JPH06242802A - Control method using neural circuit model - Google Patents

Abstract

PURPOSE:To provide a control method with which a mathematical model is able to perform the precise control even to an unknown linear control subject and also to compensate the error of a controlled variable caused by disturbance. CONSTITUTION:In the 1st process, the present time or a future target controlled amount are supplied to a 1st neural circuit model 2 which performs the learning the past target controlled amount and the past manipulated amount to a manipulator 1 as an input signal and a teacher signal, respectively, and a virtual manipulated amount of the present time is calculated. In the 2nd process, the virtual manipulated amount obtained in the 1st process and the controlled amount of the present time are supplied to a 2nd neural circuit model 3 which already performed the learning to estimate the action of the manipulator 1 in order to calculate an estimated controlled amount. In a 3rd process, the error between the estimated controlled amount and the target controlled amount is calculated. In a 4th process, the correction value of the virtual manipulated amount is calculated by the reverse transmission calculation of the model 3 using the error obtained in the process 3 and the virtual manipulated amount is corrected. In a 5th process, the virtual manipulated amount acquired through repetition of processes 2-4 is supplied to the manipulator 1 as its manipulated amount.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control method for controlling a non-linear controlled object such as a manipulator,
Particularly, it relates to a control method using a neural circuit model.

[0002]

2. Description of the Related Art Conventionally, as a method of controlling a non-linear control object such as a manipulator, a method combining nonlinear compensation and feedback control, or a manipulated variable is calculated from a mathematical model of the non-linear control object (especially a mathematical model of dynamic characteristics). The inverse problem method etc. are proposed. These methods cannot perform accurate control unless the mathematical model of the controlled object is known.

On the other hand, some control methods have been proposed so far for controlling a non-linear controlled object using a neural circuit model. According to this control method using the neural network, even if the mathematical model of the controlled object is unknown, the neural circuit model can perform control by acquiring the dynamic characteristics of the controlled object by learning. However, since the conventional control methods using the neural network model are all feedforward controls that calculate the manipulated variable from the target controlled variable, it is not possible to perform compensation when the controlled variable deviates from the target due to disturbance.

[0004]

As described above, as a conventional control method for a non-linear controlled object, a method that combines non-linear compensation and feedback control or a method based on an inverse problem method uses a mathematical model of the non-linear controlled object. If unknown, accurate control is difficult, and the method using the neural network model has a problem that the deviation of the control amount due to disturbance cannot be compensated.

The present invention eliminates the above-mentioned problems of the prior art, can perform accurate control even on a non-linear control object whose mathematical model is unknown, and can control the deviation of the control amount due to disturbance. An object of the present invention is to provide a control method using a neural circuit model that can perform compensation.

[0006]

In order to achieve the above object, the present invention provides a control method for controlling a non-linear controlled object, comprising: (a) inputting a past target controlled variable for the controlled object and inputting a past manipulated variable. A first step of inputting a current or future target controlled variable into a first neural circuit model that performs learning as a teacher signal to obtain a virtual manipulated variable at the current time, and (b) predicting the behavior of the controlled object A second step of obtaining a predicted controlled variable by inputting the virtual manipulated variable and the controlled variable at the current time obtained in the first step to the second neural circuit model that has been pre-learned to c) a third step of obtaining an error between the predicted control amount and the target control amount obtained in the second step, and (d) the second nerve using the error obtained in the third step. A correction amount of the virtual operation amount is obtained by back propagation calculation of the circuit model, and A fourth step of correcting the virtual operation amount, and (e) performing the second to fourth steps a predetermined number of times, or an error of the predicted control amount obtained in the fourth step is not more than a predetermined value And a fifth step of inputting to the control target the virtual operation amount obtained by repeating the above operation as the operation amount.

[0007]

In the control method of the present invention described above, the virtual operation amount at the current time obtained by the first neural circuit model,
The control amount at the current time is input to the second neural network model, and the second
The control amount can be predicted by the forward calculation of the neural circuit model of. In addition, the correction amount of the virtual operation amount can be calculated from the error between the predicted value of the control amount and the target control amount by the error back propagation calculation of the second neural circuit model.

