JPH03168835A - Fuzzy inference device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] A. INDUSTRIAL APPLICATION FIELD The present invention relates to a fuzzy inference device, and more particularly to a novel device with an added forgetting function.

B. Summary of the Invention The present invention incorporates a forgetting coefficient into the working rule when performing fuzzy inference, and changes the weight of each rule by changing it depending on the date, time, location, etc. when the working rule is generated and used. This makes the knowledge base very meaningful and allows practical solutions to be developed.

C. Problems to be Solved by the Prior Art and the Invention The concept of human ambiguity is vague but sophisticated thinking. For example, plant operators have a wealth of know-how regarding how to operate the plant, but most of this is based on qualitative experience.

To utilize this experience, fuzzy inference has been proposed, in which rules are expressed using fuzzy quantities, or fuzzy sets, and inference is performed based on this, and recently it has been applied to automatic dredging control and tunnel excavation equipment. ing.

By the way, in a system that uses fuzzy inference,
The most important issue was to acquire knowledge (rules) and treat them equally.

On the other hand, skilled operators are skilled at date and time. When a place or situation changes, our senses change slightly, and we make decisions that are appropriate to the situation. This means that if we focus on knowledge, some knowledge fades due to forgetting, while other knowledge is reproduced and used as is. However, in conventional fuzzy reasoning, humans tend to forget knowledge. Functions such as reproduction were ignored and became one of the flaws when building a human-friendly system.

The present invention was made under these circumstances.
Its purpose is to forget knowledge. The aim is to take into account the function of reproduction, realize inference using only a very meaningful knowledge base, and aim for more realistic solutions.

D. Means for Solving the Problems The present invention is based on the timing of rule generation for each work rule. It has a forgetting database that stores information I representing the location and/or situation, and the current time. A forgetting element calculation unit that takes in information ■, representing a position and/or situation, and calculates a value corresponding to the difference between the information 1, and γ as a forgetting element for each work rule; A forgetting coefficient estimating unit is provided, which has a forgetting rule that associates a coefficient with a forgetting element, and estimates a forgetting coefficient for each working rule based on the forgetting element and the forgetting rule obtained by the forgetting element calculation unit. Become.

E. For example, if the membership function is corrected by multiplying the membership function of each condition part and conclusion part of a working rule by the forgetting coefficient estimated by the action forgetting coefficient estimating part, when applying a certain working rule, If the generation time is quite a long time ago, the weighting will be low, and if the generation time is new, the weighting will be high.

F. Embodiment Hereinafter, an embodiment in which the present invention is applied to the control of a shield machine for excavating a tunnel will be described with reference to FIG.

In FIG. 1, lO is a working inference section, and the horizontal deviation distance SX+. between the planning line and the shield machine. Vertical deviation distance x, . It has a function of inferring the target value of the stroke amount of each jack by fuzzy inference using the input value as an input value, and its components include a work rule storage section 1, a membership function storage section 2, a forgetting coefficient storage section 3, and It is equipped with a fuzzy inference mechanism 4.

For example, n work rules, which are fuzzy rules, are stored in the work rule storage room 1, and these work rules have the above-mentioned X+, Xl as human power values, and jacks 1, 2, etc.
This rule uses the strokes fJ1y+, yt, . . . as inference values, and an example thereof is expressed as follows.

IF XI=A1% )(,=A,}Rule 1 TIIEN 'J + = R r,...
・・yr+=R, IF XI=A3, xt=A= }Rule n TIIEN y + = R 3, “””””In=
R4 where Ak (k is a natural number) and Rk are fuzzy labels. For example, the fuzzy labels Ak and Rk shown in FIG. 2(a). Membership functions as shown in (b) are assigned to each member, and these are stored in the membership function storage section 2. The forgetting coefficient storage unit 3 also stores a forgetting coefficient G1 estimated by a forgetting coefficient estimating unit to be described later.
~G n is stored.

Reference numeral 20 in FIG. 1 is a forgetting element calculation unit, which has a forgetting database 5 that stores the position of the shield machine at the time of rule generation for each working rule of the working inference unit 0, and the current position information of the shield machine. It has a function of acquiring the difference between the position information stored in the forgetting database 5 and the current position information, that is, the distance between these positions, normalizing the reciprocal of this difference, and outputting the value as a forgetting element.

