KR20000051941A - Real time learing fuzzy controller - Google Patents

Abstract

PURPOSE: A real time learning controller is provided to construct an optimum controller for a changed system while performing identification between real time control and real time plant on line without constructing any off-line data base even though a parameter of the system is changed and improve the learning speed by using a neural emulator. CONSTITUTION: First and second comparators(10,20) compare a target speed and a real speed of an induction motor(1) to output error items. A time delay part(30) delays the error item output from the first comparator. A third comparator(40) compares the error item of the first comparator with the delayed item for outputting a change amount of the error item. A self-learning fuzzy controller(50) outputs a reference current value of torque component for real time plant identification to control the speed of the induction motor according to the error items output from the first and third comparators. A fourth comparator(60) compares an estimation speed with the real speed for outputting an error item. A neural emulator(70) outputs the estimation speed according to the reference current value and the error items of the first and third comparators while changing the estimation speed closely to the real speed, and controls error of the reference current value of the torque component.

Description

REAL TIME LEARING FUZZY CONTROLLER}

The present invention relates to a real-time learning type controller, and more particularly, to a real-time learning type controller applicable to a time varying system.

In particular, the present invention relates to a speed controller of an induction motor applied to the speed control of an induction motor as a real time learning type controller applicable to a time varying system.

Induction motors with severe nonlinear characteristics are difficult to select the control gain when using the PI controller and to obtain satisfactory control characteristics in case of parameter variation even at the optimum gain. have.

Fuzzy theory is applied in the real industrial field because it is relatively simple to implement even in the case that it is difficult to solve by the existing control method or the system model is complicated.

However, the fuzzy controller has a problem in that a general rule for determining control rules, membership functions, normalization and denormalization constants is not organized, and trial and error controller design is dependent on expert intuition or experience.

Due to these problems, efforts have been made to make up for the shortcomings by combining fuzzy theory and neural network theory since the late 1980s.

The similarity between fuzzy inference and neural network is that the min-max operation of fuzzy inference is equivalent to the product-sum operation of neural network, and also the sigmo The function deduces complex nonlinearities throughout the system.

And while fuzzy reasoning deals with logic, neural networks have a complementary relationship with learning.

The conventional PID controller having a fixed gain has a problem in that the gain cannot be modified by itself when a parameter of the system is changed, and when the gain is changed, a database must be offline beforehand.

In order to solve the problems of the conventional off-line learning type fuzzy-neural network control system, the present invention enables real-time control and identification of a real-time plant (eg induction motor), learning speed is fast, and transient characteristics during learning. We want to implement this excellent real-time learning type controller.

In particular, the self-learning fuzzy control system of the present invention is composed of a fuzzy neural network controller for speed control and a neural network emulator for simulating an induction motor system.

In order to use the fuzzy-neural network as an on-line real-time controller, an error term at the controller output stage must be calculated in real time at every sampling time. For this purpose, the input / output relationship of the plant output stage is simulated and the error term of the plant output stage is reversed. It can be configured as a neural network emulator.

In addition, the real-time speed control of the induction motor requires a short learning time, excellent transient characteristics during learning, and a simple structure controller. In the present invention, a neural network emulator for real-time plant identification for speed control of an induction motor is possible. A self-learning fuzzy controller is used to learn the membership function and control rules of fuzzy controller in real time.

The present invention is a method of applying the learning theory of neural networks and reconfiguring the optimal controller to the changing system when the parameters of the system change.

The conventional PID controller having a fixed gain has a problem in that the gain cannot be modified by itself when a parameter of the system is changed, and when the gain is changed, a database must be offline beforehand.

In order to solve this problem, the neural network emulator can be configured by identifying the plant on-line and configuring the optimal controller for the changed system by itself, without having to build the database offline even if the system parameters change. It provides a real-time learning type controller that can learn the neural network controller in real time to enable real-time plant identification, and can also incorporate fuzzy theory to add robustness.

1 is a block diagram of an embodiment in which the real-time learning type controller of the present invention is applied to a speed controller of an induction motor.

Figure 2 is a flow chart showing the operation of the embodiment applied to the speed controller of the induction motor of the real-time learning type controller of the present invention.

3 is a view showing the structure of a neural network emulator (Neural Emulator) applied to the present invention.

Figure 4 is a view showing the structure of a self-learning fuzzy controller (Self-Learning Fuzzy Controller) applied to the present invention.

1 is a block diagram of an embodiment in which the real-time learning type controller of the present invention is applied to a speed controller of an induction motor.

