JPH01126761A - Trouble diagonosis system - Google Patents

Abstract

PURPOSE:To improve the speed and accuracy of trouble diagnosis by automatically preparing a decision tree, which diagnoses the symptoms of a high frequency from an experience in the past, using this decision tree at first at a diagnosing time and trying the identification of a trouble spot. CONSTITUTION:A simplified diagnosing part 23 causes the symptoms of the high frequency to be the decision tree from the experience in the past and the procedure of the processing is as follows. The route node is a decision tree is obtained as an N. Whether the N is a leaf to be condensed trouble causes or not is decided. In the case of the leaf, control is moved to a processing 7. The symptoms of the N to be questioned is transferred to an input and output part 15 and whether the symptoms are measured or not is questioned to a user. When there is an answer that the symptoms are measured, the control is moved to a processing 5 and in the case of the answer of no measurement, the control is moved to a processing 6. Next, with proceeding to the left leaf of the N, the node is obtained as the new N and the operation is returned to a processing 2. Then, after proceeding to the right branch of the N, the node is obtained as the new N and the operation is returned to the processing 2. When the N is the leaf, it is obtained as the trouble causes.

Description

[Detailed Description of the Invention] [Summary] A computer predicts the behavior based on the structural model of the target for failure diagnosis, and the components related to the constraint conditions that led to results inconsistent with the symptoms are identified as candidates for the cause of the failure. Regarding the fault diagnosis method.

The purpose of this invention is to enable simple and fast diagnosis of frequently occurring failures, and to provide a means to automatically enhance the scope and accuracy of the diagnosis.

A conflict recorder that records minimal conflicts determined in response to each symptom, and a decision tree generator that extracts those with a predetermined frequency or more from among the recorded minimal conflicts and generates a decision tree regarding failure cause candidates. and a simple diagnosis section that performs an initial diagnosis using the decision tree at the time of fault diagnosis.

[Industrial application field]

The present invention relates to a fault diagnosis method in which a computer predicts the operation based on a structural model of a target for fault diagnosis, and uses components related to constraint conditions that lead to results inconsistent with symptoms as candidates for the cause of the fault.

[Conventional technology]

Currently, the most advanced fault diagnosis technology is 9, for example, XER.
Mr. de Kleer of OX and Wi ] 1 iams
Mr.

“ARTIFICIAL IN LIGENCE Magazine 32
Paper published in Volume rDiagnosing Mult
Method GDE reported in iple FaultsJ
There is.

FIG. 7 is a block diagram of this GDE.

In FIG. 7, 10 is a processing unit including a CPU and memory, 11 is a conflict analysis unit that determines minimal conflict for a given symptom, and 12 is a structural model that describes the structure of the device to be diagnosed.

13 is a behavior prediction unit that predicts the behavior of the target device using the structural model 12; 14 is a candidate generation unit that generates failure cause candidates from the minimal conflicts obtained so far; 15 is in charge of conversation with the user. This is the input/output section.

Here, the structural model 12 is given as a set of descriptions regarding each component of the device to be diagnosed. A "competition" for a certain symptom is a subset of the structural model 12, and a behavior that can be predicted only from this subset is inconsistent with the symptom. A conflict that does not include other conflicts as a subset is called a "minimal conflict."

At this time, candidates for a given set of symptoms are a subset of the structural model 12 and share at least one element with the minimal competition for each symptom. The purpose of fault diagnosis is to find all candidates for a given set of symptoms.

For example, (C1, C2, C3
, C4, C5, C6, C7, C8).

Assume that a symptom S1 is given to this. CI, C2, . . . in this structural model 12 may be considered to correspond to elements or components such as an AND circuit, an OR circuit, or an adder-1 multiplier, taking a digital circuit as an example.

When the symptom S1 is given, the competition analysis unit 11 shown in FIG.
represents all subsets of the structural model 12 as the empty set (φ
) in ascending order of magnitude, and the motion prediction unit 13 is activated. Then, the motion predicted by the motion prediction unit 13 and the symptom S1
Find a minimal subset X1 that is inconsistent with and define it as a minimal conflict. now.

what if.

Let X1=(CI, C3, C6, C8).

The candidate generation unit 14 generates a set of symptoms (
A set Z1 of candidates for Sl) is found from Xl. In this example.

Z1= ((C1), (C3), (C6), (CB)
).

Then, given symptom S2, do the same thing.

The conflict analysis unit 11 determines the minimum conflict X2. what if.

Let X2=(C1, C3, C4).

The candidate generation unit 14 generates a symptom set (S
A set Z2 of candidates for L, S2) is determined. In this example.

Z2 = ((C1), (C3), (C4, C6).

(C4, C3)) It is. Furthermore, due to symptom S3.

When X3=(C1,,C6) is found, the set of candidates is.

Z3= ((CI), (C3, C6), (C4, C61
).

This suggests that there are three possible causes: "component CI is abnormal,""components C3 and 06 are both abnormal," and "components C4 and 06 are both abnormal."

