JPH0319031A - Inference processing system - Google Patents

Abstract

PURPOSE:To enable an expert to unnecessitate the description of an FF rule and to improve the forming efficiency of an expert system by extracting automatically the relation between the symtoms which are unconsciously described into a diagnostic rule by an expert. CONSTITUTION:An FF rule extracting part 1 consists of a factor hypothesis comparing part 2 which compares the factor hypotheses of the 1st and 2nd diagnostic rules with each other and decides whether both hypotheses are con flicting with each other or not and an FF rule producing part 3 which produces a symptom rule (FF rule) defined between the symptoms of both diagnostic rules when the factor hypotheses of both rules are conflicting with each other. When a diagnostic rule is inputted by an expert, this rule is collated with the knowledge included in a knowledge base. If such a rule with which the factor hypoteheses of a conclusion part are not consistent with each other is detected, this rule is registered into the knowledge base as an FF rule. Thus the FF rule should be unconsciously described into the diagnostic rule. Then it is not required for an expert to extract the FF rule nor to describe the relation be tween the symtoms.

Description

[Detailed Description of the Invention] [Summary] This invention relates to knowledge representation and inference methods for diagnostic expert systems in the field of artificial intelligence.

Symptom-symptom rules (FF rules) are supposed to be unconsciously described in diagnostic rules, so the purpose is to extract them and make it unnecessary for experts to describe the relationships between symptoms.

In an expert system that has a knowledge base and an inference engine, the diagnostic rules input by experts and the diagnostic rules in the knowledge base are each a symptom and an OK cause list that lists candidate causes that are definitely not the cause of the symptom. ,
candidate causes of the symptom, and a cause for comparing two diagnostic rules to see whether all candidate causes of the second diagnostic rule are included in the OK cause list of the first diagnostic rule. If all candidate causes of the second diagnosis rule are included in the OK cause list of the first diagnosis rule by the hypothesis comparison unit and the cause hypothesis comparison unit, if the symptom is the symptom of the second diagnosis rule, the first diagnosis is determined. The present invention is configured to include an FF rule derivation unit including an FF rule creation unit that generates a symptom-symptom rule that a symptom that is negative of the symptom of the rule occurs.

[Industrial application field]

The present invention relates to knowledge representation and reasoning methods for diagnostic expert systems in the field of artificial intelligence.

In recent years, the development of expert systems equipped with expert knowledge and reasoning has become popular in the field of artificial intelligence. especially,
Diagnostic expert systems that infer causes based on input symptoms are now widely used for diagnosing machine failures and diagnosing diseases in the medical field.

[Conventional technology]

In conventional diagnostic expert systems, experts input diagnostic rules, which are rules for the relationship between symptoms and cause hypotheses, and store them as a knowledge base.

Based on these knowledge bases, the cause of a certain symptom was diagnosed using an inference engine.

The operation of the inference engine in a conventional diagnostic expert system will be explained below.

Now, in the case of automobile failure diagnosis, consider the following example of a diagnosis rule.

if Light Gatsu < then Hypothesis A - Rule 1
) if the battery is normal then Hypothesis B (Rule 2) Furthermore, the expert says, ``The light is on and the battery is normal.''
Suppose you enter the following symptom and investigate its cause.

At this time, the inference engine fires Rule 1 and Rule 2 to combine Hypothesis A and Hypothesis B.

However, in this case, there were the following problems.

Here, in the condition part of rule 1, the battery is implicitly normal. This condition is assumed. That is, if you write it correctly.

if the light and (and the battery is normal)
This means that n hypothesis A. Therefore, it is correct to infer hypothesis A for the symptom ``The light is on and the battery is normal.'' However, the inference engine fires at the same time as rule 1 and rule 2, resulting in the combination of hypothesis A and hypothesis B. In this case, if the combination of hypotheses occurs, an incorrect hypothesis inference will be made.

Below, I will explain why it is correct to adopt only the hypothesis of rule l, and why combining the hypotheses leads to incorrect conclusions.

