JP2923000B2 - Servo control device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION The present invention relates to a servo control device, and relates to a neural network, such as a robot or a rolling mill, in which a drive characteristic of a controlled object has a large fluctuation and a load disturbance is easily applied. The present invention relates to a servo control device that adaptively controls learning by using a servo control.

[Conventional technology]

It is known that a neural network (neural network) has an ability to learn characteristics of a dynamic system (dynamic system) including a nonlinear system by a relatively simple algorithm. Proceedings of the Automatic Control Confere is a method of applying this neural network to control and constructing a servo controller that is robust to the fluctuations in the characteristics of the controlled object.
nce, vol. 1, pp. 910-915 (1989) has been proposed. This is because, as shown in FIG. 3, forward the provided neural network 2 to the control object 1, calculates the manipulated variable u by inputting the target command value x _r to the servo control system to the neural network 2 I do. Here, the neural network 2 previously outputs the output x of the system from the input u and the output x to the control target 1.
Of the input u (generally, this characteristic is called an inverse system).

Therefore, the operation amount u which is calculated by the neural network 2 is a manipulated variable for matching the output x of the controlled object 1 in the target command x _r. Here, the learning of the inverse system of the controlled object 1 is to correct the connection relationship between the units constituting the neural network so that the output when x is input to the neural network 2 matches the actual input u. Is executed by

The operation from the input to the output in the neural network is executed by a relatively simple algorithm, and the characteristics of the system are expressed by the connection state between a large number of units constituting the neural network. For this reason, it is known that there is more redundancy than a method of identifying the characteristics of the system using a linear relational expression. By using a neural network with these features for the servo control system,
Since the manipulated variable can be calculated adaptively with respect to the characteristic variation of the controlled object, application to a system in which the influence of parameter variation is remarkable, such as a robot or a servo system of a rolling mill, is being studied.

[Problems to be solved by the invention]

However, in the above-described conventional technique, in the process in which the neural network 2 learns the inverse system of the control target 1, the deviation between the actual inverse system and the characteristics acquired by the neural network is large. Therefore, when control is performed using an output value from a neural network provided in series with the target command value, an error between the output of the control target and the target command value increases, and the control system may become transiently unstable. is there. For this reason, it is necessary to learn a neural network using a simulated teacher signal and then connect it to a control target. Furthermore,
In addition to the initial learning process described above, even when the dynamic characteristics of the controlled object fluctuates transiently or when a sudden disturbance acts on the control system, the actual output in the learning process of the neural network is And the target value increases. Also, when using a neural network as a feedforward control element,
Although the followability to the target command value is improved, there is also a problem that the effect of suppressing the disturbance of the system is not sufficient.

An object of the present invention is to construct a servo control system that is stable even in a process in which a neural network initially learns the characteristics of a system, and also against sudden changes in the characteristics of the system and disturbance inputs. .

[Means for solving the problem]

In order to achieve the above object, a linear control system is provided in parallel with the neural network, and a linear control system calculated from an operation amount calculated by the neural network with the target command value as an input and the output of the target command value and the control target. Is controlled as an operation amount for the control target. Here, the neural network learns the inverse system of the control target from the input and output to the control target. Since this learning is performed independently of the linear control system, the neural network and the linear control system can be set individually.

[Operation] Here, the linear control system provided in parallel with the neural network functions to prevent the control system from becoming unstable in the process of learning the characteristic variation of the control target by the neural network. In addition, by using a linear control system together, a neural network that is a feedforward compensation system can configure a robust control system even for a disturbance with little suppression effect. Here, the neural network learns the inverse system from the input and output to the control target. For this reason,
The linear control system can set control characteristics independently of feedforward control by a neural network. For this reason, there is an advantage that the constants of the linear control system and the configuration of the neural network can be simplified as compared with a method of learning the neural network using the output of the linear control system.

〔Example〕

Hereinafter, an embodiment of the present invention will be described with reference to FIG. In FIG. 1, 1 is a control object, 2 is a neural network, 3 is an output error operation of the neural network, 4 is a linear control operation, and 5 is an adder for the output of the neural network and the output of the linear control. The portion 6 (indicated by a broken line) represents the operation contents of the servo control device. Reference numeral 7 denotes a switch for switching an input signal to the neural network, and reference numeral 8 denotes a switch for switching an output signal from the neural network.

