Failure analysis system, failure analysis method, and failure analysis model generation method

Abstract

To extract a past similar case similar to a newly detected failure with higher accuracy.SOLUTION: A failure analysis system 1 learns a past case which is a document which describes past failure or knowledge which is the document in which items related to a monitoring object are integrated into rules, performs similarity degree determination with an input case in which the failure generated in the monitoring object is described, and stores a plurality of learning models which output similar documents which include a similar case similar to the input case or the related knowledge related to the input case. The failure analysis system 1 acquires the similar document outputted with a score showing a similarity degree with respect to the input case from each of the plurality of learning models. The failure analysis system 1 calculates a total score obtained by totalizing scores acquired from each of the plurality of learning models for each similar document, and outputs the similar document with the total score. In the plurality of learning models, determination results of the similarity degree determination are different, according to a viewpoint or an index of the similarity degree determination.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to a fault analysis system, a fault analysis method, and a fault analysis model generation method. [Background technology]

[0002] Systems have been proposed that support efficient fault response by using machine learning models to extract the cause of the fault and countermeasures from similar past cases when a fault occurs in a monitored object.

[0003] For example, Patent Document 1 discloses a failure response support system that presents and recommends a response method for a newly detected failure. That is, the failure response support system includes "a failure email server 30 that sends a failure email 31, a failure response support server 40 that acquires response details based on the failure message and sends a response detail email 44 to a person in charge 2, and a machine learning server 50 that acquires response details for the failure message using an analytical model 53, the failure email server 30 sends the failure email 31 to the person in charge 2 and the failure response support server 40, the failure response support server 40 requests the machine learning server 50 to acquire response details for the failure message, and the machine learning server 50 acquires response details using a first analytical model that classifies the failure message into whether or not it is necessary for the person in charge 2 to contact a customer 3 and a second analytical model that extracts candidate response details from similar past failure responses" (see the abstract of Patent Document 1). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2018-136656 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the above-mentioned conventional technology has a problem in that it has low accuracy in extracting past similar cases that are similar to a newly detected fault. When the accuracy of similar case extraction is low, not only are similar cases that match the user's experience not extracted, but many similar cases that clearly contradict the user's experience are extracted, or similar cases are extracted that are biased towards specific products or fields.

[0006] The present invention has been made in view of the above-mentioned problems, and aims to extract past similar cases that are similar to a newly detected fault with higher accuracy. [Means for solving the problem]

[0007] In order to achieve the above object, the present invention provides a fault analysis system that supports analysis of faults that occur in a monitored object, the fault analysis system having a processor, a memory, and a storage unit, the storage unit stores a plurality of learning models that learn past cases, which are documents describing the past faults, or knowledge, which are documents that define rules for matters related to the monitored object, and that perform similarity determination with an input case, which is a document describing the fault that occurred in the monitored object, to output similar documents that include similar cases, which are past cases similar to the input case, or related knowledge, which is knowledge related to the input case, the processor inputs the input case to each of the plurality of learning models, acquires the similar documents output from each of the plurality of learning models together with scores that represent similarity to the input case, calculates a total score by adding up the scores acquired from each of the plurality of learning models for each similar document, and outputs the similar documents together with the total score, and the plurality of learning models are characterized in that the results of the similarity determination differ depending on the viewpoint or index of the similarity determination. [Effects of the Invention]

[0008] According to the present invention, it is possible to extract past similar cases that are similar to a newly detected fault with a high degree of accuracy. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram (part 1) showing an overview of a fault analysis system according to a first embodiment. [Figure 2] FIG. 2 is a diagram (part 2) showing an overview of the failure analysis system according to the first embodiment. [Figure 3] FIG. 3 is a diagram (part 3) showing an overview of the fault analysis system according to the first embodiment. [Figure 4] FIG. 1 is a configuration diagram of a fault analysis system according to a first embodiment. [Figure 5] FIG. 2 is a diagram showing the configuration of a document format definition according to the first embodiment. [Figure 6] FIG. 2 is a diagram showing items and uses of past cases and knowledge according to the first embodiment. [Figure 7] 4 is a flowchart showing a data collection process according to the first embodiment. [Figure 8] 4 is a flowchart showing a learning data preparation process according to the first embodiment. [Figure 9] 10 is a flowchart showing article division processing according to the first embodiment. [Figure 10] FIG. 2 is a diagram showing a division pattern according to the first embodiment. [Figure 11] 1 is a flowchart showing a learning model generation process according to the first embodiment. [Figure 12] FIG. 2 is a diagram showing a corresponding division pattern corresponding to a learning model according to the first embodiment. [Figure 13] 10 is a flowchart showing a similar document acquisition process according to the first embodiment. [Figure 14] FIG. 1 is a diagram (part 1) showing an output screen according to the first embodiment. [Figure 15] FIG. 10 is a diagram (part 2) showing an output screen according to the first embodiment. [Figure 16] FIG. 10 is a diagram showing an overview of a fault analysis system according to a second embodiment. [Figure 17] FIG. 10 is a configuration diagram of a fault analysis system according to a second embodiment. [Figure 18] 10 is a flowchart showing advance data preparation processing for product category analogy according to the second embodiment. [Figure 19] 10 is a flowchart showing a process of creating learning data for product category analogy according to the second embodiment. [Figure 20] 10 is a flowchart showing a product category extraction process according to the second embodiment. [Figure 21] FIG. 10 is a diagram showing a product category extraction result screen according to the second embodiment. [Figure 22] FIG. 10 is a diagram showing a product category extraction result screen according to a comparative example. [Figure 23] A diagram showing the hardware configuration of a computer. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, the description and drawings have been simplified or omitted as appropriate for clarity. Furthermore, not all of the combinations of features in the following description are necessarily required for the solution of the invention. Furthermore, in the following description, descriptions of previously described embodiments will be omitted as appropriate later. Furthermore, embodiments that combine some or all of the embodiments in the following description are also included in the present invention.

[0011] [Embodiment 1] The first embodiment focuses on the similarity or relevance between a newly occurring fault case (input case) and past cases that have occurred in the past and knowledge used in system operation. Then, in order to prevent future faults from occurring, a learning model trained by machine learning or the like is used to extract similar cases that are similar to the input case from past cases, and to extract related knowledge related to the input case from knowledge.

[0012] Here, "cases" refer to events such as failures, events, and incidents that occur in the monitored system, as well as documents that describe these events. "Knowledge" refers to documents such as user guides, manuals, checklists, know-how, and development information that are rules or manuals on matters related to the system.

