JPH0831037B2 - Inference method using knowledge base - Google Patents

Description

The present invention relates to an inference method using a knowledge base, and particularly to a knowledge base suitable for constructing a highly reliable diagnostic / control system using ambiguous knowledge. Inference method using.

[Conventional technology]

A conventional inference method using ambiguous knowledge is, for example,
As discussed in "Measurement and Control" vol.22, No.9 (1983) pp.774-779, MYCIN method, subjective Bayes method, Cempster & Shaper theory based method, fuzzy logic based Methods etc. are known.

In all of these methods, the vague knowledge possessed by humans can be used by using certainty factor, probability, membership function, etc., whereby inferences similar to human thought patterns can be made. It is possible to do.

[Problems to be solved by the invention]

However, the above-mentioned conventional technique has a problem that no consideration is given to the following points.

The first problem is that it cannot handle the certainty (ambiguity) of negative and positive. When determining whether an event is true, human beings do not handle ambiguity with one probability measure from “0” to “1” that indicates whether the event holds or does not hold. Inference is performed by separately processing the scale related to the establishment of and the scale related to the failure of the event.

 This can be demonstrated with the following example.

The success or failure of an event is represented by the probability of "0" to "1". Using some vague knowledge, when Event A is found to be "not likely to hold", for Event A,
For example, give a probability of 0.2.

It is also assumed that similar knowledge is obtained by using different knowledge. In this case, the probability of event A is 1-0.8 x 0.8 = 0.36, which is not intuitive. This is because the probability "0" is not satisfied.

In order to eliminate the deviation from the above intuition, MYCIN
In the method (1), the non-establishment-establishment is mapped to -1 to 1, and calculation based on the probability theory is performed within the interval of -1 to 0, 0-1. For example, if event A is unlikely to hold true, setting it to -0.2 gives -0.36 when such a result is obtained again, which is consistent with intuition.

Subjective Bayes method, Fuzzy theory is 0 to 1
Although it is a measure of, in order to eliminate the deviation from the above intuition,
Even if the same event is derived using different knowledge, the probability and membership value of the event are combined by the MAX operator.
However, the probability of event A is 0.8,0 using different knowledge.
Even when it was obtained as 9, MAX (0.8, 0.9) = 0.9, which is out of the intuition in another sense.

Although the method based on the Dempster & Shaper theory can handle ignorance, if the condition part and the conclusion part of a rule cannot be represented by a set of events, the calculation is based on the probability theory that the failure is "0" and the success is "1". It will be the same, and inevitable deviation from intuition will occur.

As described above, in the conventional fuzzy inference methods other than the MYCIN method, the ambiguity of negative and positive cannot be treated so as to match human intuition.

Also in the MYCIN method, if the results of inferring an event using different knowledge are +0.8 and -0.5, the certainty factor of this event is 0.8 + (-0.5) = 0.3. It is calculated as follows, and it expresses on one scale.

In other words, there is a problem that it does not have two scales, negative and affirmative, and in the end, it is expressed by a single scale of -1 to 1, and in reality, it handles the certainty factors of negative and positive. It is not possible.

The second problem is the lack of consideration for ensuring the correctness and reliability of ambiguity knowledge in an inference system that handles ambiguity.

It is very difficult to guarantee the correctness and reliability of knowledge in the reasoning that deals with ambiguity as compared with the logic that represents the truth by the binary value of 0,1.

However, conversely, by using ambiguity, it is possible to improve the correctness and reliability of inference using ambiguity knowledge.

In conventional reasoning systems, no consideration is given to the improvement of correctness and reliability using this ambiguity, and the application of control systems using fuzzy reasoning systems to fields requiring safety It was a big obstacle.

On the other hand, in the conventional reasoning system, since there is no semantic check of the knowledge base, there is a situation that the safety of the knowledge base system is said to have a problem.

The present invention has been made in view of the above circumstances, and an object of the present invention is to solve the above problems in the inference method using the conventional knowledge base, improve the reliability of the knowledge base system, and improve the safety. It is to provide an inference method using a knowledge base, which enables application to the fields required.