The servo controller sequentially executes learning of the neural network, calculation of the operation amount from the neural network, and linear control calculation at each predetermined control time. FIG. 2 shows a control operation at an arbitrary k point. In FIG. 2, at 601, an operation amount u (k) to the control target, an output x (k) of the control target, and a target command value x _r (k) to the control system at the time point k are fetched. Then 60
In 2, the switches 7, 8 on the input side and the output side of the neural network 2 are closed to the (a) side, and x (k) is input to the neural network to make the temporary output value u for learning the neural network.
(K) is calculated.

Next, at 603, an output error e (k) of the neural network is calculated from the provisional output value (k) and the actual input value u (k), and the respective units constituting the neural network are calculated accordingly. Modify the binding state of. The operation of the output signal from the input signal in these neural networks and the correction operation of the neural network from the output error of the neural network are described in, for example, DE Rumelhart,
"Learning Inte" by GE Hinton, and RJ Williams
rral Representations by Error Propagation ", Paralle
l Distributed Processing MIT Press, 1986, pp. 318-36
Perform by the method described in 2.

Next, at 604, the switches 7 and 8 are closed to the (b) side, and the target command value _xr (k) is input to the neural network and the manipulated variable u
_a (k) is calculated. At 605, the target command value x
_An operation amount u _l (k) is obtained from _r (k) and the output x (k) of the control object by a linear control operation. Finally, the manipulated variable u _a (k) from the neural network and the manipulated variable u from the linear control unit operation
_l (k) is added and output as an operation amount for the control target. This updated operation amount is represented by u (k + 1).
By performing the above control calculations at predetermined control cycles,
Servo control is performed.

According to the method of the present embodiment, even during the period in which the neural network learns the initial and transient characteristic fluctuations of the controlled object, linear control is performed using the target command value to the servo control system and the output of the controlled object. The control system can be stabilized. For this reason, it is possible to configure a servo control device that is stable in all operation regions without making the learning loop of the neural network more responsive than necessary. Further, according to the method of the present embodiment, since the learning characteristics of the neural network can be set independently of the linear control operation unit, there is an advantage that the control system can be easily constructed.

Next, a description will be given of a second embodiment of the present invention in which the present invention is applied to a motor speed control device. The configuration is shown in FIG. The control target includes a servo amplifier 101, a motor 102, a mechanical load 103 driven by the motor, a motor current detector 104, and a motor speed detector 105. Here, the motor current is fed back to the servo amplifier 101, and the generated torque of the motor is controlled according to the torque command value. The output u from the servo controller 6 is a torque command value for the servo amplifier, the output x from the controlled object is the motor speed, and the target instruction x _r to the servo control system.
Corresponds to the motor speed command value.

The servo control device 6 executes the following control calculation every predetermined control cycle. First, a motor speed detection value x (k) and a speed command value _xr (k) are taken. Next, switches 7 and 8 (a)
Then, x (k) is input to the neural network 2 and a provisional output value (k) is calculated. The neural network of this embodiment has one input and one input.
The output is composed of three layers: an input layer, a middle layer, and an output layer.
The configuration is shown in FIG. The input layer consists of one unit and one
The intermediate layer is composed of n units and one fixed output unit, and the output layer is composed of one unit. Now, the output of each unit of the input layer is x _α (α = 1, 2), the output of each unit of the intermediate layer is x _β (β = 1 to n + 1), and the output of each unit of the output layer is x _γ (γ = 1), the signal propagation from the input layer to the intermediate layer and from the intermediate layer to the output layer can be expressed by the following equation.

Here, wβα and wγβ are respectively from the input layer to the intermediate layer,
In the connection count from the hidden layer to the output layer, ₁ and ₂ are functions that specify the relationship between the weighted sum of input variables and the output value. Here, the following functions are used as ₁ and ₂ .

Also, among the x _alpha, Yunitsuto of beta = n + 1 of alpha = 2 of Yunitsuto and x _beta output value is 1 fixed output Yunitsuto a constant.

By executing the operations of the equations (1) to (4) in the direction from the input layer to the output layer, a provisional output (k) when the input x (k) is input is obtained.

Next, the connection counts wβα and wγβ of the neural network are corrected from the provisional output (k) and the actual input value u (k). Here, an error is back-propagated using the following formula.

Here, e is the output error of the neural network, δ _wγβ , δ
_wβα is a correction amount of each coupling coefficient, η is an adjustable parameter that determines a learning characteristic of the neural network, and ₁ ′ and ₂ ′ are derivatives of ₁ and ₂ , respectively. In this embodiment, (3),
Since the function of the equation (4) is used, ₁ ′ (ξ) = ₁ (ξ) = ξ (8) ₂ ′ (ξ) = ₂ (ξ) ｛1− (ξ)｝ (9) Therefore, the equations (6) and (7) are simplified as follows.