[0013] An "input case" is a case from which similar cases or related knowledge are extracted. A "past case" is a case that occurred in the past. Note that the case from which similar cases or related knowledge are extracted is not limited to a case that occurred most recently, but can be a case from any timing.

[0014] (Outline of fault analysis system 1 according to embodiment 1) 1 to 3 are diagrams showing an overview of a fault analysis system 1 according to embodiment 1. The fault analysis system 1 is a system that supports the analysis of a fault that occurs in a monitoring target (such as a target system 2 described below).

[0015] (Outline of Fault Analysis System 1 (Part 1)) First, an overview (part 1) of the fault analysis system 1 will be described with reference to Fig. 1. The fault analysis system 1 uses accumulated past cases and knowledge N11 as learning data, and generates multiple learning models that produce different similarity judgment results depending on the viewpoint or index of similarity judgment based on multiple patterns. In this embodiment, three learning models, a first learning model M11, a second learning model M12, and a third learning model M13, are generated, but this is not limited to this as long as there is more than one.

[0016] The first learning model M11 is a learning model that determines similarity using important keywords, such as TF-IDF. The second learning model M12 is a learning model that determines the semantic similarity of documents, including specific terms, such as Doc2Vec. "Specific terms" are unique terms related to the target system. The third learning model M13 is a learning model that determines the semantic similarity of documents, excluding specific terms, such as Sentence BERT.

[0017] The document formats of the input examples IN11, past examples, and knowledge N11 are diverse, making it difficult to extract similar examples and related knowledge with high accuracy using only a specific extraction method. Therefore, this embodiment focuses on the fact that there are differences in the tendencies of each learning model regarding the perspectives and indices used in similarity determination as described above. For example, the order of magnitude of the influence of "important keywords" on similarity determination is TF-IDF > Doc2Vec > Sentence BERT. Furthermore, the order of magnitude of the influence of "semantic information" of documents excluding "specific terms" on similarity determination is Sentence BERT > Doc2Vec > TF-IDF.

[0018] Similar cases and related knowledge OUT11 are extracted from past cases and knowledge N11 by the first learning model M11 because they have similar important keywords to the input case IN11. Similarly, similar cases and related knowledge OUT12 are extracted from past cases and knowledge N11 by the second learning model M12 because they have similar contexts to the input case IN11 from an individual perspective (including specific terms). Similarly, similar cases and related knowledge OUT13 are extracted from past cases and knowledge N11 by the third learning model M13 because they are similar to the input case IN11 from a general perspective (excluding specific terms).

[0019] In this embodiment, multiple similar cases extracted by learning models with different characteristics are combined, or multiple related knowledge is combined. Points of view for similarity determination include, for example, "proper nouns (information specific to products or services)" and "semantic information (context and general topics)" in a document. Indicators include, for example, "features for each paragraph, sentence, clause, phrase, and character" in a document.

[0020] That is, in this embodiment, the final extraction result is the intersection or union of similar case / related knowledge OUT11, similar case / related knowledge OUT12, and similar case / related knowledge OUT13. This makes it possible to utilize the advantages of each method, which has different perspectives and indicators, to obtain more comprehensive, multifaceted, and highly accurate extraction results of similar cases and related knowledge.

[0021] (Outline of Fault Analysis System 1 (Part 2)) Next, an overview (part 2) of the fault analysis system 1 will be explained with reference to Figure 2. The fault analysis system 1 divides the accumulated past cases N21 into "events" and "causes," and generates a learning model M21 that has learned the events and a learning model M22 that has learned the causes. The input cases 21 are also divided into events and causes, and similar cases OUT21 with similar events and similar cases OUT22 with similar causes are extracted.

[0022] "Cases with similar events" can be used to suggest that "in order to prevent events (events that can lead to serious accidents), it is necessary to not only take measures for the events in the input case, but also to take measures that broaden the perspective to include other cases to prevent recurrence." "Similar cases with similar causes" can be used to suggest that "if measures are not taken for the causes of the input case, it is necessary to identify the causes that could lead to the occurrence of similar or similar events, not just this event, but without being caught up in superficial events."

[0023] By combining "similar cases with similar events" and "similar cases with similar causes," it is possible to extract "cases with similar events but different causes," "cases with similar events but different causes," "cases to review from a different perspective," "cases to avoid underestimation," etc.

[0024] "Event" and "Cause" are examples of multiple pieces of subdivided information obtained by subdividing the granularity of the information of input cases and past cases. In Figure 2, the granularity of the information (past case N21, input case IN21) is subdivided into two, "event" and "cause," but the granularity of the information may be subdivided into three or more.

[0025] (Outline of Fault Analysis System 1 (Part 3)) Next, an overview (part 3) of the fault analysis system 1 will be described with reference to FIG. 3. The fault analysis system 1 divides knowledge N31 used for learning into the processes to which it is applied (for example, requirements definition, design, testing, and operation) and generates a learning model for each process. The learning models for each process include a learning model M31 that has learned requirements definition, a learning model M32 that has learned design, a learning model M33 that has learned testing, and a learning model M34 that has learned operation. Then, for each process, the learning model for each process is applied to the input example IN31 to extract related knowledge for each process. The related knowledge for each process includes related knowledge OUT31 from the perspective of requirements definition, related knowledge OUT32 from the perspective of design, related knowledge OUT33 from the perspective of testing, and related knowledge OUT34 from the perspective of operation.

[0026] If a learning model that learns knowledge without dividing the processes is used, knowledge of processes other than the process in which the cause of the input example actually exists may not be extracted and may be buried. However, by using a learning model that learns knowledge for each process, knowledge for each process is extracted rather than knowledge for only a specific process, so knowledge for improving the quality of all processes can be acquired.

[0027] (Configuration of fault analysis system 1 according to embodiment 1) 4 is a configuration diagram of a fault analysis system 1 according to embodiment 1. One or more target systems 2 that are targets of monitoring for fault occurrence are connected to the fault analysis system 1 via a network (not shown). The fault analysis system 1 is also connected to an output device 3 such as a display that outputs information to a user.

[0028] The fault analysis system 1 includes an information collection unit 10, a learning information extraction unit 20, a learning model generation unit 30, a similar document acquisition unit 40, a result output unit 50, and a memory unit 60. The memory unit 60 is a storage device such as a memory or storage, and stores a document format definition 61, past cases 62, knowledge 63, and output data 64.

