JPH07168605A - System identifying device - Google Patents

Abstract

PURPOSE:To enable the system identifying device which identifies operation parameters of an object system to identify a linear and a nonlinear system and also to make the device into hardware by using a neural network. CONSTITUTION:This system identifying device is so constituted as to calculate and output an output signal corresponding to an input signal according to a setting changeable data processing function, and is equipped with a data processing means 101 which outputs the identification result of the object system and a setting changing data output means 102 which outputs data for a setting change for changing the data processing function of the data processing means 101 on the basis of the input/output relation of the object system.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system identification apparatus for representing characteristics of a system as a mathematical model with sufficient accuracy in the field of learning control such as robot manipulators and various plants, and in particular The present invention relates to an effective system identification device when modeling is difficult, and when the characteristics of a target system change with time.

[0002]

2. Description of the Related Art When performing control by applying a conventional control method to a system to be controlled, it is necessary to represent the characteristics of the system as a mathematical model with sufficient accuracy. For this purpose, an algorithm of a system identification method has been developed for linear systems and non-linear systems.

On the other hand, in recent years, researches on recognition and robot control using the learning ability of neural networks have been actively conducted. This neural network system is a technology expected to be able to express a target mathematical model by repetitive learning even if the accuracy of the system model is not good or unknown.

However, in a control system using a neural network, learning data as a pair of input and output is usually required for learning. That is, for a certain input, learning of the neural network is performed using known output data corresponding to this as a teacher signal. Therefore, in the case of using a neural network as a learning control device in the field of control, etc., the application is limited because a signal to be a teacher can not be obtained or the generation method thereof is unknown. Is the current situation.

In the following, as a specific example, the inverse kinematics problem of a multi-degree-of-freedom manipulator is described as an example. Numbering joints from the pedestal side of the manipulator, joints 1, 2, 3 ...
It is assumed that n. Further, the joint displacement angle of the ith joint is set to q _i correspondingly. These are put together and a joint displacement vector is

[0006]

[Equation 1]

Here, t indicates a transposed vector. The same is true thereafter. Also, a vector representing the position of the hand of the manipulator is

[0008]

[Equation 2]

The relationship between these is non-linear and can generally be written as follows.

[0010]

[Equation 3]

By the way, if a joint displacement outside 1 (hereinafter referred to as vector q) is given, the hand

[0012]

[Extra 1]

The determination of the above-mentioned coordinate outside 2 (hereinafter referred to as vector x)

[0014]

[2]

Although it can be as simple as possible, the instructions are given by hand position coordinates when working. Therefore, it is necessary to reverse the equation (1) to obtain the vector q corresponding to the indicated vector x. Formally,

[0016]

[Equation 4]

We can write: The problem of solving this equation (2) is called inverse kinematics problem, and the hand position vector x
The joint displacement angle vector q corresponding to does not necessarily exist, and it has the property that even if it exists, it is not necessarily unique.

Thus, when the end position of the robot manipulator is given, it is necessary to solve the relational expression of equation (2) in order to obtain the joint displacement angle which realizes it. However, although this method is feasible for a robot manipulator with a simple structure, complex analysis processing is required for a robot manipulator with a complicated structure that has a large number of joints. It is impossible in reality. From this, for a robot manipulator having a complicated structure, a teacher signal group for realizing the control purpose can not be obtained, so that it is impossible to construct, for example, an adaptive control device configured by an adaptive data processing device. become.

Target hand position outside robot manipulator 3 (hereinafter, vector x _d

[0020]

[3]

[0021] to) when given, to realize this target position, i.e. in order to the actual hand position vector x matches the vector x _d is the input of the deviation between the vector x _d and the vector x Therefore, a controller is required to provide the manipulator with joint angle displacement vector q such that vector x matches vector x _d . When this controller is configured by, for example, a neural network, it is necessary to provide the neural network with a teacher signal that matches the solution of the inverse kinematics problem.

As described above, a conventional example in which a neural network is applied to solve the inverse kinematics problem, in other words, to identify the system and obtain the operation parameters of the system, and to realize the control purpose inherent to the system is described below. Explain to.

First of all, as an application to identification of operating parameters of a system, there is an example where a Hopfield network is applied (Reference: ISCIE May, 22-24, 199).
1). However, this example can not be used for identification of non-linear systems and is not mentioned for control purposes.

Next, for application to control purposes, FIG.
This will be described with reference to FIG. In these figures, the adaptive controller is realized by a neural network,
A learning processing device for giving a control input error signal to cause the adaptive control device to learn a data conversion function is represented by a straight line obliquely pushing the adaptive control device. Further, N represents an adaptive data processing device prepared to realize learning processing of the data conversion function of the adaptive control device.

Prior art shown in FIG. 12 (reference document / D. Psal
tis, A. Sideris and AA Yamamura: A Multilayered Neu
ral Network, IEEE Control Systems Magazine, Vol. 8, N
o.2, pp. 17-21 (1988). In the following document 1), an adaptive controller 401 which receives a control target quantity vector x _d as an input
The adaptive data processing apparatus N to which the control state quantity vector x output from the control target 403 is input when the control operation quantity vector q output from the control target 403 is given to the control target 403
402 is provided. Then, the vector q which is the output signal of the adaptive controller 401 and the adaptive data processor N4
So that it matches the vector outside 4 (hereinafter referred to as vector q ′) that is the output signal of 02.

[0026]

[Outside 4]

A control operation amount vector for realizing the control specified by the control target amount vector x _d more faithfully by learning the data conversion function of the adaptive control device 401 and the adaptive data processing device N 402. The adaptive controller 401 is constructed to obtain q and execute it while adding it to the control target 403.

