JPH04218840A - Fuzzy inference method and data evaluating method for fuzzy inference - Google Patents

Abstract

PURPOSE:To perform reliable inference even in the case of dropout of a part of input data by reducing the membership function in a condition part or a conclusion part in the inference process in accordance with the degree of importance to the conclusion part of dropped-out data. CONSTITUTION:When an operator 300 inputs '0' to an input picture or inputs nothing (NULL) in the case of drop-out of input data, this information is recorded in an answer frame 60 through a user interface 400. A drop-out data detecting part 800 detects dropped-out data and changes all membership functions for this data, and data is substituted with a reference value. That is, the detecting part 80 changes membership functions for dropped-out data during inference so that they are '1' to a domain, and the value of this data is set to an arbitrary value of the domain. The detecting part 80 reduces membership functions in accordance with the degree of importance to the conclusion part of dropped-out data.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] The present invention relates to an inference method when a part of input data is missing in fuzzy inference, and a data evaluation method for the inference, especially when reliability of inference such as life diagnosis of metal materials is a problem. This paper relates to fuzzy inference methods and data evaluation methods when

0002

[Background Art] In recent years, with the significant improvement in the processing power of computers, specialized knowledge has been used to perform some of the advanced intellectual decision-making tasks such as clinical diagnosis by doctors, equipment failure diagnosis, legal consultation, and program design. Expert systems that utilize and support computers are attracting attention. Among the inference methods of expert systems, fuzzy inference is a method of inference based on vague human thoughts such as ``the temperature is quite high'' and ``the speed is quite slow'', and it is possible to reduce the number of rules in the expert system. It has characteristics such as being able to make comprehensive judgments even when there are contradictory things in the knowledge.

FIGS. 39 and 40 show a general fuzzy inference method, taking brake control in driving a car as an example. In this example, the following distance and traveling speed are input values (
The brake pressure (conclusion part) for the input value is determined by fuzzy reasoning, and the membership functions of the condition part and conclusion part are defined for the two fuzzy rules (rule 1 and rule 2), and their compatibility is calculated using fuzzy reasoning. The representative value of brake pressure (center of gravity of overall compliance) is calculated by integrating the degrees. Fuzzy inference is often used for control because it infers a representative value for an input value and feeds back that value. Recently, it has come to be used for diagnosis as well. For example, in order to evaluate creep damage to metal materials, we need information that cannot be converted into damage rates, such as changes in hardness or changes in metal structure, or information such as ``the outside diameter of the tube is swollen'' or ``crystal grains are deformed.'' The creep damage rate is calculated by performing fuzzy inference based on information based on human subjectivity.

[0004] Furthermore, it is well known that in equipment used for long periods of time under high temperature and high pressure, such as in thermal power plants and chemical plants, the materials used undergo creep, fatigue, or age damage during operation, resulting in deterioration of the materials. ing. This kind of material deterioration is controlled by the temperature, applied stress, and usage time of the material, and in thermal power boilers, it usually
Designed to have a lifespan of 1,000,000 hours. but,
Many boilers have been used for more than 100,000 hours, and even within 100,000 hours, the material has deteriorated due to abnormal material deterioration due to an increase in metal temperature due to uneven flow of combustion gas or segregation in the material. There have also been accidents where the equipment has been damaged. Against this background, technology to extend the life of a plant by accurately predicting the remaining life of materials and performing planned partial replacements and repairs has become important.

[0005] Techniques for directly detecting the remaining life of a material are broadly classified into destructive methods and non-destructive methods, but if the remaining life of a material can be detected non-destructively, evaluation time can be shortened and costs can be reduced. Furthermore, it is extremely effective because the same location can be monitored periodically. As a non-destructive method, a replica method is widely used in which the metallographic structure of a material is transferred onto a replica film, observed, and evaluated accordingly. For example, methods for evaluating creep damage include a method of observing the degree of deformation of crystal grains due to creep (Japanese Patent Application Laid-Open No. 63-228062), or a method of quantifying cavities (Thermal and Nuclear Power Generation Vo.
l. 39, No. 6, p75-86). Also,
Recently, a method has been developed that adds information on changes in metal structure and hardness to the above method for comprehensive evaluation (Japanese Patent Laid-Open No. 1-31126
No. 8).

[0006]

As described in the above-mentioned prior art, when fuzzy inference is used for control, the fitness at the time of inference does not matter, and control is performed by feeding back representative values. However, when used for diagnosis, diagnostic reliability may be important. In this case, reliability uses goodness of fit in fuzzy inference. The degree of conformity will be high if the data in the condition section show the same tendency (for example, data showing a tendency for the condition section to be damaged in a lifespan diagnosis), and conversely, it will be low if there is data with a different tendency. Therefore, the goodness of fit is an index indicating the reliability of the input.

By the way, when using fuzzy inference for control, the measured values that serve as input data are measured online,
Since the inference results (representative values) are fed back, no input values are lost. However, when used for diagnosis, measurements are not necessarily carried out online. For example, when diagnosing the lifespan of metal materials such as boilers, it is necessary to collect replicas, observe them under a microscope, and measure their hardness in order to collect data.

[0008] However, due to time constraints and the like, not all data can be measured, and input data may be missing. In such cases, inference is impossible. Furthermore, if the same membership function is used to force inference, the fitness tends to increase as shown in FIGS. 41 and 42. This figure shows the case where all data a, b, and c are input for conditions A, B, and C (Fig. 41), and the case where all data a, b, and c are input for conditions A, B, and C.
, C, the fuzzy inference results are compared when data C for condition C is missing (FIG. 42).

As a means of calculating the degree of conformity of the conditional part, the logical product method,
Although the algebraic product method is used as a weighting method for the conclusion section, this tendency remains the same even if other methods are used. In general, when input data is missing, the amount of information decreases and the reliability decreases, but if the same membership function is used as described above, the fitness increases and the apparent
Increased reliability.

[0010] Furthermore, as described in the above prior art, when fuzzy inference is used for control, the inference results must be fed back immediately, so only the representative value of the inference result is important, and the reliability of the inference is It's not a problem. However, when used for diagnosis, unlike in the case of control, immediate response is not required, but the reliability of inference is an issue.