[Means for solving problems]

The above object of the present invention is to control the controlled object in a system for controlling the controlled object by analyzing the state of the controlled object using a knowledge base expressed in the form of rules with certainty factors. For an event derived in the inference process for, a negative certainty factor that is a certainty factor that the event does not hold and a positive certainty factor that is a certainty factor that the event holds are stored, and On the other hand, the result of the inference is derived using both the certainty factor of the negative and the certainty factor of the positive, and the certainty factor of the conclusion event concerning the result of the inference is stored in the negative factor stored by reflecting the inferred result. It is achieved by an inference method using a knowledge base, which is characterized by updating the certainty factor and the positive certainty factor.

[Action]

In the inference method using the knowledge base according to the present invention, for each event guided in the inference process, the ambiguity between two scales, negative and positive, is maintained, and inference using each scale independently. Since it becomes possible, it is possible to realize fuzzy reasoning that is closer to the human thought form.
As a result, the vague knowledge possessed by humans can be easily ported to a computer.

In addition, in the process of making inferences independently using two scales, negative and affirmative, there are contradictions such as large values for the ambiguity represented by the negative and affirmative scales for each situation that is guided by inference. Whether or not there is a contradiction is detected as a result of monitoring, and if knowledge base construction is performed, the knowledge base builder is informed of the knowledge that has caused the contradiction, and You can improve your own credibility.

In addition, by applying the inference method using the knowledge base according to the present invention, in a certain situation, the phenomenon that can be guided,
It is possible to know both the event and the event that cannot be guided, and when an event belongs to both, by considering this as a contradiction, it becomes possible to easily detect the contradiction of the knowledge base system.

As a result, when a contradiction is detected in the control system, etc. during actual operation, the operator is warned and the inconsistent event is used by not making any inference regarding the event in which the contradiction is detected. It is possible to avoid outputting an incorrect inference result.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram of an ambiguous inference system showing an embodiment of the present invention. This embodiment uses various sensor information installed in the process factory and knowledge expressed in the form of rules with certainty factors to infer the state of the process from one viewpoint and inform the operator. System.

 This system has two usage patterns.

One is a usage form in which a knowledge base builder inputs knowledge to be used for Proseca diagnosis into a computer and verifies its validity. The other one is a usage mode in which a state of a process is inferred using a knowledge base created by a knowledge base builder, and a process operator gives a control instruction based on the inference result.

The knowledge base builder creates a knowledge base using the console 1 and modifies the knowledge base based on the inference result by the created knowledge base.

The process operator obtains an inference result by the knowledge base created by the knowledge base builder from the console 2 and issues a process control instruction based on the inference result.

Before giving a detailed explanation below, in this system,
Rule description, event certainty factor, and certainty factor calculation method are explained.

First, the rule description will be described. Ambiguous knowledge is
It is stored in the knowledge base 5 shown in FIG. 1 and is represented by the rule 20 with certainty factor as shown in FIG.
One rule is rule name 21, condition part 22, conclusion part 23,
It consists of a certainty factor 24 that is greater than “0” and less than or equal to “1”.
For example, rule 20 is D if the events of A, B, C are correct.
It represents the knowledge that it is possible to judge that the event of is correct with a probability of 0.8.

A "-" description before each event represents an undefined event. For example, rule 25 is knowledge that if the events of B and E are false, it can be determined with a certainty of 0.5 that the event of D is also false. This rule can be used to derive the confidence of the event.

Next, the certainty factor of an event will be described. The certainty factor of an event is represented by two scales, negative and affirmative. The negative certainty factor of event A is represented by Cf _neg (A), and the positive certainty factor is represented by Cf _pos (A), and its range is “0. "More than" 1 "or less.

For example, Cf _neg (A) = 0.8 indicates that it is possible to determine that event A is not established with a probability of 0.8, and Cf _pos (A) = 0.5 indicates that event B is It shows that it can be determined with a certainty of 0.5. The certainty factor of each event is held in the working memory 6 shown in FIG.