[Problem that the invention seeks to solve]

In fault diagnosis using the conventional method described above, even for faults that occur relatively frequently and for which the cause can be easily identified, the diagnosis is performed in the same manner as for faults that occur only rarely, requiring an extremely large amount of calculations. Requires.

As a solution to this problem, 2 For example, if a particular symptom is typical, in order to avoid recalculation related to it, first memorize a set of input symptoms and a candidate set that is the diagnosis result. It is conceivable to search for the same input among these experience examples. However, this requires a search time that is of the linear order of magnitude proportional to the number of stored experiences.

The present invention aims to solve the above-mentioned problems, and provides a means to easily and quickly diagnose frequently occurring failures, and to automatically enhance the scope and accuracy of the diagnosis. The purpose is to

[Means for solving problems]

FIG. 1 is a block diagram of the principle of the present invention.

In FIG. 1, the same symbols as in FIG. 7 correspond to those shown in FIG. 7, and 20 represents a competition recording section, 21 represents a decision tree generation section, 22 represents a decision tree, and 23 represents a simple diagnosis section. .

The processing device 10 has a structural model 12 that describes the structure of the device that is the target of failure diagnosis, inputs symptoms through the input/output section 15, and sends the symptoms to the conflict analysis section 11. Motion prediction unit 13. The candidate generation unit 14 performs the diagnosis as explained in FIG.
Outputs candidates for failure causes.

In particular, in the present invention, the conflict recorder 20 records the minimal conflicts determined by the conflict analyzer 11 in response to each symptom. The decision tree generation unit 2I includes a competition recording unit 2
This process extracts those with a predetermined frequency or more from among the minimum conflicts recorded by 0 and automatically generates a decision tree 22 regarding failure cause candidates.

When the decision tree 22 has been generated for fault diagnosis, the simple diagnosis section 23 uses the conflict analysis section 11 and the operation prediction section 1.
3. Before the candidate generation unit 14 performs a deep search, an initial diagnosis is performed using the decision tree 22.

[Effect]

In the present invention, based on past experience, a decision tree 22 for diagnosing frequently occurring symptoms is automatically generated. At the time of one diagnosis, this decision tree 22 is first used to try to identify the fault location, and when identification fails, the decision tree 22 is automatically generated. Diagnosis using the usual method is performed only in cases where:

In the simple diagnosis using the decision tree 22, for example, frequently occurring symptoms are diagnosed based on the following assumptions.

+a) There was only one failure.

(bl) All symptoms that include the failure location in minimal contention can be observed.

According to the simple diagnosis using the decision tree 22, there is no need to perform an exhaustive search for the cause of the failure, so it becomes possible to identify the failure in a logarithmic order of time for the type of failure. In addition, since the decision tree 22 is automatically generated by the conflict recorder 20 and the decision tree generator 21, as one diagnosis experience is gained, the decision tree 22 is strengthened and the types and accuracy of faults that can be easily diagnosed are improved. It will improve.

〔Example〕

FIG. 2 shows an example of a system for implementing the present invention, FIG. 3 shows an example of minimal competition record information according to an embodiment of the present invention, and FIG. 4 shows an example of a decision tree generated in an embodiment of the present invention.

FIG. 5 shows an example of processing by the decision tree generation section, and FIG. 6 shows an example of processing by the simple diagnosis section.

The present invention is implemented by, for example, an inference system as shown in FIG. In FIG. 2, 30 is an input device, 31
3 is an inference engine that interprets knowledge and executes inference; 32 is a knowledge base that stores knowledge for solving problems; and 33 is an output device such as a display. A structural model I2 that describes the structure of a device to be diagnosed is treated as a knowledge base 32.

In this embodiment, failure diagnosis is performed by inputting symptoms from an input device 30 such as a keyboard.

By using the input device 30 as various sensors for abnormality detection, failure detection and diagnosis can also be automated.

The information recorded by the competition recording section 20 shown in FIG.

For example, as shown in FIG. 3, there are a symptom and a minimal conflict corresponding to the symptom determined by the conflict analysis section 11. Also, frequency information regarding how many times the event occurred is updated and recorded each time.

In the example shown in FIG. 3, the minimum competition determined for the symptom Sl is fcl, C3, C6, C8), and this symptom has occurred n1 times. Also, the minimal competition found for symptom S2 is (CI.

C3, C4), which has occurred n2 times.

Furthermore, the minimal competition for symptom S3 is (C1, C6), and this symptom has occurred n3 times. The competition recorder 20 shown in FIG.
Information as shown in FIG. 3 is accumulated.

The decision tree generation unit 21 uses this recorded information.

For example, a binary decision tree 22 as shown in FIG. 4 is generated. As shown in FIG. 4, each node of the decision tree 22 consists of the causes of failure that have been narrowed down so far and the symptoms that should be questioned, and the "leaves" consist of the causes of failure that have been narrowed down.