Therefore, the forward characteristic calculation and the error back-propagation calculation of the second neural circuit model as described above are repeated to correct the virtual manipulated variable, and the result is used as the manipulated variable of the non-linear control object, so that the dynamic characteristic is improved. Correct control is possible even for an unknown control object or a control object having a strong nonlinearity of dynamic characteristics.

In the forward calculation of the second neural circuit model, the control amount observed at the current time and the virtual operation amount are input to predict the control amount so that the predicted control amount approaches the target control amount. Since the correction amount of the virtual operation amount is calculated by the error back propagation calculation, even if the control amount is deviated due to the disturbance, the virtual operation amount is corrected so as to eliminate the deviation.

Further, with respect to the past input of the target control amount,
If the learning of the first neural circuit model in which the past manipulated variable is used as the teacher signal progresses, the first neural circuit model becomes the manipulated variable that realizes the target controlled variable with respect to the input of the current or future target controlled variable. Since a close value is output, an accurate amount of operation can be obtained with a small number of iterations of the forward calculation and the error back propagation calculation of the second neural circuit model.

[0011]

1 is a block diagram showing an embodiment of a control device to which the control method of the present invention is applied. In this embodiment, a two-joint manipulator 1 moving on the xy plane shown in FIG. 2 is considered as a non-linear control target, and a target joint angle trajectory for T _f seconds from time 0 to time NΔt (= T _f ) seconds. A case will be described in which the trajectory tracking control is performed so as to track.

The controller shown in FIG. 1 has a first neural circuit model 2, a second neural circuit model 3, a feedback loop 4, a time delay element 5 of the feedback loop 4, an integrator 6, switches 7 and 8 and an adder. The adder 9 generates an error signal input used for learning the feedforward controller.

The manipulator 1 is shown, for example, in FIG.
So that the first link 10, the second link 11, the first joint 1
2nd and 3rd joint 13 and the 1st link 10 is x-axis
The angle formed by the first joint angle θ₁, An extension of the first link 10
The angle formed by the second link 11 is the second joint angle θ.₂And First
Joint torque τ₁Is to the first joint 12 and the second joint torque
τ₂Respectively to the second joint 13 in the positive direction of each joint angle.
Added. This manipulator 1 is the operation amount vector
Le τ_n= (Τ_{1, n}, Τ_{2, n})^T  Is input, and the controlled variable vector θ_n= (Θ_{1, n}, Θ_{2, n}, Θ_{1, n}, Θ_{2, n})^T  To measure. The equation of motion of the manipulator 1 is θ_{n + 1}= F (θ_n, Τ_n)

It is expressed as Where τ _{1, n} is time nΔt
, Τ _{2, n} represents the second joint torque at time nΔt, and the manipulated variable vector τ _n represents a vector having these components. Further, θ _{1, n} is the first joint angle at time nΔt, θ _{2, n} is the second joint angle at time nΔt, and θ _{1, n} is time n.
Δt represents the first joint angular velocity, Θ _{2, n} represents the second joint angular velocity at time nΔt, and the control amount vector θ _n represents a vector having these components.

Next, the first neural circuit model 2 will be described. FIG. 3 shows a first neural circuit model 2 configured by a three-layer neural circuit model. The first neural network model 2 inputs the target control amount d _{n + 1} after the target manipulated variable d _n and the sampling time of the current time n.DELTA.t, the virtual operation amount by forward calculation _{τ * n = H (d n} , d _{n + 1} , w2)

Is output. This virtual manipulated variable τ _{* n} is input to the integrator 6 as an initial value. Then, in the first neural circuit model 2, the manipulated variable τ _n input from the integrator 6 to the manipulator 1 at the sampling time is used as a teacher signal,
That is, learning is performed by the error back propagation method using τ _n −τ _{* n} as an error signal. That is, the broken line in FIG.
Is a learning error signal (τ _n −τ _{* n} ) path that is input to the first neural network model 2 from, and the connection weight vector w1 that is a parameter in the first neural network model 2 is according to this error signal. Will be fixed. Details of the calculation of the error back propagation method will be described later.