The first and second comparators 10 for comparing the target speed {Wre (k)} and the actual speed {W (k)} of the induction motor 1 and outputting an error term x1 {e (k)}. 20;

A time delay unit (30) for delaying the error term (x1) compared and output by the first comparator (10);

A third comparator 40 comparing the error term x1 output from the first comparator 10 with a previous error term delayed by the time delay unit 30 and outputting a change amount x2 of the error term x1; ;

The current reference value of the torque component {u (k) to enable real-time plant identification for speed control of the induction motor 1 according to the error terms (x1) (x2) output from the first and third comparators 10 and 40. Self-Learning Fuzzy Controller (50) for outputting;

A fourth comparator 60 comparing the estimated speed {y ^ (k)} and the actual speed {W (k)} to simulate driving of the induction motor 1 and outputting the error term {e ^ (k)} Wow;

The induction motor 1 according to the error term (x1) (x2) output from the first and third comparators 10 and 40 and the current reference value {u (k)} of the torque component output from the self-learning fuzzy controller 50. Outputs an estimated speed {y ^ (k)} that simulates the driving, and calculates the estimated speed {y ^ (k)} according to the error term {e ^ (k)} compared by the fourth comparator 60. Variable to be similar to the actual speed {W (k)} of the induction motor 1,

Self-learning purge the error term {e (k)} compared by the second comparator 60 so that the actual speed {W (k)} of the induction motor 1 is made equal to the target speed {Wref (k)}. A neural network emulator 70 which backpropagates to the controller 50 and controls the torque component current reference value {u (k)} error of its output; It consists of.

Here, as shown in FIG. 4, the self-learning fuzzy controller 50 learns the membership function and the control rule in real time while changing the connection strength Wj using the neural network emulator 70. It is composed of an input layer (P) corresponding to the fuzzy, a hidden layer (Q) of a plurality of {Aj (x1)} {Bj (x2)} constituting the control rule, and an output layer (R) corresponding to the non-fuging. In addition, the input layer (P) is composed of a vertical type output layer (R) of a singleton type.

As shown in FIG. 3, the neural network emulator 70 has connection strengths Wi1, Wi2, and Wj1 in a direction of decreasing an error term {e (k)} compared by the second comparator 60. (Wj2) (Wj3) is adjusted and back propagated in the upper layer, and the connection strength Wj of the self-learning fuzzy controller 5 is adjusted again based on the lower layer.

This means that the error term x1 (x2) output from the first and third comparators 10 and 40 and the current reference value {u (k)} of the torque component output from the self-learning fuzzy controller 50 are input. One output layer neuron that outputs an input layer neuron (L), an estimated speed {y ^ (k)} that simulates the actual velocity {W (k)} of the five hidden layer neurons (M) and the induction motor 1 ( N).

2 is a flowchart showing the operation of the embodiment applied to the induction motor speed controller of the real-time learning type controller of the present invention.

First, initializing the entire system of the real-time learning type controller and preliminarily learning the neural network emulator 70;

When the pre-learning of the neural network emulator 70 is completed by the above steps, an arbitrary value is applied as the output of the self-learning fuzzy controller 50 to output the estimated value of the neural network emulator 70 and the induction motor 1 at each sampling time. Calculating an error term in a direction that minimizes the output error

After adjusting the connection strength of the neural network emulator 70 according to the error term calculated by the above step, the self-learning fuzzy controller 50 back propagates the error term between the target speed and the actual speed of the induction motor 1 and performs self-learning. Calculating an output error term of the fuzzy controller 50;

Repeating the entire process by learning and returning a control rule and a membership function of the self-learning fuzzy controller 50 according to the output error term calculated by the step; It is going to be.

Referring to Figures 1 to 4 attached to the operation of the embodiment applied to the induction motor speed controller of the real-time learning type controller of the present invention configured as described above are as follows.

First, after initializing the entire system of the real-time learning type controller, the neural network emulator 70 is preliminarily trained.

That is, the neural network emulator 70 is applied in a direction to minimize an output error of the neural network emulator 70 and the induction motor 1 at every sampling time by applying an arbitrary value as the output of the self-learning fuzzy controller 50. We will proceed to preliminary learning.

Subsequently, when the preliminary learning is completed, the self-learning purge of the error term x1 comparing the target speed {Wre (k)} and the actual speed {W (k)} of the induction motor 1 in the first comparator 10. Output to the controller 50 and the neural network emulator 70,

The error term x1 output by the first comparator 10 is compared with the previous error term delayed through the time delay unit 30 through the third comparator 40, and then the amount of change (x2) of the error term x1. ) Is output to the self-learning fuzzy controller 50 and the neural network emulator 70.

At this time, the self-learning fuzzy controller 50 calculates the current reference value {u (k)} of the torque component for driving the induction motor 1 according to the error term (x1) (x2) inputted in comparison with the induction motor (1). And output to the neural network emulator 70,

In the induction motor 1, the induction motor 1 is driven according to the current reference value u (k)} of the torque component.

In addition, in the neural network emulator 70, an error reference value x 1 (x 2) output from the first and third comparators 10 and 40 and a current reference value of the torque component output from the self-learning fuzzy controller 50 {u According to (k)}, the learning of the emulator is terminated by outputting the estimated speed {y ^ (k)} in the direction of minimizing the error with the actual speed {W (k)} of the induction motor 1.

