Threshold setting method, similar character string determination method, threshold setting device, and program

Abstract

To provide a threshold setting method capable of calculating criteria for determining a similarity degree of strings.SOLUTION: A threshold setting method includes: a step of calculating a string vector of each of multiple strings; a step of creating multiple combinations for selecting two strings out of the multiple string vectors and calculating a distance and a cosine similarity degree of the two string vectors selected for each of the combinations; a step of setting a label for determining whether the two selected string vectors are similar for each combination, and creating pieces of teacher data whose number is equal to the number of combinations; and a step of calculating a threshold of the distance and a threshold of the cosine similarity degree for determining the similarity degree between the strings on the basis of the teacher data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a threshold setting method, a similar string determination method, a threshold setting device, reference and programs. [Background technology]

[0002] A probabilistic risk assessment (PRA) model (hereinafter referred to as a PRA model) is sometimes used for risk assessment of nuclear power plants (see Patent Document 1). To maintain the accuracy of risk assessment, it is necessary to verify the validity of the cutsets of top events generated based on the PRA model. A cutset is expressed as a string of characters combining codes representing the plant's systems, equipment, and types of failures. The number of cutsets of top events output from a PRA model of a nuclear power plant can reach tens of thousands to hundreds of thousands per plant, and a method for efficiently verifying a huge number of cutsets is required. For example, a method can be considered in which a group of cutsets whose validity has already been verified is considered correct, and a group of cutsets to be evaluated is compared with a group of confirmed cutsets to evaluate them. In this method, if a cutset similar to the group of cutsets to be evaluated is found by comparing it with the group of confirmed cutsets, the cutset is deemed valid. However, if a cutset present in the group of confirmed cutsets is not found in the group of cutsets to be evaluated, it is determined that the group of cutsets to be evaluated may be missing a cutset that should have been present. By checking while referring to a group of confirmed cut sets, the validity of the group of cut sets to be evaluated can be efficiently confirmed. When checking in this way, it is necessary to determine whether the confirmed cut sets and the cut sets to be evaluated are similar. Patent Document 2 discloses a method for determining the relevance (similarity) of character data using a feature vector generated from the character data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 5480033 [Patent Document 2] Japanese Patent Application Laid-Open No. 2014-112358 Summary of the Invention [Problem to be solved by the invention]

[0004] In order to determine whether character strings are similar, it is necessary to define a criterion for determining the degree of similarity.

[0005] The present disclosure provides a threshold setting method, a similar string determination method, and a threshold setting device that can solve the above problems. reference and programs. [Means for solving the problem]

[0006] The threshold setting method of the present disclosure includes: The computer calculating a string vector for each of a plurality of strings; The computer creating a plurality of combinations of selecting two of the plurality of character string vectors, and calculating the distance and cosine similarity between the two character string vectors selected for each of the combinations; The computer a step of setting a label indicating whether the two selected character string vectors are similar or not for each of the combinations, and creating training data for the number of combinations; The computer and calculating, based on the training data, a threshold value for the distance and a threshold value for the cosine similarity for determining the similarity of the character strings.

[0008] The threshold setting device of the present disclosure also includes a means for calculating a character string vector for each of a plurality of character strings, a means for creating a plurality of combinations by selecting two from the plurality of character string vectors and calculating the distance and cosine similarity between the two selected character string vectors for each combination, a means for labeling each of the combinations to indicate whether the two selected character string vectors are similar or not, and creating training data for the number of combinations, and a means for calculating the distance threshold and the cosine similarity threshold for determining the similarity of the character strings based on the training data.

[0010] The program disclosed herein also causes a computer to execute the steps of: calculating a character string vector for each of a plurality of character strings; creating a plurality of combinations by selecting two from the plurality of character string vectors, and calculating the distance and cosine similarity between the two selected character string vectors for each combination; labeling each of the combinations to indicate whether the two selected character string vectors are similar or not, and creating training data for the number of combinations; and calculating, based on the training data, the distance threshold and the cosine similarity threshold for determining the similarity of the character strings. [Effects of the Invention]