The combination of hypotheses has conventionally been based on DS theory. This theory is applied when the causal hypotheses of symptoms are expressed as probabilistic quantities, and only when the symptoms are independent of each other, the connections between the hypotheses for each symptom are the combinations of all the combinations between the hypotheses. We need to find the product of Therefore, as shown above, rule 1
Since the symptom "Light and (") and the symptom "Battery is normal" in Rule 2 are not independent of each other, combining the hypotheses will lead to incorrect inferences. Let pa be the probability that it is the cause,
In hypothesis B, there is a correlation between rule 1 and rule 2, so there is a case where the same event X is the cause, and let Pb be the probability.

Combining the hypotheses, the probability that the event X is the cause of the symptom "the light is on and the battery is normal" is calculated by doubling the probability of the event X being caused by the rule 2. It becomes the sum of paXpb and a certain probabilistic quantity, and it becomes different from the probability pa caused by the event X obtained by firing only the hypothesis of rule l. According to DS theory, the more hypotheses that are caused by a certain event are combined, the more the probability that the event is the cause increases by doubly adding up.Therefore, the light If it sticks, the buffteri is normal. Because of this relationship, when diagnosing 9, if a symptom of a light coming on is input.

It is preferable to infer that the battery is normal and not ask questions about the battery in order to avoid double accumulation of causal probabilities. This is because the probability that this is the cause becomes greater than normal reasoning.

Therefore, there is a need for a function that handles the relationship between symptoms that are correlated with each other, such as the symptom ``the light is on'' and the symptom ``the battery is normal.''

In the conventional diagnostic-type engineered cattle subato system, nine or more problems were solved by the following means.

(1) Have an expert explicitly write the condition that the battery is normal in the condition part of rule l. By setting the inference engine's conflict resolution to fire only the one with more conditions, only Rule 1 fires even in the above situation.

(2) In the case of the above example, ask an expert to write a symptom-symptom rule (FF rule) that indicates that the battery is normal if the light is on.The inference engine checks only the FF rule and writes only rule l. set a fire. This method is F
(1) This is a more general-purpose method in the sense that the F rule can also be applied to other diagnostic rules.

[Problem to be solved by the invention]

As described above, unless the FF rule is written, the correct rule cannot be fired.

However, such FF rules are very common sense for experts, so they do not want to write them down, and it is easy to omit them when writing the rules.

If the description is omitted, there is a problem that correct inference cannot be made as mentioned above.

Since the symptom-symptom rule (FF rule) should be unconsciously described in the diagnostic rule, the present invention extracts it and eliminates the need for experts to describe the relationship between symptoms. purpose.

[Means to solve the problem] FIG. 1 is a diagram explaining the principle of the present invention.

In the figure, l is an FF rule derivation unit, a cause hypothesis comparison unit 2 that compares the cause hypotheses of the first diagnosis rule and the second diagnosis rule to determine whether they are contradictory, and An FF rule that creates a symptom-symptom rule (FF rule) that holds true between the symptoms of the first diagnostic rule and the symptoms of the second diagnostic rule when the causal hypothesis of the rule and the causal hypothesis of the second diagnostic rule contradict each other. It consists of a creation section 3.

(effect) The principle of the present invention will be explained below.

When a diagnostic rule is input by an expert, it is compared with the existing knowledge. If a rule is discovered in which the causal hypothesis in the conclusion part is incompatible (contradictory), it can be seen that the premise part of that rule does not hold at the same time.
This is registered in the knowledge base as a symptom-symptom rule (FF rule). That is, the following holds true.

Assume that the following two rules hold.

if A then B (Rule 1) if
Cthen not B (Rule 2) If the above rules hold true, then the following rule 3 holds true.

if A then not C (Rule 3) This rule 3 is registered in the knowledge base as an FF rule.

〔Example〕

A configuration diagram of an embodiment of the present invention is shown in FIG.

In the figure, 21 is a knowledge base in which one diagnosis rule is stored. 23 is a diagnostic rule input section.

This is where the diagnostic rules input by experts are accepted. 2
4 is an FF rule derivation module which derives a symptom-symptom rule (FF rule) from the 1 diagnosis rule and the diagnosis rule.