_{δw γβ = -η · e · x} γ · x β ... (10) δw βα = -η · e · x γ · x β (1-x β) x α ... (11) i.e., (5), (10 ) And (11) in this order.
The neural network is learned by modifying the coupling coefficient.

Then, close the switch 7 and 8 (b) side, and calculates a speed command value x _r to the neural network that fix the coupling coefficient (k) enter the by manipulated variable u _a (k). This is performed similarly to the calculation of the provisional output (k).

Finally, in the linear control system, u _l (k) is calculated from the speed command value x _r (k) and the detected speed value x (k). In this embodiment, proportional integral compensation is used. The calculation at this time is executed as in the following equation.

_{Δx (k) = x r (} k) -x (k) ... (12) u l (k) = u lP (k) + u lI (k) ... (13) here, It is. Where K _P is proportional gain, T _I is integration time constant, Ts
There represents a control period of the linear control, respectively, u _lP (k) is a proportional term, u lI _(k) represents the integral term.

By controlling the sum of u _a (k) and u _l (k) obtained in the above procedure as a torque command for the motor,
Motor speed control is executed.

As described in detail above, according to the present embodiment, the neural network learns the inverse system of the characteristic of the motor speed with respect to the torque command of the drive system including the motor and the mechanical load. The manipulated variable u _a (k) obtained by inputting the motor speed command value x _r (k) to the neural network is obtained by converting the motor speed x (k) into the speed command value x
_{This is a} feedforward term for controlling to follow _r (k). Since the neural network learns the fluctuation of the moment of inertia of the drive system and the resonance characteristics of the drive system, it can always calculate the optimal feedforward term. Furthermore, a linear control system that performs proportional-integral control based on a motor speed and a speed command value can suppress the effects of disturbance torque acting on the drive system.

In the present embodiment, directly to the neural network has learned the inverse system, the operation amount by inputting the motor speed command value x _r (k) u _a
(K) was calculated. Now, a precompensator for providing an ideal response characteristic of the system is provided, and a motor speed command value is input to the precompensator. The output is input to the neural network and manipulated variable u
By calculating _a (k), the characteristics of the feedforward control loop by the neural network can be optimized. As a characteristic of the pre-compensator, a first-order lag characteristic or the like is used so that the feedforward system becomes stable even with a step-like change in the speed command value.

In the neural network applied to the motor control in the second embodiment described above, only the speed command value _xr (k) is input,
The manipulated variable u _a (k) was calculated.
The configuration of the neural network as shown in the figure may be adopted. That is, this is to calculate the manipulated variable u _a (k) using the motor speed x (k) as an input in addition to the motor speed command x _r (k), and the neural network uses two inputs as shown in FIG. It is composed of one output system. This is the same as FIG. 5 except that the unit of the input layer is increased by one. First, in the learning of the system, the switching switches 7, 8 are simultaneously closed to the side (a). As inputs, the current speed x (k) and the previous speed x (k)
-1) (where x (k-1) = z ^-1 x (k)). Thereby, the motor speed is changed from x (k-1) to x
The torque (k) required to transition to (k) is calculated. Next, after completing the same correction as described above from (k) and u (k), the switches 7, 8 are closed to the (b) side. The input to the neural network is the current speed command value _xr (k)
And the current speed detection value x (k) are input. The operation amount u _a (k) calculated as a result is the current motor speed x (k)
From x to the command value _xr (k). Thus, in addition to the motor speed value x _r (k) speed detection value x (k) also by calculating the amount of operation input to the neural network can be taken into account fed back control characteristics to the neural network side Therefore, it is possible to configure a servo control device having better stability.

In the second embodiment described above, the case where the neural network is used for the motor speed control system is described. Next, a reference example in the case where the neural network is used for the position servo system of the motor will be described.

In the reference example shown in FIGS. 7 to 9, the configuration is the same as that of the speed control system shown in FIG.
3,404 have been added. Here, x ₁ is the detected value of the motor speed, x ₂ represents the detected value of motor position, respectively, also x _r is a target command value, here, the position command value.
In the reference example of FIG. 7, the neural network 2 has a motor speed.
x _1, using the position x ₂ and the torque command value u, to learn the inverse system characteristics of the motor position with respect to a torque command of the motor drive system. Obtaining a manipulated variable u _a by inputting the position command x _r in the neural network. On the other hand, the linear control operation has a configuration in which a speed control unit is provided in a minor loop of the position control unit. Here, proportional control is used for position control, and proportional integral control is used for speed control. The final manipulated variable operation value of linear control is
u _l . As described above, even when the method of the present invention is used in a control system having a speed control system in a minor loop of position control, the configuration can be the same as that of the motor speed control device.