[0029] The information collection unit 10 has a past case collection unit 11 and a knowledge collection unit 12. The past case collection unit 11 collects past cases for each target system 2 from each target system 2. The past case collection unit 11 stores the collected past cases 62 in the memory unit 60. The knowledge collection unit 12 collects knowledge for each target system 2 from each target system 2. The knowledge collection unit 12 stores the collected knowledge 63 in the memory unit 60.

[0030] The learning information extraction unit 20 has an article extraction unit 21, an information difference extraction unit 22, a learning information extraction unit 23, a morphological analysis processing unit 24, and an information division unit 25. The article extraction unit 21 divides the read past cases and knowledge into individual articles in accordance with a document format definition 61. The information difference extraction unit 22 updates the stored article according to the difference between the already stored article and the currently acquired article, or stores the currently acquired article as a new article.

[0031] The learning information extraction unit 23 extracts parts that will become learning data necessary for learning the learning model from articles updated or newly accumulated by the information difference extraction unit 22 in accordance with the document format definition 61, and assigns tags to classify the articles. The morphological analysis processing unit 24 performs morphological analysis on the learning data extracted by the learning information extraction unit 23. The information division unit 25 groups the articles for each combination pattern of all tags and saves them.

[0032] The learning model generation unit 30 generates a plurality of learning models (fault analysis models) that have different tendencies in determining similarity with past cases 62 and knowledge 63 according to tag combination patterns. In this embodiment, the learning model generation unit 30 generates, as an example, learning models for word occurrence count 31, Doc2Vec 32, TF-IDF 33, and Sentence BERT 32, and stores them in the learning model generation unit 30.

[0033] The similar document acquisition unit 40 extracts documents similar to the input case IN11 from the past cases 62 and / or knowledge 63 using multiple learning models (Doc2Vec32, TF-IDF33, Sentence BERT34) that have different tendencies in determining similarity with past cases 62 and knowledge 63 depending on the tag combination pattern.

[0034] The result output unit 50 processes documents similar to the input example IN11 extracted from the past examples 62 and / or knowledge 63 by the similar document acquisition unit 40 to generate output data 64, and outputs it to the output device 3.

[0035] The fault analysis model generation system that generates the fault analysis model includes at least the learning information extraction unit 20 and the learning model generation unit 30 of the fault analysis system 1.

[0036] (Configuration of document format definition 61 according to the first embodiment) 5 is a diagram showing the configuration of a document format definition 61 according to the first embodiment. The document format definition 61 provides information for extracting past cases and knowledge documents from an original document for each file type. By processing the original document according to the definition described in the document format definition 61, the target document can be extracted.

[0037] (Use of the items of past cases 62 and knowledge 63 according to the first embodiment) 6 is a diagram showing the items and uses of the past cases 62 and knowledge 63 according to the first embodiment. As shown in FIG. 6, each item of the past cases 62 and knowledge 63 is classified into report items and analysis items according to use. Report items are items for creating a report including the corresponding item. The report is, for example, an output screen D1 (FIG. 14), an output screen D2 (FIG. 15), etc.

[0038] The analysis items are items for performing analysis using the corresponding items. The analysis items are data items to be processed in, for example, the data collection process (FIG. 7), the learning data preparation process (FIG. 8), the article segmentation process (FIG. 9), the learning model generation process (FIG. 11), and the similar document acquisition process (FIG. 13) described below. The analysis items are also data items to be processed in the advance data preparation process (FIG. 18) for product category analogy in the first embodiment described below, and the learning data creation process for product category analogy (FIG. 19).

[0039] A table as shown in Figure 6 is provided for each document type of past cases 62 and knowledge 63, and whether each item corresponds to a report item or analysis item is indicated by "X" (applicable) or "-" (not applicable).

[0040] (Data collection process according to the first embodiment) 7 is a flowchart showing the data collection process according to embodiment 1. In the data collection process, past cases and knowledge are collected from the target system 2 and accumulated.

[0041] First, in step S11, the information collection unit 10 acquires past cases and knowledge. The past case collection unit 11 and knowledge collection unit 12 of the information collection unit 10 hold the location (the URL (Uniform Resource Locator) of the target system 2) where the past cases and knowledge used for learning are published and the access method. The past case collection unit 11 periodically acquires past cases and stores them in the memory unit 60. The knowledge collection unit 12 also periodically acquires knowledge and stores it in the memory unit 60.

[0042] Next, in step S12, the article extraction unit 21 of the learning information extraction unit 20 reads a document format definition 61 corresponding to the file format of the past cases and knowledge acquired in step S11. Next, in step S13, the article extraction unit 21 reads the past cases and knowledge acquired in step S11 in accordance with the document format definition 61 read in step S12. Next, in step S14, the article extraction unit 21 uses the document format definition 61 to acquire articles from the file read in step S14. An "article" is a past case or knowledge divided into individual cases.

[0043] Next, in step S15, the article extraction unit 21 assigns to the article acquired in step S14 an article ID for uniquely identifying the article, a document format ID for uniquely identifying the document format definition, and the date the original information was registered in the fault analysis system 1.

[0044] In steps S12 to S15, past cases and knowledge items that serve as learning data are divided into article units. In steps S12 to S14, information for learning is extracted from the files collected by the information collecting unit 10, excluding parts that are not necessary for learning, thereby improving the learning accuracy.

[0045] Next, the information difference extraction unit 22 of the learning information extraction unit 20 repeats steps S16 to S23 for the number of articles acquired in step S14. First, in step S16, the information difference extraction unit 22 selects one unselected article from the articles acquired in step S14.

[0046] Next, in step S17, the information difference extraction unit 22 reads the list of stored articles. Next, in step S18, the information difference extraction unit 22 determines the degree of match between the title and text of the stored article and the currently acquired article selected in step S15. If there is a stored article with a degree of match exceeding a predetermined degree (step S19 YES), the information difference extraction unit 22 proceeds to step S20, and if there is not (step S19 NO), the information difference extraction unit 22 proceeds to step S23.

[0047] In step S20, the information difference extraction unit 22 compares the date and time of the stored article with the date and time of the article acquired this time. Next, in step S21, the information difference extraction unit 22 discards the article with the older date and time among the articles compared in step S20, and saves the article with the newer date and time. Meanwhile, in step S22, the information difference extraction unit 22 accumulates the article acquired this time. Next, in step S23, if there is an article that has not been selected from the articles acquired in step S14, the information difference extraction unit 22 returns the process to step S16. On the other hand, if there is no article that has not been selected from the articles acquired in step S14, the information difference extraction unit 22 ends this data collection process.

[0048] In step S16, which is moved from step S23, the information difference extraction unit 22 selects one unselected article from the articles acquired in step S14.

