JPH03182910A - Digital adaptive controller - Google Patents

Abstract

PURPOSE:To attain highly precise positioning control by arranging a repetitive learning unit and a multiple layer neural circuit model-type learning unit so as to constitute the learning unit of two steps. CONSTITUTION:A position detector 5 which can highly precisely measure a mechanical load four axes end position thetaa is provided, a position error ea with a command value thetad is detected and a repetitive learning controller 8 first learns a highly precise positioning command sequence for respective control sampling based on error information. Then, a multiple layer neural circuit model 9 is caused to learn with the command sequence as a teacher signal so as to generate an error occurrence dynamic model. Since learning control is constituted so that it is executed in two steps, adaptive and highly precise positioning control which is difficult to be realized in a single learning unit can be executed.

Description

[Detailed Description of the Invention] [Field of Industrial Application] This invention provides highly accurate positioning control that compensates for disturbances caused by the external environment and motion errors inherent in the machine in position control devices for machine tools, robots, etc. This invention relates to a digital adaptive control device.

[Conventional technology]

Conventional known techniques similar to the present invention include, for example, Seto et al.'s ``Manipulator creation using an inverse dynamics model learned within a multilayer neural network'' presented at the Institute of Electronics, Information and Communication Engineers' study group on medical electronics and bioengineering. "Control" (IEICE Technical Report MBE87-135.1987). FIG. 2 shows the control device disclosed in this known document, in which (1) is a command value generation section, (2) is a servo controller, (3) is a motor drive section, and (4) is a servo controller. ) is the mechanical load, (
9) is a multilayer neural circuit model, (13) and (18) are servo gains, (14) is an adder, (15) is a driving torque, and (17) is a teacher signal.

Next, the operation will be explained. The position command value θ generated by the command value generation unit (1) is given to the servo controller (2) every T for each control sampling. A -order differentiator (11), a second-order differentiator (12), and a multilayer neural circuit model (9) for the command value θd are added to the servo controller (2). The feature of this servo controller (2) is that it has a learning function. First above) layered neural circuit model (9)
If the servo controller (2) has not yet learned the dynamics of the servo system, the servo controller (2) is configured to operate as a normal feedback servo control system. In other words, the position command value θ from the command value generator (1) is compared with the rotational position θa of the shaft end of the drive servo motor that drives the mechanical load (4), and the position command value θ is calculated based on the deviation em (−〇, −θm). Then, the servo * calculation (in this case, only proportional control) is performed, and the torque command value T ((-K
(*e b) (1B) is determined.

The multilayer neural circuit model (9) is provided to simulate this feedback control system through learning, and it calculates the position command value θ6 and its negative derivative, the speed command value O.
d and the acceleration command value Od, which is a second derivative value, is input, and the torque command value T to the mechanical load drive actuator is the sum of the torque command value T1 of the feedforward servo system and the torque command value T8 of the feedback servo system. It is a three-layer persebutron-type neural circuit model with more output. This multilayer neural circuit model (9) is configured to use the torque command value Tf (1B) of the feedback servo system as a teacher signal, so that the torque command value T of the feedback servo system is Learning progresses as if it were to be output. In the above-mentioned known document, a well-known hack propagation method is used as a learning algorithm for a multilayer neural circuit model.

[Problem to be solved by the invention]

Since the conventional control device is configured as described above, learning or progress is performed by using the feedback signal T (161) as a teacher signal, since the teacher signal is only the torque command value Tf of the feedback control system. Even if the learning is completely completed, it will only provide a substitute function for the servo controller (2).Since the feedback control system is generally a control system with a delay element, positioning accuracy is guaranteed after a certain period of time. , it has the disadvantage of not being able to improve the transient dynamic characteristics.Therefore, the alternative function of the feedback control system cannot improve these dynamic characteristics.Naturally, the position of the mechanical load shaft end cannot be compensated either. A so-called feedforward control method is often used to improve transient dynamic characteristics and mechanical errors by observing dynamic characteristics in advance and compensating for anticipated errors.However, the above method Now, while configuring the feedforward control system using the multilayer neural circuit model (9), the error ea (=
θd-08) cannot be reduced with high precision. In order to eliminate this drawback, there are methods to configure a full-closed feedback servo system by installing a detector at the end of the mechanical load shaft, or to observe the position signal θ8 at the end of the mechanical load shaft to obtain a command value error e8 (=θd It is conceivable to construct a multilayer neural circuit model using -θa) as a teacher signal, but in the former method, the gain cannot be increased because the dynamics of the machine is included in the servo loop, and high response cannot be achieved. In the latter method, the dynamics of 83 machine load (4) exists in the output T of the multilayer neural circuit model (9) and the observation position signal 08 questions, so in this case, the teacher signal for learning cannot be created using ea, and the hack propagation method, which is a numerical learning algorithm, cannot be applied.