In the prior art (document 1) shown in FIG.
As shown in (a), a random control operation vector q
Is applied to the control target 501 to actually drive it, and the control state quantity vector x of the control target 501 at that time is determined and input to the adaptive control unit 502, which is an output vector of the adaptive control unit 502 at that time. q 'and controlled object 5
Amount of difference with control operation amount vector q given to 01 5
(Hereafter, vector e

[0029]

[Outside 5]

The data conversion function of the adaptive controller 502 is learned so that Then, when actually controlling the control target 501, as shown in (b), the control target amount vector x _d of the control state amount is input to the adaptive control device 502 having the learned data conversion function, and The control operation amount vector q output from the adaptive control device 502 at this time is applied to the control target 501 to control so that the control state amount of the control target 501 becomes a target.

The prior art (document 1) shown in FIG. 14 is an adaptive controller 601 which receives a control target quantity vector x _d as an input.
The calculator 603 is provided to calculate the inverse Jacobian of the control target when the control operation amount vector q output by the control target is supplied to the control target 602. And by this calculator 603,
An amount of control operation difference that changes corresponding to the amount of difference between the control target amount vector _xd and the control state amount vector x of the control target 602

[0032]

[Extra 6]

Calculate the torque Δq). And
A control operation amount vector that more closely realizes the control specified by the control target amount vector x _d by learning the data conversion function of the adaptive controller 601 so that the control operation difference amount vector Δq becomes smaller. The adaptive controller 601 is constructed to obtain q and execute it while adding it to the control target 602.

The prior art (document 1) shown in FIG.
It is a control method which combined 3 and FIG. Second adaptive controller 70 having control target quantity vector x _d as an input
2 when the control operation amount vector q outputted is given to the control target 703, and the inverse Jacobian of the control target 703
A calculator 704 is provided to calculate. Then, the calculator 704 calculates a control operation difference amount Δq that changes corresponding to the difference amount between the control target amount vector _xd and the control state amount vector x of the control object 703, and this control operation difference amount vector The data conversion function of the second adaptive controller 702 is learned so that Δq becomes smaller. Also,
The output vector of the first adaptive controller 701 q ′
The data conversion function of the first adaptive control device 701 is learned so that the control operation amount vector q given to the control target 703 becomes equal. In this case, the difference amount Δq ′ between the vector q and the vector q ′ is calculated by the subtractor 705, and this is supplied to the first adaptive controller 701.
In this method, since the Jacobian is originally a relationship between small quantities, when the difference between the vector x _d and the vector x is large, the method of FIG. 14 is not necessarily appropriate.
The difference between the vector x _d and the vector x is reduced by the following method.

Prior art shown in FIG. 16 (reference document / M. Kawa)
to, Y. Uno, M. Isobe and R. Suzuki: Aheararchical neural
network model for voluntary movement with application
tionto robotics, IEEE Control Systems Magazine 8, 8
-16 (1988).) Is the control target quantity vector x _d and the control target 8
A control operation amount vector output from an adaptive control device 801 having a fixed gain K as an input and a control target amount vector _xd as an input while providing a difference gain with the control state amount vector x of 03 as an input An amount of addition of q and an error amount output from the feedback controller 802 is provided to the control target 803. Then, by learning the data conversion function of the adaptive controller 801 so that the error amount output by the feedback controller 802 becomes smaller, the control specified by the control target amount vector x _d is realized more faithfully The adaptive control device 801 that outputs the control operation amount vector q is constructed. In this case, the feedback controller 802
The error amount is generated based on the inverse Jacobian of the control object 803.

Prior art shown in FIG. 17 (reference document / M. Jord
an: In ref. 4. (ref. 4 / Neural Networks for Control:
ed. W. Thomas Miller, III et. al. (1990).)) is provided with an adaptive data processing device N 902 having the control operation quantity vector q output from the adaptive controller 901 as its input. The data processing device N 902 learns to be a forward system having the same input / output characteristics as the control target 903. Then, after sufficiently learning, the adaptive data processing apparatus N90
Control state quantity 2 output 2 (below,

[0037]

[7]

An input error is obtained by back-propagating the amount of error between the target value vector x _d and the control target amount vector x _d with the internal state value of the data conversion function of the adaptive data processor N902 fixed. Adaptive controller that executes output of the control operation amount vector q that realizes the control target amount vector x _d by the adaptive control device 901 learning the data conversion function using this input error as a learning signal 901
To build

[0039]

SUMMARY OF THE INVENTION However, FIG.
In the prior art shown in FIG. 6, in learning the data conversion function of the adaptive controller 401, the adaptive data processor N40
Since it is necessary to learn the data conversion function 2 in advance, there is a problem that it becomes offline learning. When the output signal of the adaptive controller 401 and the output signal of the adaptive data processing device N 402 coincide with each other at the unlearned stage, the adaptive controller 40
There is a problem that learning of the data conversion function 1 can not be performed.

Next, in the prior art shown in FIG. 13, there is a problem that it is offline learning. Then, in order to make the data conversion function of the adaptive controller correct in the inverse system, it is necessary to give a large number of random control manipulated variable vectors q to the control target 501 to perform learning. There is. Furthermore, there is a problem that when the parameters of the control object, such as the link length of the robot arm, change, it is necessary to perform learning again.

In the prior art shown in FIGS.
Since it is necessary to calculate the inverse Jacobian in real time by the calculators 603 and 704, there is a problem that the learning becomes offline. And the calculator 60
There is a problem that inverse Jacobian's knowledge of the control object 602, 703 is needed to provide 3,704.

The prior art shown in FIG. 16 has the advantage of being able to realize on-line learning. However, in order to realize learning of the data conversion function of the adaptive controller 801, the fixed gain K of the feedback controller 802 needs to be appropriately set.
There is a problem that inverse Jacobian's knowledge of the control object 803 is required for setting.

