JP7411473B2 - Validation method, validation system and program - Google Patents

Description

The present disclosure relates to a validation method, validation system, and program for a probabilistic risk assessment model.

A probabilistic risk assessment (PRA) model (hereinafter referred to as the PRA model) is sometimes used for risk assessment of nuclear power plants (Patent Document 1). To maintain the accuracy of the risk assessment, engineers may manually review the cut set of crest events generated based on the PRA model.

Patent Document 2 discloses that the feature vectors of training data are clustered, the feature vectors of data to be recognized are interpolated using the clustered feature vectors, and then input into a speech recognition model to optimize the speech recognition model. Techniques for adjustment are described.

Patent No. 5480033 International Publication No. 2018/005858

The number of cut sets of top events output from a PRA model of a nuclear power plant can range from tens of thousands to hundreds of thousands per plant, so it is difficult to manually check all of them.

The present disclosure provides a validation method, validation system, and program that can solve the above problems.

The validation method of the present disclosure is a validation method executed by a computer, and includes the steps of calculating feature amounts of a cut set output from a PRA model, and a hierarchy based on the feature amounts for the cut set. a step of performing hierarchical clustering, a step of extracting one or more of the cut sets included in each class generated by the hierarchical clustering, and evaluating the validity of the extracted cut sets. The method includes a step of obtaining a result, and a step of receiving correction information for the PRA model when the evaluation result is not valid, and in the step of calculating, a cut set is configured for one of the cut sets. A value that quantifies the number of times a word that appears in the text data indicating the basic event appears in the cut set and the tendency of whether the word appears in other cut sets is calculated as the feature quantity, and the value is calculated as the feature quantity. The risk evaluation results include evaluation results by engineers as to whether there are any errors in the basic events included in the cut set and whether there are any errors in the combination conditions of the basic events, and The correction information includes information created by the engineer for correcting the error when the validity evaluation result indicates that there is an error.

Further, the validation system of the present disclosure includes a feature calculation unit that calculates feature amounts of a cut set output from a PRA model, and a hierarchical clustering unit that performs hierarchical clustering on the cut set based on the feature amount. an extraction unit that extracts one or more of the cut sets included in each class generated by the hierarchical clustering; and an extraction unit that extracts one or more of the cut sets included in each class generated by the hierarchical clustering, and a validity evaluation result for the extracted cut sets that is not valid. In this case, the feature amount calculation unit includes a modification reception unit that receives modification information of the PRA model, and the feature value calculation unit calculates, for one of the cut sets, words appearing in text data indicating base events constituting the cut set. A value that quantifies the number of occurrences in the cut set and a tendency of whether or not the word appears in other cut sets is calculated as the feature amount, and the validity evaluation result is the number of occurrences in the cut set. The modification information for the probabilistic risk assessment model includes the evaluation results of the validity evaluation, including the results of an engineer's evaluation as to whether there are any errors in the base events to be used, and whether there are any errors in the combination conditions of the base events. If the evaluation result indicates that there is an error, information created by the engineer to correct the error is included.

Further, the program of the present disclosure causes a computer to calculate a feature amount of a cut set output from a PRA model, perform hierarchical clustering on the cut set based on the feature amount, and perform hierarchical clustering on the cut set based on the feature amount. extracting one or more of the cut sets included in each class generated by clustering; obtaining a validity evaluation result for the extracted cut set; If the PRA model is not valid, the program executes the step of accepting correction information of the PRA model, and in the step of calculating, for one of the cut sets, text data indicating basic events constituting the cut set is processed. A value that quantifies the number of occurrences of the word appearing in the cut set and a tendency of whether the word appears in other cut sets is calculated as the feature quantity, and the validity evaluation result is The correction information for the probabilistic risk assessment model includes the results of an evaluation by an engineer as to whether there are any errors in the base events included in the cut set and whether there are any errors in the combination conditions of the base events, and the modification information for the probabilistic risk assessment model includes the If the validity evaluation result indicates that there is an error, information created by the engineer to correct the error is included.

According to the validation method, validation system, and program described above, the validity of the PRA model can be efficiently confirmed.

FIG. 1 is a block diagram illustrating an example of a validation system according to an embodiment. It is a figure showing an example of validity check processing concerning an embodiment. FIG. 3 is a diagram for explaining a cut set of top events according to an embodiment. It is a figure showing an example of a cut set concerning an embodiment. FIG. 2 is a diagram illustrating hierarchical clustering according to an embodiment. FIG. 2 is a diagram illustrating non-hierarchical clustering according to an embodiment. FIG. 1 is a diagram showing an example of a hardware configuration of a validation system according to an embodiment.

<Embodiment>
The validation system of the present disclosure will be described below with reference to FIGS. 1 to 7.
(System configuration)
FIG. 1 is a block diagram illustrating an example of a validation system according to an embodiment.
The validation system 1 supports validation of a cut set of top events that constitute a PRA model. As illustrated, the validation system 1 includes a PRA model management device 10 and a validation support device 20. The PRA model management device 10 and the validation support device 20 are communicably connected.