This invention was made to solve the above-mentioned problems.Mechanical load (4)! A position detector that can measure the ilh end position with high precision detects the position error from the command value, and a repeating learning controller based on this error information and a desired servo command value series created by this learning device. The object of the present invention is to obtain a digital adaptive control device capable of highly accurate positioning control by configuring a two-stage learning device by arranging a multilayer neural circuit model type learning device using the following as a teacher signal.

[Means to solve the problem]

The digital adaptive control device according to the present invention is a control device that moves a controlled object to a target position commanded from the outside, and detects a position error between the target position commanded from the outside and the actual position of the controlled object. a position error detection means; a repeat learning device that applies a repeat learning control law to the position error and calculates a position command correction value corresponding to the position command value for each control sampling; a storage means for storing all values; an error propagation layered neural circuit model using the position command value as an input and the position command correction value as a teacher signal; and a function of the iterative learning device upon completion of learning of the neural circuit model. The system is equipped with means for stopping the operation.

[Effect]

The digital adaptive control device according to the present invention first learns a highly accurate positioning command sequence for each control sampling τ using an iterative learning controller, and then trains a multilayer neural circuit model using this command sequence as a teacher signal. By generating an error generation dynamics model,
It is possible to reduce the memory capacity and effectively handle the teacher signal of the neural circuit model learning device.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings. 1st
In the figure, (1) is a command generation unit that generates a movement command value per 76 unit samples from an operation program such as an NG program, (2) is a motor servo control unit, and (2) is a motor servo control unit.
3) is a motor drive unit, (4) is a mechanical load connected to the servo motor, (5) is a position detector that precisely measures the position of the mechanical load, (6) is a repetitive learning controller, and (7)
) is a storage device that stores the command correction value sequence learned by the iterative learning controller; (9) is a multilayer neural circuit model that learns to output the command correction sequence using the command correction sequence as a teacher signal; (10) is a flow determination mechanism that compares the command correction sequence that is the output of the iterative learning device with the output of the multilayer neural circuit model and determines whether learning of the multilayer neural circuit model is completed; Learning device (
6) and the storage device (7).

Next, the operation will be explained. Generally speaking, the operation of an NG machine tool is expressed in standard NG language.

This operation program is decoded by the NC device, and the movement amount per specific control cycle ΔT built into the NC device is calculated by the command generation unit (1) to control the servo motor (3) connected to the mechanical load (4). It is given to the servo control system to control. The mechanical load (4) is driven by the servo system and operates as specified in the NG program.

When performing high-precision machining, the actual machining trajectory θ8 of the mechanical load (4) axis end becomes a problem. In a normal digital servo control device, there is an error in the machining trajectory θ8 at the end of the mechanical load (4) axis due to delays in the servo system relative to the command value, inertia of the mechanical load, nonlinear disturbances such as friction, and the like. This error becomes a machining error, which becomes a serious problem in a control device for a machine tool intended for high-precision machining.

Generally, in order to eliminate such errors, delays in the servo system are dealt with by extremely slowing down the machining speed, and nonlinear disturbances such as those mentioned above are dealt with by pre-determined corrections. The solution was to add the value to the command value.

Here, this error is calculated using a position detector (5) such as a linear scale or laser length measuring device that detects the mechanical load shaft end position θ8.
), and is determined by taking the difference from the command value θd from the command generation unit (1).

The iterative learning control mechanism (8) is a part that performs control to reduce the machining trajectory error e, , (=θ, -0a) obtained in this way. Iterative learning control mechanism (
8) has a repetitive learning control calculation section (6) and a large capacity storage device (7) for storing time-series output data for each control sampling time ΔT of the learning control calculation section (6), and has an operation command program. A correction data string corresponding to each time-series command value is stored in this mass storage device (7).

FIG. 3(a) is a block diagram showing the learning operation for each trial of the iterative learning control mechanism (8).Physically, the elements required for one trial (memory (7), servo It consists of a gain (1B), a motor drive section (3), a mechanical load (4), and a learning control calculation section (6).

The concept of the iterative learning control mechanism configured as above is to drive the servo system in response to a given trajectory command θ6 (e.g. circular arc command) and detect the actual trajectory θ8 of the mechanical load end at that time. Then, the error ea (=θd−θ
Based on a), estimate how much the command value θd should be corrected, store the error correction series m in memory, and repeat the process many times to finally make the trajectory error ea zero. That is.