Next, the second neural circuit model 3 will be explained.
Reveal In this embodiment, the second neural circuit model 3 is shown in FIG.
And a three-layer neural circuit model as shown in FIG.
It This second neural circuit model 3 has the first neural circuit model 3 at time nΔt.
Output from the neural network model 2 of
Feeling operation amount τv_n= (Τv_{1, n}, Τv_{2, n})^T  And was fed back by field loop 4
Control amount θ_n= (Θ_{1, n}, Θ_{2, n}, Θ_{1, n}, Θ_{2, n})^T  , And the predictive control amount z after the sampling period_{n + 1}= (Z_{1, n + 1}, Z_{2, n + 1}, Z_{1, n + 1}, Z_{2, n + 1})
^T  And the predicted joint angle z after two sampling cycles^pos _{n + 2}= (Z_{1, n + 2}, Z_{2, n + 2})^T  Forward calculation (z_{n + 1} ^T , Zpos_{n + 2} ^T )^T = G (θ_n, Τ_n, W
1)

The predicted value (z _{n + 1} , zpos _{n + 2} ) obtained by this forward calculation and the target control amount (d _{n + 1} , dpos)
_{n + 2} ) as the error signal (Δθ _{n + 1} , Δpos _{n + 2} ) and based on this error signal (Δθ _{n + 1} , Δpos _{n + 2} ), it is a part of the input signal during forward calculation. The error backpropagation calculation shown in the following equation for obtaining the correction amount Δτ _n of the virtual operation amount is performed.

[0019]

[Equation 1]

4 and 5 both show the second neural circuit model 3. FIG. 4 shows the input / output relationship at the time of forward calculation and FIG. 5 at the time of error back propagation calculation. These forward calculation and error back propagation calculation will be described later.

The feedback loop 4 feeds back the measured control amount θ _{n + 1} of the manipulator 1 to be controlled to the second neural circuit model 3 via the time delay element 5.

The integrator 6 accumulates the signal input via the switch 7 at any time. Switch 8 at each sampling
Is closed, and the output of the integrator 6 is input to the manipulator 1 as the controlled object as the manipulated variable. After that, the output of the integrator 6 is reset to 0, and the switch 7 is switched to the output side of the first neural circuit model 2. That is, the first input signal of the integrator 6 is the virtual manipulated variable τ _{* n} output by the first neural circuit model 2, and then the switch 7 is switched again to output the correction signal output from the second neural circuit model 3. The quantity Δτ _n is input to the integrator 6 through the switch 7.

Next, the calculation procedure of the entire control device of FIG. 1 will be described with reference to the flow charts shown in FIGS. In this calculation procedure, the same procedure is repeated at each sampling time, and therefore the procedure for obtaining the manipulated variable at time nΔt is shown here. In addition, in the following description, FIGS.
Only the main processing indicated by (a) to (h) will be described. d is the target trajectory, d _n is the target control amount at time nΔt, T
_f is the final time. Further, itmax is the maximum number of times of repeating the prediction and correction for each sampling time, E _th is the threshold value of the error function, and if the number of iterations becomes the maximum number of times itmax or the value of the error function becomes smaller than E _th , Finish the repeated calculation. (a) First, the target control amount after the time nΔt and the sampling period _{Δt d n = (θd 1,} n, θd 2, n, Θd 1, n, Θd 2, n) d n + 1 = (θd 1, n ₊₁ , θd _{2, n + 1} , Θd _{1, n + 1} , Θd
_{2, n + 1)} is input to the first neural network model 2, the first forward calculation of the neural circuit model 2, the virtual manipulated variable _{τ * n = H (d n} , d n + 1, w2) is Is output.