[0012] The threshold setting method, threshold setting device, and program disclosed herein can calculate a threshold (criterion) for determining the similarity of character strings. The similar character string determination method, similar character string determination device, and program disclosed herein can determine the similarity of character strings. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a block diagram illustrating an example of a string classification device according to an embodiment. [Figure 2] FIG. 10 is a diagram for explaining a cutset of a top event according to an embodiment. [Figure 3A] FIG. 1 is a first diagram showing an example of a cut set according to an embodiment. [Figure 3B] FIG. 2 is a second diagram showing an example of a cut set according to the embodiment. [Figure 3C] FIG. 10 is a third diagram showing an example of a cut set according to the embodiment. [Figure 4A] FIG. 10 is a diagram illustrating an example of distribution of character string vectors in a vector space according to the embodiment. [Figure 4B] 10A and 10B are diagrams illustrating an example of distribution of inter-vector distances and cosine similarities according to the embodiment. [Figure 5] FIG. 1 is a diagram illustrating an overview of a learning process according to an embodiment. [Figure 6] FIG. 10 is a diagram illustrating an example of a classification result according to the embodiment. [Figure 7] FIG. 10 is a diagram illustrating an example of a character string classification process according to the embodiment. [Figure 8A] FIG. 10 is a diagram illustrating an example of a group of cut sets to be processed according to the embodiment. [Figure 8B] FIG. 8B is a diagram showing an example of a classification result of the cutset group illustrated in FIG. 8A. [Figure 9] FIG. 1 illustrates an example of a hardware configuration of a string classification device according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] <Embodiment> The string classification device of the present disclosure will be described below with reference to FIGS. (composition) FIG. 1 is a block diagram illustrating an example of a string classification device according to an embodiment. The string classifier 10 determines whether a string is similar to another string. Furthermore, by determining the mutual similarity or dissimilarity of multiple strings, the string classifier 10 can classify multiple strings into groups of similar ones. The string classifier 10 can be used to confirm the validity of the cut sets of the top events that make up the PRA model.

[0015] The string classification device 10 includes a string acquisition unit 11, an input reception unit 12, a control unit 13, and a storage unit 18.

[0016] The character string acquiring unit 11 acquires a character string (text data). For example, the character string acquiring unit 11 acquires a cut set extracted from the PRA model to be evaluated.

[0017] Here, cutsets will be explained. FIG. 2 is a diagram for explaining cutsets of top events. FIG. 2 shows an example of a fault tree (FT) included in the PRA model. In FIG. 2, circles A to E indicate primitive events, and squares P to Q indicate intermediate events. The OR gate indicates that a top event occurs when any of primitive events A, C, and intermediate event P occurs. The upper AND gate indicates that intermediate event P occurs when primitive event B and intermediate event Q occur. The lower AND gate indicates that intermediate event Q occurs when primitive events D and E occur. In the case of the fault tree (FT) of FIG. 2, the cutsets that cause a top event are (1) primitive event A, (2) a combination of primitive events B, D, and E, and (3) primitive event C. The character string acquisition unit 11 acquires, for example, cutset 1 (primitive event A), cutset 2 (primitive event B, D, and E), and cutset 3 (primitive event C) extracted based on the fault tree (FT) of FIG. 2.

[0018] Examples of cutsets are shown in Figures 3A to 3C. The cutset shown in Figure 3A is an example of a cutset consisting of one primitive event. The text data representing one primitive event has semantic delimiters every three characters, two characters, two characters, and four characters. Each delimited string represents, from the beginning, the system, the faulty equipment, the type of failure, the equipment number, and other information (e.g., common cause failure, human error, etc.). In strings constructed based on these rules, similar strings have similar meanings. Figure 3B shows two cutsets that indicate the same equipment in the same system is failing, but the type of failure is different. For example, cutset 1 in Figure 3B is a combination of the primitive events "AFWOC01003" and "IASPQADAE1A," and cutset 2 is a combination of the primitive events "AFWOC01003" and "IASPQCSAE1A." The only difference between cutset 1 and cutset 2 is the third delimiter character of the second primitive event ("AD" and "CS"). In such cases, from an engineer's perspective, these cutsets are considered to be of the same type, and are therefore classified into the same group. On the other hand, when multiple devices fail due to a failure caused by a common cause, the suffix "-ALL" is added to the string, as shown in cutset 1 in Figure 3C. Cutset 1 in Figure 3C, which has this meaning, and cutset 2 in Figure 3C, which shows failures caused by individual, uncommon causes, are considered to be different types of cutsets from an engineer's perspective, and are therefore classified into different groups. The character string acquisition unit 11 acquires cut sets such as those illustrated in FIGS. 3A to 3C as examples of character strings.

[0019] The input receiving unit 12 receives (acquires) various setting information input using an input device such as a keyboard, a mouse, a touch panel, or a button, instruction information for instructing the execution of processing, etc. The input receiving unit 12 records the received information in the storage unit 18 or outputs it to the control unit 13. For example, the input receiving unit 12 receives a setting that cut sets 1 and 2 in FIG. 3B are similar, and a setting that cut sets 1 and 2 in FIG. 3C are dissimilar.

[0020] The control unit 13 executes processes such as calculating a threshold value for determining whether character strings are similar, determining whether a character string to be evaluated is similar to another character string (e.g., a reference character string that serves as a standard) using the threshold value, and classifying multiple character strings (e.g., the character string to be evaluated and the reference character string) into groups of similar characters. The control unit 13 includes a feature calculation unit 14, a learning unit 15, an evaluation unit 16, and an output unit 17.