(Configuration of PRA model management device)
The PRA model management device 10 includes a cut set output section 11, a modification reception section 12, and a storage section 13.
The cut set output unit 11 analyzes the PRA model and outputs a cut set of the top event. A culmination event is, for example, a serious accident event in a nuclear power plant. A cutset of a summit event is a combination of one or more base events that causes the summit event. The base event is an individual accident event that leads to a serious accident event, such as failure to start a pump or failure to open a check valve. Hereinafter, the cut set of the top event will be simply referred to as a cut set. For example, in the case of a nuclear power plant, the number of cut sets output from a PRA model can range from tens to hundreds of thousands. A plurality of cut sets output from the PRA model is called a cut set list.
The modification reception unit 12 receives modification instructions for the PRA model and modifies the PRA model.
The storage unit 13 stores the PRA model. The PRA model includes FT (Fault Tree) and ET (Event Tree).

(Configuration of validation support device)
The validation support device 20 includes a cut set acquisition section 21 , a feature value calculation section 22 , a hierarchical clustering section 23 , an extraction section 24 , a non-hierarchical clustering section 25 , a sorting section 26 , and an output section 27 and a storage unit 28.
The cut set acquisition unit 21 acquires the cut set output by the PRA model management device 10.
The feature amount calculation unit 22 calculates the feature amount of each cut set acquired by the cut set acquisition unit 21. Specifically, the ID (identification information) of the basic event included in the cut set is coded according to a predetermined rule, but the coded text data is segmented into predetermined units and the segmented character string is is treated as a word, the cut set is treated like a document, and feature amounts based on the words appearing in the cut set are calculated using text analysis such as tfidf and word2vec. The feature amount of the calculation result is expressed as a vector. Each element of this vector is a value obtained by text analysis of a character string (word) obtained by segmenting the base event ID.

The hierarchical clustering unit 23 executes hierarchical clustering based on the feature values of the cut sets calculated by the feature value calculation unit 22, and classifies all cut sets generated from the PRA model into similar cut sets. .
The extraction unit 24 extracts a cut set sample to be checked by an engineer from each class generated by hierarchical clustering.

The non-hierarchical clustering unit 25 performs non-hierarchical clustering such as the k-means method on a set of cut sets that have been confirmed to be valid by an engineer.
The selection unit 26 selects cut sets that are likely to be inappropriate from among the cut sets based on the results of non-hierarchical clustering.

The output unit 27 displays the results of hierarchical clustering, the results of non-hierarchical clustering, etc. on a display device, or outputs them as electronic data.
The storage unit 28 stores a cut set list, data being processed, and the like.

(motion)
Next, the operation of the validation system 1 will be explained using FIGS. 2 to 6.
FIG. 2 is a diagram illustrating an example of validation processing according to the embodiment.
The operation of the validation system 1 is to unbiasedly extract cut sets to be checked using text analysis and hierarchical clustering, and to accumulate valid cut sets without bias by requesting confirmation from an engineer (described later). The process consists of two steps: (S1 to S8), and a process of learning accumulated appropriate cut sets and selecting inappropriate cut sets based on the learning results (S9 to S13, which will be described later). These processes will be explained below.

First, an engineer instructs the PRA model management device 10 to output a cut set list. The cut set output unit 11 reads the PRA model from the storage unit 13 and outputs the cut set. Reference is now made to FIG.

FIG. 3 is a diagram for explaining a cut set of top events according to the embodiment.
FIG. 3 shows an example of FT included in the PRA model. The circles A to E shown in FIG. 3 indicate basic events, and the squares P to Q indicate intermediate events. Indicates that an event will occur. The AND gate above indicates that when base event B and intermediate event Q occur, intermediate event P occurs. The AND gate further below shows that when base event D and base event E occur, intermediate event Q occurs. In the case of FT in FIG. 3, the cut set that causes the top event is (1) a combination of base event A, (2) base event B, base event D, and base event E, and (3) base event C. The cut set output unit 11 of the PRA model management device 10 outputs cut set 1 (base event A), cut set 2 (base event B, base event D, base event E), cut set 1 (base event A), cut set 2 (base event B, base event D, base event E), etc. based on the FT of FIG. A cut set list including set 3 (base event C) is output (S1).

In the validation support device 20, the cut set acquisition unit 21 acquires the cut set list output by the cut set output unit 11 (S2). The storage unit 28 stores a cut set list. The validation support device 20 provides a function that makes the work of validating the cut sets included in the acquired cut set list more efficient and labor-saving. See FIG. 4.

FIG. 4 is a diagram illustrating an example of a cut set according to the embodiment.
FIG. 4 shows a cut set based on the FT of FIG. 3. ID indicates the ID of the cut set, occurrence probability indicates the probability that an event of the cut set occurs, and base event indicates the base event included in the cut set. In the example of FIG. 4, ID=1 corresponds to cut set 1 explained in FIG. 3, the probability is X1%, and the base event constituting this cut set is base event A. Although the method of calculating the probability will not be described because it is publicly known, it can be calculated based on the FT shown in FIG. 3 and the ET (not shown), for example. For example, in the case of a cut set with ID=2 (corresponding to cut set 2 explained in FIG. 3), the probability that the event of cut set 2 occurs is X2%, and base event B, base event D, and base event E occur. Calculated based on probability. Probability indicates the magnitude of risk. Here, it is assumed that X1%>X3%>X2% holds true. The number of cut sets illustrated in Figure 4 is three, so the engineer can check all of them, but there are tens of thousands of cut sets in total, and the probability of cut set 1 is x1% and the probability of cut set 3 is x3%. Suppose that the cut set 2 is ranked in the top (for example, within the top 100) out of tens of thousands in terms of probability, and the probability X2% of cut set 2 is ranked in the bottom out of the tens of thousands. A cut set with a large probability is important in the sense that it is more likely to cause a summit event (accident). Therefore, the engineer, for example, checks the validity of the cut sets in order of probability (high risk).