(B) At this time, the switch 7 is tilted to the output side of the first neural circuit model 2, the virtual manipulated variable τ _{* n} is input as an initial value to the integrator 6, and the output value of the integrator 6 is changed. τv
_n becomes τv _{* n} . τ v _n = τ v _{* n}

(C) Virtual manipulated variable τ output from integrator 6
The v _n and the controlled variable θ _n measured through the feedback loop 4 are input to the second neural circuit model 3, and the predicted controlled variable z _{n + 1} after the sampling period of the manipulator 1 to be controlled and 2 sampling periods. Later predicted joint angle z
pos _{n + 2} is output. This calculation is performed by the forward calculation of the second neural circuit model 3. (Z _{n + 1} ^T , Zpos _{n + 2} ^T ) ^T = G (θ _n , τ v _n , w
1) Difference between predicted control amount z _{n + 1} obtained in (d) and (c) and target value d _{n + 1 for} it Δθ _{n + 1} = d _{n + 1} −z _{n + 1} Δθpos _{n + 2} = dpos _{Using n + 2} −zpos _{n + 2} , the error function E is E = (Δθ _{n + 1} ^T Δθ _{n + 1} + Δθ pos _{n + 2} ^T Δθ pos
_It is defined as _{n + 2} ) / 2.

(E) Next, the correction amount Δτ _n shown in [Equation 1] of the virtual operation amount for reducing the error function E is obtained by the back propagation calculation of the second neural circuit model 3. The back propagation calculation of the second neural circuit model 3 will be described later.

(F) The correction amount Δτ _n obtained in (e) is input to the integrator 6 through the switch 7. The virtual operation amount τv _n , which is the output signal of the integrator 6, has a value to which the correction amount Δτ _n is added.

(G) The procedures of (c) to (f) are repeated until the value of the error function E becomes equal to or less than the determined threshold value E _th or until the determined maximum number of iterations itmax. As a result, the corrected virtual manipulated variable τ _n is input to the manipulator 1 as the manipulated variable by closing the switch 8 at the sampling time nΔt.

[0029] (h) In the first neural network model 2, the target control quantity d _n, as inputs d n _{+ 1,} the operation amount tau _n teacher signal (τ _n -τ _{* n} to be input to the manipulator 1 Error signal)
As a result, learning is performed by the error back propagation learning method. The above calculations (a) to (h) are repeated at each sampling time from time 0 to the final time T _f .

Next, the forward calculation of the first neural network model 2 and the learning by the error back propagation method will be described. The calculation performed in the neural circuit model until an input signal is input and an output signal is output is called a forward calculation. As the input signal, the current time and the target controlled variable after the sampling period are input to the input layer unit. xi ₁ = θd _{1, n} xi ₂ = θd _{2, n} xi ₃ = Θd _{1, n} xi ₄ = Θd _{2, n} xi ₅ = θd _{1, n + 1} xi ₆ = θd _{2, n + 1} xi ₇ = Θd _{1, n + 1} xi ₈ = Θd _{2, n + 1} xi _i represents the input value of the i-th unit in the input layer. If the input / output function of the input layer unit is an identity function, the output of the input layer unit is expressed as follows. yi _i = xi _i (i = 1, ..., 6)

Next, the input value xh to the intermediate layer unit
_{j (j = 1, ...,} Nh 1) is the connection weight w11 _j of the input layer, hidden layer and output values yi _i of input layer _units, the sum weighted by _i, by subtracting the threshold th _j of the intermediate layer unit It becomes a value. Nh ₁ is the number of units in the intermediate layer and is represented by [Equation 2].

[0032]

[Equation 2] When the input / output function of the intermediate layer unit is a sigmoid function f (x) = {2 / (1 + exp (x)} − 1, the output value yh _j of the intermediate layer unit is the input value xh.
Using _j , yh _j = f (xh _i ) (j = 1, ..., Nh ₁ )

Can be expressed as Input value xo to output layer unit
_k (k = 1,2) is a value _obtained by subtracting the threshold th _k of the output layer unit from the sum of the output value yh _j of the intermediate layer unit weighted with the coupling load w12 _{k, j} between the intermediate layer and the output layer,
It is expressed as in [Equation 3].