[0021] The feature calculation unit 14 calculates the feature of a character string. For example, if the character string is a cutset, the base events included in the cutset are coded according to a predetermined rule as shown in FIG. 3A. The coded text data is segmented into predetermined units, and the segmented character strings are treated as words and the cutsets as documents. Using text analysis techniques such as TFIDF and Word2Vec, feature calculations are performed based on the words that appear in the cutset. The feature calculation results are expressed as a vector. This vector is called a character string vector. If the character string is a cutset, each element of the character string vector is a value obtained by text analysis of the character strings (words) obtained by segmenting the text data of the base events.

[0022] FIG. 4A shows the distribution of character string vectors in a vector space. In this example, the character string vectors are three-dimensional vectors. Each point in FIG. 4A represents one character string vector. The character string vector enclosed by g1 is a character string vector generated from character strings with similar meanings. The same applies to g2 to g4. In other words, character string vectors generated from character strings with similar meanings exist in close positions in the vector space. Using this property, the feature calculation unit 14 calculates a feature that represents the proximity of the positions of two character string vectors. Specifically, the feature calculation unit 14 calculates the inter-vector distance and cosine similarity between the two character string vectors. As a simple example of two-dimensional vectors, let character string vector 1 be (a1, a2) and character string vector 2 be (b1, b2). The inter-vector distance between character string vector 1 and character string vector 2 can be calculated using the following formula (1).

[0023]

number

[0024] The cosine similarity between character string vector 1 and character string vector 2 can be calculated using the following formula (2).

[0025]

number

[0026] Figure 4B shows an example of the distribution of inter-vector distance and cosine similarity between character string vectors of semantically similar character strings and the distribution of inter-vector distance and cosine similarity between character string vectors of dissimilar character strings. The vertical axis of Figure 4B represents cosine similarity ("cosθ between vectors"), and the horizontal axis represents inter-vector distance. Points p1 to p3 represent inter-vector distance and cosine similarity between character string vectors generated from two semantically similar character strings, while points p4 to p6 represent inter-vector distance and cosine similarity between character string vectors generated from two semantically dissimilar character strings. When two character string vectors completely match, the inter-vector distance is "0" and the cosine similarity is "1." For example, point p1 in Figure 4B represents two character string vectors in such a relationship. It can be seen that the inter-vector distance is relatively small and the cosine similarity tends to be close to 1 for the other points p2 and p3 when the meanings are similar. On the other hand, for points p4 to p6 when the meanings are dissimilar, point p4 has a cosine similarity close to "1" but a relatively large inter-vector distance. Point p5 does not have a very large inter-vector distance, but its cosine similarity is farther from "1" than points p1 to p3 when the meanings are similar. Point p6 has a large inter-vector distance and a cosine similarity farther from "1". From these facts, if the cosine similarity of two character string vectors is close to "1" and the inter-vector distance is close to "0", it is considered that the meanings of the two character string vectors are likely to be similar. In this embodiment, the learning unit 15, which will be described next, learns the cosine similarity and inter-vector distance of the character string vectors of two character strings that are similar in meaning and the cosine similarity and inter-vector distance of the character string vectors of two character strings that are dissimilar in meaning by machine learning, and calculates the boundary (threshold) of the inter-vector distance and cosine similarity that separates character string vectors that are similar in meaning from those that are not.

[0027] Furthermore, the feature calculation unit 14 calculates the standard deviation of the inter-vector distance and the standard deviation of the cosine similarity. For example, if there are 10 character strings to be processed, the feature calculation unit 14 calculates character string vectors for the 10 character strings, and calculates the standard deviation of the cosine similarity based on these 10 character string vectors. 10 C2 combinations (i.e.,10 The feature calculation unit 14 calculates the inter-vector distance and cosine similarity of the 45 inter-vector distances and the standard deviation of the 45 cosine similarities. The inter-vector distances and cosine similarities calculated by the feature calculation unit 14 are labeled as "similar" or "dissimilar" and used as training data for calculating thresholds for determining whether two character strings are similar to each other at certain inter-vector distances and cosine similarities. This is because the degree of variation in inter-vector distances and cosine similarities of all samples (10 character strings in this example) may affect the calculation of the thresholds for inter-vector distances and cosine similarities. The feature calculation unit 14 calculates character string vectors for the reference character string (character string used as training data) and the character string to be evaluated, and calculates the inter-vector distance, cosine similarity, standard deviation of inter-vector distances, and standard deviation of cosine similarities of the two character string vectors.

[0028] The learning unit 15 learns training data in which the inter-vector distance, cosine similarity, standard deviation of inter-vector distance, and standard deviation of cosine similarity calculated for a reference character string are labeled as "similar" or "dissimilar," and calculates thresholds for determining that two character strings are similar. The thresholds include the inter-vector distance, cosine similarity, standard deviation of inter-vector distance, and standard deviation of cosine similarity. Figure 5 shows how the cosine similarity and inter-vector distance thresholds are calculated from the training data by machine learning. For example, threshold Th1 is the inter-vector distance threshold, and threshold Th2 is the cosine similarity threshold.