A symptom-symptom rule (FF rule) is created by the FF rule deriving module 4 from the diagnostic rule accepted by the diagnostic rule input unit 23 and the diagnostic rule in the knowledge base 21 .

Next, the general form of the nine diagnostic rules is shown.

(def-rule ″((1 pause, answer) (question, answer) −・−) OK cause list ° ((candidate cause, CF value) (candidate cause, CF value)・−111…)) Here The question is about the symptoms.

The OK cause list is a list of things that absolutely cannot be the cause of two or more symptoms. A candidate cause is a candidate that is the cause when there are 5 or more symptoms, and the CF value is the probability that it is the cause.

In other words, when a certain symptom occurs, things that cannot be the cause are shown in the list of OK causes, indicating that the cause is one of the candidate causes. .

Next, FIG. 2 shows an example of the description of the diagnostic rule.

This example is a diagnostic rule regarding the questions "Does the engine start?" and "Does the starter turn?"

In the figure, 21 is the first diagnosis rule, 22 is the second diagnosis rule, 23 is the third diagnosis rule, and 24
is the fourth diagnostic rule.

Next, details of the processing procedure of the symptom-symptom rule (FF rule) derivation module 4 will be explained.

FIG. 3 is a processing flow of the FF rule derivation module 4.

In the figure, a diagnosis rule (Rule 1) is manually entered by an expert in the diagnosis rule input unit 3 at Sl, and the diagnosis rule has the following form.

if A then Causal hypothesis H1 (Rule 1)
Next, in S2, it is determined whether all the diagnostic rules in the knowledge base 1 have been compared with the rule 1. If you compare it, S
Finish the process with IO. If there is a diagnostic rule that has not been compared, that diagnostic rule (rule 2) is extracted from the knowledge base 1 in S3.

Rule 2 has the following form:

if B then Causal hypothesis H2 (Rule 2)
Next, in steps 34 to S7, it is determined whether or not the cause hypothesis H1 and the cause hypothesis H2 are compatible, and if they are compatible, the process returns to steps S2 and subsequent steps. If they are not compatible, the following rules are added to the knowledge base 1 as FF rules in S8, and the process returns to steps 32 and below.

if A then not Bif B
then not A Here, cause hypothesis H1 and cause hypothesis H2 are incompatible (contradictory) when all candidate causes that appear in one rule are included in the OK cause list of the other rule. . That is, rule 2
And all the things that are causing a certain symptom.

This is a case where it is absolutely not the cause of the symptoms according to Rule 1. The explanation will be given below by returning to FIG.

In the figure, in S4, it is determined whether the candidate causes of rule 1 and the OK cause list of rule 2 exist, and if they do not exist, 3
If there are any candidate causes, it is determined in S5 whether or not all candidate causes of rule 1 are included in the OK cause list of rule 2. If they are included, the process goes to step S8, and if not, S6 It is determined whether the candidate causes for rule 2 and the OK cause list for rule 1 exist, and if they do not exist, proceed to step S2.
Returning to the process below, if there are any, it is determined in S7 whether all candidate causes of rule 2 are included in the OK cause list of rule 1, and if they are not included, returning to the process of 32 and below,
If included, S9. Go to processing.

In S8, FF rule 1 is added to the knowledge base, which states that if the symptom of rule 1 is negative, the symptom of rule 2 is negative, and the process returns to step 32 and below. In S9, FF rule 2 is added to the knowledge base, which states that if the symptom of rule 2 is the negative of the symptom of rule l, the process returns to step 32 and below.

Next, the process of checking for contradictions in causal hypotheses and deriving FF rules will be explained using an example of the second diagnostic rule.

The candidate causes in the fourth diagnostic rule 24 are the following five.

■Insufficient battery charge ■Poor contact between the cords between the starters and starters ■The starter is broken - ■Disconnection between the starters and the starters ■Over-discharge of the battery All of these candidate causes are listed in the OK cause list in the first rule 21. included in. Therefore, the causal hypothesis of the first rule and the causal hypothesis of the fourth rule are incompatible (contradictory).

From this, 9 If the starter turns? -It doesn't turn then
Does the engine start? - Does not cost or.

