JPH06186064A - System modeling device - Google Patents

Abstract

PURPOSE:To provide a system modeling device which can accurately describe the identification model for the input and output characteristics of a subject system throughout an input space. CONSTITUTION:A system modeling device 11 is for describing the identification model for the input and output characteristics of a subject system 10 using an IF-THEN type fuzzy rule and includes an antecedent computing portion 12 for calculating adaptation w1,..., wN related to the antecedent of the fuzzy rule for the input (x) of the subject system 10; a consequent computing portion 13 for outputting function vales y1,..., yN of the consequent of the fuzzy rule for the input (x); a rule synthesis portion 14 adapted for synthesizing a rule according to the outputs of the antecedent computing portion 12 and the consequent computing portion 13 and comprising a BP type neural network for giving the identification result for the output (y) of the subject system 10; and a correcting portion 15 for correcting the coupling load of the network of the rule synthesis portion 14 so that the identification error (epsilon) of the rule synthesis portion 14 decreases.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system modeling apparatus for describing an identification model of input / output characteristics of various controlled systems, and particularly to identify input / output characteristics by using an IF-THEN fuzzy rule. The present invention relates to a system modeling device.

[0002]

2. Description of the Related Art A system modeling device describes the input / output characteristics of various systems as an identification model, and is useful when designing and operating a controller in a control system. As such a system modeling device, in recent years, a system modeling device describing an identification model by a fuzzy rule that can quantify ambiguous human knowledge and an identification model by a neural network that can learn and store the input / output characteristics of the system. Furthermore, a system modeling device has been proposed in which fuzzy rules are realized by a neural network and the features of both are combined.

In such a fuzzy rule or an identification model in which a fuzzy rule and a neural network are combined, a plurality of IF-THEN type fuzzy rules are generally used to describe the input / output characteristics of the target system. Specifically, in the conventional system modeling device of this type, the antecedent part computing part storing the antecedent part of the IF-THEN format fuzzy rule computes the suitability of the antecedent part of the fuzzy rule to the input value. At the same time, the consequent part calculator calculates the function value of the consequent part of the fuzzy rule with respect to the input. Then, the matching result obtained by the antecedent part computing part is used as a weighting coefficient, and the calculation result of the consequent part computing part is weighted and added to obtain the identification result of the output of the target system.

[0004]

The above-mentioned IF-
In the system modeling device using the THEN fuzzy rule, the input space of the target system is divided into a plurality of partial input spaces by the antecedent part membership function in the antecedent part arithmetic part. Each of these partial input spaces is associated with an individual fuzzy rule, and each fuzzy rule is constructed to describe the input / output characteristics of the target system in the corresponding partial input space. However, in the vicinity of the boundary of the sub-input space, the suitability of the antecedent part of any rule becomes low, and even in such a case, if the weighted average of the operation results of the antecedent part is obtained to obtain the identification result, May give inaccurate results.

It is an object of the present invention to provide a system modeling device capable of accurately describing the input / output characteristics of the target system for any input within the input space of the target system.

[0006]

In order to solve the above problems, the present invention is an IF-THEN comprising an antecedent part and a consequent part.
In a system modeling device for describing an identification model of input / output characteristics of a target system by a fuzzy rule of a format, an antecedent part computing means for performing an operation on the antecedent part based on an input of the target system, and based on the input A consequent part computing means for performing a computation on the consequent part, a rule synthesizing part for obtaining an identification result of the output of the target system based on the computation results of the antecedent part computing part and the consequent part computing part, and the target Correction means for correcting the rule synthesizing means so as to reduce the error of the identification result with respect to the output of the system.

More specifically, the rule synthesizing means is, for example, a back-propagation type neural network, and the correcting means is configured to correct the connection weight in the neural network.

[0008]

As described above, according to the present invention, by modifying the rule synthesizing means so that the error of the identification result becomes small, it is possible to perform the rule synthesizing operation having higher flexibility than the mere weighted average. Even if the input value does not conform to the fuzzy rule as in the case where the input value is included near the boundary of the sub-input space, the input / output characteristics of the target system can be accurately identified. Also,
It is also possible to additionally adjust the rule synthesizing means after once determining the identification model.

[0009]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing the configuration of a system modeling apparatus according to an embodiment of the present invention.