30 is a forgetting coefficient estimation unit, which includes a forgetting rule store 6, a membership function store 7, and a fuzzy inference mechanism 8. In the forgetting rule storage room 6, there is a working inference section 10.
A forgetting rule is stored for each working rule, and each forgetting rule is composed of a fuzzy rule whose condition part is the forgetting element calculated by the forgetting element calculating section 20 and whose conclusion part is the forgetting coefficient. To give a specific example, the forgetting rule corresponding to work rule 1 is expressed as follows.

I P i +=C+ THEN (;+=
D + However, 11 is the forgetting element regarding work rule 1,
C,,D. is a fuzzy label. Membership functions as shown in FIGS. 3(a) and 3(h) are assigned to these fuzzy labels C, , D, respectively, and these are stored in the membership function storage section 2.

Next, the operation of the above embodiment will be explained. First, the current position information of the shield machine is taken into the forgetting element calculation unit 20, and the difference between this position information and the position information at the time of creation of each work rule in the forgetting database 5 is determined. For example, if the current position of the shield machine is PI and the position of the shield machine when generating the work rule l is P, then the above difference (distance!) is IP, P, ! becomes. Then, the reciprocal of this distance is taken and normalized, and the normalized value is output to the forgetting coefficient estimation unit 30 as a forgetting element. As a result, the forgetting coefficient estimation unit 30 has forgetting elements corresponding to the working rule I = n! I-f n will be captured. Next, the fuzzy inference mechanism 8 applies the forgetting element to the forgetting rule for each working rule to perform fuzzy inference and obtain the forgetting coefficient. For example, when calculating the forgetting coefficient G of working rule 1, if i is 0.5, then the membership value of the corresponding membership function of C1 (0.75) is calculated as shown in Figure 3. is obtained, and the member bump function of D is cut using this value, and the centroid value of the lower part thereof is obtained as a value of 01. The forgetting coefficients G1 to Gn obtained in this way are stored in the forgetting coefficient storage unit 3 of the working inference unit 10, and the fuzzy inference mechanism 4 calculates each membership l of the condition part and conclusion part of each working rule 1 to n. The input values X (horizontal deviation distance), x
Apply t (vertical deflection distance) to infer the stroke ffi y +~yn of each jack 1-n.

As a method for correcting membership functions using forgetting coefficients, this embodiment employs a method in which each membership function is multiplied by a forgetting coefficient. In this case, the above fuzzy rule can be specifically described as follows.

IF Xl=AI-Gl x,=A,・G }Rule l TIIEN yI R. -G. , ・−・−・−yn−=Rt−GlIF
Xl=A3”GII Xt=Ah・Gr1}Rule n TIIEN y+=Rt・Gll, .+++++M
n=R,・G1 To briefly describe the process of applying these rules, for rule 1, the membership value of A1・G1 corresponding to x1 and the membership value of A,・G1 corresponding to
, with the smaller membership value of BG,−R,・
Cut the membership function of C1.

Such processing is performed for all rules, and y1 to y. ，
For each of , the cut portions of all rules are combined and the center of gravity position of the union portion is taken as the inference result of y1 to y0.

In the above, the forgetting coefficients G1 to Gn are values indicating the degree of forgetting, that is, the weighting, of the working rule, respectively.The state of forgetting will be explained with reference to FIG. 4, taking the working rule 1 as an example. However, for convenience of explanation, only the membership function for X I+ y I is shown in FIG. 4.

First, when the position of the shield machine that generated the rule l is close to the current position, the forgetting coefficient G is, for example, l, and the membership functions for X++ and y+ of the rule at the time of inference are shown in Figure 4 (a). ). As shown by the solid line (1) in (b), it is the original shape of A,,11,,
The generated rule 1 is reproduced as is.

However, if the distance becomes large, 01 becomes, for example, 0.5
Then, each membership function is shown in Figure 4(a).
, the height is halved as shown by the dotted line (2) in (b).

In other words, half of rule l has been forgotten. When the shield machine advances further and the separation distance becomes considerably large, G becomes O, and each membership function becomes as shown in Fig. 4 (
a). As shown in (3) of (b), it becomes 0 for any input value. In other words, Rule 1 is completely forgotten.