In other words, after comparing the estimated speed {y ^ (k)} and the actual speed {W (k)} to simulate the driving of the induction motor 1, the error term {e ^ (k)} is compared with the fourth comparator 60 ) Is output from

In the neural network emulator 70, the estimated speed {y ^ (k)} is calculated according to the error term {e ^ (k)} compared by the fourth comparator 60, and the actual speed {W ( k)}.

In more detail, the neural network emulator 70 uses a delta rule generalized to minimize the error between the actual speed {W (k)} and the estimated speed {y ^ (k)} of the induction motor 1. Driven bar,

The neural network emulator 70 has connection strengths Wi1 (Wi2) (Wj1) (Wj2) to reduce the slope of the error term {e ^ (k)} compared and output by the fourth comparator 60. By adjusting (Wj3), it is calculated while varying the error term of one output layer neuron (N) outputting the estimated speed {y ^ (k)} which simulates the actual speed {W (k)} of the induction motor 1.

In the neural network emulator 70, since the map output values are not known in the case of the five hidden layer neurons M, the error terms of the hidden layer neurons M are used by using the known error terms in the output layer neurons N. Calculate

For faster learning speed, the momentum is added and weighted to complete the learning.

Then, when the learning of the neural network emulator 70 is completed, the target speed {Wref (k)} and the actual speed {W (k)} of the induction motor 1 for the learning of the self-learning fuzzy controller 50. Is compared in the second comparator 20 and the error term {e (k)} is input to the neural network emulator 70.

In the neural network emulator 70, the error term {e compared by the second comparator 60 such that the actual speed {W (k)} of the induction motor 1 is equal to the target speed {Wref (k)}. (k)} is backpropagated to the self-learning fuzzy controller 50.

At this time, the self-learning fuzzy controller 50 calculates the torque component current reference value {u (k)} error term of the output according to the error term {e (k)} which is back propagated.

That is, the self-learning fuzzy controller 50 corresponds to the fuzzy to learn the membership function and the control rule in real time while changing the connection strength Wj using the neural network emulator 70 as shown in FIG. 4. The input layer P is composed of a plurality of {Aj (x1)} {Bj (x2)} hidden layers Q constituting a control rule, and an output layer R corresponding to a non-fuge.

By obtaining the belonging function value of the fuzzy variable for the error term (x1) (x2) input to the input layer (P) by the first and third comparators (10) and (40), and passing it to the hidden layer (Q), The hidden layer Q calculates this by multiplication {Aj (x1)} × {Bj (x2)} and gives a weight Wj to the output layer R to perform the defocusing by the center of gravity method. will be.

Therefore, the control rule and the belonging function learning of the self-learning fuzzy controller 50 are performed in real time through the error term of the backpropagation calculated according to the non-fuzzy operation performed in the output layer R.

Real-time plant identification is possible for speed control of the induction motor 1.

As described above, the present invention provides a neural network emulator for simulating a fuzzy-neural network controller and an induction motor system for speed control, and a real-time learning type controller for learning a neural network controller in real time using the neural network emulator. In addition, it has the effect of real-time learning type controller that enables the real-time control and plant identification, speeds up the learning speed, and has excellent transient characteristics during learning.

Claims (3)

First and second comparators for comparing a target speed and an actual speed of the induction motor and outputting an error term; A time delay unit for delaying the error terms compared and output by the first comparator; A third comparator for comparing an error term output from the first comparator with a previous error term delayed by the time delay unit and outputting a change amount of the error term; A self-learning fuzzy controller for outputting a current reference value of a torque component to enable real-time plant identification for speed control of an induction motor according to the error terms output from the first and third comparators; A fourth comparator for comparing the estimated speed to simulate the driving of the induction motor and the actual speed, and outputting an error term; According to the error terms output from the first and third comparators and the current reference value of the torque component output from the self-learning fuzzy controller, an estimated speed that simulates driving of the induction motor is output, and estimated according to the error terms compared by the fourth comparator. The speed is varied to be similar to the actual speed of the induction motor, A neural network emulator for propagating the error term compared by the second comparator to a self-learning fuzzy controller to control the torque component current reference value error of its output such that the actual speed of the induction motor is equal to the target speed; Real-time learning type controller, characterized in that consisting of. 2. The self-learning fuzzy controller according to claim 1, wherein the self-learning fuzzy controller uses an neural network emulator to change the connection strength while learning the belonging function and the control rule in real time, and a plurality of hidden layers constituting the control rule. Real-time learning type controller, characterized in that configured as an output layer corresponding to the de-fuzzy. 2. The neural network emulator of claim 1, wherein the neural network emulator adjusts the connection strength in a direction of reducing the error term compared by the second comparator and back propagates in the upper layer to adjust the connection strength of the self-learning fuzzy controller based on the lower layer. so, Outputs an estimated speed that simulates the actual speeds of three input layer neurons, five hidden layer neurons, and induction motors, into which the error terms output from the first and third comparators and the current reference value of the torque component output from the self-learning fuzzy controller are input. Real-time learning type controller, characterized in that composed of one output layer neurons.