The calculation of the error correction series m is performed by the learning control calculation section (6). A storage device (
7) is configured as shown in Fig. 3(b), and the error correction amount for the command value θd(J) (j=1°..., n) is set as mk(j). It is constantly rewritten with each trial.

Therefore, when the learning progresses and the error e8 becomes zero, the series of command values θd(1) is stored in the storage device (7).

..., error correction series m(1) corresponding to θd(n),
... m (n) will be stored.

This error correction series is set for the program p, and the final values obtained by the iterative learning control mechanism (8) are mp□(1)9mpm(2),...
2mpm(n).

Next, the calculation of the error correction series by the iterative learning control mechanism will be specifically described.

When executing the operation program Fp(X) of program No. p, the command value series θ is calculated for each control sampling Δ
d(1), θd(2), ... θd(n) are generated, but at the same time, an error correction series mpk corresponding to this operation program is generated for this one operation program trial.
(1) , -rllp'(2) l ·mPk(n) is generated. Here, k represents the number of times the program is executed.

Also, n is the total time T required to execute the operating program fp.
Then, n=776 (1) This is an integer value expressed.

Further, the correction series mpk is usually determined by the following arithmetic expression.

m p'(i) = K L, * e a'
(j) 10 K LD*(e, k(J) −e
ak(j-1)) ten, *Σeak(F)
(2) Here, KLP is the proportional gain of the iterative learning control mechanism, KLD is the differential gain, and KLI is the integral gain.

Next time, when executing the operation program f, (x), the previous error correction sequence mpk is added to the operation program command value θd and is given to the servo system as a new command value uk++. That is, uk+I (i) = θd(i) + m pk(+)
.

i=1. ..., n (3) Therefore,
It is expected that the trajectory error will be reduced the next time the operating program is executed, but in general, servo systems and mechanical load systems have dynamics and follow-up delays to commands, so the trajectory error does not necessarily decrease significantly. However, if the above operation is repeated, the trajectory error decreases and finally converges to a certain value or less, and the error correction sequence mpII at that time also stops. This is the learning operation of the iterative learning control mechanism (8).

Providing the iterative learning device (6) in the conventional control system in this way makes it possible to reduce trajectory errors, but the correction series mpk that must be stored for command value correction is
must have a series corresponding to each operation program number and each time operating conditions such as machining speed change, so a large-capacity storage device is required in practice. Therefore,
In the present invention, a multilayer neural circuit model (9) is added to the iterative learning control mechanism (8). Multilayer neural circuit model (9)
As is well known, there is a so-called flow effect that self-organizes the dynamics of the system through learning based on the relationship between input and output. Here, we introduce a multilayer neural circuit model (
9), we self-organize the correction command value generation system using the already obtained iterative learning data mpm(1), ... mpm(n), and perform general-purpose feedforward control. Build a mechanism.

The multilayer neural circuit model used in the present invention is a bercebutron type neural circuit model that has a layered structure of three or more layers in which neurons are arranged in layers. FIG. 4 shows 3 according to an embodiment of the present invention.
A layered bersebutron type neural circuit model is shown. In the figure, (24a), (24b), ... (25a)
, (25b) s", (26a), (26b
)-- are input layer cells, (27a), (27b),-,
(28a), (28b), -(29a), (29b),
- is a hidden layer cell, and (30) is an output layer cell. In this embodiment, the command value θd, speed θd, and acceleration θd are transmitted through delay elements (21a), (21b), (22a),
(22b), (23a) and input to the neural circuit model via (23b). Here, if the number of inputs of position command value θ4, speed command value θd, and acceleration command value θd is N1, N2, and N3, this network has θ, 1(i)
, θd(i-1) θd(12) ,...θd(j-N
+-+), θ-(i), θd(i-1)θ, h (
i-2),...θd(i-N2-1), θ, h
(i) , e d(i-1) e d(i-2) ,・
...θd(i-N3-+) is input. Furthermore, the human output characteristics of each cell in the neural circuit model are modeled by a sigmoid function. The sigmoid function g(x) is expressed as g(x)=2/(1+exp(-x))-1(4), where the input is X. Here, input X is the coupling coefficient between neuron i and neuron 1 as W+, +(k), and output aj(k) of neuron j, then k)
(5) shall be expressed as

It is assumed that each cell is connected by a link with a changeable weighting coefficient, and there is no connection between the same layers. It is also assumed that the connections between each cell are made sequentially for each layer. In other words, each cell output of the input layer becomes the cell input of the hidden layer,
It is assumed that the cell output of the hidden layer becomes the input of the output layer.

Therefore, in one embodiment of the present invention, the command value generation section (1)
The command value θd from is sequentially given to the input layer of the neural circuit model (9), and after performing a connection operation between each layer, a command correction output is output from the output layer.