Further, in the prior art shown in FIG. 17, since it is necessary to learn in advance the data conversion function of the adaptive data processing device N902 in learning the data conversion function of the adaptive control device 901, offline learning is possible. There is a problem of becoming

To summarize the problems of the prior art described above, the prior art in which system identification is performed using a neural network has a problem that it is not possible to identify a non-linear system. In addition, in the method of obtaining a mathematical model of a system, when the system to be controlled has non-linear redundant degree of freedom, it is difficult to model the system accurately, and furthermore, the solution of inverse dynamics in the mathematical model is mathematical There is a problem that it can not be determined uniquely.

When a neural network is applied to adaptive control of a robot or the like, many conventional examples use an off-line learning method, which says that it is necessary to re-learn the network if the characteristics of the control object change. There was a problem.

Further, in the system capable of on-line learning, that is, in the control system described with reference to FIG. 16, a fixed gain is used for the feedback controller 802, and if this gain value is not properly designed, the control target There was a problem that could not be achieved. In particular, in systems with strong non-linearity, setting of the fixed gain becomes more difficult.

The first object of the present invention is to make it possible to realize hardware by identifying system parameters of not only linear systems but also nonlinear systems and further realizing an identification apparatus using a neural network. . A second object of the present invention is to construct a highly adaptive adaptive controller by creating a teacher signal to a neural network used in a control system using the output result of the system identification device.

[0048]

FIG. 1 is a block diagram of the principle of the present invention. This figure is a block diagram showing an example of the basic configuration of an identification apparatus for a system or the like equipped with data processing means for outputting the identification result of a control target or the identification result of a system parameter.

In FIG. 1, the data processing means 101 has a configuration for calculating and outputting an output signal corresponding to an input signal in accordance with a data processing function which can be changed in setting, and is, for example, a hop field neural network.

The setting change data output means 102 operates the data processing means 10 based on the input / output relationship of the target system.
Setting change data for changing the setting of the data processing function to 1 is given, and is constituted by, for example, a correlator.

[0051]

In the present invention, as the identification result of the target system, for example, Jacobian or inverse Jacobian for the robot manipulator is output from the data processing means 101, for example, the Hopfield neural network.

A correlator constituting the setting change data output means 102 is connected to the hop field neural network, and the correlator is adapted to determine the hop field · · · based on the relationship between the input and the output of the target system.
The external input to each neuron constituting the neural network and the coupling coefficient between each neuron are given as an output.

As the input / output relationship of the target system, for example, the relationship between the minute displacement of the joint angle of the robot manipulator and the corresponding minute displacement of the hand position of the robot is given, and output from the correlator according to the relationship External inputs and coupling coefficients are provided to the Hopfield neural network. The operation of the input to the correlator of the minute input and the corresponding minute output, and the operation of the correlator to the external input and the coupling coefficient output to the network are repeated until the output of the network becomes constant, and the output becomes constant. When this happens, Jacobian or reverse Jacobian will be output from the Hopfield neural network.

Furthermore, in the present invention, a hierarchical neural network is provided in parallel with the correlator and the hop field neural network, and an input signal to the target system is provided as an input signal to the hierarchical neural network. A system that can cope with non-linear systems and with changes in system characteristics by learning the hierarchical neural network so that the outputs of the neural network and the hierarchical neural network coincide with each other. An identification device can be configured.

Further, for example, in the adaptive controller described in FIG. 16, a hierarchical neural network is used instead of the fixed feedback gain K, and a teacher signal for this hierarchical neural network is created using the system identification apparatus of this embodiment. By doing this, it is possible to configure a control device with excellent adaptability.

[0056]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS As an embodiment of the system identification apparatus of the present invention, an example adapted to the inverse kinematics problem of a robot arm will be described as a specific example. In the following description, the vector x is the position of the hand of the robot, and the vector q is the vector of the joint angle of the robot.
x) is a minute displacement of the hand position of the robot,

[0057]

[Ex. 8]

The vector Δq is a minute displacement of the joint angle,
These minute displacements should be measurable. Between these vector Δx and vector Δq, vector q
Assuming that the Jacobian of is 9 (hereinafter, referred to as vector J), there is a relationship of the following equation (3).

[0059]

[外 9]

[0060]

[Equation 5]

This relationship is non-linear. Therefore, in order to be able to treat as a linear system, consider the value of a specific joint angle 10 (hereinafter referred to as vector q ₀ ),

[0062]

[Extra 10]

Suppose that the Jacobian is identified with respect to the articulating angle vector q ₀ . That is, the vector J (q ₀ ) is used as a system parameter to be identified. Now consider the neural network shown in FIG. This neural network is a hop field network in which each operation element, that is, neurons are mutually connected. The computing elements are each shown in FIG.
It is realized by 00 ₁ to 200 ₄ . The number N of operation elements is 4 here, and the input to each operation element of the network is I
_i (i = 1,..., N), the output from each operation element of the network is V _i (i = 1,..., N), and the coupling coefficient between each operation element is T _ij (i, It is assumed that j = 1,..., N).

In FIG. 2, with respect to the operation element, the sum of the inputs is u _i

[0065]

[Equation 6]

The output V _{i of the} operation element is

[0067]

[Equation 7]

It can be expressed as: The dynamics of this network are as follows. First, the energy of the circuit is

[0069]

[Equation 8]

It will be. _Assuming that V _k has changed by ΔV _{k, the} change in energy is

[0071]

[Equation 9]

It becomes. Here, for example, δ _ki is a Kronecker delta symbol, and its value is 1, when k = i, k ≠ i
It becomes 0 at the time of. From this relationship, T _kk 0, 0 (k = 1,
..., N), the second term of the equation (7) becomes 0 or negative whatever ΔV _k is. As for the first paragraph,

[0073]

[Equation 10]

It is always 0 or negative by changing it. That is, the relationship between the output V _i of each operation element and the internal potential u _i is as follows, as described above.