The process of the decision tree generation unit 21 that generates this decision tree 22 is as follows.
For example, this is done as shown in FIG. Decision tree generation unit 21
2 is automatically activated when a predetermined number of times has been reached, for example, when the number of diagnoses reaches 10,000. Alternatively, it is activated by a user's instruction. ■ to ■ in the following description correspond to processes ■ to ■ shown in FIG.

(1) Based on specified criteria, symptoms that have occurred frequently in the past, for example, 5% or more of the number of past diagnoses, and their minimal conflicts are extracted from the conflict recorder 20.

■ Perform the following processes (al to (C1) for all extracted symptoms St.

La) If there is already another condition with the same minimal competition as Sr, discard that Si.

(b) If there is only one type of minimal competition for Si, let that minimal competition be Xi.

(If there are two or more types of minimal competition for cl St, let their intersection set be Xi.

■ Create a root node and use the union of all Xi as the narrowed down cause of the root node.

■ Push the root node onto the stack.

■ Next, determine whether the stack is empty. If the stack is empty, the process ends.

■ Pump a node from the stack and set it as N.

(2) Determine whether there is a symptom St that satisfies the following conditions (1) to (3).

Condition (1): At any node on the path from the root node to N, the symptom Si is not a symptom that should be questioned.

Condition f21: The intersection set of N narrowed down causes and Xi is not empty.

Condition +31: The narrowed down cause of N is not a subset of Xi.

When there is no symptom Sj that satisfies such conditions.

Control is returned to process (■) and the process is repeated in the same manner.

■ If there is a symptom St that satisfies conditions (1) to (3), N
Extend the left branch of , create a new node, make the narrowed down cause the intersection of N narrowed down causes and Xi, and push the new node onto the stack.

■ Extend the right branch of N and create a new node.

The narrowed down causes are the remaining narrowed down causes of N, excluding Xi, and a new node is pushed onto the stack. After that, control is returned to process (2), and the process is repeated in the same way until the stack is empty.

When an abnormality occurs, the simple diagnosis section 23 shown in FIG.
It is activated automatically or by a user's instruction before the competition analysis unit 11 performs a diagnosis.

The simple diagnosis section 23 executes 9, for example, the processes 1 to 2 shown in FIG. 6 based on the decision tree 22 shown in FIG.

■ First, let N be the root node of the decision tree 22.

■ Determine whether N is a leaf. If it is a leaf,
Transfer control to processing ■.

■ Transfer the symptoms that N should ask about to the input/output unit 15.

Ask users if they observe the symptoms.

■ If the answer is ``observed,'' control is transferred to process ■; if the answer is not observed, control is transferred to process ■.

■ Go to the left branch of N and make that node a new N.

After that, control is returned to process (2).

■ Go to the right branch of N and make that node a new N.

After that, control is returned to process (2).

■ If N is a leaf, narrow down the cause of N.

It is transferred to the human output unit 15 and treated as one answer.

Note that instead of asking the user questions about symptoms, observed events may be checked using input from a sensor for detecting an abnormality.

In addition, frequency information based on past experience is set in the nodes of the decision tree 22, and this frequency information is output together with the cause of failure at the time of one answer, so that it can serve as a guideline for certainty through simple diagnosis. You can.

〔Effect of the invention〕

As explained above, according to the present invention, by using a decision tree, it is possible to quickly diagnose relatively frequent symptoms by omitting behavior prediction using a structural model. .

Moreover, this decision tree is automatically strengthened, and the range and accuracy of two diagnoses improve according to the number of times of two diagnoses.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the principle of the present invention. FIG. 2 shows an example of a system for implementing the present invention. FIG. 3 is an example of minimal competition record information according to an embodiment of the present invention. FIG. 4 is an example of a decision tree generated in one embodiment of the present invention. FIG. 5 shows an example of processing by the decision tree generation section. FIG. 6 shows an example of processing by the simple diagnosis section. FIG. 7 shows a block diagram of the conventional system. In the figure, 10 is a processing device, 11 is a competition analysis unit, 12 is a structural model, 13 is a motion prediction unit, 14 is a candidate generation unit, 15 is an input/output unit, 20 is a competition recording unit, 21 is a decision tree generation unit, 2
2 represents a decision tree, and 23 represents a simple diagnosis section.

Claims (1)

[Claims] The system has a structural model (12) to be diagnosed, and uses operation prediction based on the structural model to find minimal conflicts of each input symptom, thereby finding minimal candidates for the cause of the failure. In the desired computer-based failure diagnosis method, there is provided a conflict recorder (20) for recording the minimal conflicts determined in response to each symptom, and a conflict recorder (20) for recording the minimal conflicts determined above in response to each symptom; A fault diagnosis method comprising: a decision tree generation unit (21) that generates a decision tree regarding cause candidates; and a simple diagnosis unit (23) that performs an initial diagnosis using the decision tree at the time of fault diagnosis.