Assume that there is an error in the FT in FIG. 3, and that the AND gate at the bottom of FIG. 3 is the correct OR gate. In this case, cut set 2 is incorrect, and the cut set list output from the correct FT includes cut set 2a consisting of base event B and base event D, and cut set 2a consisting of base event B and base event E. The cut set 2b, the cut set 1, and the cut set 3 described above are included. Let the probability of cut set 2a be X2a% and the probability of cut set 2b be X2b%. For example, if probability X2a% is ranked high, originally, high-risk cut set 2a should be This is the cut set that should be confirmed. However, if the probability of a cut set is calculated to be small due to an error in the PRA model (for example, FT), there is a possibility that it will not be checked by an engineer. If the PRA model is correct and cut set 2a is output, attention will be directed to base event B and base event D related to cut set 2a, and the probability of occurrence of the top event will be reduced from the perspective of risk information utilization. It becomes possible to take measures to prevent this. However, if the cut set 2a with a high probability of occurrence is not output due to an error in the PRA model, the cut set 2a may be overlooked, and as a result, no measures may be taken to reduce risk. The validation support device 20 extracts cut sets and presents them to the engineer so that validity can be checked for a wide range of cut sets, regardless of the high or low probability of the cut sets.

Returning to FIG. 2, the processing of the validation support device 20 will be explained. First, the feature calculation unit 22 executes "(1) vectorization of cut sets by text analysis" for each cut set in the cut set list acquired by the cut set acquisition unit 21 (S3). Basic events can be represented as coded text data. For example, the ID of the base event A is "Aa1bbb1052A" as illustrated in FIG. For example, the first three digits "Aa1" indicate the plant system ID, the next four digits "bbb1" indicate the failure mode ID, and the last four digits "052A" indicate the equipment number related to the failure. In this way, the basic event ID includes information regarding the equipment, system, and type of failure in which the failure event occurred. Therefore, we consider a cut set as a sentence, segment the basic event ID of each basic event included in the cut set into meaningful elements (words) such as system ID, failure mode ID, equipment number, etc. The importance and meaning of a cut set are analyzed using text analysis, and the analysis results are treated as features of the cut set. Based on this idea, the feature calculation unit 22 performs text analysis on the text data of the basic event included in the cut set, and converts the cut set into numerical values and vectors. As described below, vectorizing the cut sets allows the cut sets to be classified based on vector proximity (text similarity).

Next, calculation of the feature amount of a cut set, that is, an example of vectorization of a cut set will be specifically explained. For convenience of explanation, the cut set list includes only cut sets 1 to 3, cut set 1 includes only base event A, cut set 2 includes only base event B, and cut set 3 includes only base event C. shall be held. The basic event IDs of basic events A to C are shown in FIG. First, the dimension of the vector of each cut set is divided into elements (words) obtained by segmenting the base event IDs of all base events (base events A to C) that exist in all cut sets (cut sets 1 to 3). number. Specifically, the ID of base event A is segmented into "Aa1", "bbb1", and "052A", and the ID of base event B is segmented into "Aa1", "ddd2", and "052A", and the ID of base event B is segmented into "Aa1", "ddd2", and "052A". Since the ID of event C is segmented into "Aa2", "ccc3", and "024B", the number of elements is "Aa1", "Aa2", "bbb1", "ccc3", "ddd2", " 052A" and "024B", and the number of dimensions of the vector is seven. For example, the feature amount calculation unit 22 calculates, for each of cut sets 1 to 3 (a text analysis value for the word "Aa1", a text analysis value for the word "Aa2", a text analysis value for the word "bbb1", a text analysis value for the word " vectors are created. Text analysis is, for example, tfidf. For example, the text analysis value using tfidf for the first vector element of cut set 1, word “Aa1”, can be calculated using the following equation.

tf (“Aa1”, cut set 1) = number of occurrences of “Aa1” in cut set 1 (“Aa1bbb1052A”)/all words of cut set 1 (“Aa1”, “bbb1”, and “052A”) Sum of the number of occurrences...(1)
idf (“Aa1”) = log (total number of cut sets (=3)/number of cut sets in which the word “Aa1” appears (=2))... (2)
Text analysis value of the first element of cut set 1 (tfidf) = tf (“Aa1”, cut set 1) × idf (“Aa1”) = formula (1) × formula (2)... (3)

A word that appears many times in a certain cut set (tf is large) but does not appear in other cut sets (idf is large) has a large text analysis value using tfidf. In addition, tfidf becomes smaller for other words. For example, a word appearing in multiple cut sets cannot be said to represent the characteristics of each cut set, so the value of tfidf is small. On the other hand, a word that appears only in a specific cut set can be said to represent the characteristics of the cut set, so the value of tfidf becomes large. Using tfidf, the importance of words in the cut set can be calculated. By creating a vector whose elements are analysis values based on tfidf having such properties, the characteristics of the cut set can be expressed from the perspective of the text (combination of words) that constitutes the basic event ID. The values of the elements of the vector may be calculated using word2vec.

