JPH05324013A - System modeling method - Google Patents

Abstract

PURPOSE:To adjust a membership function corresponding to variation in characteristics of an object system by individually correcting networks according to the cause of error occurrence when the output of a system model and the output of the controlled system have an error. CONSTITUTION:A fuzzy room when it is once given determines the constitution of respective subordinate networks and stores and infers N fuzzy rules on the whole. An antecedent network part 2a consists of (m) networks which are as many as dimensions (m) of an input variable (x). A consequent network 2c consists of (n) networks and when a fuzzy rule is given, the structure of a system model is automatically determined. In initial learning, knowledge regarding an object system 1 is given in the form of the fuzzy room and a network for input/output data clustering is learnt. Then a consequent network 2c performs a learning process. The antecedent network 2a and an inference network 2b are connected to generate a network for classification.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system modeling method used in designing and operating a controller in a control system, and more particularly to a system modeling method in which a fuzzy modeling method and a modeling method using a neural network are integrated.

[0002]

2. Description of the Related Art Generally, system modeling is an important technique when designing a control system. In such system modeling, the method of describing the target system by a mathematical formula was the mainstream, but in recent years, a fuzzy modeling method for describing the system by fuzzy rules,
The modeling method by the neural network which makes the neural network learn the input and output data is proposed, and those methods are also diversified.

Under such circumstances, a combination of a fuzzy modeling method and a method using a neural network has been proposed, and it is expected that a modeling method having both features will be obtained.

[0004]

However, the conventional method combining the fuzzy modeling and the neural network cannot perform the system modeling adapted to the change of the input / output characteristics of the target system, so that the estimation is performed with high accuracy. There was a problem that I could not.

The present invention has been made in consideration of such a conventional situation, and is capable of performing system modeling adapted to changes in the input / output characteristics of the target system and capable of performing accurate estimation. It is intended to provide a method.

[0006]

That is, a system modeling method according to the present invention is a system modeling method for describing input / output characteristics of a target system by a fuzzy rule, and includes a plurality of antecedent networks, an inference network, and a plurality of inference networks. And a consequent network, each of the antecedent networks roughly storing a plurality of membership functions based on existing knowledge of the input to the target system, and the inference network including each of the members. A logical relationship between a ship function and an input space that is vaguely divided by the membership function is stored, and the input / output data of the target system prepared in advance is used as the output of the antecedent network as the input to the inference network. By connecting to a classification network connected so that I / O data classified by each input space that is divided ambiguously, and input / output data classified by each input space that is ambiguously divided are stored in each consequent network, and each of the above based on the output of the classification network. Based on a deviation between an output obtained by giving an input in the input / output data to the system model and an output in the input / output data, the system model is configured by connecting the outputs of the consequent network to synthesize the output. In the system model, the output value to be output by the antecedent network is obtained, the antecedent network is modified to output the output value, and the input / output characteristics of the target system are stored. When an error occurs in the output of the system model, the output of the classification network for the input of the system model corresponding to the output, Input of the serial system model, in the ambiguously classified input space,
Divide into a first case that is mostly contained in a certain space and a second case that is ambiguously contained in multiple spaces,
In the case of the first case, the specific input is selected so that the output of the input / output data is converted into the output of the consequent network based on the output of the classification network to obtain a desired output. It is desirable that the consequent network corresponding to the space is modified, and in the case of the second case, the output in the input / output data is converted into the output of the inference network based on the output of the classification network. It is characterized in that the inference network is modified so that the output becomes.

Further, the system modeling method according to claim 2 further obtains an output of the antecedent network that obtains an output in the input / output data, and obtains the output. Is fine-tuned, and the system model is adaptively modified according to the change of the input / output characteristics of the target system.

[0008]

In the system modeling method of the present invention having the above configuration, when constructing the model of the target system, first, knowledge about the target system is given in the form of fuzzy rules, and the antecedent network and the inference network are learned, Create a network for input / output data clustering.

Next, using the input / output data of the target system clustered by this data clustering network, the consequent network is learned to construct a system model.

If there is an error between the output of the system model and the output of the target system, the networks that cause the error are individually corrected according to the error. Thereby, the adaptive system modeling method can be realized by adjusting the membership function using the input / output data of the target system in response to the change in the characteristic of the target system.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the system modeling method of the present invention will be described in detail below with reference to the drawings.