Next, the certainty factor calculation method will be described with reference to FIG.

As shown in the description 40 of the rule, the condition part has a condition event sequence 41,
It is assumed that the conclusion part is composed of the conclusion event sequence 42, and the confidence factor of this rule itself is Cf _r 43.

The condition event and the conclusion event are A, B, C shown in Fig. 2 earlier.
, -B, -E, -D, consisting of events with a "-" sign (called "negative events") and events without a "-" sign (called "positive events"). ing.

The Cf (conditional part) 44 is a value indicating to what extent the condition part of the rule is satisfied, and is the min value of Cf (conditional event i) as shown by the equation 45 shown in FIG.

The value of Cf (conditional event i) is a negative event, that is, Cf _neg (conditional event i) is a positive event if it is an event with a "-" sign, as shown by the equation 46 shown in FIG. An event, that is, an event without a "-" symbol is _Cfpos (conditional event i).

For example, Cf (conditional part) of the rule of IF A, −B, C THEN ... Is represented by min (Cf _pos (A), Cf _neg (B), Cf _pos (C)). C above
f _pos (A), Cf _neg (B) and Cf _pos (C) are 0.3,
If 0.5 and 0.2, Cf (conditional part) becomes 0.2.

Cf (conclusion part) 47 is a rule for the conclusion event described in the conclusion part as shown by the equation 48 shown in FIG.
It is the certainty when inference is performed using 40. The confidence level of the conclusion event is updated by this value based on the probability theory as shown in Expression 49. In Equation 49, C
f _neg and Cf _pos are the negative and positive certainty factors immediately before inference is performed using this rule, and these are Cf ′ _neg and Cf ′.
It is updated to the value represented by _pos .

A specific example is shown in FIG. It is assumed that the current value of the working memory 6 shown in FIG. 1 is shown in Table 50.
In this situation, inference using Rule 51 yields
52,53 gives Cf = 0.24. next,
From Expressions 54 and 55, Cf _pos (D) becomes 0.468 and Cf _neg (E) becomes 0.772, and the contents of the working memory 6 are updated to 0.468 and 0.772, respectively, as shown in Table 56.

Returning to FIG. 1, the function of the system of this embodiment will be described in detail.

Programs 7, 8 and 9 are inference mechanism using only negative certainty factor, inference mechanism using only positive certainty factor,
It is an inference mechanism that uses both negative and positive certainty. These inference mechanisms perform inference using the rules with certainty factor stored in the knowledge base 5 and store the events derived in the inference process in the working memory 6. That is, the knowledge base 5 stores the rule group shown in FIG. 2, and the working memory 6 stores the negative and affirmative certainty factors of each event derived by inference in the table shown in FIG. It is stored as 50.

The program 10 collects information on various sensors installed in the process factory 11. The inference mechanisms 7, 8 and 9 described above can know the sensor information required for inference from the program 10.

 The programs 7, 8 and 9 will be described below.

The program 7 is an inference mechanism that uses only the negative certainty factor, and its processing flowchart is shown in FIG.

First, the working memory 6 is cleared, and the certainty factors concerning all events are erased (step 60). Next, only the rules consisting of only negative events are read from the knowledge base 5 and stored (step 61). A rule consisting only of negative events means that all the events that make up the conditional part and the conclusion part, such as IF-A, -B THEN-C with 0.8, are
It is a rule that is an event with a "-" sign.

However, the event related to the sensor data may be a positive event. For example, the event of the first condition of the IF (sensor 01 value is 65.2 or higher), -B, -C THEN ... rule: A sensor 01 value of 65.2 or higher is a positive event, but it relates to sensor data. This rule is also selected because it is an event.

Next, the program 10 is queried for the truth of the event regarding the sensor data appearing in the selected rule (step 62). This value is a binary value of true / false (1,0).