When the feature value calculation unit 22 calculates a vector representing each cut set, the hierarchical clustering unit 23 then performs “(2) hierarchical clustering” on the vectors of all cut sets to make the vectors similar. The cut sets are classified into the same class (S4). The hierarchical clustering unit 23 repeats the process of forming a set (cluster) of cut sets whose vectors are close to each other until all the cut sets are finally integrated into one cluster. This process is performed in order to categorize the cut set, assuming that the text analysis result of the words included in the cut set represents the degree of similarity between the combinations of basic events. FIG. 5 shows an example of the results of hierarchical clustering.

Next, the extraction unit 24 divides the cluster group generated by hierarchical clustering at an arbitrary hierarchy (threshold L1 in FIG. 5) and divides it into classes (S5). Each of C1 to C20 described below the dendrogram in FIG. 5 is a class generated when dividing by the threshold L1. The threshold L1 may be predetermined, or the number of classes after division may be predetermined, and the extraction unit 24 may set the threshold L1 according to this number of classes. Regarding the divided classes, the extraction unit 24 "(3) extracts some cut sets from each class" (S6). Each class has a collection of cut sets with similar vectors, so by extracting cut sets from all classes, it is possible to unbiasedly extract representative samples from tens of thousands of cut sets. can. For example, the extraction unit 24 may extract one or more cut sets completely arbitrarily from each class, or may extract a combination of cut sets whose vector proximity is lower than a predetermined threshold in order to extract cut sets evenly. may be extracted. Alternatively, when extracting only one cut set from each class, the extraction unit 24 may calculate the centroid of the vector of each class and extract the cut set closest to the centroid. In addition to the threshold L1, the extraction unit 24 also sets a threshold L2 (FIG. 5) for extracting cut sets, and extracts, for example, one cut set at a time (or more than one) from the classes divided by the threshold L2. You may. Further, the extraction unit 24 may extract cut sets from each of the classes C1 to C20 based on the probabilities illustrated in FIG. 4. For example, the extraction unit 24 may extract a predetermined number of cut sets in order from the highest probability within each class, extract a predetermined number of cut sets in order from the lowest probability, or extract any cut set from those with probabilities within a predetermined range. You may also extract cut sets. The extraction unit 24 also extracts a predetermined number of cut sets by combining probabilities and vectors, for example, starting from the one with the highest probability in each class, on the condition that the proximity of the vectors is a predetermined distance or more. You can. The extraction unit 24 records the ID of the extracted cut set in the storage unit 28. This is to prevent previously extracted cut sets from being extracted when the process of S6 is performed again later.

When the extraction unit 24 extracts a cut set, the output unit 27 outputs the extracted cut set to a display device or the like. Next, perform "(4) Validity check by an engineer." The engineer manually checks the cut set extracted from each cluster. For example, the engineer checks whether there are any errors in the basic events included in the cut set, whether any unnecessary basic events are included, whether there are any errors in the combination conditions (OR, AND) of the basic events, etc. . If there is an error, the engineer inputs the correction instruction information into the PRA model management device 10, and the PRA model management device 10 executes “(5) If an inappropriate cut set is identified, correct the PRA model.” (S7). Specifically, in the PRA model management device 10, the modification reception unit 12 acquires modification instruction information of the PRA model, and modifies the PRA model in the storage unit 13 based on the modification instruction information. If there is no error in the cut set, the engineer performs "(6) Input to the validation support device 20 if no inappropriate cut set is identified" (S8). In the validation support device 20, the cut set acquisition unit 21 records the cut set that has been verified by the engineer in the storage unit 28 with a mark indicating that the cut set has been verified.

The above is a reasonable cut set learning process using text analysis technology. The validation system 1 repeatedly performs the processes of S1 to S8 in "(4) Validation by engineer" until no inappropriate cut sets are identified. In the second and subsequent processing, the extraction unit 24 does not extract the cut set that has been extracted once again, but extracts other cut sets. When the processes of S1 to S8 are repeated, cut sets that have been confirmed to be not inappropriate are accumulated in the storage unit 28. The accumulated confirmed cutsets are cutsets extracted from each class generated by hierarchical clustering, so they are considered to be cutsets that represent various cutsets that exist among all the cutsets. .