[0029] The evaluation unit 16 determines whether the character string to be evaluated is similar to the reference character string based on the threshold calculated by the learning unit 15. Alternatively, the evaluation unit 16 classifies the character string to be evaluated and the reference character string into groups of similar character strings. Taking the case where the character string vectors are two-dimensional as an example, an example of a classification result in which multiple character strings are classified into groups of character strings with similar meaning is shown in FIG. 6. In the vector space, the evaluation unit 16 determines that a character string vector within a threshold range calculated by the learning unit 15 is a character string vector with similar meaning to a given character string vector. For example, in the example of FIG. 6, character string vectors v1 to v2 within range e1 are similar to each other, character string vectors v3 to v7 within range e2 are similar to each other, and character string vectors v8 to v10 within range e3 are similar to each other. Furthermore, character string vectors v11 and v12 are character string vectors that are dissimilar to character string vectors v1 to v10. If an engineer checks and determines that the classification by the evaluation unit 16 is inappropriate, the engineer can again appropriately label the combinations of character strings that were not properly classified as "similar" or "dissimilar," provide this as training data, and have the learning unit 15 learn from it, thereby improving the accuracy of the threshold value.

[0030] The output unit 17 displays various information, such as the result of determining whether the character string to be evaluated is similar to a reference character string, and the result of classifying multiple character strings into similar groups as shown in Figure 6, on a display device or outputs it as electronic data. The storage unit 18 stores character strings such as cut sets, data being processed, and the like.

[0031] (operation) Next, the operation of the string classification device 10 will be described using an example in which the strings are cut sets, with reference to FIGS. 7 to 8B. FIG. 7 illustrates an example of string classification processing according to the embodiment. First, an engineer classifies confirmed cut sets into cut sets with similar meanings. FIG. 8A illustrates an example of a group of cut sets to be processed. Assume that the PRA models for plants 1 to 3 have already been confirmed. That is, cut sets 11 to 13 for plant 1, 21 to 23 for plant 2, and 31 to 33 for plant 3 are reference strings. The engineer classifies the reference strings into cut sets with similar meanings. For example, the engineer classifies cut sets 11, 12, 21, 22, 31, and 32 into one group, and cut sets 13, 23, and 33 into another group. FIG. 8B illustrates an example of the classification result. The engineer inputs the classification result to the string classification device 10. The input receiving unit 12 acquires information on the classification result of strings (cut sets) with similar meanings (step S1) and records it in the storage unit 18.

[0032] Next, the engineer gives an instruction to calculate a threshold value. The input receiving unit 12 receives this instruction and outputs instruction information for calculating a threshold value to the control unit 13. Then, in the control unit 13, the feature amount calculation unit 14 calculates character string vectors, inter-vector distances, similarities, etc., and creates training data (step S2). First, the feature amount calculation unit 14 calculates character string vectors. The feature amount calculation unit 14 uses a text analysis technique such as word2vec to generate character string vectors for each of cut sets 11 to 13, 21 to 23, and 31 to 33. The feature amount calculation unit 14 records the calculated character string vectors in the storage unit 18. Next, the feature amount calculation unit 14 calculates the inter-vector distance and cosine similarity between the character string vector of cut set 11 and the character string vector of cut set 12. Similarly, the feature calculation unit 14 calculates the inter-vector distance and cosine similarity between the character string vector of cut set 11 and each of the character string vectors of cut sets 13, 21-23, and 31-33 (hereinafter referred to as cut sets 13-33, etc.). The feature calculation unit 14 also calculates the inter-vector distance and cosine similarity for combinations of character string vectors of the other two cut sets. In addition, the feature calculation unit 14 calculates the standard deviation of all the calculated inter-vector distances, and calculates the standard deviation of all the calculated cosine similarities. The feature calculation unit 14 records the various calculated data in the storage unit 18.

[0033] Next, the control unit 13 assigns a label of "similar" or "dissimilar" to the inter-vector distance, cosine similarity, standard deviation of inter-vector distance, and standard deviation of cosine similarity calculated for the two character string vectors based on the classification results by the engineer. For example, the control unit 13 assigns a label of "similar" to the character string vectors of cut sets 11 and 12, and the character string vectors of cut sets 11 and 32. For example, the control unit 13 assigns a label of "dissimilar" to the character string vectors of cut sets 11 and 13, and the character string vectors of cut sets 11 and 33. This generates training data. The training data includes the inter-vector distance, cosine similarity, standard deviation of inter-vector distance, standard deviation of cosine similarity, and labels. Training data is generated for each combination of cut sets 11 to 33.

[0034] The control unit 13 (feature calculation unit 14) may perform the above process (calculating inter-vector distances, labeling them as "similar" or "dissimilar," and generating training data) multiple times for the same group of cut sets (e.g., cut sets 11 to 33). When generating character string vectors using word2vec or similar, different character string vectors may be generated each time the process is performed, even if the original character strings are the same. If different character string vectors are generated, the calculated inter-vector distances and cosine similarities will also vary, resulting in different variations. Since variations in inter-vector distances and cosine similarities across the entire data may affect the calculation of the thresholds for inter-vector distances and cosine similarities, the standard deviations of inter-vector distances and cosine similarities are calculated and included in the training data. By performing the above process multiple times for the same group of cut sets, the training data can be increased.