Equation 47 showing Cf (conclusion part) of all rules in Fig. 3
The rule that is not executed and Cf (conclusion part)> 0 is selected (step 63), and the certainty factor of the negative event described in the conclusion part is calculated by the formula 49 shown in FIG. It is updated and stored in the working memory 6 (step 64).
This is repeated until there is no rule that has not been executed and Cf (conclusion part)> 0.

Due to the above-described inference mechanism, only the certainty factor of negative event is stored in the working memory 6.

Program 8 is an inference mechanism that uses only positive certainty factors. The process flow chart is as shown in FIG.
61 is changed to “read a rule consisting of only positive events and events related to sensor data from the knowledge base”. In other words, in the program 8, the inference is performed using only the rules including only the events related to the sensor data and the positive events, such as IF- (the value of the sensor 01 is 65.2 or more), B, C THEN .... By this inference, the information stored in the working memory 6 is only the certainty factor of the affirmation of the event.

The program 9 is an inference mechanism that uses both negative and positive certainty factors. The processing flowchart is such that step 61 in FIG. 5 is changed to “read all rules from the knowledge base”. That is, in the program 9, the inference is performed using all the rules including the rules in which positive and negative events are mixed such as IF-A, B, -C THEN-D .... The information stored in the working memory 6 by this inference is the positiveness of the event and the positive certainty factor.

The inference methods have been described above.Next, the main program of the fuzzy inference system, program 3,
4 will be described.

Program 3 has a dialogue with the knowledge base builder,
Using programs 7, 8 and 9, instructing inference, notifying the knowledge base builder of the contents of working memory 6, and monitoring the negative and positive certainty of each event in working memory 6 to detect inconsistencies. In this case, the rule that caused the contradiction is displayed and the inference is paused.

 FIG. 6 shows a flowchart of this processing.

First, one of the inference using only the negative certainty factor, the inference using only the positive certainty factor, and the inference using both the negative certainty factor and the positive certainty factor is inquired (step 7
0).

In the case of inference using only the negative certainty factor, the program 7 is activated (steps 71 and 72), waits until the inference is completed (step 75), and the contents of the working memory 6 are displayed (step 83).

In the case of inference using only positive certainty factor, similarly, the program 8 is activated (steps 73 and 74), waits until it is completed (step 75), and the result is displayed (step 8).
3).

In the inference that uses both the negative certainty factor and the positive certainty factor, first, the reference value for determining contradiction is inquired (step
76). In this system, the contradiction C _{n + r} (A) of the event A is C _{n + r} (A) = max (Cf _neg (A) + Cf _pos (A) -1,0). This is the broken plane 90 shown in FIG. 7 (A).
Becomes This function takes a value of “0” at 91 and a value of “0” or more and “1” or less at 92 shown in FIG. 7 (B).
_{When neg} (A) = 1 and Cf _pos (A) = 1, C _{n + r} (A) = 1,
Takes the maximum value. That is, when Cf _neg (A) = 1 and Cf _pos (A) = 1, the maximum contradiction is determined.

In the step 76, an inquiry is made as to whether or not the degree of inconsistency calculated by this formula becomes more than or equal to the inconsistency. Next, the program 9 is instructed to start (step 77), and during inference, C _{n + r} (•) is constantly monitored for all the events in the working memory 6 (step
78, 79), C _{n + r} (•) is larger than the value instructed in step 76, the program 9 is instructed to pause, and the event that detected the inconsistency and the rule that led to this event are displayed. (Step 80), it is inquired whether to restart the inference (Step 81).

When restarting, the restart instruction is given to the program 9 (step 82), and the processing from step 78 is performed. When finished, display the contents of working memory 6 (step 8
3), end.

Next, the program 4 for performing dialogue with the process operator and inference management will be described. Program 4
Basically performs the same processing as the processing flowchart shown in FIG.