[0034]

[Equation 3] The input / output function of the output layer unit is an identity function. yo _k = xo _k (k = 1, 2) The output value of this output layer unit is the torque. τ _{* 1, n} = yo ₁ τ _{* 2, n} = yo ₂

Next, learning of the first neural network model 2 by the back propagation method will be described. First, the first output value tau _{* 1} neural network model _{2, n,} and tau _{* 2, n,} the operation amount is input to the final manipulator 1 tau _{1, n,} the error function E from tau _{2, n} Define. E = {(τ _{1, n} −τ _{* 1, n} ) ^T (Τ _{1, n} −τ _{* 1, n} ) + (τ
_{2, n −} τ _{* 2, n} ) ^T (Τ _{2, n} −τ _{* 2, n} )} / 2

Here, the correction amount of the connection weight value w1 which is a parameter of the first neural circuit model 2 is obtained. First, the corrected value of the output value of the output unit and the middle layer unit is
It is calculated as in [Equation 4].

[0037]

[Equation 4] Combined load w1 between the middle layer and output layer using these values
2 _kj , the correction amount of the coupling load w11 _ij between the input layer and the intermediate layer is
It is calculated as in [Equation 5].

[0038]

[Equation 5]

Next, the forward calculation of the second neural network model 3 and the error back propagation calculation will be described. The forward calculation of the second neural network model 3 is based on the first neural network model 2
Similar to the case of, the calculation is from the input of the input signal to the output of the output signal. The forward calculation of the second neural circuit model 3 can be expressed by the following equation. First, the input values xi of 6 units in the input layer of the second neural network model 3
_The virtual operation amount τ _n and the measured control amount θ _n are input as _i (i = 1, ..., 6). xi ₁ = τ v _{1, n} xi ₂ = τ v _{2, n} xi ₃ = θ _{1, n} xi ₄ = θ _{2, n} xi ₅ = Θ _{1, n} xi ₆ = Θ _{2, n} where the input layer unit input The output function is the identity function. The output value yi _i (i = 1, ..., 6) of the input layer unit can be expressed as yi _i = xi _i (i = 1, ..., 6).

Next, the input value xh to the intermediate layer unit
_j (j = 1, ..., Nh _id ) is the output value y of the input layer unit.
coupling weight w21 _j of the input layer, hidden layer and i _i, from the sum weighted by _i, is the value obtained by subtracting the threshold th _j of the intermediate layer unit is expressed by the following equation 6]. Nh ₂ represents the number of units in the intermediate layer.

[0041]

[Equation 6] When the input / output function of the intermediate layer unit is a sigmoid function f (x) = {2 / (1 + exp (x))}-1, the output value yh _j of the intermediate layer unit is the input value xh.
Using _j , it can be expressed as yh _j = f (xh _i ) (j = 1, ..., Nh ₂ ).

Input value to the output layer unit xo _k (k = 1,
, 6) is a value _obtained by subtracting the threshold th _k of the output layer unit from the sum of the output value yh _j of the intermediate layer unit weighted by the coupling load w22 _{k, j} between the intermediate layer and the output layer. Is represented as

[0043]

[Equation 7]Output layer unit input / output function is identity function yo_k= Xo_k(K = 1, ..., 6). The output value of this output layer unit is the sampling frequency.
Predictive value of control amount after phase, joint angle prediction after 2 sampling cycles
It is a measured value. z_{1, n + 1}= Yo₁ z_{2, n + 1}= Yo₂ Z_{1, n + 1}= Yo₃ Z_{2, n + 1}= Yo_Four z_{1, n + 2}= Yo_Five z_{2, n + 2}= Yo₆ z_{n + 1} = (Z_{1, n + 1}, Z_{2, n + 1}, Z_{1, n + 1}，
Z_{2, n + 1})^T  zpos_{n + 2}= (Z_{1, n + 2}, Z_{2, n + 2})^T  The above is the forward calculation. Calculated by prospective calculation
Predicted value and target controlled variable d_{n + 1}= (D_{1, n + 1}, D_{1, n + 1}, D_{1, n + 1}, D_{2, n + 1})
^T  dpos_{n + 2}= (D_{1, n + 2}, D_{2, n + 2})^T  The error function E is defined as follows from the difference between and. Δθ_{n + 1}= D_{n + 1}-Z_{n + 1} Δθ pos_{n + 2}= Dpos_{n + 2}-Zpos_{n + 2} E = (Δθ_{n + 1} ^T Δθ_{n + 1}+ Δθpos_{n + 2} ^T Δθ pos
_{n + 2}) / 2