[0035] Furthermore, the control unit 13 (feature calculation unit 14) may generate training data by performing the above-mentioned processing (calculating the distance between vectors, labeling the cutsets as "similar" or "dissimilar," and generating training data) on different groups of cutsets. For example, training data may first be generated based on the cutsets of plants 1 to 3, and then a group of cutsets of plants 4 to 5 may be added, training data may be generated based on the cutsets of plants 4 to 5, and these training data may be added. Furthermore, when adding the cutsets of plants 4 to 5, the cutsets of plants 4 to 5 may be added to the original plants 1 to 3, and the above-mentioned processing may be performed on all the cutsets of plants 1 to 5 to generate training data.

[0036] Next, the learning unit 15 learns the created training data using an arbitrary machine learning method and calculates thresholds (step S3). For example, the learning unit 15 calculates a threshold Th1 for inter-vector distance, a threshold Th2 for cosine similarity, a threshold Th3 for standard deviation of inter-vector distance, and a threshold Th4 for standard deviation of cosine similarity using a decision tree to determine whether two character string vectors are similar. The learning unit 15 records the calculated thresholds in the storage unit 18.

[0037] Next, an engineer inputs cut sets A1 to A3 of plant A to be evaluated into string classifier 10. Input receiver 12 acquires the strings to be evaluated (cut sets A1 to A3) (step S4) and records them in memory 18. Next, the engineer instructs that the strings to be evaluated be classified into groups of similar strings. Control unit 13 then classifies the string vectors of cut sets A1 to A3 and cut sets 11 to 33 to be evaluated into groups of similar strings based on the threshold calculated in step S3 (step S5).

[0038] First, the feature calculation unit 14 calculates character string vectors for the cut sets A1 to A3 to be evaluated and the reference cut sets 11 to 33 using word2vec or the like, and stores the calculated vectors in the storage unit 18. For the cut sets 11 to 33, previously calculated character string vectors may be used. Next, the feature calculation unit 14 calculates the inter-vector distance and cosine similarity between the character string vector of the cut set A1 and the character string vectors of the cut sets A2 to A3 and the cut sets 11 to 33. Similarly, for the character string vectors of the cut sets A2 and A3, the feature calculation unit 14 calculates the inter-vector distance and cosine similarity between these character string vectors and the character string vectors of the other cut sets A1, etc. and the cut sets 11 to 33. Furthermore, the feature calculation unit 14 calculates the standard deviation of the inter-vector distance and the standard deviation of the cosine similarity across all of the cut sets A1 to A3 and 11 to 33. For the inter-vector distance and cosine similarity of the cut sets 11 to 33, previously calculated values ​​may be used.

[0039] Next, the evaluation unit 16 determines whether the character string vector of the cutset A1 is similar to any of the other character string vectors of the cutsets A2 to A3 and the cutsets 11 to 33, based on the threshold calculated in step S3.

[0040] For example, suppose that the inter-vector distance between the character string vectors of cutset A1 and cutset 11 is r1, the cosine similarity is θ1, the standard deviation of the inter-vector distances of all samples is d1, and the standard deviation of the cosine similarity is d2. For example, the evaluation unit 16 determines that the character string vectors of cutset A1 and cutset 11 are similar if the inter-vector distance r1 is equal to or smaller than a threshold Th1, the cosine similarity θ1 is equal to or larger than a threshold Th2, the difference between the standard deviation d1 of the inter-vector distances and the threshold Th3 is within a predetermined range, and the difference between the standard deviation d2 of the cosine similarity and the threshold Th4 is within a predetermined range. (Since it may not be appropriate to judge similarity using thresholds calculated using training data in which the standard deviations of vector distance and cosine similarity are too different, the standard deviation of vector distance and the standard deviation of cosine similarity are used to determine whether it is appropriate to judge the similarity of the strings of the cut sets being evaluated using the current thresholds.) The evaluation unit 16 calculates the vector distance and cosine similarity calculated for all combinations of cut sets A1 to A3 and 11 to 33 (the vector distance and cosine similarity between cut sets 11 to 33 may be calculated again in step S5, or the ones calculated in step S2 may be used), and for each combination, the vector distance, cosine similarity, standard deviation of vector distance, and standard deviation of cosine similarity are judged based on the above thresholds Th1 to Th4, and similar items are grouped together.

[0041] The output unit 17 outputs the classification results to a display device or the like (step S6). The display device displays the classification results, for example, as shown in FIG. 6. The engineer looks at the classification results, checks the groups into which the cut sets A1 to A3 have been classified, and confirms whether the classification is appropriate (step S7). If the classification results are not appropriate (step S7; No), the engineer correctly classifies the inappropriately classified character strings (step S8). For example, in FIG. 6, if the character string vector v12 (assuming v12 is the character string vector to be evaluated) is derived from a cut set whose meaning is similar to that of the character string vectors v8 to v10, the character string vectors v8 and v12, the character string vectors v9 and v12, and the character string vectors v10 and v12 are each labeled as "similar." These data are added as new training data. Then, the process from step S2 is repeated to recalculate the threshold.