When inappropriate cut sets are no longer identified, the engineer instructs the validation support device 20 to execute extraction processing for inappropriate cut sets. Then, the non-hierarchical clustering unit 25 executes “(7) clustering using the k-means method” on the cut sets that are confirmed to be not inappropriate and are stored in the storage unit 28 (S9). The non-hierarchical clustering unit 25 uses the k-means method to classify the confirmed cut sets into a preset number of classes based on the proximity of the vectors of the confirmed cut sets, and calculates the centroid of each class and the centroid. Calculate the range in which the vectors of the confirmed cut set centered are distributed. In addition, the non-hierarchical clustering unit 25 uses the k-means method for cut sets not extracted by the extraction unit 24 and cut sets other than those confirmed to be inappropriate (these are referred to as unconfirmed cut sets). Perform evaluation based on each class generated by . FIG. 6 shows an example of a class generated by the k-means method. Shown in FIG. 6 is the vector space of cut sets. For convenience, the vector space is shown in two dimensions and only one class is shown. The star point shown in FIG. 6 is the center of gravity of this class. The circled points indicate vectors of cut sets that have been confirmed by the engineer. Both the x-marked points and the triangular-marked points indicate vectors of unconfirmed cut sets. A circle 61 indicates a range including the confirmed cut set around the center of gravity. The non-hierarchical clustering unit 25 identifies the cut set with the farthest distance from the center of gravity among the confirmed cut sets classified into classes, and creates a circle 61 whose radius is the distance from the identified cut set to the center of gravity. stipulates. A circle 62 indicates an allowable range for determining that an unconfirmed cut set is an appropriate cut set. The non-hierarchical clustering unit 25 defines an allowable range, for example, by a circle 62 whose radius is the length calculated by multiplying the radius of the circle 61 by a predetermined coefficient. The non-hierarchical clustering unit 25 calculates the distance from the center of gravity for each of all unconfirmed cut sets, and determines whether the cut set is included in the circle 61 or the circle 62.

The non-hierarchical clustering unit 25 determines that the unconfirmed cut set included in the circle 62 is an appropriate cut set, and the unconfirmed cut set located outside the circle 62 is determined to be an inappropriate cut set (S10). The non-hierarchical clustering unit 25 records the determination result for each unconfirmed cut set in the storage unit 28. The non-hierarchical clustering unit 25 performs calculations such as the centroid and allowable range for all the classes generated in S9, and makes the determination in S10. If the engineer confirms all unconfirmed cut sets that have been determined to be inappropriate, the accuracy of the PRA model can be improved, but if the number of unconfirmed cut sets that have been determined to be inappropriate becomes large. , the engineer may not be able to fully confirm the problem. On the other hand, if you can confirm at least one cut set related to an incorrect element (basic event or logical formula) included in the PRA model and correct the incorrect element, you can correct the cut set related to the incorrect element. All cutsets are corrected correctly. Therefore, if you can accurately select one of the cut sets related to the incorrect element in the PRA model and ask the engineer to confirm it, the engineer only needs to confirm a small number of cut sets. , the PRA model can be modified efficiently.

Based on the above idea, the selection unit 26 selects cut sets that are actually likely to be inappropriate from the cut list determined to be inappropriate in S10 (S11). Specifically, the sorting unit 26 performs "(8) classification based on the distance from the center of gravity of each class". (a) For example, the sorting unit 26 selects a predetermined number of cut sets (triangular marks in FIG. 6) that are determined to be inappropriate in order of distance from the center of gravity, and actually uses the selected cut sets. The cut set is likely to be inappropriate. (b) Alternatively, the selection unit 26 uses the hierarchical clustering unit 23 to perform hierarchical clustering on the cut set determined to be inappropriate. Then, from each class generated by hierarchical clustering, a predetermined number of cut sets are selected in descending order of distance from the center of gravity, and these are set as cut sets with a high possibility of actually being inappropriate. By categorizing cut lists determined to be inappropriate through hierarchical clustering, it is possible, for example, to classify cut sets generated due to errors present in the PRA model into one or more specific classes. Then, by selecting a class that is far from the center of gravity among the classified classes, it can be expected that the accuracy of selecting a cut set that is likely to be actually inappropriate can be increased. Furthermore, by using hierarchical clustering, it is possible to unbiasedly select cut sets that are likely to be inappropriate. Note that the method for selecting a cut set that is likely to be inappropriate from a class may be the same as that described in S6. (c) Also, for example, the sorting unit 26 selects all cut sets whose distance from the center of gravity is a predetermined value or more from among the cut sets determined to be inappropriate, and selects these cut sets that are actually likely to be inappropriate. May be set. By selecting a cutset that is far away from the vector of the validated cutset, it is possible to increase the possibility of selecting a cutset associated with an erroneous element of the PRA model described above. The sorting unit 26 performs the sorting process in S11 for all classes generated in S9.

The sorting unit 26 “(9) extracts some cut sets” from all the cut sets that are actually likely to be inappropriate, sorted in S11, and outputs them to the output unit 27 (S12). If the number of cut sets that are actually likely to be inappropriate is less than or equal to a predetermined number, the sorting unit 26 may output all the cut sets that are actually likely to be inappropriate to the output unit 27 . The output unit 27 outputs part or all of the selected cut sets to a display device or the like. Next, "(10) Validity check by engineer" is performed. The engineer checks the validity of the cut set that has a high possibility of actually being inappropriate. The PRA model management device 10 "(11) corrects the PRA model when an inappropriate cut set is identified" (S13). Specifically, the modification reception unit 12 acquires modification instruction information for the PRA model, and modifies the PRA model in the storage unit 13.

When an engineer manually checks hundreds of thousands of cut sets, he or she may preferentially check the cut sets with the highest risk.

In contrast, according to the present embodiment, by performing hierarchical clustering on all cut sets, it is possible to uniformly extract representative cut sets from the viewpoint of combinations of basic events from the whole, and to check the validity of the cut sets. I can do it (S1-S8). As a result, regardless of the level of risk, it is possible to check the cut set without any bias. Further, by repeating the processes S1 to S8, the accuracy of the PRA model can be improved.