Next, the error learning algorithm (backpropagation algorithm) of the multilayer neural circuit model (9) will be explained. Now, focusing on the output layer of the multilayer neural circuit model (9), this output is a. (k), the desired output is m□(K) learned by the iterative learner (6), so the output error δ of the multilayer neural network model (9). (k) is δo(k)”', (m, (k) −a o(k))
*(1-a, (k)2)/2 (6). Using this error, the amount of change ΔWJ(k) in the coupling coefficient between the neuron j in the hidden layer and the neuron in the output layer is calculated as ΔWhJ(k)=τ*δ. (k)*a
Let hj(k) (7). Here, τ is a time constant that governs the speed of learning progress, and ahJ indicates the hidden layer neuron j output.

Next, consider error propagation from the hidden layer to the input layer.

The error δhj(to) of neuron j in the hidden layer is δy(k)
= ((1a hj(K)2/2) *Σδ.(k
) * Wh (k) (8), so the hidden layer neuron j and the input layer neuron i
The amount of change in the coupling coefficient ΔW+,, (k) is ΔWiJ (k) = T *δhj(k) * a
+j(k) (9). Here, 81 indicates the output of the input layer neuron j. By successively changing the coupling coefficients as described above, the layered neural circuit model (9)
As the learning progresses, it becomes possible to output the same output as the teaching signal mm.

The flow determination mechanism (10) uses the learning data m(k) from the iterative learning device (6) and the output a of the multilayer neural circuit model (9).
. (k) is constantly monitored, and when this error becomes less than a predetermined value, the multilayer neural network model (9)
It is assumed that the learning has been completed, and the operation of the iterative learning control mechanism (8) is stopped. Therefore, at this time, the data in the mass storage device required for repeated learning control becomes unnecessary.

As described above, the operation of determining that learning of the multilayer neural circuit model (9) is completed by the flow determination mechanism (10) is to first set the switch of the flow determination mechanism (lO) to the AC connection state; The iterative learning control mechanism (8) is operated to create a correction value series m, (x) for each program.

At this time, the iterative learning control mechanism (8) observes the actual error ea, performs learning control calculations based on this error ea, and finally generates a large number of NO command programs fl(X),
...Correction value series m 1m(X) corresponding to fp(×)
, m2m(x) , ・, m++m(X)
get. Here, m +-(X) = (m +-(1
) 1m + m(2) , ...m+m(n) ).

Next, set the switch of the flow determination mechanism (lO) to free,
In this state, training is performed using this correction value series data as a teacher signal so that the multilayer neural circuit model (9) can output the already learned correction value series m□(X). Once this training is completed, the multilayer neural circuit model (9)
outputs a correction value series for the command value θd, so the flow determination mechanism (10) connects the switch AB and takes the correction input for the command value from the multilayer neural circuit model (9). Make it.

In this way, the flow function of the multilayer neural circuit model (9) makes it possible to output an appropriate correction value series even for unlearned and inexperienced NC programs, and the iterative learning control mechanism (8) It is no longer necessary to store correction value series for each program, and the capacity of the storage device can be saved.

In the above embodiment, a high-precision NC machine tool has been described, but the same effects as described above can be obtained even if the present invention is applied to other mechatronic equipment systems such as robots.

〔Effect of the invention〕

9 As described above, according to the present invention, since the learning control is configured to be performed in two stages, it is possible to perform adaptive and high-precision positioning control, which was difficult to achieve with a single learning device. There is.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a digital adaptive control device according to an embodiment of the present invention, FIG. 2 is a block diagram showing the configuration of a conventional digital control device, and FIG. 3(a) is a block diagram showing the configuration of a conventional digital control device. FIG. 4(b) is a block diagram showing the control operation, FIG. 4(b) is a detailed view of the storage device (7), and FIG. In the figure, (5) is a position detector, (6) is a learning control calculator, (7) is a storage device, (8) is a repetitive learning control mechanism, (9) is a multilayer neural circuit model, and (10) is a This is a flow determination mechanism. Note that the same reference numerals in the figures indicate the same or corresponding parts.

Claims (1)

[Claims] In a control device for moving a controlled object to a target position commanded from the outside, a position error detection means for detecting a position error between the target position commanded from the outside and the actual position of the controlled object; an iterative learning device that applies an iterative learning control law and calculates a position command correction value corresponding to the position command value at each control sampling; a storage means that stores the position command correction value across the external command values; An error propagation type layered neural circuit model that receives a command value as an input and uses the position command correction value as a teacher signal, and means for stopping the function of the iterative learning device upon completion of learning of the neural circuit model. Features a digital adaptive control device.