[0075]

[Equation 11]

[0076] Here, the update amount of u _k according to the value of u _k delta
Define u _k as follows.

[0077]

[Equation 12]

As a result, since the relation of ΔV _k to Δu _k is obtained, the first term is also zero or negative and the energy decreases. The following description of the identification device makes use of this property of neural networks.

Next, an identification apparatus for outputting the Jacobian will be described. The actual displacement of the arm hand by the vector Δx, outside 11 (hereinafter, the vector Δx ′

[0080]

[11]

Let】) be the displacement calculated using the output V _{k of the} network 17. That when V _k to correspond with _{Jacobian: V k = J ij (q} 0). Consider the following as an evaluation function.

[0082]

[Equation 13]

The above equation can be embedded in the energy equation of the network 17. 2 arms moving in 2D plane
In the robot of

[0084]

[Equation 14]

[0085]

[Equation 15]

This can be achieved by FIG. 3 is a block diagram of an identification apparatus for identifying system operation parameters, here, the Jacobian. In the figure, the identification device is a minute displacement vector Δ of the input of the target system 301.
q and a small displacement vector Δx at the output are input, and a matrix T of coupling coefficients in the Hopfield network 303
And an external input vector 12 for each neuron in the network 303 (hereinafter referred to as vector I)

[0087]

[12]

It comprises a correlator 302 for output, and the above-mentioned Hopfield network 303 for receiving the input from the correlator 302 and outputting the Jacobian as the output V from each neuron inside.

FIG. 4 is a block diagram showing the detailed configuration of correlator 302. Referring to FIG. The correlator 302 includes a T setting unit 304 that outputs a coefficient matrix T, and an I setting unit 305 that outputs a vector external input matrix I.

Next, the operation of the identification device is performed in the following order. 1. Give a minute input to the system and observe the corresponding displacement of the output. It is also conceivable to use noise as an example of a minute input.

2. By the correlator using the observation results,
The coefficient matrix T of the network 303 and the input vector I are obtained and set in the network 303. 3. Until the output V of the network 303 becomes constant,
1. And 2. Repeat the process of

4. The output V that has become constant is the Jacobian of the robot arm. Among the operations of the identification device described above, the operation of the correlator will be described in more detail.
As described above, a minute input to a system, that is, a control target, for example, a joint angle displacement vector Δq, and a corresponding minute output, for example, a hand tip displacement vector Δx are observed using an appropriate sensor or the like.

The correlator 302 of FIG. 4 implements multiplication of vector elements. The setting section of T matrix giving coupling coefficients, ie, T setting section 304 in FIG. 4, calculates the simultaneous correlation between joint angular displacements when obtaining the Jacobian, and the result is
Assign as an element of a matrix. Also, when determining the inverse Jacobian, the simultaneous correlation between the hand displacements is calculated, and the result is assigned as an element of the T matrix.

The setting unit of the input vector I to the network 303, ie, the I setting unit 305 of FIG. 4 calculates the simultaneous correlation between the minute input and the minute output, ie, the vector elements of the vector Δq and the vector Δx by multiplication. The multiplication result is assigned as each element of the input vector I.

Here, the correlation of vectors means a certain time,
That is, this means simultaneous correlation between the minute input vector Δq (n) and the minute output vector Δx (n) at time n, that is, multiplication at time n. In general, correlations at different times n and m are called correlation functions, and the correlation functions are defined in the form <Δq _i (n) · Δq _i (m)>. Here, <...> means a kind of time average.

In the present embodiment, the way of setting the coefficient matrix T and the input vector I is the definition of the correlation. Consider the expansion of the above expression of the evaluation function E. Jacobian J is a quadratic square matrix in this example, and its element is the output vector of Hopfield neural network 303.
It becomes each element of (it is hereafter set as the output vector V). The

[0097]

[13]

That is, the output vector V is given by the following equation.

[0099]

[Equation 16]

Therefore, T setting unit 30 for setting coefficient matrix T
The function of 4 is to obtain the coupling coefficient T _ij and set it as the weighting coefficient of the network.

[0101]

[14]

Let the vector Δq ^t ) = (Δq ₁ , Δ
When q ₂ ), the coefficient matrix T is determined by the following equation.

[0103]

[Equation 17]

Here, vector Δq · vector Δq^tIs 2
The next square matrix, and 0 is the second 0
T is a fourth-order square matrix because
Ru. Similarly for the coefficient vector I, the vector I = 2
(Vector Δx · vector Δq ^tDefine as)
It is possible.

Among the operations of the above-mentioned identification apparatus, the operation 3. That is, 1. until the output of the network 303 becomes constant. And 2. It will be explained in detail how long the process is repeated.

As described above, the input to each neuron constituting the Hopfield neural network is

[0107]

[Equation 18]

This value is given by the internal potential (internal state) of the neuron. The value at time n + 1 of this internal potential u is given by the following equation.

[0109]

[Equation 19]

That is, the update amount Δu (n) at time n of the internal potential u is

[0111]

[Equation 20]

This update amount is proportional to the sum of the inputs. Then, the internal potential is updated according to the update amount, and the output V = g in the equilibrium state where the internal potential u does not change any more.
(U) is required. Therefore, the time required for the output of the network 303 to become constant corresponds to the time when the update amount of the internal potential becomes zero.

