and the J2 axis. The input layer is 3 units, the intermediate layer is 4 units, and the output layer is There are 2 units.

The calculation of the input layer is performed by the following equation (1).

Z1i = {θd2, θd3, θd5} (i = 1 to 3) (1) Then, the calculation of the intermediate layer is performed by the following equation (2).

U2j = φ2j1 · z11 + φ2j2 · z12 + φ2j3 · z13-th2j (j = 1 to 4) (2) z2j = 1 / {1 + exp (-u2j)} (j = 1 to 4) (3) and output The calculation of layers is performed as in the following equation (3).

Fnk = φ3k1 · z21 + φ3k2 · z22 + φ3k3 · z23 + φ3k4 · z24-th3k (k = 1 to 2) (4) where z1i is the output of the input layer i unit and z2j is the output of the intermediate layer j unit. , Th2j are threshold values of j units in the intermediate layer, and th3k are threshold values of k units in the output layer. Further, φ2 represents the weight of the coupling between the units of the input layer and the intermediate layer, and φ3 represents the weight of the coupling between the units of the intermediate layer and the output layer. For example, φ2j1 represents the unit of the input layer 1 and the unit of the intermediate layer j. It represents the weight of connection between them.

The number of units in the input layer and the intermediate layer is determined in consideration of the amount of calculation and required accuracy.
It is not limited to the above.

If the natural frequency is calculated by the neural network as described above, it is possible to accurately model the characteristics of a robot having a strong non-linear characteristic without using complicated analysis and arithmetic expressions. It is possible to calculate with a simple formula. Therefore, the device can be realized without specially providing a CPU, and the cost can be reduced.

Now, FIG. 4 shows a case where the variable notch filter operation unit 4 of FIG. 1 is configured by a digital filter.

In FIG. 4, D represents a process of delaying the signal by one sampling. Also, a1, a2, b0, b
1 and b2 represent the coefficients of the notch filter, which is a function of the natural frequency of each axis.

Therefore, the variable notch filter calculation unit 4
Then, first, the natural frequency fnk of each axis, which is the output of the neural network rendering unit 5, is used to calculate the coefficients a1, a2, b0, b1, b2 of the notch filter, and then the notch filter is calculated. The notch filter coefficient calculation is performed as in the following equations (5) to (11). However, hereinafter, "S (X)" is defined as "X squared".

A1 = (− 2 · S (ωnk) +2) / A (5) a2 = (− 1 + ωnk / Q−S (ωnk)) / A (6) b0 = (1 + S (ωnk)) / A (7) b1 = (2 · S (ωnk) −2) / A (8) b2 = (1 + S (ωnk)) / A (9) A = 1 + ωnk / Q + S (ωnk) (10) ωnk = 2 · π · fnk (k = 1 to 2) (11) As described above, the coefficients a1, a2, b0, b1 of the notch filter 4 are determined according to the natural frequency fnk accurately obtained by the neural network 5. b2 is sequentially calculated, whereby the center frequency fN of the variable notch filter 4 is sequentially changed. For this reason, it is possible to reduce vibration in all postures of the robot.

Next, the process of learning the neural network will be described with reference to FIG. The apparatus of FIG. 5 is an apparatus for learning the neural network when the robot 13 is shipped, and the teaching point position storage unit 11 and the speed command calculation unit 12 are specifically robot controllers. The data processing device 15 is specifically a personal computer. Also, the accelerometer 1
4 is attached to the arm of the robot 13 as shown in FIG. The learning by the device of FIG. 5 may be performed once for each robot 13.

In the apparatus of FIG. 5, first, the target position of each axis of the robot 13 is read from the teaching point position storage unit 11,
The robot 13 is positioned at the position. Next, the speed command calculation unit 12 generates a stepwise speed signal for exciting the robot 13 to vibrate, and this is applied to the motor 7 (FIG. 1). As a result, vibration is excited in the robot 13.

The vibration generated by the robot 13 is measured by the accelerometer 14, and the measurement signal is sent to the data processor 15.
Entered in. In the data processing device 15, the natural frequencies f1 and f2 of the axes J1 and J2 are calculated based on the vibration measurement data, and the calculated frequencies f1 and f2 are used as the "teaching signal" of the neural network 5. .

Further, a position different from the position read as described above is read from the teaching point position storage unit 11, and the same measurement is repeated.

Here, the more the position data in the storage unit 11 is, and the different the respective positions are, the wider the identification range as the model of the neural network 5 is.