[0011] In particular, the ``damage diagnosis system for metal materials'' described in the prior art uses human subjective opinions such as ``the outer diameter of the tube is swollen'' and ``the metallographic structure has changed significantly'' as input data. In addition, although hardness is measured, it may lack reliability as it is measured on-site. These factors may reduce the reliability of the inference. However, conventionally there was no function to check the validity of input data.

Furthermore, the expert system described in the above prior art is based on the assumption that non-experts will quickly diagnose creep damage on-site, and the input data shown in FIG.
As shown in Figure 4, replica observation results (state of metal structure, degree of deformation of crystal grains, degree of cavity formation), hardness, and degree of change in tube outer diameter are used, and things that can be measured such as hardness are Qualitative measurements that cannot be measured are ranked on the screen according to the degree of damage, and the operator is required to input this rank. These input values need to be re-entered each time the diagnosis site is changed on the operation data input screen shown in FIG. 35, which takes time.

Furthermore, the above-mentioned "damage diagnosis system for metallic materials" uses stress-dominated data (deformation of crystal grains, cavity formation, changes in pipe outer diameter) and age-dominated data (hardness, changes in metallographic structure). , weighting of creep damage is carried out according to the shape of the membership function, but since the weighting of creep damage in age-dominated data differs depending on the damage rate, it is difficult to infer accurately over the entire damage range. It was hot.

A first object of the present invention is to provide an inference method that can be used when part of input data is missing in fuzzy inference where reliability is an issue.

A second object of the present invention is to provide a method for checking the validity of input data in fuzzy inference when the input data is uncertain, such as human subjectivity.

A third object of the present invention is to provide an evaluation data inference system that eliminates the drawbacks of the prior art described above and allows fuzzy inference just by checking the default value.

A fourth object of the present invention is to provide a fuzzy inference system that eliminates the drawbacks of the prior art described above and can accurately evaluate creep damage over the entire damage area.

[0018]

[Means for solving the problem] (1) The first purpose is to lower the membership function for other input data (conditional part) in the inference process according to the importance of the missing data to the conclusion part. Alternatively, this can be achieved by lowering the membership function of the conclusion.

Missing data is determined by not inputting that data or by inputting "0" (NULL) and detecting it. When it is determined that data is missing, a frame with knowledge of the importance of the missing data for the conclusion section is called, and the rule section is changed to change the membership function of the condition section for the input data, or the membership function of the condensation section is changed. Decrease function. Also, since the conditions decrease, the condition part of the fuzzy rule also changes,
To deal with this, either modify the membership function for missing data during the inference process so that it does not affect the inference results, and then set the missing data to an arbitrary value in the domain, or deal with the missing data. Prepare rules for each individual.

(2) The above-mentioned second object is achieved as follows. First, fuzzy inference is performed using input data for multiple diagnostic parameters, and the reliability of the inference is evaluated. The integrated fitness is used as an index to express the reliability of the inference, and if the integrated fitness is lower than a preset value, one of the input data is changed between the domains,
Perform fuzzy inference on each numerical value. In this case, input values are used except for the data to be changed. The above procedure is repeated by sequentially replacing the data to be changed, and the representative value and integrated fitness in fuzzy inference are determined respectively. Compare these representative values and integrated fitness with the representative values and integrated fitness in fuzzy inference performed on the input data, and find if the difference in representative values is smaller than any value and the integrated fitness is sufficiently high. Select the conditions and record the combination of those data.

Next, among diagnostic parameters, if the data range can be inferred from surrounding information, the data range is inferred. It infers a data combination that matches this data range with the data combination in the fuzzy inference described above, outputs a comment based on the inference result, and requests the operator to reconsider.

(3) The above problem can be solved by inferring the estimated value of the input data on the evaluation data input screen shown in FIG. 34 and displaying the estimated value in the input area as a default value. can be achieved. In order to infer the estimated value of the input data, if the part has been previously diagnosed, data on the part is first searched from a database that stores the previous diagnosis data. The metal temperature is estimated based on the data, and input data for evaluation is inferred from the metal temperature, operating time, and pressure, and displayed as a default value in the input area. Furthermore, in the case of a first diagnosis, estimated values of each input data are inferred from the operation data (temperature, pressure, operation time) input on the operation data input screen shown in FIG. 35 and displayed. The operator checks the value displayed as the default value and continues inference if there is no problem. Furthermore, if the measured value or the operator's judgment differs from the displayed value, only that portion needs to be corrected.

(4) The fourth object described above is achieved by the following steps. First, the creep damage rate is calculated by fuzzy inference using data with a high degree of importance for the conclusion part, that is, stress-dominated data (deformation of crystal grains, formation of cavities, changes in tube outer diameter). Next, the amount of correction for the creep damage rate is determined by fuzzy reasoning using data that is less important to the conclusion, that is, age-dominated data (hardness, metallographic changes), and the creep damage rate described above is corrected. . In addition, the data check is performed using the degree of integrated fitness of each fuzzy inference as an index.

[0024]

[Operation] According to the present invention, fuzzy inference is possible even when a part of the input data is missing, and the membership function of the conditional part or the conclusion part can be changed in the inference process according to the importance of the missing data to the conclusion part. As a result, the degree of fitness depends on the number of inputs, and this eliminates the contradiction that in inference using degree of fit (reliability), the degree of fit increases when data is missing.

As mentioned above, reliability evaluation in fuzzy inference uses integrated fitness as an index. This is due to the following reason. If the integration fitness is high, it means that the trends in the input data are in the same direction; conversely, if the integration fitness is low, it means that the input data contains different trends. It means that there is. Therefore, a low integrated fitness means that the input data contains foreign matter, and the reliability of the inference is low.

Furthermore, when the integrated fitness is low, fuzzy inference is performed while changing one of the input data, and data that is close to the representative value of the fuzzy inference result with the original input data and has a high integrated fitness is obtained. By selecting combinations, it is possible to detect heterogeneous data that causes low inference reliability. In this case, only one of the input data is changed, and the other input data are used as they are, so the input data is reflected in the result. In addition, some input data can be inferred from surrounding information to a certain extent, although it is not certain. For example, in a ``metallic material damage diagnosis system,'' changes in metal structure and hardness can be roughly estimated from operating time and metal temperature if the plant has no abnormal operating history.

