JPH04143808A - Diagnostic device - Google Patents

Abstract

PURPOSE:To obtain an exact rule in a short time by producing and correcting automatically a rule containing a trouble tree form. CONSTITUTION:A learning block 10 includes an induction 11 which is needed when a knowledge acquired previously is automatically learnt in a trouble tree form. An inference block 20 where the rule produced from the induction 11 is down-loaded is added together with a signal analyzing block 40 which analyzes the induction data, and a display device 60 which displays the diagnos tic result of the induction data together with a trouble tree. In other words, the technology of the induction 11 is applied as a learning function. Thus the relations set among various parameters, i.e., the relation between the input data (indication) and the output data (trouble factor) are described in a trouble treee form and also learnt and produced automatically as an IF-THEN produc tion rule for display of a diagnostic process for output of the output data. As a result, a rule is exactly obtained in a short time.

Description

Detailed Description of the Invention [Industrial Application Field] The present invention relates to a diagnostic device used for the health of systems and equipment in general industries such as power plants, factories, and buildings, or for medical treatment. It concerns a short and reliable diagnostic device.

[Prior Art] Generally, when a failure occurs in equipment that is in operation, some symptoms appear, so in order to diagnose equipment abnormalities, the relationship between symptom data and the cause of the failure must be determined using diagnostic inference rules (hereinafter simply referred to as rules). ), and based on this rule, the cause of the failure can be identified retroactively from the symptom data actually obtained.Normally, this type of rule is
A production rule of the form -THEN is used.

Traditionally, the relationship between symptom data and failure causes has been expressed using a failure tree (
There are equipment diagnosis systems on the market that have expert systems organized in the format of FT diagrams and described using production rules. A method for improving diagnostic accuracy by reflecting this in a fault tree is described. However, in this case, manually creating a fault tree requires a great deal of effort and time, and as the number of related parameters increases, it becomes difficult to accurately grasp and describe the relationships between the parameters.

In addition, examples of diagnostic devices with automatic learning functions include:
Although there is a method described in Japanese Patent Application Laid-open No. 1-270623, it does not consider learning in the form of a fault tree, so it is not practical due to poor man-machine performance.

Furthermore, a learning method using a neural network as shown in FIG. 7 has also been proposed. In Figure 7, (1) is the input layer, (2) is the intermediate layer, (3) is the output layer, and (1a
) to (1f) are input nodes corresponding to symptom data, (3
^) ~ (3D) are output nodes corresponding to the cause of failure.

In this case, input nodes (1a) to (1
f) and the output nodes (3^) to (3D) indicating the cause are related by statistical weights, but the process concluded as the cause becomes a complete black box. Therefore, in a diagnostic device using a neural network, since the reason for deriving the cause is not presented, the user is not very man-machine, and it is impossible to judge whether or not the answer (cause) can be trusted.

[Problems to be Solved by the Invention] As mentioned above, conventional diagnostic devices have the problem of requiring a lot of effort and time because they are trained manually when using a fault tree. . Furthermore, when a neural network is used, the process of deriving the cause of a failure becomes unclear, resulting in a problem of poor man-machine performance.

This invention was made to solve the above-mentioned problems, and by automatically creating and modifying rules in the form of a fault tree, it improves man-machine performance by associating symptoms and causes. Another object of the present invention is to obtain a diagnostic device that can generate rules accurately and in a short period of time.

[Means for Solving the Problems] A diagnostic device according to the present invention includes a learning block including a rule induction for automatically learning knowledge obtained in advance in the form of a fault tree, and a learning block including a rule induction for automatically learning knowledge obtained in advance in the form of a fault tree. This system includes an inference block to which rules are downloaded, a signal analysis block to analyze symptom data, and a display device to display diagnostic results of the symptom data together with a fault tree.

[Operation] In this invention, by introducing rule induction technology as a learning function, the relationship between various parameters, that is, the relationship between input data (symptoms) and output data (failure causes), is expressed in the form of a failure tree. In addition to describing I
It is automatically learned and generated as a production rule of F-THEN, and allows the user to examine the diagnostic process in deriving output data.

[Example] Hereinafter, an example of the present invention will be described with reference to the drawings. 1st
The figure is a block diagram showing one embodiment of the present invention.

In the figure, (10) is a learning block into which failure cause information J based on expert knowledge is input;
Rule induction (
11) outputs a rule R based on the failed woodwork and the failed tree T, and also outputs the failure cause information J or the analysis result of the symptom data (
(described later), the fault tree is automatically learned.

(20) is display knowledge! (21) and analysis procedure knowledge (22)
is an inference block in which is stored, and a learning block (10
) The latest rule R generated from the rule induction (11) is downloaded to the analysis procedure knowledge (22).