[0049] In steps S16 to S22, duplicate articles are eliminated. On the other hand, since past cases and know-how are regularly added and updated, the articles are checked for updates, and if there are at least some updates, they are replaced with new information. The presence or absence of updates is determined based on the percentage of overlap in the article text and whether the article titles match.

[0050] (Article division process according to the first embodiment) 8 is a flowchart showing the learning data preparation process according to embodiment 1. In the learning data preparation process, the learning data is processed.

[0051] First, in step S21, the learning information extraction unit 23 of the learning information extraction unit 20 acquires the document format definition 61 based on the document format ID assigned to the article. Next, in step S22, the learning information extraction unit 23 extracts the parts necessary for learning from the article in accordance with the document format definition 61 acquired in step S21.

[0052] The articles acquired by the information collection unit 10 contain information that is not necessary for learning, such as the name of the department responsible for the relevant past case. Therefore, in steps S21 and S22, only the parts necessary for learning are acquired. The document format definition 61 describes what information the collected documents are made up of and which parts of the information are necessary for learning. Therefore, the learning information extraction unit 23 can retrieve the document format definition 61 using the document format ID assigned to the article and acquire the information necessary for learning according to the information described in the document format definition 61.

[0053] Next, in step S23, the learning information extraction unit 23 generates learning data by assigning classification tags to the articles from which the necessary parts were extracted in step S22. For example, articles include past cases and knowledge. Past cases are tagged with "case" and knowledge with "knowledge." Past cases include product-related cases and service-related cases, and learning is performed separately for each. Therefore, to distinguish between articles, a "product" or "service" tag is assigned. Each piece of knowledge may have a process for which utilization is recommended, and the recommended process is assigned as a tag. Information regarding which tag to assign to which article in which file is held by the learning information extraction unit 23.

[0054] Next, in step S24, the morphological analysis processing unit 24 of the learning information extraction unit 20 Unicode-normalizes the training data generated in step S23. Next, in step S25, the morphological analysis processing unit 24 morphologically analyzes the training data Unicode-normalized in step S24. Next, in step S26, the morphological analysis processing unit 24 converts abbreviations included in the result of the morphological analysis of step S25 into formal names. Next, in step S27, the morphological analysis processing unit 24 generates compound words from the morphemes obtained by converting the abbreviations into formal names in step S26.

[0055] Next, in step S28, the morphological analysis processing unit 24 removes unnecessary words from the compound word generated in step S27. Next, in step S29, the morphological analysis processing unit 24 assigns the article ID and document format ID of the article before processing to the compound word from which unnecessary words have been removed in step S28.

[0056] In steps S24 to S29, morphological analysis and removal of unnecessary words are performed on the article text for each group of articles divided by tags. Morphological analysis and removal of unnecessary words are processes generally required for Doc2Vec and TF-IDF, which are used to generate learning models. In addition, preprocessing required to improve accuracy includes unifying full-width / half-width characters, converting device-dependent characters (Unicode normalization), converting abbreviations, and processing compound words that would be split using simple morphological analysis.

[0057] (Article division process according to the first embodiment) 9 is a flowchart showing the article division process according to embodiment 1. In the article division process, articles are grouped according to the combination patterns of tags assigned to them.

[0058] First, in step S31, the information division unit 25 of the learning information extraction unit 20 acquires a combination of division pattern tags (a combination of tags 1 to 3) from the division pattern T1 (FIG. 10) of the learning data. As shown in FIG. 10, the division pattern T1 is a list of combination patterns of, for example, three tags. Articles are grouped according to the tag combination patterns. The division pattern T1 is held by the information division unit 25.

[0059] To divide cases into "events" and "causes," tag 3 "events" and "causes" are used for the pattern where tag 2 of division pattern T1 is "case." To divide knowledge into "requirements definition," "design," "testing," and "operations," tag 3 "requirements definition," "design," "testing," and "operations" are used for the pattern where tag 2 of division pattern T1 is "knowledge."

[0060] Next, in step S32, the information division unit 25 acquires articles to which all of the tags of the combination acquired in step S31 have been assigned. Next, in step S33, the information division unit 25 saves the acquired group of articles in a predetermined storage area (such as the storage unit 60). Next, in step S34, the information division unit 25 acquires articles having all of the tags specified in step S31 from the compound words from which non-words have been removed and article IDs, etc. have been assigned in step S29. Next, in step S35, the information division unit 25 assigns the article IDs and document format IDs that were assigned to the articles before processing to the articles acquired in step S34, and then saves them in a predetermined storage area (such as the storage unit 60) in units of tag groups.

[0061] In steps S31 to S35, articles are divided into groups based on the combination of tags assigned to the articles. By learning articles for each tag group, it becomes possible to extract similar articles and related knowledge from within the user's expected range. For example, if you want to find similar articles from past product cases, you can find the desired article by searching for similar articles from the learning model of articles tagged with "product" and "case."

[0062] In the learning model generation process (Fig. 11) described later, there are two types of learning processes: one that uses morphologically analyzed data as input, and one that uses raw data that has not been morphologically analyzed as input. Therefore, for each division pattern based on a tag, both a group with morphologically analyzed data and a group with raw data are stored.

[0063] (Learning model generation process according to the first embodiment) 11 is a flowchart showing a learning model generation process according to embodiment 1. In the learning model generation process, articles are grouped according to the combination patterns of tags assigned thereto, and the articles are learned in units of groups.

[0064] First, in step S41, the learning model generation unit 30 reads the learning data that has been subjected to the article segmentation process (FIG. 9). Next, in step S42, the learning model generation unit 30 calculates the trends in the number of appearances and appearance rate of keywords in the learning data read in step S41.

[0065] In step S42, the number of times a keyword appears for each unit period (year or term) is calculated based on the date and time when the case described in the past case was recorded from the learning data read in step S41. The trend in the rate of increase or decrease in the number of times a keyword appears for each unit period is also calculated. As shown in FIG. 12, the data used to calculate the trend in the number of times a keyword appears and the rate of increase or decrease is the data of pattern A of the division patterns in FIG. 10. FIG. 12 is a diagram showing a corresponding division pattern T2 corresponding to the learning model according to the first embodiment. The corresponding division pattern T2 manages the type of model and learning method to be applied depending on the type of document.

[0066] Next, in step S43, the learning model generation unit 30 generates a learning model. In step S43, learning is performed for each division pattern (FIG. 10) created based on the tags assigned to the article data.