From the above two inferences, it is possible to check the validity of input data in fuzzy inference. Note that the only output is a display indicating that there is a possibility that the input data is incorrect; of course, the input data may be correct, so the operator makes that determination.

The degree of progression of creep damage is determined by temperature, stress, and time. If these values are known, the input data shown in FIG. 34 can be estimated. In general, operating time and pressure can be accurately determined, but temperature (metal temperature) is often uncertain. Therefore, if a diagnosis has been made previously, that information is searched from the database, and the previous evaluation input data and operating data are compared to estimate the metal temperature. Based on the estimated metal temperature, operating time, and pressure, the value of the evaluation data is inferred. Furthermore, in the case of a first diagnosis, inferences are made in the same way based on the input operating data (same as the design data).

Furthermore, as described above, the creep damage rate is mainly determined by stress-dominated data, and the importance of age-dominated data on the creep damage rate differs depending on the damage rate. That is, as shown in FIG. 34, when the damage rate is small, the decrease in creep strength is large due to material deterioration due to aging, but as the damage rate increases, the effect becomes smaller. The present invention was made based on the results of this research, and the creep damage rate is mainly inferred from stress-dominated data, and the age-dominated data is calculated based on the idea of Figure 34, in which the creep damage rate is corrected. Use with. This makes it possible to accurately infer the damage rate for all types of damage. Furthermore, the stress-dominated data and the aging-dominated data change independently, and in the present invention, the data is checked separately for each, and the above-mentioned problems can be solved.

[0030]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to the drawings. FIG. 1 shows the overall configuration of an expert system using fuzzy inference according to the present invention. This system 100 includes an input screen display section 40, an output screen display section 50, a fuzzy inference section 20, a judgment section 30, and a control section 10 that controls them, which are expressed in production rules (IF~THEN~ format).
The frame includes a fuzzy inference starting frame 70 and an answer frame 60 as frames for describing facts and performing procedural processing. A missing data detection section 80 is attached to the answer frame 60. Further, input and output are performed by the operator 300 on the display device 200 via the user interface 400.

Next, the functions of each component will be explained along the flow of processing, taking boiler heat exchanger tube creef life diagnosis as an example. First, when inference is started, the input screen display section 40
A command is issued to display an input screen as shown in FIG.
00 is displayed. In this example, the input data for life evaluation are: 1) degree of deformation of crystal grains, 2) degree of cavity formation, 3) degree of outer diameter change, 4) degree of structural change, 5)
Five pieces of data, such as the Vickers hardness, are input, but other information such as the color of the tube, the state of adhesion of ash, etc. may also be input. In this embodiment, regarding 1) to 4), four ranks are displayed on the display device 200 according to the degree of damage, and the operator 300 inputs the numerical value of the corresponding rank. Regarding 5), the measured value of Vickers hardness is input.

[0032] The input data is input to the user interface 400.
is temporarily recorded in the answer frame 60. Next, the input data recorded in the answer frame 60 is sent to the fuzzy inference activation frame 70. In addition to recording input data, the fuzzy inference starting frame 70 also records fuzzy inference results and names of fuzzy knowledge to be applied. The fuzzy inference unit 20 performs fuzzy inference using the input data recorded in the fuzzy inference activation frame 70 to calculate the creep damage rate, but in the present invention, inference is possible without inputting all data. Moreover, inference can be made with a degree of fitness (reliability of inference) according to the number of input data,
This will be discussed later.

The results of the fuzzy inference are sent to the judgment section 30, and if the degree of suitability (reliability of the inference) is low, a comment requesting reexamination or detailed evaluation is added, and the result is displayed on the output screen display section 50.
Pass the diagnosis results to. In the output screen display section 50, the judgment section 30
The results sent from the user interface 400 are displayed on the display device 200 via the user interface 400.

Next, fuzzy inference will be explained. Fuzzy knowledge is necessary to perform fuzzy inference, and fuzzy knowledge consists of fuzzy rules expressed as production rules and membership functions that give the fitness of various propositions. As shown in Figure 3, a fuzzy rule consists of a condition part proposition and a conclusion part proposition. Write a proposition with ambiguous expressions that humans can judge, such as [increase the creep damage rate] for the part. In this example, Figure 4
The following three rules are used. A membership function is defined for each of these propositions.

FIG. 5 shows an example of a membership function of a conditional proposition and a membership function of a conclusion proposition. The vertical axis is the degree of fitness, the horizontal axis is the rank (hardness value in the case of Vickers hardness) in the condition part membership function, and the creep damage rate in the conclusion part membership function. Since there are multiple propositions in the conditional part, it is necessary to calculate the degree of conformity of the condition. Generally speaking, there are the conjunctive method and the algebraic product method.The conjunctive method selects the minimum value among the degrees of fitness for each proposition, and the conjunctive method selects the product of the degrees of fitness of each proposition. is the algebraic product method. These means can be selected depending on the phenomenon.

The conclusion part is weighted based on the degree of conformity of the condition part obtained in this way. This method also includes the logical product method and the algebraic product method. The membership function of the conclusion part is F(X),
Each of these is expressed by the following equations, where α is the degree of conformity of the conditional part. Conjunctive product method conclusion part suitability = Min [F(X), α] Algebraic product method conclusion part suitability = α×F(X) These means can also be selected depending on the phenomenon.

As shown in FIG. 4, the subject of the conclusion part of the three fuzzy rules is the same, and the degrees of conformity for each rule determined above are integrated to arrive at the final conclusion. As shown in FIG. 6, weighted membership functions for each rule are synthesized (thick solid line) and their center of gravity is determined. This is a representative value, and is the creep damage rate in this example. Furthermore, as shown in the figure, the degree of conformity to the representative value can be determined, and this value becomes the reliability of the inference for the input data.

Next, a fuzzy inference method when a part of input data is missing, which is the key point of the present invention, will be described. Missing input data can be detected by the operator 300 inputting "0" on the input screen or by not inputting anything (NULL), the information is displayed in the user interface 4.
00 and is recorded in the answer frame 60. The missing data detection unit 80 detects missing data from the data recorded in the answer frame 60, and determines all the membership functions for the missing data (for example, if the rank is large,
Three membership functions (medium and small) Figure 7
Change the data as follows and replace the data with the reference value. The reference value can be any value as long as it is within the domain of the membership function; here, it is set to "1" for items 1) to 4), and 5).
The items were replaced with "100".