Next, from the teaching point position storage section 11, among the target positions of the respective axes of the robot 13 thus positioned,
J2, J3 and J5 axis target positions θd2, θd3 and θ
d5 is input to the neural network 5, and the natural frequencies fn1 and fn2 of the axes J1 and J2 are calculated as described above. Then, these calculated natural frequencies fn1 and fn2
And the error ε between the natural frequencies f1 and f2 which are the above-mentioned teacher signals
1 and ε2 are calculated respectively. As described above, this error is obtained a number of times measured a plurality of times. Then, the weight of the connection between the units of the neural network 5 is reduced until ΣS (ε1) + ΣS (ε2), which is the sum of the square sums ΣS (ε1) and ΣS (ε2) of these errors, becomes equal to or less than a predetermined threshold value. Is changed and the process of calculating the error between the teacher signal and the output of the neural network is repeated.

As a method of changing the weight of the neural network 5, for example, an algorithm called back propagation can be used.

It is sufficient to carry out the above learning once at the time of shipment, but when a secular change occurs, it may be carried out by the customer. In addition, since learning is all automatic, adjustment is extremely easy.

Although the input to the neural network is the target position of each axis of the robot in the embodiment, it may be the output θ of the position detector 8 or the target trajectory of the arm tip.

In the embodiment, the center frequency fN of the notch filter is changed according to the natural frequency fn in order to reduce the vibration according to the robot posture, but a variable low pass filter may be used. Alternatively, the vibration may be reduced by changing the cutoff frequency of the low-pass filter according to the natural frequency fn.

[0046]

As described above, according to the present invention,
By using a neural network, it is possible to accurately model the characteristics of a robot with strong non-linear characteristics without requiring complicated analysis and formulas, and to accurately determine the natural frequency of each axis of the robot. The position of the robot is input to the neural network, the natural frequencies of each axis of the robot are sequentially calculated, and the frequency component corresponding to the calculated natural frequency is removed from the control signal. Even in this case, the vibration generated by the robot can be surely reduced.

Further, since the neural network can be operated by a simple formula, the device can be realized without providing a special CPU and the cost can be reduced.

As a result, it is possible to provide the market with a vibration reducing apparatus which can surely reduce the vibration generated during the operation of the robot without increasing the cost.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of a vibration reducing apparatus for a robot according to the present invention.

FIG. 2 is a diagram showing a configuration of a neural network operation unit shown in FIG.

FIG. 3 is a diagram showing a configuration of a neural network operation unit shown in FIG.

FIG. 4 is a block diagram showing a configuration of a variable notch filter calculation unit shown in FIG.

5 is a block diagram showing a configuration of an apparatus for performing learning of the neural network operation unit shown in FIG.

FIG. 6 is a perspective view showing the appearance of a 6-axis robot applied to the embodiment.

[Explanation of symbols]

 4 Variable notch filter calculator 5 Neural network calculator

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI technical display location G05B 19/18 G05D 3/12 305 V // F16F 15/02 Z 9138-3J

Claims (3)

[Claims] 1. A robot axis drive source which is provided for each axis of the robot and drives the robot axis according to an input control signal, and a frequency component corresponding to the natural frequency of the robot axis is removed from the control signal. A vibration reducing device for a robot for reducing vibration generated in the robot axis by applying a control signal processed by the signal processing means to the robot axis drive source. By inputting the current position of each axis, a natural frequency of each axis of the robot is calculated and output, and a neural network for adding this to the signal processing means is provided, and the natural frequency of the robot axis output from the neural network. A vibration reduction device for a robot, wherein a frequency component corresponding to is removed from the control signal. 2. The vibration reducing apparatus for a robot according to claim 1, wherein the signal processing means is a notch filter, and the notch frequency is changed according to the natural frequency of the robot axis output from the neural network. 3. A positioning means for positioning the robot shaft at a predetermined driving position, and a control signal for exciting vibrations in the robot shaft at the driving position positioned by the positioning means. By generating a signal generating means for outputting to the drive source, and measuring the vibration generated in the robot axis,
A detection unit that detects the natural frequency of the robot axis, and an error between the natural frequency that is arithmetically output by the neural network using the drive position positioned by the positioning unit as an input and the detected natural frequency of the detection unit. , A calculating means for calculating each axis of the robot, and a weight of coupling between the units of the neural network is changed until the sum of squares of the errors calculated by the calculating means for a plurality of different driving positions becomes a predetermined value or less. The vibration reduction apparatus for a robot according to claim 1, further comprising: a control unit configured to repeatedly perform the process.