[0067] Furthermore, the learning model created for each division pattern differs for each pattern, as shown in Figure 12. Note that the "patterns" in Figure 12 correspond to the "patterns" in Figure 10. Figure 12 shows that a model of "trends in keyword appearance counts and appearance rates" is generated using the learning data of "pattern A." It also shows that the learning data of "pattern B" generates the following models: "Doc2Vec," "TF-IDF," "Sentence BERT," and "Labeled LDA."

[0068] Next, in step S44, the learning model generation unit 30 stores the generated learning model in a predetermined storage area (such as the storage unit 60).

[0069] (Similar Document Acquisition Process According to the First Embodiment) 13 is a flowchart showing the similar document acquisition process according to embodiment 1. In the similar document acquisition process, similar cases similar to the input case of the target system 2 or related knowledge related to the input case are extracted as similar documents.

[0070] First, in step S51, the similar document acquisition unit 40 of the fault analysis system 1 accepts input of an input example to be compared input by a user. Next, in step S52, the similar document acquisition unit 40 performs Unicode normalization, morphological analysis, formal name conversion, compound word generation, and non-word removal on the input example document accepted in step S51. The Unicode normalization, morphological analysis, formal name conversion, compound word generation, and non-word removal are the same as steps S24 to S28 of the training data preparation process (FIG. 8), respectively.

[0071] Next, in step S53, the similar document acquisition unit 40 assigns a tag of "case" or "knowledge" to the document of the input case after various processes have been performed in step S53. The tag of "case" is assigned when it is desired to extract similar cases to the input case, and the tag of "knowledge" is assigned when it is desired to extract related knowledge of the input case.

[0072] Next, the similar document acquisition unit 40 repeats steps S54 to S59 for the learning models held by the learning model generation unit 30. In this embodiment, the learning models held by the learning model generation unit 30 are three, Doc2Vec32, TF-IDF33, and Sentence BERT34, and therefore steps S54 to S59 are repeated for each of these three learning models.

[0073] In step S54, the similar document acquisition unit 40 selects one learning model for which steps S54 to S59 have not yet been executed.

[0074] Next, in step S55, the similar document acquisition unit 40 uses the learning model selected in step S54 to acquire the top (N+α) similar cases or related knowledge with the highest scores from the learning data that fall into the same division pattern (FIG. 10) as the input case, where N is an integer of 2 or greater, and α is an integer of 1 or greater.

[0075] Next, in step S56, the similar document acquisition unit 40 calculates the similarity between the top (N+α) similar cases or related knowledge and the top (N+α-1) similar cases or related knowledge acquired in step S55. Next, in step S57, the similar document acquisition unit 40 re-evaluates the score S of the top (N+α) similar case or related knowledge by using the score of the similar case or related knowledge with the highest similarity among the top (N+α-1) similarities calculated in step S56. The re-evaluation may involve, for example, reducing the score S by a predetermined percentage.

[0076] Next, in step S58, the similar document acquisition unit 40 ranks the top N similar cases or related knowledge by the score S after reevaluation in step S57, and returns the rankings together with the document ID and the score S after reevaluation.

[0077] Next, in step S59, if there is an unselected learning model, the similar document acquisition unit 40 returns the process to step S54. On the other hand, if there is no unselected learning model, the similar document acquisition unit 40 moves the process to step S60.

[0078] In step S53, which is moved from step S58, the similar document acquisition unit 40 selects one learning model for which steps S53 to S58 have not yet been executed.

[0079] In step S60, the similar document acquisition unit 40 calculates a total score by adding up the scores S after re-evaluation in step S58 for Doc2Vec32, TF-IDF33, and Sentence BERT34, and returns the document IDs of the similar cases or related knowledge and the total score. The result output unit 50 transmits the document IDs and scores of the similar cases or related knowledge returned by the similar document acquisition unit 40 to the output device 3. The output device 3 configures and displays an output screen D1 ( FIG. 14 ) or an output screen D2 ( FIG. 15 ) based on the document IDs and scores of the similar cases or related knowledge received from the result output unit 50.

[0080] In steps S51 to S60, a learning model is used to obtain a document vector for the input text, and the similarity with accumulated past cases and knowledge is calculated. Data with high scores is then presented to the user as similar cases or related knowledge. The learning models used include Doc2Vec, TF-IDF, Sentence BERT, etc., and the results of learning models with different characteristics are aggregated to calculate a score. By using multiple learning models with different characteristics in this way, weaknesses can be complemented by other learning models, and comprehensive and highly accurate results can be expected to be returned.

[0081] In addition, when calculating similarity with accumulated past cases and knowledge, candidate past cases and knowledge items are reordered using MMR (Maximal Marginal Relevance). MMR evaluates how diverse a document is compared to documents already selected. In other words, documents that are similar to documents already selected are adjusted to have a lower score. This allows results to be returned that are balanced between similarity and diversity.

[0082] In steps S54 to S60, when cases are divided into "events" and "causes," similar cases are obtained for each "event" and "cause" using the learning models Doc2Vec32, TF-IDF33, and Sentence BERT34. The scores S obtained from the learning models for each "event" and "cause" of these multiple learning models are then summed for each "event" and "cause" and for each similar case to calculate and output a total score.

[0083] Furthermore, in steps S54 to S60, when knowledge is divided into multiple processes such as "requirements definition," "design," "test," and "operation," related knowledge is acquired for each process using the learning models Doc2Vec32, TF-IDF33, and Sentence BERT34. Then, the scores S of the related knowledge acquired from the learning models for each process of these multiple learning models are summed for each process and each related knowledge to calculate and output a total score.

[0084] (Output screens D1 and D2 according to the first embodiment) 14 and 15 are diagrams showing output screens D1 and D2 according to embodiment 1. The output screen D1 has display areas D11, D12, D13, D14, and D15. The output screen D2 has display areas D21 and D22.

[0085] The display area D11 displays documents related to input cases that occurred in the target system 2 and were input to the fault analysis system 1. These input cases are the targets for searching for similar cases or related knowledge. Specific examples of these input documents are input cases IN11 (FIG. 1), IN21 (FIG. 2), and IN31 (FIG. 3).

[0086] In the display area D12, a score indicating the relevance to the input example is displayed for each topic obtained by topic analysis of the document related to the input example displayed in the display area D11.

[0087] Display area D13 displays a score indicating the degree of relevance and a time series of the number of occurrences for each keyword included in similar cases extracted as being similar to the input case displayed in display area D11.

[0088] Display area D14 displays related knowledge for each viewpoint extracted as being related to the input example displayed in display area D11. Specific examples of related knowledge for each viewpoint include related knowledge OUT31, OUT32, OUT33, and OUT34 (FIG. 3).