By changing the membership function in this way, this missing data no longer affects the inference results. At the same time, it has another frame knowledge of the importance of missing data on the inference results, and issues an instruction to change the membership function for the input data according to the importance. In this example, among the input items, 1), 2), and 3) have high importance because they are parameters that directly correspond to creep damage, and items 4) and 5) change due to exposure to high temperatures. This parameter has low importance because it does not directly correspond to creep damage.

FIG. 8 shows an example in which the membership function is changed, and in this embodiment, it is moved at a constant ratio with respect to the vertical axis as shown in this figure. This ratio is 0.5 times the original value if data in 1) to 3) is missing, 4), 5
) is missing, it is multiplied by 0.75, but this value is determined by inferring various cases. After changing the membership function and input value corresponding to the missing data and the membership function for other data, the data is sent to the fuzzy inference starting frame 70,
A fuzzy inference unit 20 performs fuzzy inference.

FIG. 9 shows the case where fuzzy inference is performed using the method of the present invention when one of the input data is missing (
The results are compared between case 1) and the case where fuzzy inference is performed without changing the membership function (case 2). Note that in the latter case, the fuzzy rules must also be changed, but corresponding fuzzy rules are created and inferred. For comparison, the results of inference obtained by inputting all data with other input values being the same are also shown (Case 3). As an inference method, the logical product method is used to calculate the fitness of the conditional part, and the algebraic product method is used as the weighting method for the conclusion part. Comparing Case 1 and Case 2, the representative values are the same, but the goodness of fit is greater in Case 2. The value indicates that all data has been entered (i.e. there is a large amount of information)
It is larger than Case 3 and there is a contradiction. Case 1
The fitness of case 2 is smaller than that of case 3.
It can be seen that the contradiction in the case of is resolved.

In the example, for missing data, the membership function for that data is changed as shown in Figure 7, and the data is set to an arbitrary value in the domain so that it does not affect the inference results. There is. Instead, by deleting the fuzzy rule condition part for the missing data, exactly the same inference result can be obtained.

FIGS. 10 to 14 are diagrams for explaining a second embodiment of the present invention. FIG. 10 is a diagram showing a display example of the initial input screen, and as shown in the diagram, the operation history of the target boiler can be input. 11 and 12 are diagrams showing input screens when the objects to be inferred are base metals and welded parts, and are slightly different from the case of FIG. 2 described above. FIG. 13 is a diagram showing a display screen of input data,
FIG. 14 is a diagram showing a display screen of diagnosis results. The other configurations, inference means, etc. are the same as those of the above-described embodiment, so their explanation will be omitted.

FIG. 15 is a diagram for explaining a third embodiment of the present invention. In the first and second embodiments, the condition part membership function for the input data is lowered according to the importance of the missing data to the conclusion part, but as shown in FIG. may be lowered. In this case, if the weighting means of the conclusion part is an algebraic product method, exactly the same inference result can be obtained. Furthermore, in each of the embodiments described above, if the degree of conformity is low, a re-examination or a detailed investigation is requested. Furthermore, if the diagnosis shows that the damage is large, a diagnosis of similar or nearby parts is requested.

Next, the plant life diagnosis apparatus including these will be explained. FIG. 16 shows an overall configuration diagram of the plant life diagnosis support device of the present invention. This device 501 includes a processing device 502, a display device 505, an input device 506, and an output device 507. The processing device 502 includes a data/knowledge base 503 and an inference calculation unit 504. This device 501
is connected to the host computer 508, and the operation history of various plants recorded in the host computer 508,
Periodic inspection records, accident history, etc. are searched and imported into the data/knowledge base section 503, which can be used as a database for this device. The inference calculation unit 504 performs inference and calculation based on the information input from the input device 506 according to the dialog screen output to the display device 505 and the data/knowledge base, and displays the results on the output device 507 and the display device 5.
Output to 05. Note that the results are sent to the host computer according to a certain format and become a new database.

FIG. 17 shows the processing flow of this apparatus. This device first creates an inspection chart. To do this, first, follow the dialogue screen output on the display device (CRT).
Data on the plant to be diagnosed, such as the plant name and inspection timing, are input from an input device (keyboard), and an inspection chart is created by inference and calculation based on this information and the data/knowledge base 525 for creating an inspection chart.

[0047] Inspection chart creation data/knowledge base 52
5 stores data on the design/structure, repair/structure, operation/history, inspection/history, accidents/damage, etc. of various plants, and allows you to search for target plants from input data, and search for similar (
Perform inferences and calculations such as searching for plants (type, capacity, and operation history), statistical processing, and accident probability calculations, and select parts with a high accident probability. A site with a high accident probability is a site that has a statistically high accident probability based on accident examples at similar plants, or a site that is estimated to be highly damaged based on operation history, etc.

As an output 511, as shown in FIG. 18, a list of parts that have a high accident probability and require inspection and an inspection method are output to a display device or an output device (printer). The inspection method selects and displays the minimum necessary inspection items from the inspection method item selection database. For example, in the case of a heat exchanger tube, visual inspection of the appearance, outer diameter measurement, etc. are displayed. In addition, the expected damage factors are also output for reference. In addition, if you select a part using an input device such as a mouse or keyboard, you can select a drawing from a database that stores drawings and show the target part in a drawing, and such drawings can also be output using an output device. I can do it. Furthermore, the inspection result record table shown in FIG. 19 can be outputted for each part, allowing smooth inspection work.

Next, in this apparatus, the inspection results are inputted from the input device at the input section 512, and damage evaluation is performed. Inspection results may be expressed in numerical values, such as hardness, or in text, such as scale conditions determined by visual inspection, but the inspection result record sheet mentioned above allows numerical and text input for inspection items. In the case of text input, the user can select from several results. Based on such input information and the damage evaluation data/knowledge base 527, inferences and calculations are performed to evaluate the lifespan. The damage evaluation data/knowledge base 527 includes data showing the relationship between evaluation parameters and the amount of damage,
It stores membership function data for inferring lifespan by fuzzy inference from several evaluation parameters including ambiguous quantities expressed in sentences.