In FIG. 1, a target system 10 is a target system for which an identification model of input / output characteristics should be described by the system modeling device 11, and in this example, it has m-dimensional inputs and one-dimensional outputs. The system modeling device 11 stores the input / output characteristics of the target system 10 and identifies the target system 10.
2, consequent part operation part 13, rule composition part 14, correction part 15
And an error calculator (subtractor) 16. Fuzzy rules R1, ..., RN handled by this system modeling device
Is the IF-THEN format as shown in equation (1).

[0011]

[Equation 1]

In such an IF-THEN format fuzzy rule, the part after If and before the then is called the antecedent part, and the part after the then is the consequent part. If the input value is "speed", for example, the fuzzy label represents "fast", "slightly fast", "slow", etc. by a membership function. The fuzzy rule shown in equation (1) is
Since the consequent part is not a fuzzy proposition, but the output value y is given in the form of a function of the input value x, it is characterized in that the input / output characteristics of the target system are excellent.

The antecedent part computing part 12 finds the minimum value of the antecedent part membership function value of each fuzzy rule with respect to the input x, and determines the minimum value of the antecedent part's suitability w1, w2 for the input value.
,, wN. The consequent operation unit 13 applies the fuzzy rules R1 and R to the input value x.
2, ..., Functions f1 (x), f2 (x), ..., Of the consequent part of RN
It is configured to output the values y1, y2, ..., YN of fN (x). The rule synthesizing unit 14 inputs 2N in this example.
Consistency parts w1, w2, ..., WN obtained by the antecedent operation unit 12 and the consequent operation unit 13 obtained by the antecedent operation unit 12 and configured by a 1-output BP (back propagation) neural network. Function values of y1, y
The inference result of the output value y is obtained from 2, ..., yN.

The correction unit 15 corrects the network connection weight in the rule synthesis unit 14 by the BP method so as to reduce the error (identification error) ε of the identification result with respect to the output of the target system 10 obtained by the error calculation unit 16. Amount Δ
W1, ΔW2, ..., ΔWN are calculated, and the rule synthesizing unit 14 is corrected based on the calculated values.

Here, the BP type neural network will be described. As is well known, the BP type network is a hierarchical type network composed of units having a first-order differentiable output function, and the structure of a K layer BP type neural network is shown in FIG. When the number of units in the k-th layer is Lk, the output of the l-th unit in the k-th layer (k ≧ 2) is given by equations (2), (3) and (4). In the first layer, the input is output as it is as shown in equation (5). In this BP neural network, signals are transmitted in one direction from the first layer to the Kth layer.
Further, the units in the same layer operate in parallel in synchronization.

[0016]

[Equation 2]

The network learning algorithm based on the BP method is widely used and has been described in many documents, so a detailed description thereof will be omitted here. What is the BP method?
Network output when input data (Equation (6)) is given (Equation (7)) and corresponding output data (Equation (8))
This is a learning method that determines the correction amount Δw of the connection weight of the network so as to reduce the evaluation function shown in equation (9) based on the error between and.

[0018]

[Equation 3] Next, the operation of the system modeling apparatus of this embodiment will be described with reference to FIG.
This will be described with reference to the flowchart shown in FIG.

First, in the antecedent operation unit 12, each fuzzy rule R1, R2, ...
.., wN are calculated and the fuzzy rules R1, R1 and
Function values y1, y2, ..., yN of the consequent part of R2, ..., RN
Is calculated (S1).

Next, the rule synthesizing unit 14 is connected to the antecedent operation unit 1.
, WN and the function values y1, y2, ..., yN obtained by the consequent operation unit 13 are input to identify the output value y of the target system 10. Is output (S2). Next, the error ε of the identification result obtained in S2 with respect to the output value y of the target system 10 is obtained by the error calculator (subtractor) 16 (S3).

Next, the correction unit 15 evaluates whether or not the identification error ε is sufficiently small. If the identification error ε is small, the rule composition unit 14 is not corrected, and the antecedent operation unit 12 and consequent operation unit at the present time are evaluated. The calculation unit 13 and the rule synthesis unit 14 are used as the identification model of the target system 10, and if not, the process proceeds to S5.

In S5, the rule synthesizing unit 14 that reduces the identification error ε based on the BP method in the modifying unit 15 is used.
Then, the correction amount ΔW of the connection weight W of the network at is obtained, and the rule synthesis unit 14 is corrected by this,
After that, the process returns to S2.