[0089] Display area D15 displays the similarity scores of similar cases extracted as being similar to the input case displayed in display area D11 for each occurrence period. The display in display area D15 allows one to see the degree of similarity and the time period in the past when similar cases occurred.

[0090] Display area D21 displays similar cases (events) extracted as being similar to the input case displayed in display area D11. Specific examples of similar cases (events) include similar case OUT21 (FIG. 2).

[0091] Display area D22 displays similar cases (causes) extracted as having causes similar to those of the input case displayed in display area D11. Specific examples of similar cases (causes) include similar case OUT22 (FIG. 2).

[0092] (Effects of the First Embodiment) According to the first embodiment, a total score is calculated for each similar document by adding up the scores obtained from multiple learning models that produce different similarity assessment results depending on the viewpoint or index of similarity assessment, and the total score is output together with the similar document. This makes it possible to take advantage of the respective advantages of multiple learning models (the influence of product- or service-specific information on similarity assessment, and the magnitude of the influence of document semantic information (context, general topics) on similarity assessment) and obtain similar documents that are more comprehensive, multifaceted, and accurate.

[0093] In this embodiment, a total score is calculated for each piece of subdivided information and for each similar case, and the similar cases are output together with the total score for each piece of subdivided information. Therefore, for example, when dividing cases into "events" and "causes," "cases with similar events but different causes" and "cases with different events but similar causes" can be extracted as "cases for reviewing causes from a different perspective" and "cases for determining the impact of customer-specific terminology," respectively. This suggests broadening your perspective without being caught up in superficial phenomena.

[0094] In this embodiment, the scores of the related knowledge obtained from each learning model for each process are summed up for each process and for each related knowledge to calculate a total score, and the related knowledge is output along with the total score for each process. This makes it possible to obtain hints (knowledge) for problem solving and quality improvement in all processes, without being biased towards a specific process.

[0095] In this embodiment, the system calculates a total score for each similar document by summing the scores reordered by MMR for each of the multiple learning models, and outputs the similar documents together with the total score. Therefore, documents similar to previously selected documents are adjusted to have lower scores, enabling the acquisition of similar documents with a good balance of similarity and diversity.

[0096] [Embodiment 2] In the second embodiment, terms appearing in similar cases extracted as being similar to an input case are matched with terms appearing in product manuals, etc. Then, using a learning model that has learned the product group to which the product belongs, the product group to which the similar case belongs is estimated based on the terms.

[0097] In the second embodiment, explanations that overlap with the first embodiment will be omitted, and the differences from the first embodiment will be mainly described.

[0098] (Outline of fault analysis system 1B according to embodiment 2) 16 is a diagram showing an overview of a fault analysis system 1B according to the second embodiment. In the fault analysis system 1B, a learning model M4 is generated that learns a product manual N41 and outputs product categories and scores for each product category using specified cases as input. By inputting an extracted case IN42 to the learning model M4, product category information OUT4 to which the case IN42 applies is obtained. The product category information OUT4 indicates the distribution of product category scores (the likelihood that the product category actually applies) for each occurrence period of similar cases similar to the input case.

[0099] In this embodiment, it is possible to determine whether the input example is a problem specific to the product or a problem common to other products using the product category information OUT4. In addition, the score of each product category according to the specified example suggests the possibility that the problem is common to other product categories.

[0100] (Configuration of fault analysis system 1B according to embodiment 2) 17 is a configuration diagram of a fault analysis system 1B according to the second embodiment. Fault analysis system 1B differs from fault analysis system 1 according to the first embodiment in that learning information extraction unit 20B includes morphological analysis processing unit 24B instead of morphological analysis processing unit 24, and further includes category analogy learning data creation processing unit 26. Fault analysis system 1B also differs from fault analysis system 1 in that learning model generation unit 30 further includes Labeled LDA (Latent Dirichlet Allocation) 35. Fault analysis system 1B also differs from fault analysis system 1 in that it further includes product category extraction unit 41.

[0101] In addition to the functions of the morphological analysis processing unit 24 in embodiment 1, the morphological analysis processing unit 24B performs morphological analysis of the text of the product manuals and creates a list of words for each product manual. The category analogy learning data creation processing unit 26 determines and labels the product categories to which the past cases are related based on the words that make up the past cases.

[0102] Labeled LDA35 is a type of classifier that determines the product category of an article as the one with the highest proportion of topics (words) that make up the article.

[0103] (Preliminary data preparation process according to the second embodiment) FIG. 18 is a flowchart showing a preliminary data preparation process for product category analogy according to the second embodiment.

[0104] First, in step S61, the morphological analysis processing unit 24B of the learning information extraction unit 20B reads a predetermined document related to a product, specifically, an index of a product manual. Next, in step S62, the morphological analysis processing unit 24B performs morphological analysis on the sentences of the index of the product manual read in step S61 and divides them into words.

[0105] Next, in step S63, the morphological analysis processing unit 24B generates pairs of "product category-word." In step S63, the "product category" has not yet been labeled, so the pairs of "product category-word" become a word list for each product manual. Next, in step S64, the morphological analysis processing unit 24B stores the pairs of "product category-word" generated in step S63 in a predetermined storage area (such as the storage unit 60).

[0106] The product category is estimated based on an analysis of which product significantly affected the events and causes of the case. For example, even if a case is reported as involving product X, if the mechanism of the event is significantly affected by product Y or has commonalities with product Y, it may be suggested that product Y needs to be reviewed and measures to prevent recurrence may be necessary.

[0107] In this embodiment, Labeled LDA, an extension of LDA, a type of topic model, is used to estimate product categories. LDA is based on the assumption that each sentence is composed of multiple topics, and is capable of determining the word groups that make up each topic and the probability of their occurrence. However, the meaning of each topic depends on the analyst's interpretation. Labeled LDA, on the other hand, is characterized by its high topic discrimination ability, as it uses label information assigned to documents during the generation process.

[0108] (Learning Data Creation Process for Product Category Analogy According to the Second Embodiment) 19 is a flowchart showing a learning data creation process for product category inference according to embodiment 2. In the learning data creation process for product category inference, a product category to which the relevant past cases and knowledge are related is determined and labeled based on the words constituting the past cases and knowledge.

[0109] First, in step S71, the category analogy learning data creation processor 26 reads article data that corresponds to a tag division pattern for creating a TF-IDF and has a TF-IDF score assigned, from the learning data used for learning in step S43 (FIG. 11). In the example shown in FIG. 12, patterns B, C, and D correspond to tag division patterns for creating a TF-IDF.