In addition, in hierarchical clustering, it is not possible to quantitatively compare the similarity of vectors between confirmed cut sets and unconfirmed cut sets, but in the processing of S9 to S12, each class is The distance in vector space between the center of gravity of and the unconfirmed cut set is determined, and if it is close, it is determined to be "appropriate", and if it is far, it is determined to be "inappropriate". As a result, based on the cut sets that the engineer judged to be appropriate in the processing of S1 to S8, the cut sets that have low similarity to the cut sets considered to be appropriate and are considered to have a high possibility of actually being inappropriate are selected. It can be automatically extracted and presented to the engineer. This allows engineers to efficiently confirm inappropriate cut sets and correct the PRA model with a small amount of work.

Further, the work of checking the cut set is a work that is continuously performed by engineers from the viewpoint of improving safety. By introducing the validity confirmation system 1 of this embodiment, it is possible to enjoy the effects of increased efficiency and labor saving in confirmation work in the long term.

In the above embodiment, it has been described how to save labor and improve the efficiency of checking a cut set of a PRA model created in a nuclear power plant. As mentioned above, in a nuclear power plant, the number of cut sets may reach tens of thousands to hundreds of thousands, and the PRA model validation method of this embodiment is effective. However, the field of application of the validation system 1 is not limited to nuclear power plants. It can be applied to any industrial field that performs risk assessment using the PRA model.
Further, specific methods of hierarchical clustering and non-hierarchical clustering are not limited to the above-described methods, and similar methods can be used.

FIG. 7 is a diagram illustrating an example of the hardware configuration of the validation system.
The computer 900 includes a CPU 901, a main storage device 902, an auxiliary storage device 903, an input/output interface 904, and a communication interface 905.
The above-described PRA model management device 10 and validation support device 20 are implemented in a computer 900. Each of the above-mentioned functions is stored in the auxiliary storage device 903 in the form of a program. The CPU 901 reads the program from the auxiliary storage device 903, expands it to the main storage device 902, and executes the above processing according to the program. Further, the CPU 901 reserves a storage area in the main storage device 902 according to the program. Further, the CPU 901 secures a storage area in the auxiliary storage device 903 to store the data being processed according to the program.

Note that a program for realizing all or part of the functions of the PRA model management device 10 and the validation support device 20 is recorded on a computer-readable recording medium, and the program recorded on this recording medium is readable by the computer system. Processing may be performed by each functional unit by loading and executing the program. The "computer system" here includes hardware such as an OS and peripheral devices. Furthermore, the term "computer system" includes the homepage providing environment (or display environment) if a WWW system is used. Furthermore, the term "computer-readable recording medium" refers to portable media such as CDs, DVDs, and USBs, and storage devices such as hard disks built into computer systems. Further, when this program is distributed to the computer 900 via a communication line, the computer 900 that received the distribution may develop the program in the main storage device 902 and execute the above processing. Further, the above program may be one for realizing a part of the above-mentioned functions, and further may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system. .
The PRA model management device 10 and the validation support device 20 may each be configured by a plurality of computers 900 that can communicate with each other. Further, the PRA model management device 10 and the validation support device 20 may be configured by one computer 900.

As described above, several embodiments according to the present disclosure have been described, but all these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and changes can be made without departing from the gist of the invention. These embodiments and their modifications are included within the scope and gist of the invention as well as within the scope of the invention described in the claims and its equivalents.

<Additional notes>
The validation method, validation system 1, and program for the probabilistic risk assessment model (PRA model) described in each embodiment are understood, for example, as follows.

(1) The method for validating a probabilistic risk assessment model (PRA model) according to the first aspect includes the step of calculating feature quantities of a cut set output from the probabilistic risk assessment model (PRA model). S3), a step (S4) of performing hierarchical clustering on the cut set based on the feature amount, and one or more of the cut sets included in each class generated by the hierarchical clustering. (S6), obtaining a validity evaluation result for the extracted cut set (S7, S8), and if the evaluation result is not valid, the probabilistic risk evaluation model ( (S7) of receiving modification information of the PRA model).

By calculating and quantifying the feature values of the cut sets, clustering can be performed for all the cut sets. By performing hierarchical clustering on all cut sets, it is possible to extract a cut set that is representative of the whole. Since the confirmation work is performed by extracting a representative cut set, the cut set can be confirmed without bias and with less effort.

(2) A validation method according to a second aspect is the validation method according to (1), in which non-hierarchical clustering is performed based on the feature amount of the cut set for which the evaluation result is validated. In the step (S9) of generating one or more second classes, calculating the centroid of the second class and the range in which the cut set can be considered valid, and the step of extracting (S6), A step (S10) of determining validity based on the range for the cut set that is not valid; a step of performing selection (S11), and a step of receiving correction information for the probabilistic risk assessment model (PRA model) if the validity evaluation result for the selected cut set is not valid (S13); It further has.

Thereby, the validity of unconfirmed cut sets can be automatically determined based on the cut sets whose validity has been confirmed by the engineer. In addition, for cut sets that are determined to be unreasonable (inappropriate), by further sorting based on the distance from the center of gravity of the class, it is possible to present cut sets that are likely to be inappropriate to the engineer. can.
The validity checking methods (1) and (2) make it possible to check the validity of a huge amount of cut sets efficiently and with fewer omissions.