If does the engine start? -It takes then
Does the starter turn? -The FF rule of spinning is derived.

Here, another method may be considered for deriving the FF rule.

Consider the second diagnostic rule and the fourth diagnostic rule.

There is a correlation between the fourth diagnostic rule, "The starter does not turn," and the second diagnostic rule, "The engine does not start." The above OFF rule exists.

At this time, the candidate cause in the fourth diagnostic rule (1) is that the starter does not turn at all due to a failure.

■Disconnection between the loose starters・−・ ■Battery overdischarge... All of these are included in the second candidate cause.

Therefore, by examining the candidate causes of the second diagnosis rule and the fourth diagnosis rule, if all the candidate causes of the fourth diagnosis rule are included in the candidate causes of the second diagnosis rule, then the fourth diagnosis rule If the symptoms are the symptoms of the second diagnostic rule. That is,
It seems possible to derive the FF rule that says, ``If the starter doesn't turn, the engine won't start.'' However, if the symptom of "starter does not turn" in the fourth diagnostic rule holds true, all of its candidate causes are included in the candidate causes in the second diagnostic rule. If any of these conditions hold true, it does not necessarily mean that the second diagnostic rule, ``engine won't start,'' will occur.

This method cannot completely guarantee the establishment of the FF rule, and it is correct to use a method of deriving the FF rule from the contradictions between the above causal hypotheses.

[Other Examples]

FIG. 5 shows a configuration diagram of another embodiment of the present invention.

In the figure, numeral 51 is a symptom input section, which receives symptoms input by experts. Reference numeral 52 denotes an extended inference engine, which searches the diagnostic rules stored in the knowledge base 53 for diagnostic rules that match the input symptoms, and processes the searched diagnostic rules with the FF rule derivation module of the above embodiment. The FF rule is derived by performing exactly the same operation. Using the OFF rule, a fault diagnosis is made from the input symptoms and the diagnostic rules stored in the knowledge base 53, and an inference is derived. 53 is the knowledge base. 2 diagnosis rules are stored.

With the above configuration, the extended inference engine 52 is activated from the symptoms accepted by the symptom input section 51.

A diagnostic rule that matches the above symptoms is searched from the diagnostic rules in the knowledge base 53, and if there is a diagnostic rule that contradicts the causal hypothesis among the searched diagnostic rules.

An FF rule is derived between the symptoms of these diagnostic rules, and this FF rule is used to combine the input symptoms with the knowledge base 5.
It performs inference from the diagnostic rules in 3 in the same way as conventional inference engines, and outputs the inference results.

〔Effect of the invention〕

As described above, according to the present invention, by automatically deriving the relationships between symptoms that experts have unconsciously described in diagnostic rules, it is no longer necessary for experts to write FF rules, and an expert system can be constructed. This greatly contributes to improving efficiency.

[Brief explanation of drawings]

FIG. 1 is a diagram explaining the principle of the present invention. FIG. 2 is a block diagram of one embodiment of the present invention. FIG. 3 is an example of one diagnosis rule. FIG. 4 is a processing flow diagram of the FF rule derivation module. FIG. 5 is a block diagram of another embodiment of the present invention. In Figure 1. l is an FF rule derivation unit. 2 is the cause hypothesis comparison section. 3 is an FF rule creation section. 4j star theory of the present invention 1

Claims (1)

[Claims] In an expert system having a knowledge base and an inference engine, the diagnostic rules input by the expert and the diagnostic rules in the knowledge base each identify a symptom and a candidate cause that is definitely not the cause of the symptom. Consisting of a listed OK cause list and candidate causes of the symptom, between the two diagnostic rules, the OK cause list of the first diagnosis rule includes all candidate causes of the second diagnosis rule. The cause hypothesis comparison unit (2) compares whether or not the first diagnosis rule is O.
When the K cause list includes all candidate causes of the second diagnostic rule, a symptom-symptom rule is generated that states that if the symptom of the second diagnostic rule is negative, a symptom that negates the symptom of the first diagnostic rule occurs. F consisting of FF rule creation part (3)
An inference processing method characterized by having an F-rule deriving section (1).