In this way, an accurate identification model of the target system 10 can be constructed by learning the rule synthesizing unit 14 which is a neural network based on the input / output data of the target system 10 by the BP method.

As another embodiment of the present invention, FIG. 4 shows a structural example of a rule synthesizing unit different from the rule synthesizing unit 14 of FIG. This rule synthesizing unit is composed of N 1-input / 1-output neural networks 211, 212, ..., 21N and integrating units 221, 222 ,.

In this case, the goodness of fit w1, of the fuzzy rules R1, R2, ..., RN output from the antecedent operation unit 12
w2, ..., wN are neural networks 211, 21
, ..., 21N, and the corrected values s1, s2, ..., sN are used as the final goodness of fit. .., 22N, and the consequent part computing unit 1 and the adaptability s1 ,.
, YN corresponding values of the function values y1, y2, ...
22N outputs are added to obtain the identification result of the output value y.

Here, the neural networks 211, 212, ..., 21N are individually trained by using the input / output data of the target system 10 given from the outside, so that accurate identification can be performed as in the previous embodiment. A model is obtained.

According to this embodiment, it is possible to construct an accurate identification model, and it is possible to take out the goodness of fit s1, ..., sN of each fuzzy rule after modification. Therefore, it is possible to selectively correct only the rule having a high degree of conformity, and there is an advantage that the time required to correct the rule synthesizing section can be shortened. The present invention is not limited to the above-described embodiments, but can be implemented with various modifications as follows.

(1) In the embodiment, the antecedent part computing part 12 obtains the minimum value of the antecedent part membership function value of each fuzzy rule, and these are the fitness w of each rule to the input value.
1, w2, ..., WN, the product of each membership function value may be obtained as the goodness of fit w1, w2 ,.

(2) The antecedent operation unit 12 is used as the target system 10
, WN of the input values of the fuzzy rules R1, R2, ..., RN with respect to the antecedent part can be configured by a neural network.

(3) Similarly, the consequent operation part 13 is used for each fuzzy rule R1, R for the input value of the target system 10.
It is also possible to configure by a neural network that outputs the function values y1, y2, ..., YN of the consequent part of 2 ,.

(4) In the embodiment, the system modeling apparatus for identifying the input / output characteristics of the multi-input / single-output target system has been described, but the present invention is a system modeling apparatus for identifying the multi-input / multi-output target system. Can be similarly applied to. In that case, the number of output dimensions of the rule synthesis unit is the same as the number of output dimensions of the target system.

(5) The input / output data of the target system used for modifying the rule synthesizing unit does not necessarily have to be the input / output data obtained online from the operating target system. It is also possible to modify the rule synthesizing part offline using.

[0033]

As described above, according to the present invention, the input / output characteristics of the target system can be accurately identified over the entire input space, and the rule synthesizing section can be additionally adjusted, thereby providing flexibility. It is possible to provide a rich system modeling device.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a system modeling apparatus according to an embodiment of the present invention.

FIG. 2 is a diagram showing a structure of a BP type neural network.

FIG. 3 is a flowchart for explaining the operation of the embodiment.

FIG. 4 is a diagram showing a configuration of a rule synthesizing unit according to another embodiment of the present invention.

[Explanation of symbols]

 10 ... Target system 11 ... System modeling device 12 ... Antecedent part operation part 13 ... Consequent part operation part 14 ... Rule synthesis part 15 ... Correction part 16 ... Error operation part

Claims (2)

[Claims] 1. An IF-THEN comprising an antecedent part and a consequent part.
In a system modeling device for describing an identification model of input / output characteristics of a target system by a fuzzy rule of a format, an antecedent part computing means for performing an operation on the antecedent part based on an input of the target system, and based on the input A consequent part computing means for performing a calculation on the consequent part, a rule synthesizing part for obtaining an identification result of the output of the target system based on the calculation results of the antecedent part computing part and the consequent part computing part, and the target A system modeling device comprising: a correction unit that corrects the rule synthesizing unit so as to reduce an error of the identification result with respect to an output of the system. 2. The system modeling apparatus according to claim 1, wherein the rule synthesizing means is a backpropagation type neural network, and the correcting means corrects a connection weight in the neural network.