In the above reference example to the position servo, the inverse characteristic of the motor position with respect to the torque command to the drive system was learned by the neural network.

Next, the reference example of FIG. 8 shows a case where feedforward control by a neural network is used for a position control system which is a major loop control system of a position servo system. Here, the same numbers as those in FIG. 7 indicate the same elements. Here, _xr is a position command value,
x ₁ is the speed detected value, x ₂ is the position detection value. Neural network 2
Learns the inverse system of the characteristics of the drive system including the speed control system. That is, to master the inverse system characteristics of x ₂ for u from the command value u and the motor position detection value x ₂ Metropolitan to the speed control system. Neural network calculates the operation amount u _a Enter the position command value x _r to the servo system. On the other hand, in the position control system, the command value
calculating a manipulated variable u _l a x _r and the detected value x ₂ Prefecture. By using the sum of u _a and u _l as the command value to the speed control system, the operation amount that compensates for the deterioration of the speed control characteristics due to fluctuations in the characteristics of the drive system can be calculated by the neural network, so that the position servo with good tracking performance The system can be configured.

Further, the reference example of FIG. 9 shows a case where feedforward control by a neural network is used for a speed control system which is a minor loop control system of the position servo system. In this case, the speed control system is the same as the speed control system in the case where the output of the position control, which is a major loop of the speed control, is set as a target command value (speed command value), and the configuration of the speed control system can be the same as that in FIG. .

In the above embodiments and reference examples, the case where the neural network is used for the speed control system and the position servo system of the motor has been described. However, the current control system of the motor or a system having this as a minor loop control system is described. The same can be applied.

〔The invention's effect〕

According to the method of the present invention, the neural network that has learned the inverse system of the control object calculates the feedforward operation amount with respect to the command value, and the linear control system calculates the feedback operation amount. Therefore, the control system can be stabilized by the linear control system even in the transient region where the neural network learns the characteristic fluctuation of the control target.

Further, in the method of the present invention, the neural network can learn the inverse system of the system independently of the linear control system by using the input / output relationship of the controlled object, so that the feedforward control by the linear control system and the neural network can be performed. There is an advantage that the system can be designed independently.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of the present invention, FIG. 2 is a flowchart showing the processing contents of FIG. 1, FIG. 3 is a block diagram of a conventional control method, and FIG. FIG. 5 is a detailed diagram of the neural network in FIG. 4, FIG. 6 is a diagram of the neural network in another embodiment, and FIGS. 7 to 9 are diagrams of the reference example. . 1 ... Control target, 2 ... Neural network, 3 ... Output error calculator of neural network, 4 ... Linear control calculator.

Continuation of front page (56) References JP-A-2-54304 (JP, A) JP-A-2-93708 (JP, A) JP-A-1-250103 (JP, A) JP-A-63-214801 (JP) , A) Toru Setoyama, et al., "Manipulator Control by Inverse Dynamics Model Learned in Multilayer Neural Network", IEICE Technical Report, IEICE, 1988 No. 87, No. 428, P.M. 249-256 Demetri Psaltis, 2 others, “A Multilayered Neural Network Controller”, IEEE CONTROL SYSTEMS MAGAZINE, IEEE CONTROL SYSTEMS. 8, No. 2, April 1988; 17-21 Hiroaki Gomi and one other, "Closed-loop learning control of unstable system by feedback error learning", IEICE Technical Report, IEICE, March 16, 2002 Vol. 89, No. 463, pp. 7-12 (58) Fields investigated (Int. Cl. ⁶ , DB name) G05B 13/00-13/04 G06F 15/18 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A neural network for learning an inverse system of a control target from an input and an output to the control target, and a linear control means for executing a linear control from a target command value to a servo control system and an output of the control target. The neural network calculates the output value for learning by inputting the output of the control object at one time point and the previous time point, and calculates the output value and the operation amount input to the control object. Learning the inverse system based on, the next, input the target command value at the one time point and the output at the one time point to calculate a first operation amount, the first operation amount, A servo control device, wherein a sum with a second operation amount obtained from the linear control means is controlled as an operation amount for a control target.