[0110] Next, in step S72, the category inference learning data creation processing unit 26 performs morphological analysis on the article read in step S71 to obtain pairs of [word-TF-IDF score]. Next, in step S73, the category inference learning data creation processing unit 26 reads the [product category-word] pairs generated in the advance data preparation process (FIG. 18) for product category inference.

[0111] Next, in step S74, the category analogy learning data creation processor 26 compares the [word-TF-IDF score] obtained in step S72 with the [product category-word] read in step S73. The TF-IDF score of the matching word is then used as the score for the product category. This process is then repeated for each word included in the relevant past case or knowledge, and the scores for each product category are totaled.

[0112] Next, in step S75, the category analogy learning data creation processing unit 26 determines the product category with the maximum product category score among the words contained in the relevant past case or knowledge as the product category of the relevant past case and knowledge.

[0113] Next, in step S76, the category analogy learning data creation processing unit 26 stores the product categories for each past case and knowledge determined in step S75 in a predetermined storage area (such as the storage unit 60).

[0114] In steps S71 to S76, for each past case and knowledge, the relevance (number of occurrences) of product categories in the article is counted using the words that make up the article and pairs of words and product categories. At this time, weighting is performed using the TF-IDF values ​​of the words that make up the article, and the product categories that are paired with key words in the article are adjusted so that the relevance is emphasized. As a result, the product category with the highest relevance is labeled as the product category to which the past case and knowledge containing that article belong.

[0115] The learning process of the Labeled LDA 35 can be performed by applying the learning model generation process (FIG. 11) according to the first embodiment to the learning data generated by the learning data creation process for product category analogy. That is, the Labeled LDA 35 is trained by taking as input past cases to which product categories are labeled based on the proportions of product categories (proportions of topics) calculated from the word distributions constituting the past cases and knowledge, and taking past cases or knowledge as input and outputting product categories (topics).

[0116] (Product Category Extraction Process According to the Second Embodiment) 20 is a flowchart showing a product category extraction process according to embodiment 2. In the product category extraction process, the result of Labeled LDA 35 (the topic (product category) with the highest proportion) is acquired for each case extracted as a similar case or related knowledge.

[0117] First, in step S81, the product category extraction unit 41 of the fault analysis system 1B acquires the product category together with the similarity using the Labeled LDA 35 for the similar cases or related knowledge extracted in the similar document acquisition process (FIG. 13) according to the first embodiment.

[0118] Next, in step S82, the product category extraction unit 41 sorts the similar cases or related knowledge of step S81 by the date of occurrence. Next, in step S83, the result output unit 50 creates a scatter plot (product category extraction result screen D3 in FIG. 21) with the date of occurrence and the similarity as the coordinate axes, and transmits the output data to the output device 3. The output device 3 outputs the product category extraction result screen D3 (FIG. 21) based on the received output data.

[0119] (Product category extraction result screen D3 according to the second embodiment) Fig. 21 is a diagram showing a product category extraction result screen D3 according to embodiment 2. Fig. 22 is a diagram showing a product category extraction result screen D4 according to a comparative example.

[0120] The data plotted on the product category extraction results screen D3 is displayed with a distinction for each product category, allowing the estimated product category to be identified. In Figure 21, the plots corresponding to Category 1 account for a certain percentage or more, indicating that the corresponding cases occur frequently in a specific product category (group). In this case, it can be said that it is suggested that the cause analysis and countermeasures should be carried out with a focus on the specific product category.

[0121] On the other hand, in Figure 22, the plots are distributed across each product category from Category 1 to Category 6, and there is no product category that accounts for more than a certain percentage. This suggests that the case in question is not a case that occurs frequently in a specific product category (group), but is based on factors common to multiple products. In this case, it can be said that it is suggested that the common factors be identified and that wide-ranging measures be implemented to prevent recurrence.

[0122] In this embodiment, the "product category" is estimated, but the estimation may be performed on any item that can classify the target system 2, not limited to the "product category".

[0123] (Effects of the second embodiment) According to the second embodiment, the most likely topic (product category) is obtained for the extracted similar cases or related knowledge using Labeled LDA. This makes it possible to recognize the product category of the similar cases or related knowledge, and determine whether the problem is specific to the product or whether similar risks are lurking in other products, thereby enabling efficient and effective determination of the scope of review.

[0124] (Hardware configuration of computer 1000) 23 is a diagram showing the hardware configuration of the computer 1000. The computer 1000 executes predetermined programs to realize various systems, servers, devices, terminals, etc. (including the fault analysis systems 1 and 1B, the target system 2, and the output device 3).

[0125] The computer 1000 comprises a processor 1001 including a CPU, a main memory device 1002, an auxiliary memory device 1003, a network interface 1004, an input device 1005, and an output device 1006, all of which are interconnected via an internal communication line 1007 such as a bus.

[0126] The processor 1001 controls the overall operation of the computer 1000. The main memory device 1002 is composed of, for example, a volatile semiconductor memory, and is used as a work memory for the processor 1001. The auxiliary memory device 1003 is composed of a large-capacity nonvolatile memory device such as a hard disk device, an SSD (Solid State Drive), or a flash memory, and is used to store various programs and data for a long period of time.

[0127] An executable program 1003a stored in the auxiliary storage device 1003 is loaded into the main storage device 1002 when the computer 1000 is started up or when needed, and is executed by the processor 1001. This realizes the various systems, servers, devices, terminals, etc. according to the above-described embodiments.

[0128] The executable program 1003a may be recorded on a non-transitory recording medium, read from the non-transitory recording medium by a media reading device, and loaded into the main memory device 1002. The executable program 1003a may also be obtained from one or more external computers via a network and loaded into the main memory device 1002. Furthermore, the executable program 1003a may be divided into multiple programs as appropriate.

[0129] The network interface 1004 is an interface device for connecting the computer 1000 to each network within the system or for communicating with other computers. The network interface 1004 is configured, for example, by a network interface card (NIC) for a wired local area network (LAN) or a wireless LAN.

[0130] The input device 1005 is composed of a keyboard, a pointing device such as a mouse, and the like, and is used by the user to input various instructions and information to the computer 1000. The output device 1006 is composed of, for example, a display device such as a liquid crystal display, and an audio output device such as a speaker, and is used to present necessary information to the user when necessary.