(30) is a signal database block into which symptom data K including various signal data S such as equipment monitoring data is input, and includes a raw data file (31), a processing result file (32), and an analysis history file (33). ) and database management (34), and is configured to generate management data U including symptom data K and analysis results Q.

(40) is a signal analysis block that analyzes the signal data S input from the signal database block (30) and generates an analysis result Q, and a smoothing block (41) that smoothes the signal data S; A differentiation block (42) that differentiates data S, an AR model (43) that gives mathematical meaning to signal data S, and an FFT (44) that performs frequency analysis of signal data S using fast Fourier transform.
and a trend (45) for analyzing the tendency of the signal data S.

(50) is the diagnosis result ■ from the inference block (20),
This is a graph display block into which the failed woodwork from the learning block (10) and the management data U from the signal database block (30) are input, and the diagnostic process display block (51) and the time series display block (52). and a spectrum display block (53).

(60) is a CRT as a display device connected to the graph display block (50);
0), symptom data K, analysis results Q, failure tree T, diagnosis results V, their processes, etc. are displayed.

Next, the general operation of the embodiment of the present invention shown in FIG. 1 will be described.

First, failure cause information J is input into the rule induction (11) in the learning block (10) based on the knowledge of the expert. Rule induction (11) automatically converts failure cause information J into failure tree T and rule R, and converts rule R into analysis procedure knowledge (22) in inference block (20).
), and the fault tree T is also downloaded to the graph display block (50). Fault tree T and rule R generated from rule induction (11)
is constantly learned and updated by management data U including new failure cause information J or analysis results Q.

The analysis procedure knowledge (22) in the inference block (20) compares the symptom data K input via the signal database block (30) with the rule R, performs diagnosis, and generates the diagnosis result V.
The 0 diagnostic result ■ that generates the graph display block (50
) and CRT (60). The management data U from the signal database block (30) includes symptom data as well as analysis results Q, etc., and is input into the analysis procedure knowledge (22).

The signal database block (30) stores the symptom data K in the raw data file (31) in the form of raw data;
This is transferred as signal data S to the signal analysis block (40). Also, under the control of database management (34),
Analysis result Q subjected to primary processing in the signal analysis block (40)
is stored in the processing result file (32>), and the history of the analysis result Q is stored in the analysis history file (33).Then, the management data U including the signal data S, the analysis result Q, etc. is stored in the learning block (10), Transfer to an inference block (20) and a graph display block (50).

The signal analysis block (40) is a smoothing block (41)
, differential block (42), AR model (43), F F
Analyzing signal data S included in symptom data K using various methods using T (44) and trend (45),
A meaning is given to the signal data S (for example, the vibration of the frequency component is "large", etc.).

The graph display block (50) displays the diagnosis result V and
The raw data (signal data S) and the primary processing data (analysis result Q) of the failed woodwork and symptom data necessary to explain the diagnosis process are displayed on the CRT (60).

Next, using the case of diagnosing an abnormality in a rotating machine in a power generation plant as an example, we will specifically explain the rule-forming operation in the learning block (10) in Figure 1 while referring to the explanatory diagrams in Figures 2 to 6. explain. In this case, rule R indicates the relationship between phenomena (symptom data) that appear in monitoring parameters when a hypothetical abnormality is the cause.

Generally, experts familiar with this type of technology can predict the phenomenon that will occur when, for example, a bearing fails, judging from the structure of the object. Therefore, for example, as shown on the horizontal axis of FIG. 2, the operator has operational knowledge between the cause of an abnormality (in this case, a bearing abnormality) and the phenomenon. Also, as shown in FIG. 5, for example, if the voltage balance collapses as a cause of an abnormality, the vibration intensity increases, a buzzing sound is generated, and the current value fluctuates.

Also, based on many years of usage experience, if the opposite phenomenon is observed, what is the cause?

They also have inductive knowledge based on maintenance inspection results.

The different knowledge in this case is 0 shown on the vertical axis of FIG. 5, for example, the knowledge that the cause of the increased vibration intensity is voltage imbalance or alignment failure.

Fragmentary knowledge such as that shown in Figure 2 or Figure 5 is entered into the learning block (10) as appropriate via a keyboard or the like. All you have to do is input the relationship "Temperature Rise" at random. The more you have this knowledge, the more reliable the Rule R and diagnosis will be. It is recommended that you enter it.

FIG. 2 shows the knowledge of only bearing abnormalities among motor abnormalities in a simplified manner. Hereinafter, with reference to Figures 1 to 4, we will discuss the causes of bearing abnormalities. A process of generating a relationship between a bearing abnormality and symptom data as a failure tree T based on lA~work, and a process of generating a rule R in a production rule format based on a failure woodwork will be described.