FIG. 20 is an example of data showing the relationship between evaluation parameters and damage rate, and is master data showing the relationship between cavity area ratio and creep damage rate, which are creep damage evaluation parameters. Based on these data/knowledge bases and input data, lifespan is evaluated through calculations using master data, fuzzy inference, statistical calculations such as extreme value statistics, etc.

As a result of the damage evaluation performed by the damage amount inference/calculation unit 513, if the reliability of the evaluation is high and the damage is large and countermeasures are required, the countermeasure method selection knowledge/base 52 is used.
Based on 8, the optimal countermeasure method such as replacement, repair, modification, etc. is inferred and output to a display device or an output device. An example of the output is shown in FIG. Further, the parts similar to the severely damaged part that are not included in the inspection chart output in step 511 are inferred based on the similar part selection data/knowledge base 529, and an inspection request is output. This can eliminate oversights.

As a result of damage evaluation by the damage amount inference/calculation unit 513, if the reliability of the evaluation is high and no countermeasures are required, the inspection schedule creation data/knowledge base 5 is
30, the next inspection time, inspection items, inspection method, etc. are output. Furthermore, if temperature measurement data is required at the next inspection, it instructs the work that must be done during the inspection period, such as requesting the installation of a thermocouple, and outputs instructions for collecting the necessary data at the next inspection.

[0053] As a result of damage evaluation by the damage amount inference/calculation unit 513, if the reliability of the evaluation is low due to lack of information or input of abnormal data, etc., the evaluation is performed with a high degree of certainty. Additional inspection parts and items are inferred and output based on the inspection item and method selection data/knowledge base 526. For example, a request is made to collect a replica for detailed evaluation and calculate the cavity area ratio. In addition to additions, if the data is presumed to be inappropriate, a re-inspection may be requested, and if it is difficult to evaluate only by inspecting the target area, inspection of adjacent areas may be requested. Note that there are cases where a sample is taken and a destructive test such as a creep test is requested, but in this case, the results will be available several months later, and the device will end the process by instructing how to replace it.

[0054] When the additional inspection results are input, damage amount estimation/
Returning to the calculation unit 513, damage evaluation is performed again, and the same procedure is repeated. In this case, the amount of damage inferred and calculated will be comprehensively evaluated, including the previously collected inspection data.

As described above, the plant life diagnosis support system according to the present invention can perform various tasks such as creating an inspection chart, assessing damage, taking countermeasures, pointing out similar parts, instructing additional inspection items and methods, and creating future inspection schedules. The data is input and output in an interactive format to support plant life diagnosis work, but after the inspection is completed, this data is stored in the host computer and used as a database.

Next, another embodiment of the present invention will be described with reference to the drawings. FIG. 22 shows the overall configuration of an expert system using fuzzy inference according to the present invention. This system 100
are an input screen display section 40, an output screen display section 50, and
It consists of a fuzzy inference section 20, a judgment section 30, an input data evaluation section 90 attached thereto, a control section 10 that controls them, and a fuzzy inference activation frame 70 and an answer frame 60 as frames for describing facts and performing procedural processing. Further, input and output are performed by the operator 300 on the display device 200 via the user interface 400.

Next, using boiler heat exchanger tube creep life diagnosis as an example, the functions of each component will be explained along the processing flow. First, when inference is started, the input screen display section 40
A command is issued to display a screen for inputting the target plant and part, as well as operating data such as operating time and metal temperature.
It is displayed on the display device 200 via the user interface 400. The operator 300 inputs the necessary items according to the screen, and when the input is completed, a command is issued to display an evaluation data input screen as shown in FIG. 200 is displayed.

In this example, the input data for life evaluation are (1) degree of deformation of crystal grains, (2) degree of cavity formation, (3) degree of outer diameter change, (4) degree of structure change, (
5) Five pieces of data, such as the Vickers hardness, are input, but other information such as the color of the tube, the state of adhesion of ash, etc. may also be input. In this embodiment, regarding (1) to (4), four ranks are displayed on the display device 200 according to the degree of damage, and the operator 300 inputs the numerical value of the corresponding rank. Regarding (5), the measured value of Vickers hardness is input.

[0059] The input data is the user interface 400.
is temporarily recorded in the answer frame 60. Next, the input data recorded in the answer frame 60 is sent to the fuzzy inference activation frame 70. In the fuzzy inference starting frame 70, in addition to recording input data, a record of fuzzy inference results and the name of fuzzy knowledge to be applied are written. The fuzzy inference unit 20 performs fuzzy inference using the input data recorded in the fuzzy inference activation frame 70 and calculates the creep damage rate.

The results of the fuzzy inference are sent to the judgment section 30, and if the integrated suitability (reliability of the inference) is low, the input data is passed to the input data evaluation section 90, and the validity of the data is checked using the method of the present invention. I do. A comment based on the result and the creep damage rate inferred by the fuzzy inference section are passed to the output screen display section 50. In the output screen display section 50, the judgment section 3
The results sent from 0 are displayed on the display device 200 via the user interface 400.

Next, fuzzy inference will be explained. Fuzzy knowledge is necessary to perform fuzzy inference, and fuzzy knowledge consists of fuzzy rules expressed as progta- tion rules and membership functions that give the fitness of each proposition.

As shown in FIG. 28, a fuzzy rule consists of a condition part proposition and a conclusion part proposition. In the condition part, a proposition such as ``the rank of the degree of change of crystal grains is large'' is written for the input data, and in the conclusion part. Then, for the condition part, write a proposition that is vague enough for a human to judge, such as "increase the creep damage rate." In this embodiment, three rules shown in FIG. 29 are used. A membership function is defined for each of these propositions. Figure 30 shows an example of the membership function of the conditional part proposition and the membership function of the conclusion part proposition, where the vertical axis is the degree of fitness, and the horizontal axis is the rank (in the case of Vickers hardness, the hardness value) of the conditional part membership function. The membership function in the conclusion section uses the creep damage rate.

Since there are multiple propositions in the conditional part, it is necessary to calculate the degree of conformity of the conditional part. Generally speaking, there are the conjunctive method and the algebraic product method.The conjunctive method selects the minimum value among the degrees of fitness for each proposition, and the conjunctive method selects the product of the degrees of fitness of each proposition. is the algebraic product method. These means can be selected depending on the phenomenon.