As described above, the change in energy of the network 303 is

[0114]

[Equation 21]

The relation between the inner potential u of the neuron and the output V is given by

[0116]

[Equation 22]

If given by the function g is a monotonically increasing function, the directions of change of Δu and ΔV are the same and the energy will be reduced. However, in the present invention, since the coefficient matrix T and the input vector I are set each time, the energy of the network 303 may not necessarily decrease with time, but if the moment of updating the internal potential u is taken, u changes Direction corresponds to the energy reduction direction. In the case of an ordinary hop field neural network, the energy generally decreases with time because the values of coupling coefficients and input vectors are constant.

Next, FIG. 5 shows the configuration of an identification apparatus capable of identifying a non-linear system. This apparatus is obtained by adding the apparatus of FIG. 3 which identifies a linear system in parallel with a learning apparatus 306 (a hierarchical neural network) capable of learning a non-linear mapping. In the identification apparatus of FIG. 3, the coefficient input to the network 303 does not have a memory capacity because the coefficient input to the network 303 is successively updated by the correlator 302, but in the apparatus of FIG. It has a memory capacity because it is configured as follows.

Also, the stored contents are corrected by the identification device when the system changes. That is, in the identification device of FIG. 3, the Jacobian could be calculated with a specific state vector q ₀ . In order to extend this to nonlinear identification, it is necessary to consider the dependence of the Jacobian on the state vector q. For this,
A neural network device 306 having a learning function is used. The learning of the neural network device 306 is performed in the following procedure.

1. A learning device 306 for the internal state vector q
To obtain an output 15 (hereinafter referred to as a vector V ′ (q)) of the corresponding learning device 306.

[0121]

[15]

2. The minute displacement vector Δq around the state vector q is added to the above-mentioned state vector q, the corresponding hand displacement vector Δx is measured, and these are added to the correlator 302. 3. From the correlator 302, the coefficient matrix T of the identification device and the input vector I are obtained.

4. Process 2 until the output V of the network 303 converges And 3. repeat. 5. After convergence, the state vector q is
The vector V '(q) corresponding to is trained. The teacher signal at the time of learning is the output of the subtractor 307.

6. Process 1 To 5. Is repeated changing the state vector q. Thus, the learning device 306 stores a set of the state vector q and the corresponding output vector V '(q). The vector V '(q) is Jacobian J (q).

The identification device shown in FIG. 5 can also learn reverse Jacobian. In this case, the following (21)
Use the evaluation function E of the equation: V _k = J ⁻¹ _ij (q ₀ ).

[0126]

[Equation 23]

Hopfield network 3 is obtained by embedding the above equation into the energy equation of the network.
The matrix T of coupling coefficients in 03 and the external input vector I for each neuron in the network are determined as follows.

[0128]

[Equation 24]

[0129]

[Equation 25]

The correlator 302 generates the matrix T of coupling coefficients and the external input vector I, and inputs them to the hop field network 303. By this, the hop field network 303 outputs V ⁻¹ which is the reverse Jacobian. The reverse Jacobian learning method is the same as the above-described Jacobian learning method except that the matrix T of coupling coefficients and the external input vector I are different, so detailed description will be omitted.

FIG. 6 is a configuration block diagram of a first adaptive learning control device using the system identification device of FIG. 5 described above. In FIG. 6, the system identification device consisting of the correlator 302, the network 303, the learning device 306 and the subtractor 307 of FIG. 5 described above is shown in block 310 and the system identification result by this identification device 310, here the output of the Jacobian The result is used for the adaptive learning controller.

The control device includes a control object (Co) 311, a feedforward control mechanism (for example, composed of a neural network Np) 312, an externally input target value vector x _d, and an output vector x of the control object (Co) 311. subtraction result, i.e. the multiplier providing the multiplication result by multiplying the subtractor 313, the output and the control deviation e _x of the identification apparatus 310 for obtaining a control deviation e _x as a teacher signal to the neural network Np312 with 3
It consists of fourteen.

The load update amount T for the neural network Np 312 corresponding to the feedforward controller in FIG. 6 is given by the following equation.

[0134]

[Equation 26]

In the equation (24), it is considered to change the load w ₁ so as to minimize the evaluation function of the equation (11). When the steepest descent method is used as the method of change, the updated amount of load can be obtained as follows.

[0136]

[Equation 27]

Here, μ is a positive constant. Details of the method of change of the equation (25) will be described later. Here, the symbol T
Is different from the coefficient matrix described in FIG. The first in FIG.
The learning processing method of the adaptive type learning control apparatus will be described.

[0138] Control deviation vector Δx (= e _x) is calculated by the difference between the target value vector x _d and output vector x. However, although the vector Δx is a minute displacement in the above description, the minute displacement corresponds to the output of the control target, and the same symbol is used. The same applies to the vector Δq.

The neural network Np 312 receives the Jacobian J (q) output from the system identification unit 310.
And the control deviation vector .DELTA.x is multiplied as a teaching signal, and learning is performed online so as to output a control operation amount vector q of the joint angle corresponding to the target value vector _xd . The vector q used to obtain the Jacobian J (q) is the same as the vector q input to the control target (Co) 311.

When learning progresses sufficiently, network N
p312 acquires the data conversion function of "vector x _d → vector q" as the inverse characteristic of the input / output characteristic of the control object 311. Also, even when the parameter of the control target (Co) 311 changes, the system identification device 3
The data conversion function of the network Np 312 is adaptively changed by re-learning of the hop field network 303 in 10 and the learning device 306 which is a hierarchical network.

FIG. 7 is a block diagram of a second adaptive learning control apparatus using the system identification apparatus of FIG. 5 described above. In FIG. 7, the system identification device consisting of the correlator 302, the network 303, the learning device 306 and the subtractor 307 of FIG. 5 is shown at block 320, and the system identification result by this identification device 320, here the output of Jacobian The result is used for the adaptive learning controller.