The different part is the processing 80 to 82 when the contradiction is detected in step 78. In the program 4, when a contradiction is detected in step 78, not in steps 80 to 82, “display the event name in which the contradiction has occurred as a warning, and in the rule stored in program 9, The related rules are searched for, erased, and the program 9 is instructed to restart inference. "

According to this embodiment, ambiguous knowledge used in plant diagnosis can be efficiently constructed and its reliability is high. In addition, when making a diagnosis in a practical operation problem using ambiguous knowledge, it is possible to know what is currently occurring and what is not occurring, and even when a contradiction occurs during ambiguous reasoning, You can get reliable inference results,
This has the effect of improving stability.

FIG. 8 is a block diagram of an inference system showing an application example of the inference method using the knowledge base according to the present invention. The present embodiment is a system that infers the state of a process and informs the operator by using various sensor information installed in a process factory and knowledge in a complementary knowledge base.

 This system also has two usage patterns.

One is a usage pattern in which a knowledge base builder inputs knowledge into a computer in the form of rules, verifies its correctness, and corrects and adds knowledge, and the other one is a knowledge base builder. Using the created knowledge base and the sensor information from the process factory, the control operation to be performed is inferred.
This is a form of use in which process control is performed automatically.

The knowledge base builder inputs, via the console 1, a knowledge base 12 that guides a control operation to be performed and a knowledge base 14 that guides a control operation that should not be performed. The complementarity checking mechanism 13 determines whether or not the knowledge base 12 is appropriate as a "knowledge base that guides a control operation to be performed", and the knowledge base 14 "knowledge base that guides a control operation that should not be performed".
It is determined whether the two knowledge bases are complementary, and an event that does not satisfy the complementarity is output to the console 1 as a monitor together with a warning.

FIG. 9 shows a rule expression method. The basic configuration is
It is the same as that shown in FIG. 2 above, and here,
It indicates that there is a positive rule indicating a control operation (D) to be performed, as shown in (A), and a negative rule indicating a control operation () that should not be performed, as shown in (B). Is.

In the knowledge base 12 described above, only the positive rule in which the X is not described in the conclusion part as shown in (A) above, and in the knowledge base 14 are both the rules shown in (A) and (B) above. A rule in the form of is described.

Next, the complementary checking mechanism will be described with reference to FIG.

It is checked whether or not there is an event in the rule description of the knowledge base 12 (step 30), and if it exists, this rule should be described in the knowledge base 14, so the message “rule in knowledge base 12 “Exists” is output (step 31) and the process ends. Next, an event described in the conclusion part and not described in the condition part is searched for, and is set as a conclusion event (step 32).

It is checked whether or not the conclusion events obtained in step 32 are all control operations (step 33). If there is something that is not a control operation, the event is a meaningless intermediate event. There is an event that does not lead to the inference of the operation ”(step 34), and the process ends.

Through the above processing, it is possible to determine whether the knowledge base 12 is appropriate as a "knowledge base that guides the control operation to be performed".

In steps 35 to 37, the knowledge base 14 is checked in the same manner as above.

Step 38 is a process for checking the complementarity between knowledge bases 12 and 14, and in the set of conclusion events of each knowledge base obtained in steps 32 and 35, an event that belongs to only one set Is output as a non-complementary event (step 39). When there is no non-complementary event, the message “OK” is output and the processing ends.

A specific example of the processing of the complementary checking mechanism 13 described above is shown in FIG.

The rule group 65 is a rule belonging to the knowledge base 12. First, a conclusion event is calculated. For events C, E, G and H in the conclusion part of all rules, check if they appear in the condition part. Since C and E appear in the conditional part of other rules, the conclusion event of rule group 65 is
It becomes {G, H}.

On the other hand, the rule group 67 is a rule belonging to the knowledge base 14, and similarly, the conclusion event is {G}.

Therefore, by comparing these sets, the event H can determine the “control operation that should be performed”, but cannot determine the “control operation that must not be performed”. It was found not complementary in H.

Next, a method of performing process control using the knowledge bases 12 and 14 which is the second usage will be described.