(3) A validation method according to a third aspect is the validation method of (2), in which, in the step of sorting, a predetermined number of the above-mentioned Select cut sets.

(4) The validation method according to the fourth aspect is the validation method of (2), in which in the step of performing the selection, hierarchical clustering is further performed to generate clusters, and each cluster is A predetermined number of the cut sets are selected in order of distance from the center of gravity.

By using the validation methods (3) and (4), it is possible to more reliably select a cut set that is likely to be actually inappropriate. Furthermore, according to the validation method (4), if there is a tendency among cut sets that are likely to be inappropriate, hierarchical clustering can be used to collect the cut sets with that tendency into one class. This makes it possible to efficiently select cut sets that are likely to be actually inappropriate.

(5) The validation method according to the fifth aspect is the validation method of (1) to (4), wherein the feature amount appears in text data indicating basic events forming the cut set. This value is vectorized by text analysis of words.

Thereby, the feature amount of the cut set can be calculated from the viewpoint of the words related to the basic event included in the text data of the basic event of the cut set.

(6) A validation method according to a sixth aspect is the validation method of (1) to (5), wherein in the step of extracting, each of the first The cut set in which the proximity of the feature amount is lower than a predetermined value is extracted from the class.

(7) A validation method according to a seventh aspect is the validation method of (1) to (5), wherein in the extracting step, each of the first The cut set is extracted from the class based on the probability of occurrence of the cut set.

By the validation methods (6) to (7), it is possible to unbiasedly extract the cut set to be checked from a wide range.

(8) The validation method according to the eighth aspect is the validation method described in (1) to (7), which is the validation method described in (1) that is determined to have no validity. The step of performing the calculation described in (2), the determining step of (2), and the sorting step of (2) are performed repeatedly until the cut set is no longer extracted. and the step of accepting the modification information described in (2).

As a result, the probabilistic risk assessment model (PRA model) is corrected while accumulating cut sets whose validity has been confirmed. When a cut set is no longer valid, learning based on validated cut sets (non-hierarchical clustering) is used to select cutsets that are likely to be inappropriate from unconfirmed cutsets, making it efficient and labor-saving. The probabilistic risk assessment model (PRA model) can be modified by finding inappropriate cut sets.

(9) The validation system according to the ninth aspect includes a feature calculation unit 22 that calculates a feature of a cut set output from a probabilistic risk assessment model (PRA model), and a a hierarchical clustering unit 23 that performs hierarchical clustering based on feature amounts; an extraction unit 24 that extracts one or more of the cut sets included in each first class generated by the hierarchical clustering; The apparatus further includes a modification receiving unit 12 that receives modification information of the probabilistic risk assessment model (PRA model) when the validity evaluation result for the extracted cut set is not valid.

(10) The validation system according to the tenth aspect is the validation system according to (9), which performs non-hierarchical clustering based on the feature amount of the cut set for which the evaluation result is validated. a non-hierarchical clustering unit 25 that generates one or more second classes by calculating the centroid of the second class and a range in which the cut set can be considered valid, and the extraction unit 24 that generates one or more second classes. A determination unit (non-hierarchical clustering unit 25) that determines the validity of the cut set based on the range, and a cut that is determined to be invalid by the determination unit (non-hierarchical clustering unit 25). It further includes a sorting unit 26 that performs sorting on the set based on the distance from the center of gravity, and the correction reception unit 12, if the validity evaluation result for the selected cut set is not valid, Receiving modification information for the probabilistic risk assessment model (PRA model).

(11) The program according to the eleventh aspect includes the step of causing the computer 900 to calculate the feature amount of a cut set output from a probabilistic risk assessment model (PRA model), and calculating the feature amount of the cut set output from the probabilistic risk assessment model (PRA model). performing hierarchical clustering based on the hierarchical clustering; extracting one or more of the cut sets included in each first class generated by the hierarchical clustering; A step of acquiring a validity evaluation result, and a step of receiving correction information of the probabilistic risk assessment model (PRA model) when the evaluation result is not valid are executed.

(12) The program according to the twelfth aspect is the program according to (11), which causes the computer 900 to perform non-hierarchical clustering based on the feature amount of the cut set for which the evaluation result is valid. a step of generating one or more second classes, calculating the center of gravity of the second class and a range in which the cut set can be considered valid; and regarding the cut set that was not extracted in the extracting step, a step of determining validity based on the range; a step of further selecting the cut set determined to be invalid in the determining step based on a distance from the center of gravity; If the validity evaluation result for the cut set is not valid, the step of receiving modification information of the probabilistic risk assessment model (PRA model) is further executed.