Further, in the present invention, as a method of correcting the membership function using the forgetting coefficients G, to Gn, the membership function may be shifted to the negative side of the membership value according to the value of the forgetting coefficient. Figure 5(a). (b) are respectively A.
,R. This is a diagram showing the situation using the following as an example. When 61 Rikikawa is used, the membership function is the original shape as shown by the length line (+), but when G1 becomes 045, it is shown as the dotted line (2). so that each member nop value is 0.5 (+-0.5
) shift to the negative side. When C, h<O, the chain line (3
), each member bump value is shifted to the negative side by + (+-0). However, the part where the membership value is negative is meaningless as a member sop function, so G1-0
When this happens, the rules have been completely forgotten.

In the present invention, the fuzzy inference mechanism 4 may be provided with a function of multiplying the membership function by a forgetting coefficient and a function of shifting the membership function according to the forgetting coefficient, and configured such that one of the functions can be selected. .

The reason for introducing the forgetting factor in the present invention is that, for example, in the case of the embodiment, if the position of the shield machine changes, the soil quality changes, and if the current position is close to the position at the time of rule generation, the soil quality is similar, so the rule can be changed as is or This is because, although it can be used with a high weighting, as the separation distance increases, the soil quality will change, so the weighting of that rule must be lowered.

In the above-mentioned embodiment, the distance between the position at which the human input value was taken at the time of rule generation and the position at which the input value was taken at the time of rule use is used as the forgetting element. (Fl, time, etc.) or a value corresponding to the difference between the information I representing the situation and the current information I can be used as the forgetting element.

Note that the data in the oblivion database 5 that stores information 1+ (corresponding to the position of Nold Man Moon in the embodiment) at the time of rule generation is updated every time a rule is generated.

G. Effects of the Invention According to the present invention, a forgetting coefficient is introduced, thereby correcting the membership functions of the conditional part and the conclusion part of the working rule, and at the time of rule ljE and the time of rule use. For example, the forgetting coefficient is determined by fuzzy inference according to the forgetting factor, which is a difference in location or situation, and the degree of forgetting, that is, the weight, of the work rule is changed accordingly.
Working rules can be used flexibly depending on the location and situation, making the knowledge base very interesting.
As a result, the know-how of experienced operators can be utilized more accurately.
7. Able to come up with realistic solutions.

[Brief explanation of the drawing]

Fig. 1 is a configuration diagram showing an embodiment of the present invention, Fig. 2(a),
(b) is a graph showing the membership function of the working rule, Figures 3 (a) and (b) are graphs showing the membership function of the forgetting rule, Figures 4 (a), (b), and 5. (a
) and (b) are explanatory diagrams showing how rules are forgotten. ! ... Working rule storage section, 2.4... Membership function storage section, 5... Forgetting database, 10...
Working reasoning unit, 20... Forgetting element calculation unit, 30...
Forgetting coefficient estimator. Figure 1 Real Heaven = yj's connection diagram Figure 2 (Q) A+'s mechanism 6 open stability (b) B+'s mechanism 0 degree Celsius Figure 3? ff,% IL-file email address +1) File opening (Q
) Membership function of C+ (b) Membership function of D+ ) 81's Mechanism/Yn\Riff 6 Sekisho Length x 1 y1 Figure 5 Explanation of Jshiru's Shizuru's Massage (Q) A+'s Mechanism ) B+'s mempa "shi\sop M% sentence X, 4) T+4

Claims (3)

[Claims] (1) In a fuzzy inference device equipped with a working inference unit that executes fuzzy inference using working rules that are fuzzy rules, for each working rule, the time point, position and/or
Or, it has a forgetting database that stores information I_1 representing the situation, imports information I_2 representing the current point in time, position and/or situation, and uses a value corresponding to the difference between the information I_1 and I_2 as a forgetting element as a work rule. and a forgetting rule that associates a forgetting coefficient and a forgetting element for each of the working rules, and a forgetting element calculation unit that calculates the forgetting factor for each work based on the forgetting factor and the forgetting rule that are calculated by the forgetting element calculation unit. a forgetting coefficient estimation unit that estimates a forgetting coefficient for each rule, and the working inference unit determines the membership of each condition part and conclusion part of the working rule according to the forgetting coefficient estimated by the forgetting coefficient estimation part. A fuzzy inference device characterized by having a function of correcting a function. (2) The fuzzy inference device according to claim (1), wherein the function of correcting the membership function is a function of multiplying the membership function by a forgetting coefficient. (3) The fuzzy inference device according to claim (1), wherein the function of correcting the membership function is a function of shifting the membership function to the negative side of the membership value according to the forgetting coefficient.