The controller is a control object (Co) 321, a feedforward control mechanism (for example, composed of a neural network Np) 323, and instead of the feedback controller 502 of FIG. that the target value vector x _d and the control object (Co) 321 subtraction result between the output vector x, i.e. to obtain a control deviation e _x subtracter 3
24, the output of the identification device 320 control deviation e _x and Neural Network the multiplication result by multiplying the Na32
Multipliers 325 and 2 to give 2 as a teacher signal
And an adder 326 for adding the outputs of two networks Na 322 and Np 323.

The load update amount T for the hierarchical neural network Na 322 corresponding to the feedback controller in FIG. 7 is given by the equation (24), similarly to that used in the first adaptive learning control device.

The learning processing method of the second adaptive control device of FIG. 7 will be described. This apparatus has a hierarchical neural network Na 322 as an alternative to the conventional example described in FIG. 5 having a feedback control apparatus with a fixed feedback gain. And this network Na 322 is an identification device 3 of the present invention.
Learning is performed using the teacher signal generated by the multiplier 325 based on the 20 outputs.

Network Na 322 receives control deviation vector Δx (= target value vector x _d −output vector x) as an input, and outputs difference amount vector Δq of the control operation amount (joint angle) corresponding to control deviation vector Δx. Learn to do. In addition, when learning, a multiplication result (= T) of Jacobian J (q) output from the system identification unit 320 and the control deviation vector Δx is received as a teacher signal.

On the other hand, the network Np 323 constituting the feedforward control mechanism is a network Na 32
The deviation vector Δq (= C) output from 2 is used as a teacher signal, and learning is performed online so as to reduce the deviation vector Δq. For the control target (Co) 321, the addition result of the outputs of the two networks Na 322 and Np 323 is given as a control operation amount.

When learning progresses sufficiently, network N
Although p323 acquires the data conversion function of "vector x _d → vector q '" as the inverse characteristic of the input / output characteristic of the control target 321, the data conversion function is always the output of the system identification unit 320 and the network Na322. It changes adaptively corresponding to the data conversion function of Compared to the first adaptive control device, the second adaptive control device has the advantage of being able to adapt quickly even when the parameter of the control target (Co) 320 changes, because the network Na 322 is present.

FIG. 8 is a block diagram of the configuration of a third adaptive learning control apparatus using the system identification apparatus of FIG. 5 described above. The system identification unit comprising the correlator 302, the network 303, the learning unit 306, and the subtractor 307 in FIG. 5 in FIG. 8 is indicated by a block 330. As a result of the system identification by the identification unit 330, here the output of Jacobian. The result is used for the adaptive learning controller.

The difference between the second adaptive learning controller and the third adaptive learning controller is that the joint displacement vector q is input to Na332 of the third adaptive learning controller, and Np333 The link length L is input.

The respective terms in the equation (25) will be described below.

[0151]

[16]

(First Term) This is an error of the hand position and is observed by a sensor or the like. (Section 2) The part related only to the input and output of the system
Partial differentiation (Jacobian) of the state q of the system output x
It is. this is,

[0153]

[Equation 28]

It is approximately given by That is, the calculation of this part can be obtained by measuring the change of the input and the output of the system without actually calculating the partial derivative of the mathematical model. (Section 3) The part that depends only on the neural network,
It is a partial differential of the load w ₁ of the internal state q. This is given at the input to the output neuron. That is, assuming that φ is the transition function of the neural network, the output of the multilayer neural network is

[0155]

[Equation 29]

It is given by _Assuming that _i0 is an initial value of the joint angle, the nth joint angle is

[0157]

[Equation 30]

Since the partial derivative of the third term is obtained as

[0159]

[Equation 31]

Here, φ ′ (u _i ) represents the derivative of the transition function. By the way, if it is assumed that the transition function is linear, this term is a constant (set as 1), so the third term depends on only V _j . Since V _j is the output of the second layer (preceding layer of the output layer), it depends on how the input is represented in the second layer (in the case of the second layer neuron in a tabular neural network) This is an output, which is “1” or “0”, so it contributes to learning only when it is “1”.

In this way, the load is updated (learning)
Thus, Na332 operates so as to be x → _xd . As a result, a map of Δx → Δq corresponding to the joint displacement vector q, that is, an inverse Jacobian is formed in Na 332 after learning. And it's done online.

In addition, Np 333 is the target position of the hand x _d
Is input. In addition, if the link length is changed by inputting the link length L etc. (if,
It is possible to output an output corresponding to the change immediately if it is the learned link length. In FIG. 8, Np 33
Assuming that the output of 3 is q ', learning of Np 333 is performed using q which is an internal state of the system as a teacher. Evaluation function, J ₂ =
If (q-q ') ^2/2, the update amount is given as follows.

[0163]

[Equation 32]

As a result, the inverse transformation x _d → q of the input and output of the controlled system is formed in Np 333. Learning method of the formula (31) to an evaluation function J ₂ can be interpreted in the following manner. Since q = Δq + q ′, minimizing J ₂ is the same as minimizing Δq with q as a virtual target value. That is, it may be said to be a learning method in which learning of Np 333 is performed so as to minimize Δq which is the output of Na332.

In the configuration shown in FIG. 8, learning is performed based on the operation of Na332 and a target is imposed on Na332 which performs adaptive operation using information from sensors such as vision and sense (joint angle) as a main information source. Proper operation of the controlled system
Np 333 obtained from 332 takes a structure added, and Np 333 functions as a feedforward adaptive controller with learning ability.

When the link length L is not added as an input to Np 333, the change can not be output correctly after the correction by N p 333 is completed.