1... Validity confirmation system 10... PRA model management device 11... Cut set output unit 12... Correction reception unit 13... Storage unit 20... Validity confirmation support device 21... Cut set acquisition unit 22...feature calculation unit 23...hierarchical clustering unit 24...extraction unit 25...non-hierarchical clustering unit 26...selection unit 27...output unit 28... -Storage unit 900...Computer 901...CPU,
902...Main storage device,
903... Auxiliary storage device,
904...I/O interface 905...Communication interface

Claims (12)

A computer-implemented validation method, comprising:
a step of calculating features of a cut set output from the probabilistic risk assessment model;
performing hierarchical clustering on the cut set based on the feature amount;
extracting one or more of the cut sets included in each first class generated by the hierarchical clustering;
obtaining a validity evaluation result for the extracted cut set;
If the evaluation result is not valid, receiving correction information for the probabilistic risk evaluation model;
has
In the step of calculating, for one of the cut sets, the number of occurrences of a word that appears in the text data indicating the base event constituting the cut set, and whether the word also appears in the other cut sets. Calculate a numerical value of the tendency as the feature amount,
The validity evaluation result includes an evaluation result by an engineer as to whether there are any errors in the basic events included in the cut set, and whether there are no errors in the combination conditions of the basic events,
The correction information for the probabilistic risk assessment model includes information for correcting the error created by the engineer when the validity evaluation result indicates that there is an error. Become,
Validation method. One or more second classes are generated by non-hierarchical clustering based on the feature amount of the cut set for which the evaluation result is valid, and the center of gravity of the second class and the cut set are determined to be valid. a step of calculating a range that can be considered;
determining the validity of the cut set that was not extracted in the extracting step based on the range;
further sorting the cut set determined to be invalid in the determining step based on the distance from the center of gravity;
If the validity evaluation result for the selected cut set is not valid, receiving correction information for the probabilistic risk assessment model;
The validation method according to claim 1, further comprising: In the step of sorting, a predetermined number of the cut sets are sorted in order of distance from the center of gravity.
The validation method according to claim 2. In the step of performing the sorting, hierarchical clustering is further performed to generate clusters, and for each cluster, a predetermined number of the cut sets are selected in order of distance from the center of gravity.
The validation method according to claim 2. The feature amount is a value obtained by vectorizing words appearing in text data indicating basic events forming the cut set by text analysis.
The validation method according to any one of claims 1 to 4. In the extracting step, extracting the cut set in which the proximity of the feature amount is lower than a predetermined value from each of the first classes generated by the hierarchical clustering.
The validation method according to any one of claims 1 to 5. In the extracting step, the cut set is extracted from each of the first classes generated by the hierarchical clustering based on the probability of occurrence of the cut set.
The validation method according to any one of claims 1 to 5. The validation method according to claim 1 is repeatedly executed until the cut set that is determined to have no validity is extracted, and then the step of calculating according to claim 2; executing the determining step as described in claim 2, the step of performing the sorting as described in claim 2, and the step of receiving the modification information as described in claim 2;
Validation method. a feature calculation unit that calculates features of a cut set output from the probabilistic risk assessment model;
a hierarchical clustering unit that performs hierarchical clustering on the cut set based on the feature amount;
an extraction unit that extracts one or more of the cut sets included in each first class generated by the hierarchical clustering;
a modification reception unit that receives modification information of the probabilistic risk assessment model when the validity evaluation result for the extracted cut set is not valid;
has
The feature calculation unit calculates, for one cut set, the number of occurrences of a word that appears in text data indicating basic events constituting the cut set, and the number of occurrences of the word in the other cut set. Calculate the numerical value of the tendency as the feature amount,
The validity evaluation result includes an evaluation result by an engineer as to whether there are any errors in the basic events included in the cut set, and whether there are no errors in the combination conditions of the basic events,
The correction information for the probabilistic risk assessment model includes information for correcting the error created by the engineer when the validity evaluation result indicates that there is an error. Become,
Validation system. One or more second classes are generated by non-hierarchical clustering based on the feature amount of the cut set for which the evaluation result is valid, and the center of gravity of the second class and the cut set are determined to be valid. a non-hierarchical clustering unit that calculates a range that can be considered;
a determination unit that determines the validity of the cut set not extracted by the extraction unit based on the range;
a sorting unit that further sorts the cut set determined to be invalid by the determining unit based on the distance from the center of gravity;
It further has
The modification receiving unit receives modification information of the probabilistic risk assessment model if the validity evaluation result for the selected cut set is not valid.
Validation system according to claim 9. to the computer,
a step of calculating features of a cut set output from the probabilistic risk assessment model;
performing hierarchical clustering on the cut set based on the feature amount;
extracting one or more of the cut sets included in each first class generated by the hierarchical clustering;
obtaining a validity evaluation result for the extracted cut set;
If the evaluation result is not valid, receiving correction information for the probabilistic risk evaluation model;
A program that executes
In the step of calculating, for one of the cut sets, the number of occurrences of a word that appears in the text data indicating the base event constituting the cut set, and whether the word also appears in the other cut sets. Calculate a numerical value of the tendency as the feature amount,
The validity evaluation result includes an evaluation result by an engineer as to whether there are any errors in the basic events included in the cut set, and whether there are no errors in the combination conditions of the basic events,
The correction information for the probabilistic risk assessment model includes information for correcting the error created by the engineer when the validity evaluation result indicates that there is an error. Become,
program. One or more second classes are generated by non-hierarchical clustering based on the feature amount of the cut set for which the evaluation result is valid, and the center of gravity of each second class and the cut set are determined to be valid. a step of calculating a range that can be considered to exist;
determining the validity of the cut set that was not extracted in the extracting step based on the range;
further sorting the cut set determined to be invalid in the determining step based on the distance from the center of gravity;
If the validity evaluation result for the selected cut set is not valid, receiving correction information for the probabilistic risk assessment model;
The program according to claim 11, further causing the program to execute.

Images