[0042] On the other hand, if the classification result is valid (step S7; Yes), the output unit 17 outputs a list of character strings having similar meanings (step S9). For example, if the cut sets A1 and A2 in Fig. 8A are classified into group 1 in Fig. 8B and the cut set A3 in Fig. 8A is classified into group 2 in Fig. 8B, a list is output in which the cut sets A1 and A2 are added to group 1 in Fig. 8B and the cut set A3 is added to group 2 as group 1.

[0043] Cutsets of confirmed plants (in this example, cutsets 11 to 33 of plants 1 to 3) that are expected to have cutsets similar to those of plant A to be evaluated (e.g., because the plant configuration is similar) are prepared, a threshold is calculated based on these (step S3), and the cutsets of plant A to be evaluated are classified based on this threshold (step S6). If the cutsets of plants 1 to 3 and plant A are all similar, only groups that contain an equal number of cutsets from each plant should be generated. For example, if a group does not contain a cutset from plant A, that cutset may be missing from the PRA model of plant A. Also, if the cutset of plant A contains a cutset that is not present in plants 1 to 3, a group containing only the cutset from plant A will be generated. By referring to the list output in step S9 and focusing on the differences in the distribution of cut sets between plant A and plants 1 to 3, engineers can find differences between the PRA model of plant A and the PRA models of plants 1 to 3, and by referring to the PRA models of plants 1 to 3, they can quickly correct any defects in the PRA model of plant A.

[0044] In the above embodiment, a cutset extracted from a PRA model was used as an example of a character string. However, this embodiment can be used to evaluate the similarity of any character string or to classify similar character strings. For example, this embodiment can be used to classify character strings with similar meanings, taking into account spelling variations. In the above embodiment, the thresholds used were the vector distance, cosine similarity, the standard deviation of the vector distance, and the standard deviation of the cosine similarity. However, a configuration using only the vector distance and the cosine similarity is also possible. Furthermore, while the training data was labeled as "similar" or "dissimilar," it may be possible to label the data according to the degree of similarity, such as "high similarity," "moderate similarity," "low similarity," or "dissimilar," and calculate the thresholds for determining "high similarity," "moderate similarity," and "low similarity" through learning. For example, when using this embodiment to confirm cutsets, it is possible to focus on cutsets determined to have "moderate similarity," thereby further streamlining the PRA model confirmation process. In addition, in the explanation of FIG. 7, the threshold value is calculated using the cut set of plants 1 to 3, but the number of plants used to calculate the threshold value may be two or less, or may be four or more.

[0045] (effect) As described above, according to this embodiment, it is possible to objectively evaluate whether a certain character string is similar to another character string. Furthermore, in order to determine the similarity between character strings, it is necessary to define a criterion for determining the similarity. According to this embodiment, it is possible to calculate a threshold value for determining the similarity between character strings. Furthermore, even for a new character string to be evaluated, it is possible to classify the character string to be evaluated by providing training data and executing the process of FIG. 7.

[0046] Cutsets output from a PRA model can sometimes be generated on a scale of hundreds of thousands per plant. It is difficult for engineers to manually verify the validity of all of these cutsets and ensure that all necessary cutsets have been obtained. Using the string classification method of this embodiment, it is possible to efficiently detect missing cutsets that should have been generated by comparing them with other plants. Furthermore, by comparing cutsets between plants, it is possible to mutually check for deficiencies in the PRA model's input conditions (FT, ET, and their configurations), thereby enabling consistent evaluation between plants. Furthermore, cutset verification is a task that engineers perform continuously from the perspective of improving safety. By introducing the string classification device 10 of this embodiment, adding cutsets extracted from a verified PRA model and updating thresholds such as inter-vector distance and cosine similarity, or by calculating thresholds such as inter-vector distance and cosine similarity for each plant type and configuration (each PRA model content), it is possible to enjoy the long-term benefits of improved efficiency and labor savings in PRA model verification.

[0047] In the above embodiment, we have described how to reduce the labor and efficiency of the work of checking cut sets of a PRA model created in a nuclear power plant. As mentioned above, in a nuclear power plant, the number of cut sets can reach tens of thousands to hundreds of thousands, and the method for checking the validity of a PRA model of this embodiment is effective. However, the application field of the string classification device 10 is not limited to nuclear plants. It can be applied to any industrial field where risk assessment is performed using a PRA model.