FIG. 9 is a block diagram of a fourth adaptive learning control apparatus using the identification result by the system identification apparatus of FIG. In FIG. 9, the correlator 302, the network 303, the learning device 306, and the subtractor 30 of FIG.
The system identification system consisting of 7 is indicated at block 340, and the system identification result by this identification system, here the output result of the inverse Jacobian, is utilized in the adaptive learning controller.

The control device is a control object (Co) 341, a feedforward control mechanism (for example, composed of a neural network Np) 342, a target value vector x
subtraction result between the output vector x _d and the control object 341, i.e. to the subtractor 343, the neural network Np342 was by multiplying the multiplication result output and a control deviation e _x of the identification apparatus 340 for obtaining a control deviation e _x It comprises a multiplier 344 which provides a training signal as a teaching signal, and an adder 345 which adds the multiplication result from the multiplier 344 and the output from the network Np 342.

The neural network Np 342 performs learning so as to reduce the teacher signal C generated based on the output of the system identification device 340. The teacher signal C is given by the following equation.

[0170]

[Equation 33]

When learning progresses sufficiently, network N
At p343, the data conversion function of "vector x _d → vector q '" as the inverse characteristic of the input / output characteristic of the control object 341 is acquired. Since there is an advantage that the deviation vector Δq can be obtained directly as compared to the second adaptive learning controller, the deviation vector Δq and the network Np 342 are obtained.
A more adaptive operation is made possible by directly inputting the control operation amount vector q of the joint angle obtained by adding the output q 'of f to the control target (Co).

In FIG. 10, the link lengths of two links are respectively 0.5 m, and the value of Jacobian J at the time of system identification is 1 to 100 repetitions.

[0173]

[Equation 34]

The number of repetitions 101 to 200 is

[0175]

[Equation 35]

The simulated result of the value of Jacobian J when the setting is set is shown. It is apparent from FIG. 10 that the value of Jacobian J converges to the correct value by repeating about 20 times.

FIG. 11 shows the change of the energy of the network 17 at the time of the simulation of FIG.
When obtaining the aforementioned equations (12) and (13), at the time of expansion of equation (11),
Although the term of (vector Δx) ² is ignored, FIG. 11 shows the energy including the term. Note that, as described above, in the case of an ordinary Hopfield network, the energy decreases with time, this (vector Δx)
^This is because the term corresponding to ² is a constant.

Referring to FIG. 11, energy convergence can be seen in a short time of about 5 unit times corresponding to FIG. The energy temporarily increases at 100 units of time when the Jacobian's value is changed, but it converges to 0 again in a short time as in the first case. Also, in the case of obtaining the value of the inverse Jacobian as the identification result, the value converges to the correct value as in the case of obtaining the value of the Jacobian, but the detailed description here is omitted.

[0179]

As described above, according to the present invention, the identification result of the target system can be obtained using a neural network, and the identification apparatus can be realized in hardware. And by using a layered neural network in combination, it can also be used as an identification device for nonlinear systems.

Also, by utilizing the identification device of the present invention to obtain a teacher signal in an adaptive controller,
It becomes possible to make the adaptive control device learn online, re-learning is unnecessary even when the characteristic of the control object changes, and the use efficiency is improved as compared with the conventional off-line learning method.

Furthermore, the present invention can be applied to non-linear systems by combining hierarchical neural networks as described above, and is not limited to solving only inverse kinematics problems, but to systems having dynamic characteristics. Is also applicable.

Brief Description of the Drawings

FIG. 1 is a block diagram showing the principle of the present invention.

FIG. 2 is a diagram showing a configuration of a neural network used to identify an operation parameter of a system.

FIG. 3 is a block diagram showing a configuration of an embodiment of a system identification device.

FIG. 4 is a block diagram showing a detailed configuration of a correlator.

FIG. 5 is a block diagram showing a configuration of a system identification apparatus that enables identification of a non-linear system.

FIG. 6 uses the system identification result according to the invention first
FIG. 12 is a block diagram showing a configuration of an adaptive learning control device of the present invention.

[FIG. 7] Second using the system identification result according to the present invention
FIG. 12 is a block diagram showing a configuration of an adaptive learning control device of the present invention.

[FIG. 8] Third using the system identification result according to the present invention
FIG. 12 is a block diagram showing a configuration of an adaptive learning control device of the present invention.

FIG. 9 utilizes the system identification result according to the present invention;
FIG. 12 is a block diagram showing a configuration of an adaptive learning control device of the present invention.

FIG. 10 is a diagram for explaining a Jacobian identification result.

11 is a diagram showing changes in energy in the simulation of FIG.

FIG. 12 is a block diagram showing a configuration of a conventional example (Part 1) of a control device using a neural network.

FIG. 13 is a block diagram showing the configuration of a conventional example (Part 2) of a control device using a neural network.

FIG. 14 is a block diagram showing a configuration of a conventional example (Part 3) of a control device using a neural network.

FIG. 15 is a block diagram showing a configuration of a conventional example (Part 4) of a control device using a neural network.

FIG. 16 is a block diagram showing a configuration of a conventional example (No. 5) of a control device using a neural network.

FIG. 17 is a block diagram showing a configuration of a conventional example (part 6) of a control device using a neural network.