For example, if a bearing abnormality occurs, the bearing temperature will rise, vibration will increase, and vibration noise called data will be generated.
Suppose that the following knowledge A is input. Next, inductive knowledge is added that when the bearing was abnormal, the ``bearing temperature was high'' and the ``vibration was large'', but ``there was a sound called data''.

Such a fragmented knowledge group is input sequentially and is configured as shown in FIG.

If the knowledge TlaA to TlaA in FIG. 2 is represented diagrammatically, it can be positioned as a set of cells in a three-dimensional space corresponding to three symptom data, as shown in FIG. Here, each axis indicates the high/low bearing temperature, the magnitude of vibration, and the tone, and the ×
indicates that the bearing is abnormal, and Q indicates that the bearing is normal.

First, in order to generate a woodwork with a failure as shown in Fig. 4, it is possible to classify the cell set shown in Fig. 3, for example, by the magnitude of vibration, and to determine whether it is possible to classify the cell set in Fig. 3 into the same result, that is, only O or only ×. Examine. In this case, if it is a vibrating vehicle, only O (bearing normal), and if the vibration is large, only × (bearing abnormal), which corresponds to (4a) and (4b) in Figure 4, respectively. Can be expanded into a tree.

If the vibration is "slightly large", O and × are mixed, so next classify the tone (data or rattling).
If the sound “data” is being generated, ○(
(4c), (4d), and (4c), respectively.
and (4e) can be expanded into a fault tree corresponding to (4e). In this way, a fault tree T as shown in FIG. 4 can be generated.

In the above classification process, it becomes clear that the bearing temperature factor is not a factor representing bearing abnormality, so it is excluded from the failure tree T.

This makes it possible to minimize only the observation data necessary to identify the event, and eliminate waste such as providing unnecessary sensors.

Also, based on the fault tree T in Fig. 4, the rule R can be easily written. For example, IFr vibration is large" →T
HEN r Bearing abnormality” IFr vibration vibration wheel → THEN r
A rule R such as "Bearing normal", "IFF vibration is slightly large"&"Rattlenoise" → THEN r Bearing abnormal" IFr vibration is slightly large"&"Datasound" → THEN r Bearing normal" is generated, and this is applied to the inference block (20
), it can be used as an actual rule for diagnosis.

Note that the process of generating failed woodwork differs depending on the method of classification; for example, classification based on sound may be followed by classification based on vibration. Therefore, the learning block (10) is
After generating fault trees by various classification procedures, (4&)
~(4e) A fault tree with the fewest branches and which can complete the determination in the shortest path is finally adopted.

With reference to FIGS. 1, 5, and 6, a case will be explained in which a faulty woodwork as shown in FIG. 6 is generated based on knowledge related to various causes of abnormalities as shown in FIG. .

Figure 5 shows the knowledge needed to identify the causes of motor abnormalities, such as voltage imbalance, loss of cooling water, poor alignment, etc.
It shows the relationship between multidimensional causes of abnormalities, such as oil whirl and poor foundations, and observed abnormal symptoms.

In this case, as described above, the first bearing temperature, second bearing temperature, vibration intensity, vibration sound, current, waveform, etc. are classified in order, and each phenomenon is narrowed down to one specific cause of the abnormality. I'll go.

First, when classifying based on the first bearing temperature, when the first bearing temperature is "normal", "voltage imbalance" and "defective alignment" are extracted, and when the first bearing temperature is "increased", "oil whirl" and "cooling water" are extracted. "Loss" is extracted, [high temperature j
When , only "oil whirl" is extracted.

Next, for example, when the first bearing temperature is normal, classification is performed according to the second bearing temperature in order to classify "voltage imbalance" or "malalignment". In this case, for malalignment, Since there are contradictory events in which the second bearing temperature is "stable" and "unstable," the second bearing temperature is used as a bus and the following classification is performed based on vibration intensity.
However, in this case as well, since there is a contradictory phenomenon in which the vibration intensity is "normal" and "large" for the oil whirl, it is similarly passed and the next classification based on vibration sound is performed.

As a result, it is possible to specify "voltage imbalance" when the vibration sound is [Dane j], and it is possible to specify "defective alignment" when the vibration sound is "rattling" or "jiggling".

Furthermore, by classifying based on the second bearing temperature, when the vibration noise is "rattling", if the second bearing temperature is "stable", it is a "bad alignment", and if it is "unstable", it is a "foundation defect". ” can be specified.

On the other hand, if the first bearing temperature is "increased", then
Classification is performed based on the second bearing temperature, and the second bearing temperature is
When the condition is "stable," it can be identified as "oil whirl," and when it is "unstable," it can be identified as "loss of cooling water."

Through the above procedure, the fault tree T corresponding to (6a) to (6j) in FIG. 6 is expanded and generated. in this case,
Vibration intensity, current, and waveform are excluded from the fault tree T and rule R because a single cause of the abnormality cannot be identified. Based on the fault tree T completed in this way, the rule R expanded in the IF-THEN format can be generated in the same way as described above.