[0048] FIG. 9 is a diagram illustrating an example of a hardware configuration of a string classification device. The computer 900 includes a CPU 901 , a main memory device 902 , an auxiliary memory device 903 , an input / output interface 904 , and a communication interface 905 . The above-described string classification device 10 is implemented in a computer 900. Each of the above-described 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, loads it into the main storage device 902, and executes the above-described processing in accordance with the program. The CPU 901 also allocates a storage area in the main storage device 902 in accordance with the program. The CPU 901 also allocates a storage area in the auxiliary storage device 903 for storing data being processed in accordance with the program.

[0049] A program for implementing all or part of the functions of the string classification device 10 may be recorded on a computer-readable recording medium, and the program may be loaded into a computer system and executed to perform processing by each functional unit. The term "computer system" as used herein includes hardware such as an OS and peripheral devices. Furthermore, if a WWW system is used, the term "computer system" also includes the homepage provision environment (or display environment). Furthermore, the term "computer-readable recording medium" refers to portable media such as CDs, DVDs, and USBs, as well as storage devices such as hard disks built into the computer system. Furthermore, if the program is distributed to the computer 900 via a communication line, the computer 900 that receives the program may load the program into the main storage device 902 and execute the above-described processing. Furthermore, the program may be for implementing part of the above-described functions, or may be capable of implementing the above-described functions in combination with a program already stored in the computer system.

[0050] As described above, several embodiments according to the present disclosure have been described, but all of 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 modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents as defined in the claims, as well as in the scope and spirit of the invention.

[0051] <Additional Notes> The threshold setting method, similar string determination method, threshold setting device, similar string determination device, and program described in the embodiments can be understood, for example, as follows.

[0052] (1) A threshold setting method according to a first aspect includes the steps of: calculating a character string vector for each of a plurality of character strings; creating a plurality of combinations of selecting two of the character string vectors, and calculating the distance (vector distance) and cosine similarity between the two selected character string vectors for each of the combinations; labeling each of the combinations to indicate whether the two selected character string vectors are similar or not, thereby creating training data for the number of combinations; and calculating, based on the training data, the distance threshold and the cosine similarity threshold for determining the similarity of the character strings. This makes it possible to calculate a standard (threshold) for determining the similarity of character strings.

[0053] (2) A threshold setting method according to a second aspect is the threshold setting method of (1), further comprising a step of calculating the standard deviation of the distance and the standard deviation of the cosine similarity for a plurality of the combinations, wherein in the step of calculating the thresholds, in addition to the distance threshold and the cosine similarity threshold, a threshold of the standard deviation of the distance and a threshold of the standard deviation of the cosine similarity are calculated. This makes it possible to determine the similarity of character strings using thresholds set by learning from training data with similar variations, depending on the degree of variation in the inter-vector distance and cosine similarity in the group of character strings used to calculate the thresholds.

[0054] (3) A similar string determination method according to the third aspect includes the steps of: calculating a string vector of a string to be evaluated; and determining whether at least one of one or more reference strings is similar to the string to be evaluated based on the string vector of the one or more reference strings calculated using the threshold setting method described in any one of (1) to (2), the threshold, and the string vector of the string to be evaluated. This allows you to determine whether the strings are similar.

[0055] (4) A similar string determination method according to a fourth aspect includes the steps of: calculating a string vector of a string to be evaluated; and classifying one or more reference strings into groups of similar strings based on the string vectors of the one or more reference strings calculated using the threshold setting method described in claim 1 or claim 2, the threshold, and the string vector of the string to be evaluated. This allows similar character strings to be classified into the same group.

[0056] (5) A similar string determination method according to a fifth aspect is a similar string determination method according to (3) to (4), further comprising the steps of: receiving an evaluation of the validity of the result of the determining step; and, based on the evaluation of the validity, performing label setting on the string to be evaluated and the combination of the string to be evaluated and the string to be evaluated for similarity to the string to be evaluated, and creating additional training data; and in the step of calculating the threshold, the threshold is calculated based on the training data and the additional training data. This makes it possible to improve the accuracy of the threshold value.

[0057] (6) A similar string determination method according to a sixth aspect is a similar string determination method according to any one of (3) to (5), wherein the one or more reference strings are strings representing a cutset extracted from an evaluated PRA (Probabilistic Risk Assessment) model, and the string to be evaluated is a string representing a cutset extracted from the PRA model to be evaluated. This allows flaws in the PRA model to be quickly identified and corrected.

[0058] (7) A similar string determination method according to a seventh aspect includes the steps of obtaining thresholds for distance and cosine similarity of string vectors for determining the similarity of strings, calculating the string vectors of a string to be evaluated and a reference string, and determining the similarity between the string to be evaluated and the reference string based on whether the difference between the string vector of the string to be evaluated and the string vector of the reference string falls within the range of the distance threshold and the cosine similarity threshold. This allows you to determine whether the strings are similar.

[0059] (8) A threshold setting device according to an eighth aspect includes: means for calculating a character string vector for each of a plurality of character strings; means for creating a plurality of combinations by selecting two from the plurality of character string vectors and calculating the distance and cosine similarity between the two selected character string vectors for each of the combinations; means for labeling each of the combinations to indicate whether the two selected character string vectors are similar or not, and creating training data for the number of combinations; and means for calculating, based on the training data, the distance threshold and the cosine similarity threshold for determining the similarity of the character strings.