Next, an error backpropagation calculation for obtaining the correction amount of the input signal so that the value of the error function E approaches 0 will be described. In the error backpropagation calculation here, the correction amount of the unit output value that reduces the error function is sequentially obtained for each layer from the output layer to the input layer, and finally the correction amount of the input signal is obtained. First, the correction amount of the output value of the output layer unit is calculated as in [Equation 8].

[0045]

[Equation 8] Next, the correction amount of the output value of the intermediate layer unit is expressed as [Equation 9] using the correction amount of the output layer unit.

[0046]

[Equation 9] Similarly, the correction amount of the output value of the input layer unit is expressed as [Equation 10] using the correction amount of the intermediate layer unit.

[0047]

[Equation 10] Here, f ′ (xh _j ) represents the differential value of the sigmoid function f, and specifically f ′ (xh _j ) = (1 + f (xh _j )) (1-f (xh
It can be expressed as _j )) / 2. Finally, the correction amount of the input signal with respect to the input from the integrator 6 is expressed as follows using the constant η. Δτ _{1, n} = −ηΔyi ₁ Δτ _{2, n} = −ηΔyi ₂ Then, this correction amount is input to the integrator 6 via the switch 7.

[0048]

As described above, according to the present invention, it is possible to accurately control a non-linear control target whose dynamic characteristic mathematical model is unknown or a non-linear control target whose dynamic characteristic nonlinearity is strong. It is possible to provide a control method using a neural circuit model that can perform the compensation and can compensate for the deviation of the control amount due to the disturbance.

[Brief description of drawings]

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

FIG. 2 is a diagram showing a two-joint manipulator that moves on a plane that is a non-linear control target in the same embodiment.

FIG. 3 is a diagram showing input / output signals at the time of forward calculation of the first neural circuit model in the embodiment.

FIG. 4 is a diagram showing input / output signals at the time of forward calculation of the second neural circuit model in the same embodiment.

FIG. 5 is a diagram showing input / output signals at the time of back propagation calculation of the second neural circuit model in the embodiment.

FIG. 6 is a diagram showing a part of a flow chart for explaining a calculation procedure in the embodiment.

FIG. 7 is a diagram showing another part of the flowchart for explaining the calculation procedure in the embodiment.

[Explanation of symbols]

 1 ... Manipulator (controlled object) 2 ... First neural circuit model 3 ... Second neural circuit model 4 ... Feedback loop 5 ... Time delay element 6 ... Integrator 7, 8 ... Switch 9 ... Adder

Claims (1)

[Claims] 1. A control method for controlling a non-linear controlled object, wherein a first neural circuit model for learning is used as a first neural circuit model for learning by using a past target controlled variable for the controlled object as an input signal and a past manipulated variable as a teacher signal. The first step of inputting the target controlled variable of (1) to obtain the virtual manipulated variable at the current time, and the first step of the second neural circuit model previously learned so as to predict the behavior of the controlled object. The second step of inputting the virtual operation amount and the control amount at the current time obtained in step 3 to obtain the predicted control amount, and the error between the predicted control amount obtained in the second step and the target control amount A third step and an error obtained in the third step are used to obtain a correction amount of the virtual operation amount by back propagation calculation of the second neural circuit model, and the virtual operation amount is calculated by the correction amount. Fourth to fix
And the virtual operation obtained by repeatedly performing the second to fourth steps a predetermined number of times or until the error of the predicted control amount obtained in the fourth step becomes equal to or less than a predetermined value. A fifth step of inputting an amount as an operation amount to the controlled object, the control method using a neural circuit model.