The conclusion part is weighted based on the degree of conformity of the condition part obtained in this way. This method also includes the logical product method and the algebraic product method. The membership function of the conclusion part is F(X),
Letting α be the degree of conformity of the conditional part, each is expressed by the following equations. Conjunctive product method conclusion part suitability = Min [F(X), α] Algebraic product method conclusion part suitability = α×F(X) These means can also be selected depending on the phenomenon.

As shown in FIG. 29, the subject of the conclusion part of the three fuzzy rules is the same, and the degrees of conformity for each rule obtained above are integrated to arrive at the final conclusion. This is Figure 31
As shown in the figure, the weighted membership functions for each rule are combined (thick solid line) and their center of gravity is determined. This is a representative value, and is the creep damage rate in this example. Further, as shown in the figure, the degree of conformity to the representative value (integrated degree of conformity) can be determined, and this value becomes an index indicating the reliability of inference with respect to the input data.

Next, the input data checking method of the present invention will be described. FIG. 24 shows the flow of the input data check method. First, when the operator inputs data, fuzzy inference is performed based on the data, and the creep damage rate (representative value) φC and integrated fitness degree f are determined and recorded. Then, a preset value fR of the integrated fitness f is set in advance.
If the integrated fitness degree f is larger than the set value fR, the inferred creep damage rate φC is output. Also, if the integrated fitness f is smaller than the set value fR, check A.
and Check B to check the validity of the data, create a comment corresponding to it, and output it together with the recorded creep damage rate. Note that in this embodiment, the set value f
The value of R was set to 0.3.

FIG. 25 shows the flow of check A. First, a domain defined by a certain evaluation parameter i (for example, hardness, degree of tissue change, etc.) is divided into Ki equal parts. In this example, the value of Ki is 6 for the parameter input by rank and 20 for hardness (the domain is 100 to 200).
). Then, while increasing the value Vi of the evaluation parameter i by equal parts of the domain Ki from the domain minimum value (Vmin)i, fuzzy inference is performed for each, and the creep damage rate φc i j and the integrated goodness of fit fi j are determined. . In this process, Vi is the domain maximum value (Vmax
) Continue until i. Note that in this case, input data is used for parameters other than the parameters to be changed. Evaluation parameter i that changes the above processing
Repeat by changing sequentially. The creep damage rate φc i j and the integrated fitness fi j obtained in this way,
The validity of the input data is checked by comparing the creep damage rate φc determined from the input data with the integrated fitness f. The validity of the data is φc and φc i j
and the integrated fitness fi j is 0.5.
This was done by selecting the above data combinations.

FIG. 26 shows the flow of check B. First, Larson-Miller parameters are calculated from the operating time and metal temperature input as input data. The Larson-Miller parameter LMP is expressed by the following equation using temperature and time parameters. There is a correlation between this Larson-Miller parameter and changes in hardness and metal structure, and these changes can almost be inferred if there is no abnormal temperature rise during operation. So Larson
The relationship between the Miller parameter, the Vickers hardness, and the rank of tissue change is created as a database, and by referring to this database, the range of ranks of the hardness and tissue change of the target region is inferred.

Check A and Check B as described above.
and infer a combination of data that matches both. An appropriate comment corresponding to this inference result is created and output together with the recorded creep damage rate, as shown in Figure 2.
An example is shown in 7. In this way, the validity of the input data is checked and a comment is output to request the operator to reconsider.

Next, still another embodiment of the present invention will be described. FIG. 32 shows the overall configuration of a creep damage evaluation expert system including the input data inference system according to the present invention. This system 100 evaluates creep damage using fuzzy reasoning, and as shown in FIG. 34, when evaluating the base material, information on the degree of deformation of crystal grains, the degree of structural change, the degree of change in tube outer diameter, and hardness is used. The fuzzy inference unit 20 performs fuzzy inference based on these data, and outputs the results to the display device 200 via the output screen display unit 50. In the input screen display unit 40, an input data inference unit 85 according to the present invention
A default value is displayed in the input area on the evaluation data input screen shown in FIG. 34, reflecting the inference result in . In addition, in the evaluation of welds, the degree of cavity formation is used instead of the degree of grain deformation, but this reflects the difference in the creep damage mechanism between the base metal and the weld.

FIG. 33 shows the flow of the inference method for evaluation input data according to the present invention. First, on the operation data input screen shown in FIG. 35, the operator inputs operation data such as the region to be diagnosed, operation time, design temperature, and pressure. Based on this data, the system searches the database 500 that stores previous diagnosis results to check whether the same region has been previously diagnosed. if,
If the same site has been previously diagnosed, the input data for evaluation at that time is imported. The metal temperature is estimated from the evaluation data, design conditions (pressure), and operating time up to that point. Input data for evaluation is inferred from this estimated metal temperature, operating time, and pressure, and is displayed as a default value of the input value. This inference uses a database 600 that stores relationships between temperature, pressure, time, and evaluation input data. In addition, if the area has not been diagnosed before, a database 6 will be created based on the input operating conditions (temperature, pressure, operating time).
Input data for evaluation is inferred with reference to 00 and displayed as a default value.

[0072] Inferring the evaluation data as described above,
It is displayed in the input area as a default value, and the operator checks this value and if there is no problem, continues the next inference.
Also, if there is a problem, you only need to correct that part.

Next, still another embodiment will be described. The overall system configuration according to the embodiment is the same as that shown in FIG. 22. Next, using boiler heat exchanger tube creep life diagnosis as an example, the functions of each component will be explained along the processing flow. first,
When inference is started, a command is issued from the input screen display unit 40 to display a screen for inputting the target plant and parts, and operating data such as operating time and metal temperature, and is displayed on the display device 200 via the user interface 400. Ru. The operator 300 inputs the necessary items according to the screen, and when the input is completed, a command is issued to display an evaluation data input screen as shown in FIG. 200 is displayed. In this example, four types of data are input as input data for life evaluation for the base material: 1) degree of deformation of crystal grains, 2) degree of change in outer diameter, 3) degree of structure change, and 4) Vickers hardness. However, other information such as the color of the tube and the state of adhesion of ash may also be used. In this embodiment, regarding 1) and 3), four ranks are displayed on the display device 200 according to the degree of damage, and the operator 300 inputs the numerical value of the corresponding rank. Further, for 4), the measured value of the Vickers hardness is input, but for 2), either the rank or the measured value is input. Note that in the welded part, the degree of cavity formation is used as input data instead of the degree of deformation of crystal grains due to the difference in the creep damage mechanism, but in the following explanation, the base material will be described.