[0131] It should be noted that the present invention is not limited to the above-described embodiments, and includes various modifications. Furthermore, the above-described embodiments have been described in detail to clearly explain the present invention, and are not necessarily limited to those having all of the described configurations. Furthermore, not only are such configurations removed, but replacement or addition of configurations is also possible. Furthermore, forms in which some or all of the above-described embodiments are appropriately combined to be consistent are also included in the embodiments of the present invention.

[0132] Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. The present invention can also be implemented by software program code that implements the functions of the embodiments. In this case, a recording medium on which the program code is recorded is provided to a computer, and a processor within the computer reads the program code stored in the recording medium.

[0133] In this case, the program code itself read from the recording medium will realize the functions of the above-described embodiment, and the program code itself and the recording medium on which it is stored constitute the present invention. Examples of recording media for supplying such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, SSDs (Solid State Drives), optical disks, magneto-optical disks, CD-Rs, magnetic tapes, non-volatile memory cards, ROMs, etc.

[0134] Furthermore, the program code that realizes the functions described in this embodiment can be implemented in a wide range of program or script languages, such as assembler, C / C++, perl, Shell, PHP, and Java (registered trademark).

[0135] In the above-described embodiment, the control lines and information lines are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. All components may be interconnected. [Explanation of symbols]

[0136] 1,1B: fault analysis system, 10: information collection unit, 20,20B: learning information extraction unit, 30: learning model generation unit, 40: similar document acquisition unit, 41: product category extraction unit, 50: result output unit, 60: memory unit, 63: knowledge, 1000: computer, 1001: processor, 1002: main memory unit, 1006: output device.

Claims

1. A fault analysis system that supports analysis of a fault that occurs in a monitored object, the fault analysis system includes a processor, a memory, and a storage unit; The storage unit storing a plurality of learning models that learn past cases, which are documents describing the failures in the past, or knowledge, which is a document that defines rules for matters related to the monitoring target, and determine the similarity between the past cases and input cases, which are documents describing the failures that have occurred in the monitoring target, and output similar cases, which are past cases similar to the input cases, or similar documents that include related knowledge, which is knowledge related to the input cases; The processor: inputting the input examples into each of the plurality of learning models; acquiring the similar documents output from each of the plurality of learning models together with scores representing similarities to the input examples; calculating a total score by adding up the scores obtained from each of the plurality of learning models for each of the similar documents; outputting the similar documents together with the total score; The plurality of learning models The result of the similarity determination differs depending on the viewpoint or index of the similarity determination. A fault analysis system comprising:

2. 2. The fault analysis system of claim 1, Each of the plurality of learning models The past cases are divided into a plurality of pieces of segmented information obtained by segmenting the granularity of information, and a learning model for each piece of segmented information is included, which has learned documents related to the plurality of pieces of segmented information; The processor: Dividing the input example into the plurality of pieces of segmented information; inputting the input examples for each piece of segmentation information into a learning model for each piece of segmentation information of each of the plurality of learning models; acquiring the similar cases output together with the scores for the input examples for each piece of subdivision information from the learning models for each piece of subdivision information of each of the plurality of learning models; calculating a total score by summing the scores obtained from the learning models for each piece of segmentation information of the plurality of learning models for each piece of segmentation information and for each similar case; For each of the segmentation information, the similar cases are output together with the total score. A fault analysis system comprising:

3. 3. The fault analysis system according to claim 2, The plurality of pieces of segmented information include at least one of the events and causes of the past cases. A fault analysis system comprising:

4. 2. The fault analysis system of claim 1, Each of the plurality of learning models The knowledge is divided into a plurality of steps, and a learning model for each step is included, which has learned the knowledge for each step; The processor: inputting the input examples into each of the learning models for each of the processes; acquiring the related knowledge output together with the score for the input example from the learning model for each process; calculating a total score by summing the scores of the related knowledge acquired from each of the learning models for each of the processes for each of the processes and for each of the related knowledge; For each step, output the related knowledge together with the total score. A fault analysis system comprising:

5. 2. The fault analysis system of claim 1, The processor: reordering the scores of the similar documents obtained for the input example from each of the plurality of learning models using Maximum Marginal Relevance (MMR) for each of the plurality of learning models; calculating a total score by summing the reordered scores for each of the plurality of learning models for each of the similar documents; Outputting the similar documents together with the total score A fault analysis system comprising:

6. 2. The fault analysis system of claim 1, The processor: constructing a classifier that learns terms that appear in predetermined documents related to products for each product group and predicts the product group based on the terms; The similar case is input to the classifier, and the product group to which the similar case belongs is estimated based on terms appearing in the similar case. A fault analysis system comprising:

7. 2. The fault analysis system of claim 1, The plurality of learning models includes at least one of Doc2Vec, TF-IDF, and Sentence BERT. A fault analysis system comprising:

8. A fault analysis method executed by a fault analysis system that supports analysis of a fault that occurs in a monitored object, comprising: the fault analysis system includes a processor, a memory, and a storage unit; The storage unit storing a plurality of learning models that learn past cases, which are documents describing the failures in the past, or knowledge, which is a document that defines rules for matters related to the monitoring target, and determine the similarity between the past cases and input cases, which are documents describing the failures that have occurred in the monitoring target, and output similar cases, which are past cases similar to the input cases, or similar documents that include related knowledge, which is knowledge related to the input cases; the processor: inputting the input examples into each of the plurality of learning models; acquiring the similar documents output from each of the plurality of learning models together with scores representing similarities to the input examples; calculating a total score by adding up the scores obtained from each of the plurality of learning models for each of the similar documents; outputting the similar documents together with the total score; The plurality of learning models The result of the similarity determination differs depending on the viewpoint or index of the similarity determination. A failure analysis method comprising:

9. A fault analysis model generation method executed by a fault analysis model generation system that generates a fault analysis model that supports analysis of a fault that occurs in a monitored object, comprising: The fault analysis model generation system includes a processor and a memory, the processor: A tag for classifying documents is assigned to past cases, which are documents describing the past failures, or to knowledge, which is a document that defines rules on matters related to the monitoring target; grouping the past cases or the knowledge for each pattern of the tag combination; A plurality of fault analysis models are generated, each of which uses a predetermined learning method corresponding to each pattern from among the plurality of fault analysis model learning methods to learn the past cases or the knowledge group corresponding to each of the patterns, and performs a similarity determination with an input case, which is a document describing the fault that occurred in the monitored object, to output a similar case, which is the past case similar to the input case, or a similar document including related knowledge, which is the knowledge related to the input case. Each process has The generated plurality of fault analysis models are The result of the similarity determination differs depending on the viewpoint or index of the similarity determination. A fault analysis model generation method comprising:

Images