The inference mechanisms 15 and 16 perform inference using the knowledge bases 12 and 14, respectively. The inference is performed by sequentially executing the rules satisfying the conditional part and storing the control operations derived in the process of execution in the files 17 and 19, respectively. Continue this operation until there are no more executable rules.

If the condition part of the rule includes an event related to sensor information, the process control mechanism 26 is inquired about its authenticity.

The inference mechanisms 15 and 16 store the inference results when the knowledge bases 12 and 14 are used in the files 17 and 19, respectively. The inconsistency detection mechanism 18 compares the files 17 and 19 and checks whether the “control operation that should be performed” in the file 17 is in the “control operation that should not be performed” in the file 19. The above "control operation to be performed" is instructed. If the control operation is a “control operation that should not be performed”, the control operation is ignored.

A specific example of the above process is shown in FIG. Knowledge base 93
Is knowledge about “control operation to be performed”, and it is assumed that {G, H} was obtained as a result of inference using this (94). That is, in this example, it is currently inferred that "the control operations G and H should be performed". Similarly,
As a result of inference using the knowledge base 95, it is inferred that the conclusion event is {H}, that is, "control operation H must not be performed" at present.

Here, the control operation H to be performed is also a control operation that should not be performed, and thus is ignored, and only the control operation {G} is performed.

According to the present embodiment, it is possible to obtain the effects that the construction of the complementary knowledge base becomes easy, and that the reliability of the control system is improved by using the complementary knowledge.

In the above embodiment, the knowledge base used for inference does not directly include a "knowledge base that guides a control operation that should be performed" and a "knowledge base that guides a control operation that should not be performed", and includes both of them. However, the present invention should not be limited to this.

〔The invention's effect〕

As described above, according to the present invention, an event derived in the inference process in a system that analyzes the state of a target, diagnoses, controls, etc. using knowledge expressed in the form of rules with certainty factors. On the other hand, since the certainty factor that the event is not established (negative certainty factor) and the certainty factor that the event is established (affirmative certainty factor) are independently handled and inference is performed, the knowledge base system The inference method using the knowledge base, which has improved reliability and can be applied to fields requiring security, can be realized, which is a remarkable effect.

[Brief description of drawings]

FIG. 1 is a block diagram of an ambiguous inference system showing an embodiment of the present invention, FIG. 2 is a diagram showing a rule configuration, FIG. 3 is a diagram explaining a method of calculating a certainty factor, and FIG. FIG. 5 is a diagram showing an example of calculation, FIG. 5 is a processing flowchart of the inference mechanism, and FIG.
FIG. 7 is a process flowchart of the monitor program, FIG. 7 is a diagram showing a contradiction function, FIG. 8 is a system configuration diagram showing an embodiment of the second configuration of the present invention, and FIG. 9 is a diagram showing an example of rule description. FIG. 10 is a processing flowchart of the complementary check mechanism, FIG. 11 is a diagram showing an example of complementary check, and FIG. 12 is a diagram showing an example of conflict detection. 1,2: Console, 3,4: Dialogue and inference management mechanism, 5,12,14:
Knowledge base, 6: working memory, 7: inference mechanism using negative certainty factor, 8: inference mechanism using positive certainty factor, 9: inference mechanism using negative certainty factor, 13: complementarity checking mechanism, 18: Contradiction detection mechanism.

Claims (2)

[Claims] 1. A system for controlling a control target in a system for controlling the control target by analyzing the state of the control target using a knowledge base expressed in the form of a rule with certainty factor. For an event derived in the inference process, a negative certainty factor that is the certainty factor that the event does not hold and a positive certainty factor that is the certainty factor that the event holds are stored, and for the event , Deriving the result of inference using both of the negative certainty and the positive certainty, and for the certainty of the conclusion event regarding the result of the inference, the negative certainty stored by reflecting the inferred result. And an inference method using a knowledge base, which is characterized by updating positive certainty. 2. The inference method using a knowledge base according to claim 1, wherein the negative certainty factor and the positive certainty factor are updated based on probability theory.