[0074] The input data is the user interface 400
is temporarily recorded in the answer frame 60 via. Next, the input data recorded in the answer frame 60 is sent to the fuzzy inference activation frame 70. In the fuzzy inference starting frame 70, in addition to recording input data, a record of fuzzy inference results and the name of fuzzy knowledge to be applied are written. The fuzzy inference unit 20 performs fuzzy inference using the input data recorded in the fuzzy inference activation frame 70 and calculates the creep damage rate.

In addition, the input data evaluation section 90 checks the input data based on the fuzzy inference results, and sends the input data together with the inference results to the judgment section 30, generates a comment based on the results, and passes it to the output screen display section 50. . The output screen display unit 50 displays the results sent from the determination unit 30 on the display device 200 via the user interface 400.

Next, fuzzy inference in the present invention will be explained. Fuzzy knowledge is necessary to perform fuzzy inference, and fuzzy knowledge consists of fuzzy rules expressed as progression rules and membership functions that give the fitness of each proposition. As shown in Figure 3, a fuzzy rule consists of a condition part proposition and a conclusion part proposition. Write a proposition with ambiguous expressions that would be judged by a human, such as ``increase the creep damage rate.''

In this embodiment, a total of six rules are used, three for stress-dominated data as shown in FIG. 36 and three for age-dominated data. A membership function is defined for each of these propositions. Figure 5 shows an example of the membership function of the conditional part proposition and an example of the membership function of the conclusion part proposition, where the vertical axis is the fitness, the horizontal axis is the rank and measured value in the conditional part membership function, and the creep damage in the conclusion part membership function. percentage (for stress-dominated data) or contribution rate (for age-dominated data).

Since there are multiple propositions in the conditional part, it is necessary to calculate the degree of conformity of the conditional part. Generally speaking, there are the conjunctive method and the algebraic product method.The conjunctive method selects the minimum value among the degrees of fitness for each proposition, and the conjunctive method selects the product of the degrees of fitness of each proposition. is the algebraic product method. These means can be selected depending on the phenomenon. The conclusion part is weighted based on the degree of conformity of the condition part obtained in this way. This method also includes the logical product method and the algebraic product method. Letting the membership function of the conclusion part be F(X) and the degree of conformity of the condition part be α, they are expressed by the following equations. Conjunctive product method conclusion part suitability = Min [F(X), α] Algebraic product method conclusion part suitability = α×F(X) These means can also be selected depending on the phenomenon.

As shown in FIG. 36, the subject of the conclusion part of the three fuzzy rules for stress-dominated data and the three fuzzy rules for age-dominated data is the same, and the goodness of fit for each rule obtained above is integrated. and draw a final conclusion. As shown in FIG. 6, weighted membership functions for each rule are synthesized (thick solid line) and their center of gravity is determined. This is a representative value, and in this example, becomes the creep damage rate or contribution rate. Further, as shown in the figure, the degree of conformity to the representative value (integrated degree of conformity) can be determined, and this value becomes an index indicating the reliability of inference with respect to the input data. In other words, the integrated goodness of fit will be a large value if the tendencies of the input data for the conclusion part match, and a small value if they are different. Therefore, it is necessary to check the input data separately for stress-dominated data and aging-dominated data. can do.

As described above, the contribution rate to the creep damage rate from the stress-dominated data and the contribution rate to the damage rate from the aging-dominated data are determined by fuzzy inference, and the aging effect is corrected based on the concept of FIG. Calculate creep damage rate. Specifically, as shown in Figure 38, for the creep damage rate obtained from stress-dominated data, the amount of correction when the contribution rate is 1 is determined based on the concept of Figure 37, and the correction amount is determined from this figure. calculating the amount. For example, if the creep damage rate determined from stress-dominated data is φc0, and the contribution rate determined from age-dominated data is k, then from FIG. 38, φc0
It is possible to obtain the correction amount k0 when the contribution rate is 1. In this case, the correction amount Δφc for the creep damage rate is Δφc
= k×k0, and the final creep damage rate φc is φc = φc
0 +Δφc.

Furthermore, if the degree of integrated fitness in each fuzzy inference is low, the input data is checked and its contents are output. Note that the present invention employs 0.4 as the limit value of the integrated fitness in this case, but this is based on the results confirmed in various cases.

[0082]

[Effects of the Invention] According to the present invention, fuzzy inference is possible even when some data is missing in fuzzy inference, and since the inference takes the missing data into consideration for reliability, the number of pieces of input data The reliability of the inference depends on the Furthermore, since these operations are automatically performed during the inference process, users of this system can use it without being aware of missing data.

Furthermore, according to the present invention, when fuzzy inference is performed based on uncertain information in diagnosis, etc., the validity of input data is checked and a comment is output according to the content of the check, thereby increasing reliability. This enables highly accurate diagnosis. Furthermore, diagnosis can be made even if the operator is not an expert, which has great utility value.

Furthermore, according to the present invention, there is no need to input evaluation input data for each diagnostic region, and in most cases, inferences can be made just by checking the default values, so the time required for input can be saved. , enabling rapid diagnosis.

Furthermore, according to the present invention, when performing fuzzy inference based on uncertain information in diagnosis, etc., data is classified into data with high and low importance for the conclusion part, and fuzzy inference is performed independently for each data. As the creep damage rate is calculated by calculating the creep damage rate, it is possible to accurately evaluate the creep damage over the entire damage area. In addition, since the validity of the input data is checked using integrated suitability for data that have the same degree of importance for the conclusion part, and comments are output according to the content of the check, even non-experts can make highly reliable diagnoses. It has become possible, and its utility value is great.

[Brief explanation of the drawing]

FIG. 1 is an overall configuration diagram of an expert system using fuzzy inference according to the present invention.

FIG. 2 is an explanatory diagram showing a display example of an input screen.