[0060] (9) A similar string determination device according to a ninth aspect includes a means for acquiring a threshold value for distance and cosine similarity of string vectors for determining the similarity of strings, a means for calculating the string vectors of a string to be evaluated and a reference string, and a means for determining the similarity between the string to be evaluated and the reference string based on whether the difference between the string vector of the string to be evaluated and the string vector of the reference string falls within the range of the threshold value for distance and the threshold value for cosine similarity.

[0061] (10) A program according to a tenth aspect causes a computer to execute the steps of: calculating a character string vector for each of a plurality of character strings; creating a plurality of combinations by selecting two from the plurality of character string vectors, and calculating the distance and cosine similarity between the two selected character string vectors for each of the combinations; labeling each of the combinations to indicate whether the two selected character string vectors are similar or not, thereby creating training data for the number of combinations; and calculating, based on the training data, the distance threshold and the cosine similarity threshold for determining the similarity of the character strings.

[0062] (11) A program according to an eleventh aspect causes a computer to execute the steps of acquiring a threshold for distance and cosine similarity of character string vectors for determining the similarity of the character strings, calculating the character string vectors of the character string to be evaluated and the reference character string, and determining the similarity between the character string to be evaluated and the reference character string based on whether the difference between the character string vector of the character string to be evaluated and the character string vector of the reference character string falls within the range of the threshold for distance and the threshold for cosine similarity. [Explanation of symbols]

[0063] 10...String classifier 11...Character string acquisition part 12 Input reception section 13 Control section 14. Feature calculation unit 15. Learning Department 16. Evaluation section 17. Output section 18...Storage section 900···Computer 901 CPU 902...Main memory 903...Auxiliary storage device 904 Input / Output Interface 905···Communication Interface

Claims

1. A method for generating a character string vector by a computer, the method comprising: a step in which the computer creates a plurality of combinations in which two of the plurality of character string vectors are selected, and calculates a distance and a cosine similarity between the two character string vectors selected for each combination; a step in which the computer sets a label for each of the combinations indicating whether the two selected character string vectors are similar or not, and creates training data for the number of combinations; a step of calculating, by the computer, the distance threshold and the cosine similarity threshold for determining the similarity of the character strings based on the training data; A threshold setting method having the following.

2. A step in which the computer calculates the standard deviation of the distances and the standard deviation of the cosine similarities for a plurality of the combinations; and In the step of calculating the thresholds, the computer calculates a threshold for the standard deviation of the distance and a threshold for the standard deviation of the cosine similarity in addition to the threshold for the distance and the threshold for the cosine similarity. The threshold setting method according to claim 1 .

3. A step in which the computer calculates a character string vector of a character string to be evaluated; a step in which the computer determines whether at least one of the character strings to be evaluated and the reference character strings is similar to one another based on character string vectors of the one or more reference character strings calculated using the threshold setting method of claim 1 or 2, the threshold, and a character string vector of the character string to be evaluated; A similar string determination method having the following.

4. A step in which the computer calculates a character string vector of a character string to be evaluated; a step in which the computer classifies the evaluation target string and the reference string into groups of similar characters based on character string vectors of the one or more reference strings calculated using the threshold setting method of claim 1 or 2, the threshold, and the character string vector of the evaluation target string; A similar string determination method having the following.

5. A step in which the computer receives an evaluation of the validity of the result of the determining step; a step of the computer setting the label for the character string to be evaluated and the combination evaluated for similarity to the character string to be evaluated based on the evaluation of validity, and creating additional training data; and In the step of calculating the threshold value, the computer calculates the threshold value based on the teacher data and the additional teacher data. The similar character string determination method according to claim 3 .

6. the one or more reference strings are strings representing cutsets extracted from an evaluated Probabilistic Risk Assessment (PRA) model; The character string to be evaluated is a character string representing a cutset extracted from a PRA model to be evaluated. The similar character string determination method according to claim 3 .

7. means for calculating a character string vector for each of a plurality of character strings; a means for creating a plurality of combinations of selecting two of the plurality of character string vectors, and calculating a distance and a cosine similarity between the two character string vectors selected for each of the combinations; a means for setting a label indicating whether the two selected character string vectors are similar or not for each of the combinations, and creating training data for the number of combinations; means for calculating the distance threshold and the cosine similarity threshold for determining the similarity of the character strings based on the training data; A threshold setting device having:

8. On the computer, calculating a string vector for each of a plurality of strings; creating a plurality of combinations of selecting two of the plurality of character string vectors, and calculating the distance and cosine similarity between the two selected character string vectors for each combination; a step of setting a label indicating whether the two selected character string vectors are similar or not for each of the combinations, and creating training data for the number of combinations; calculating the distance threshold and the cosine similarity threshold for determining the similarity of the character strings based on the training data; A program that executes the following.

Images