Fault tree T generated by rule induction (11)
and rule R are downloaded to the graph display block (50) and the inference block (20), respectively, in the same manner as described above. Then, during actual diagnosis, symptom data K (
It is used as software for diagnosing and displaying diagnostic results (actual data).

Further, the knowledge data as shown in FIG. 5 is filed in the learning block (10) as a database for learning the rule R to be carried out in the next stage.

In addition, if there is no knowledge about the second bearing temperature, the broken woodwork branches (6e) and (6f) and (
6h) and (61) become impossible. Therefore, if the first bearing temperature is ``normal'' and the vibration sound is ``rattling,'' it is considered either ``defective alignment'' or ``foundation failure,'' and if the first bearing temperature is ``increased.'' teeth,
The malfunctioning tree T will be stopped as either "oil whirl" or "cooling water loss".

In addition, if the knowledge regarding the "foundation defect" in Figure 5 has not been input, there is no need to classify "rattling noise" in relation to "alignment defect", and "foundation defect" due to the second bearing temperature is not required. There will be no separation between the two.

On the other hand, whenever new knowledge is input, the learning function is activated and the classification regarding all factors is performed again. Therefore, if the symptom data K and alarm input etc. from the signal database block (30) are continuously input to the learning block (10), and these simultaneous triggers are applied, the related data will be automatically input. It is also possible to learn by

In this way, rule induction (1
By introducing the technology 1), it is possible to describe the relationship between various parameters, that is, the relationship between input data (symptoms) and output data (fault causes) in the form of a fault tree, and also Input data (symptoms) associated with tree T
Since the production rule format rule R is automatically modified and generated based on the output data (causes), the user can check the validity of the diagnosis process in the derivation by following the process of establishing the derived diagnosis result V. can be examined closely.

In addition, by sorting out the complicated cause-and-effect relationships shown in Figure 5 and extracting only the data necessary for diagnosis, it is possible to automatically create woodwork with the minimum number of failures as shown in Figure 6 and rule R based on this. By generating and downloading to the inference block, accurate rules R can be created and learned in a short time.

Furthermore, it is possible to easily modify or add causal relationships.

Furthermore, by downloading the failed woodwork to the graph display block (5o), the failed tree T can be drawn accurately in a short time, and the man-machine interface can be improved.

In the above embodiment, a case has been described in which equipment abnormality diagnosis is performed, but the present invention can be applied to the relationship between patient's symptoms, test data, and disease name, and the same effect can be achieved as a medical diagnostic device.

In addition, if it shows a causal relationship, such as organizing the relationship between the yen rate or stock price and the factors that are affected by it,
Similarly, it can be expressed as a rule and applied to questions to be studied.

Further, in FIG. 1, the learning block (10) is configured integrally with the diagnostic device, but it may be separated from the diagnostic device and used as a stand-alone block off-line.

[Effects of the Invention] As described above, according to the present invention, a learning block including a rule induction for automatically learning knowledge obtained in advance in the form of a fault tree, and a rule generated by the rule induction are provided. An inference block to be downloaded, a signal analysis block that analyzes symptom data, and a display device that displays diagnostic results of symptom data together with a fault tree are provided, and input data (symptoms) and output data (failure causes) are provided.
In addition to automatically generating rules based on the fault tree by describing the relationship between the
This has the effect of improving man-machine performance by associating symptoms and causes, and providing a diagnostic device that can generate rules accurately and in a short period of time.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is an explanatory diagram showing an example of the knowledge group input to the learning block in FIG. 1, and FIG. 4 is an explanatory diagram showing a fault tree generated based on the knowledge group in FIG. 2, and FIG. 5 is an explanatory diagram showing the knowledge group input to the learning block in FIG. 1. FIG. 6 is an explanatory diagram showing an example of a fault tree generated based on the knowledge group in FIG. 5. FIG. 7 is an explanatory diagram showing a neural network used in a conventional diagnostic device. . (10)...Learning block (11)...Rule induction (20)...Inference block (50)...Signal analysis block (60)...CRT (display device) J...Failure cause Information K...Symptom data T...
・Fault tree R...Rule ■...Diagnosis result In the diagram, the same reference numerals indicate the same or equivalent parts.

Claims (1)

[Scope of Claim] A diagnostic device that determines the cause of an abnormality based on previously obtained knowledge and actual symptom data, comprising a learning block that includes a rule induction for learning the knowledge in the form of a fault tree; An inference block for downloading rules generated by the rule induction, a signal analysis block for analyzing the symptom data, and a display device for displaying diagnostic results of the symptom data together with the fault tree. Characteristic diagnostic equipment.