FIG. 1 shows the configuration of an embodiment of the present invention. In FIG. 1, 1 is a target system with m inputs and r outputs, 2
Stores the input / output characteristics of the target system 1
The system model 3, which functions as an estimation model of, is an error cause estimator that estimates the network that is the cause of the error from the output error. Also, system model 2
Determines the output of the system model 2 by linear combination of the output of the antecedent network unit 2a, the inference network unit 2b, the antecedent network unit 2c, the inference network unit 2b, and the output of the antecedent network unit 2c. And a linear coupling section 2d for

In this embodiment, the target system 1 is described by the system model 2, and when an error occurs between these outputs, the error cause estimator 3 causes the system model 2 to be described.
A case will be described in which the target system 1 is adaptively described by estimating the network that causes the error and then individually correcting the network that causes the error.

First, the system model (FNN) 2 will be described.

FIG. 2 shows a detailed structure of the system model 2. The fuzzy rules R ¹ , ..., R ⁿ handled in this system model have the following format.

[0016]

[Equation 1] The fuzzy rule expressed by the equation (1) is characterized in that the consequent part is not a fuzzy proposition, but the output is given in the form of a function, and is excellent in the description ability of the system.

Given the fuzzy rules, the configuration of each sub-network is largely determined. That is, N fuzzy rules are stored and inferred as a whole. The antecedent network unit 2a has a number of m equal to the dimension m of the input variable x.
, INN ^m , the antecedent network (INN ¹ , ..., INN ^m ), and the consequent network section 2c includes n consequent network (TN) that is equal in number to the number n of functions of the consequent section.
N ¹ , ..., TNN ⁿ ). Therefore, when the fuzzy rules R ¹ , ..., ^RN are given, the structure of the system model is automatically determined. Here, all the sub-networks are of the so-called BP (Error Back Propagation Learning) type.

Next, the input / output characteristics of the system model will be described. The input / output characteristic is given by the following equation.

[0019]

[Equation 2] Next, a method of constructing a system model will be described.

The initial learning is performed by the following procedure.

First, knowledge about the target system is given in the form of fuzzy rules as shown below (step 1).

[0022]

[Equation 3] Next, a network (temporary learning network) for input / output data clustering of the target system is learned (step 2).

[0023]

[Equation 4] Next, the consequent network is learned (step 3).

That is, first, the antecedent network IN
N ¹ , ..., INN ^m and the inference network RNN are connected for classification (for clustering) as shown in FIG.
Create a network CNN. Then, x of the input / output data (x, y) of the target system is input to the classification network to receive w. Weights w _j ≧ epsilon become data, learning data for the consequent network TNN ^j (x ^j,
y ^j ). Then, the variation of the extracted learning data is adjusted.

Based on the learning data (x ^j , y ^j ) extracted by clustering, the consequent network TNN ^j ,
Learning j = 1, ..., N (BP method). The consequent network TNN ^j has m × X _T × r dimensions. Here, the dimension X _T of the intermediate layer is experimentally determined.

Next, the entire system is constructed (step 4).

The classification for the network, the consequent network TNN ^1, ..., and connect the TNN ^n, constituting the entire system as shown in FIG. 2 (system model).

Then, all the input / output data (x, y) of the target system are used as learning data, and the antecedent network I
Further learning (adjustment of membership function) of NN ¹ , ..., INN ^m is performed to improve the inference accuracy as a whole.

Antecedent network INN ¹ , ..., INN
The additional learning (adjustment of the membership function) is not considered to be correct because the inference network RNN is correct in learning the entire system model.

Consequent network TNN ¹ , ..., TNN
Do not correct because ⁿ is correct.

It is realized by assuming that This procedure is shown below.

First, by solving the inverse problem of the inference network RNN, the antecedent network INN ¹ , ..., I
Calculate the membership grade (shown below) that NN ^m should output (step 1).

[0033]

[Equation 5] Next, using the obtained membership grade as learning data, the antecedent network INN ¹ ,
…, Correct INN ^m .

Next, the adaptive system modeling method will be described.

In FIG. 1, it is assumed that the input / output characteristics of the target system 1 are completely described by the system model 2. When an error occurs between the output of the target system 1 and the output of the system model 2 due to the change in the input / output characteristics of the target system 1, this system model 2 has the antecedent part membership function-the antecedent part network and the consequent part. It has a one-to-one correspondence network such as function-consequent network, inference method-inference network. Therefore, the system model 2 is modified by the following procedure so that the input / output characteristics of the target system 1 are completely described again.