FIG. 3 is a basic explanatory diagram of fuzzy rules.

FIG. 4 is a concrete explanatory diagram of a fuzzy rule according to an embodiment.

FIG. 5 is an explanatory diagram of membership functions of condition part propositions and conclusion part propositions.

FIG. 6 is a characteristic diagram of goodness of fit.

FIG. 7 is a characteristic diagram of membership functions for missing data.

FIG. 8 is a characteristic diagram of goodness of fit when the membership function is changed.

FIG. 9 is a comparison diagram of fitness characteristics in each case.

FIG. 10 is an explanatory diagram showing a display example of an initial input screen.

FIG. 11 is a diagram showing an input screen when the object to be inferred is a base material.

FIG. 12 is a diagram showing an input screen when the object to be inferred is a weld.

FIG. 13 is a diagram showing a display of input data.

FIG. 14 is a diagram showing a display screen of diagnosis results.

FIG. 15 is an explanatory diagram showing a third embodiment of the present invention.

FIG. 16 is an overall configuration diagram of a plant life diagnosis support device.

FIG. 17 is a processing flowchart of the device.

FIG. 18 is a diagram showing a list of inspection parts.

FIG. 19 is a diagram showing an inspection result record table.

FIG. 20 is a characteristic diagram showing the relationship between evaluation parameters and damage rate.

FIG. 21 is a diagram illustrating an example of countermeasure inference results output to a display device.

FIG. 22 is an overall configuration diagram of a fuzzy inference system according to another embodiment of the present invention.

FIG. 23 is a diagram of an evaluation data input screen.

FIG. 24 is an overall flowchart of checking the validity of input data.

FIG. 25 is a flowchart of Check A therein.

FIG. 26 is a flowchart of Check B therein.

FIG. 27 is a diagram of an output screen when checking the validity of data.

FIG. 28 is a format diagram of fuzzy rules.

FIG. 29 is an explanatory diagram of a fuazi rule in an example.

FIG. 30 is a diagram showing an example of a membership function.

FIG. 31 is a diagram showing an example of an integrated membership function.

FIG. 32 is an overall configuration diagram of an expert system according to another embodiment.

FIG. 33 is a flowchart of the evaluation input data inference system.

FIG. 34 is a diagram showing an input screen for evaluation input data.

FIG. 35 is a diagram of an operation data input screen.

FIG. 36 is an explanatory diagram of a fuzzy rule according to this embodiment.

FIG. 37 is a diagram illustrating the effect of aging on creep damage.

FIG. 38 is a diagram showing the amount of correction of creep damage rate due to aging.

FIG. 39 is an explanatory diagram of fuzzy inference when brake pressure is inferred using the inter-vehicle distance and traveling speed as input values.

FIG. 40 is an explanatory diagram of fuzzy rules at that time.

FIG. 41 is a fitness characteristic diagram when all data is input.

FIG. 42 is a fitness characteristic diagram when data is missing.

[Explanation of symbols]

10 Control section 20 Fuzzy Reasoning Department 30 Judgment Department 60 Answer frame 80 Missing data detection section

Claims (12)

[Claims] [Claim 1] In the process of inference, the membership function for missing data is changed so that the fitness is "1" for the entire domain, and then the value of the missing data is changed to any value in the domain. A fuzzy inference method characterized in that the degree of fitness of a membership function for existing data is lowered in accordance with the importance of missing data for a conclusion section. [Claim 2] In the process of inference, the membership function for missing data is changed so that the fitness is "1" for the entire domain, and then the value of the missing data is changed to any arbitrary value in the domain. A fuzzy inference method characterized in that the membership function of the conclusion part is lowered according to the importance of missing data to the conclusion part. [Claim 3] In the process of inference, the fuzzy rule condition part for missing data is deleted, and the fitness of the membership function for existing data is reduced according to the importance of the missing data for the conclusion part. Features a fuzzy inference method. 4. Fuzzy inference, characterized in that in the process of inference, the fuzzy rule condition part for missing data is deleted, and the membership function of the conclusion part is lowered according to the importance of the missing data to the conclusion part. Method. Claim 5: A judgment unit that evaluates the reliability of the inference and an input data evaluation unit that evaluates the validity of the input data when the reliability is low, and comments corresponding to the results inferred by each inference unit are provided. A data evaluation method in fuzzy inference that is characterized by creating and outputting. 6. In the fifth aspect, the determining unit evaluates reliability by comparing the integrated fitness in fuzzy inference with a preset value, and the input data evaluating unit sequentially changes the input data. while converting the data between domains and performing fuzzy inference on each domain, there is a part that infers the combination of data that increases the integrated fitness, and a part that infers the input data range from surrounding data. A data evaluation method in fuzzy inference characterized by evaluating validity. [Claim 7] In a system that evaluates the remaining life of metal materials from diagnostic data such as metallographic changes, hardness, and tube outer diameter, input data for evaluation is inferred based on operating conditions and previous diagnosis results. , a fuzzy inference method characterized by displaying the value as a default value in an input area. [Claim 8] In a fuzzy inference method in which a plurality of input data have different degrees of importance with respect to the conclusion part, the fuzzy inference results based on the data with a high degree of importance are classified into data with a high degree of importance and data with a low degree of importance. A fuzzy inference method characterized by calculating a final result by correcting it using fuzzy inference results based on data with a small degree. 9. The fuzzy inference method according to claim 8, characterized in that the validity of the input data is checked based on the integrated fitness of fuzzy inferences based on data with high importance and data with low importance. 10. In a system for evaluating the creep life of metal materials using fuzzy inference, the data is classified into stress-dominated data and age-dominated data, and the fuzzy inference results from the stress-dominated data are compared to the age-dominated data. A fuzzy inference method characterized by calculating a creep damage rate by correcting it using fuzzy inference results based on data. 11. The fuzzy inference method according to claim 10, characterized in that the validity of the input data is checked based on the integrated fitness of the respective fuzzy inferences based on stress-dominated data and age-dominated data. . 12. In claim 10 or claim 11, the stress-dominated data is the degree of deformation of crystal grains, the degree of cavity formation, and the degree of change in the outer diameter of the tube, and the age-dominated data is the degree of metallographic structure. A fuzzy inference method characterized by changes in hardness and changes in hardness.