[Description of the code]

 101 Data processing means 102 Data output means for changing settings

── ── ── ── ──続 き Continuation of the front page (51) Int.Cl. ⁶ Identification code Office serial number FI Technical display location // G06G 7/60

Claims (21)

[Claim of claim] 1. A system identification apparatus for outputting an identification result of an operation parameter of a control target system, comprising: setting change data for changing a setting of a data processing function based on an input / output relationship of the control target system. System identification characterized by: data output means for setting change, and data processing means for calculating and outputting the identification result using the data for setting change output from the data output means for setting change apparatus. 2. The input / output relationship of the controlled system
The system identification device according to claim 1, wherein the system identification apparatus has a correspondence relationship between a minute input to the control target system and a minute output of the control target system for the minute input. 3. The data processing means is constituted by a hop field neural network in which a plurality of neurons are mutually connected, and the setting change data output means is used as the setting change data for the neural network. The external input signal to the neuron and the coupling coefficient between the neurons are given while being slightly changed until the output of the neural network becomes constant, and the neural network determines the constant output value as the identification result. The system identification device according to claim 1, wherein the system identification information is output. 4. The system identification device according to claim 2, wherein noise is used as the minute input. 5. The setting change data output means sets the input / output relationship of the controlled system as a relationship between a minute displacement for a specific input value and a minute displacement of an output of the controlled system for the input minute displacement. The internal potential of each neuron is updated by changing the coupling coefficient between the external input signal to each neuron constituting the neural network and the corresponding neuron while making the specific input value minutely displaced successively. The system identification device according to claim 3, wherein the output of the neural network is changed. 6. The system identification device further includes a learning device to which an input signal to the control target system is input, wherein the learning device matches the output of the own device with the output of the data processing means. The system identification apparatus according to claim 1, wherein the system identification apparatus performs the learning and stores the identification result of the operation parameter of the control target system obtained as a result of the learning. 7. The system identification device according to claim 6, wherein the learning device stores the identification result of the operation parameter for the non-linear control target system. 8. The control target system is a multi-joint robot, and the input / output relation of the control target system is a relation between a minute displacement of a joint angle of the multi-joint robot and a micro displacement of a hand position of the robot. The data processing means may output, as the identification result, a Jacobian which is a matrix for converting minute displacements of joint angles of the articulated robot into minute displacements of hand positions. System identification device. 9. The control target system is a multi-joint robot, and the input / output relation of the control target system is a relation between a minute displacement of a joint angle of the multi-joint robot and a micro displacement of a hand position of the robot. The data processing means outputs, as the identification result, a reverse Jacobian which is a matrix for converting a minute displacement of the end position of the articulated robot into a minute displacement of a joint angle. System identification device as described. 10. An adaptive learning control apparatus having a feedforward control mechanism for a control target system, comprising: setting change data for changing a setting of a data processing function based on an input / output relationship of the control target system; Data output means for outputting setting change data, data processing means for calculating and outputting an identification result using the data for setting change output from the data output means for setting change, target value and actual value for the system to be controlled An adaptive learning control device comprising: a multiplier for giving the feedforward control mechanism a multiplication result of a control deviation between the output value of the data processing means and the output value of the data processing means. 11. The feedforward control mechanism according to
The adaptive system according to claim 10, comprising: a neural network, wherein an output from the multiplier is input as a teacher signal, and on-line learning is performed so that an output of the feedforward control mechanism becomes an appropriate amount. Learning control device. 12. An output from the data processing means is a Jacobian calculated from a minute input to the control target system and a minute output of the control target system for the minute input. 10. The adaptive learning control apparatus according to 10. 13. A feedforward control mechanism for a control target system and a control deviation between a target value and an actual output value for the control target system are input, and an output corresponding to the control deviation is the feedforward control mechanism. An adaptive learning control device including a feedback control mechanism that adds the result to the output of the control target system to the control target system, the feedback control mechanism comprising a neural network having a learning function, A setting change data output unit that outputs setting change data for changing the setting of the data processing function based on an input / output relationship of the system; and the setting change data output from the setting change data output unit Data processing means for calculating and outputting an identification result using It multiplies the control deviation, the adaptive learning control apparatus characterized by comprising: a multiplier to give as a teacher signal of the result to the neural network. 14. The feedforward control mechanism according to
The adaptive learning control apparatus according to claim 13, wherein an output to the control target system is changed such that the control deviation becomes gradually smaller by on-line learning based on the teacher signal. 15. The feedforward control mechanism according to
A neural network comprising: a neural network, wherein an output from the feedback control mechanism is input as a teacher signal, and on-line learning is performed so that an output of the feedforward control mechanism becomes an appropriate amount. Adaptive learning control device. 16. The adaptive learning control apparatus according to claim 15, wherein the feedback control mechanism further receives an addition result input to the control target system. 17. The controlled system is an articulated robot, and the input / output relation of the controlled system is a relation between a minute displacement of a joint angle of the articulated robot and a minute displacement of an end position of the robot. The adaptive learning control device according to claim 16, wherein the feedforward control mechanism further receives a link length of the robot. The output from the data processing means is a Jacobian calculated from a minute input to the system to be controlled and a minute output of the system to be controlled for the minute input. The adaptive learning control apparatus according to 13. 19. An adaptive learning control apparatus having a feedforward control mechanism for a control target system, comprising: setting change data for changing a setting of a data processing function based on an input / output relationship of the control target system; Data output means for outputting setting change data, data processing means for calculating and outputting an identification result using the data for setting change output from the data output means for setting change, target value and actual value for the system to be controlled A multiplier giving the result of multiplication of the control deviation between the output value of the data processing means and the output of the data processing means to the feedforward control mechanism, and adding the output of the feedforward control mechanism and the output of the multiplier , An adder for giving the result of the addition to the control target system as a control operation amount, and adaptive learning Control device. 20. The feedforward control mechanism according to
20. The adaptive learning control device according to claim 19, comprising: a neural network, wherein the multiplication result is given as a teacher signal to the neural network. 21. The apparatus according to claim 1, wherein the output from the data processing means is a reverse Jacobian calculated from a minute input to the control target system and a minute output of the control target system for the minute input. The adaptive learning control device according to Item 19.