Step 1 In the error cause estimator 3, the classification network CNN is used.
Based on the outputs W1, ..., Wn of the input, the input x to the system model is divided into case 1 and case 2 shown below.

Case 1: When the input is mostly contained in a specific division space IS ⁱ .

Here, the inputs included in the specific divided space IS ⁱ are the outputs W1, ..., Wn of the classification network.

[Equation 6] Is the input.

Case 2: When a plurality of divided spaces are ambiguously included (when they are included in the boundary area, that is, when they are not Case 1).

Step 2 Modify the network according to the classification of the input.

In case 1, input / output data (x, y)
The output y is the output of the classification network W1, ..., W
The consequent network is learned by the BP method so that the output of the consequent network corresponding to the divided space IS ⁱ corresponding to n becomes a desired output.

In case 2, input / output data (x, y)
The output y is the output of the classification network W1, ..., W
The inference network is learned by the BP method so that the output converted from the output of the inference network according to n becomes a desired output.

Step 3 Further, as fine adjustment, the output W1 of the classification network,
, Output of the output y of the input / output data (x, y) converted to the output of the inference network according to Wn,
The antecedent network is learned by the BP method so as to solve the inverse problem of the inference network, find the corresponding input, and make it the desired output of the antecedent network.

As described above, according to the system modeling method of this embodiment, the system modeling adapted to the change of the input / output characteristic of the target system 1 can be performed, and the accurate estimation can be performed. In addition, it has a structure that reflects the structure of the rules, and the knowledge of the expert can be utilized to the maximum.
Furthermore, a clustering method for input / output data is unnecessary.

Although the above embodiment is a system modeling method using the input / output data of the target system online, offline adjustment for modeling the system using the input / output data prepared in advance is also possible. Is.

[0046]

As described above, according to the system modeling method of the present invention, system modeling adapted to changes in the input / output characteristics of the target system can be performed.
Accurate estimation can be performed.

[Brief description of drawings]

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

FIG. 2 is a diagram showing a configuration of a system model according to an embodiment of the present invention.

FIG. 3 is a diagram showing a configuration of a classification network according to an embodiment of the present invention.

[Explanation of symbols]

 1 Target system 2 System model 2a Antecedent part network part 2b Inference network part 2c Consequent part network part 2d Linear combination part 3 Error cause estimator

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Internal reference number FI technical display location G06G 7/60

Claims (2)

[Claims] 1. A system modeling method for describing input / output characteristics of a target system by a fuzzy rule, comprising a plurality of antecedent networks, an inference network, and a plurality of antecedent networks, each antecedent network , Roughly stores a plurality of membership functions based on existing knowledge of inputs to the target system, and the inference network includes each membership function and inputs that are vaguely divided by the membership function. By storing the logical relationship of the space and giving the input / output data of the target system prepared in advance to the classification network connected so that the output of the antecedent network becomes the input to the inference network, the ambiguous The input space is divided into The input / output data classified for each input space is stored in each consequent network, and the output of each consequent network is connected based on the output of the classification network. The input value in the input / output data is given to the system model, and the output value to be output by the antecedent network is obtained based on the deviation between the output obtained in the input / output data and the output in the input / output data. If the antecedent network is modified to output an output value and the system model is configured to store the input / output characteristics of the target system, and if an error occurs in the output of the system model, the output corresponding to the output According to the output of the classification network for the input of the system model, the input of the system model is the fuzzy classified input. In the space, it is divided into a first case that is almost included in a specific space and a second case that is ambiguously included in a plurality of spaces. In the case of the first case, The output of the consequent network is converted based on the output of the classification network so that a desired output is obtained, and the consequent network corresponding to the specific input space is modified. In the second case, the inference network is modified so that the output of the input / output data is converted into the output of the inference network based on the output of the classification network to obtain a desired output. Modeling method characterized by. 2. The system modeling method according to claim 1, further comprising obtaining an output of the antecedent network that obtains an output in the input / output data, and obtaining the output. Is corrected and finely adjusted, and the system model is adaptively corrected in accordance with the change of the input / output characteristics of the target system.