A Compressor Fault Diagnosis Method Based on an Event Logic Graph

Abstract

The present invention relates to the field of artificial intelligence technology, and in particular to a method for diagnosing compressor faults based on a causal graph, including constructing a compressor fault diagnosis ontology knowledge model; constructing a compressor fault diagnosis causal graph; performing relationship extraction on the collected compressor fault diagnosis corpus, and storing the extracted causal graph in the form of event entity-relationship-event entity and event entity-attribute-attribute value triples into a graph database; performing normalization on questions; performing question parsing on questions asked by users; converting questions input by users into structured query statements; generating standardized answers, and generating natural language answers through database query statements. The present invention stores and presents compressor faults and related processing knowledge in the form of a causal graph, and obtains the causes and processing methods for users to deal with compressor failures in a human-computer interactive question-answering manner, which helps to improve the work efficiency of technicians and reduce economic losses.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and in particular to a compressor fault diagnosis method based on an event logic graph. Background Art

[0002] In traditional search engines, queries are matched mainly in the form of words and phrases, and there are some problems. For example, when a user inputs a long, complex and detailed sentence, the machine cannot understand the user's true meaning, and the amount of information returned to the user is large, making it difficult for the user to screen and search.

[0003] At present, due to the increasing number and types of compressors, the types of their faults are also increasing, and the fault causes and solutions are also different. Compressor fault diagnosis mainly relies on technical experts to conduct diagnostic reasoning manually based on past fault cases and empirical knowledge. Therefore, there are problems such as limited number of experts, slow response speed, incomplete coverage of diagnostic knowledge, and high operation and maintenance costs, making it difficult to cope with rapid fault analysis and disposal. How to structurally express the unstructured fault diagnosis cases or empirical texts accumulated in the industry for a long time and realize the autonomous calculation and reasoning of knowledge in the diagnostic process is a common need in the industry at present. Summary of the Invention

[0004] In view of the deficiencies of the existing algorithms, the present invention stores and presents compressor faults and related processing knowledge in the form of an event logic graph, and in the form of human-computer interactive Q&A, provides users with the reasons and processing methods for dealing with compressor faults, which helps to improve the work efficiency of technicians and reduce economic losses.

[0005] The technical solution adopted by the present invention is: a compressor fault diagnosis method based on an event logic graph includes the following steps:

[0006] Step 1: Construct an ontology knowledge model for compressor fault diagnosis, including: constructing the categories and attributes of the compressor fault diagnosis ontology model; among them, the categories of the compressor fault diagnosis ontology model include compressor fault modes, fault causes, equipment structures, fault impacts, maintenance measures, and fault diagnosis; the attributes of the compressor fault ontology model include the data attributes of each category itself and the object attributes between each category.

[0007] Further, it specifically includes:

[0008] Build the relationships between various categories through the object attributes between various categories of the compressor fault diagnosis ontology model; among them, the words of the relationships include: occur, include, cause, need, relationship, and generate.

[0009] Centered on the compressor fault event mode, the compressor fault event occurs in the equipment structure (HappentedAt), which includes equipment, systems, components, and parts (Contain); the fault impact and consequences caused by the compressor fault event (LeadTo); the fault processing or measures required by the compressor fault event (Need) and the relationship between the fault event and the fault cause (cause);

[0010] Step 2: Construct a compressor fault diagnosis event graph, including: collecting and processing relevant corpus for compressor fault diagnosis, performing named entity recognition on it, extracting system entities, component entities and part entities in the compressor required in the compressor event graph to expand the event graph, and automatically updating the entity part in the event graph; performing relationship extraction on the collected compressor fault diagnosis corpus, and storing the event graph in the form of event entity-relationship-event entity and event entity-attribute-attribute value triples into a graph database;

[0011] Further, it specifically includes:

[0012] Firstly, named entity recognition is performed on the collected compressor fault related corpus, which includes system entity recognition, component part entity recognition and event entity recognition.

[0013] Secondly, the BIO sequence annotation framework is used to annotate event entities, the BILSTM algorithm is used for training, and then the CRF algorithm is used for decoding. The CRF algorithm is used to decode sentences without causing disorder.

[0014] Secondly, the relationship categories are annotated for the compressor fault diagnosis related corpus to obtain the relationship category training set; the CasRelation algorithm is used to extract the relationship to extract the triples of the event graph, and the CasRelation algorithm is used to extract and form a set of event entity-relationship-event entity triples;

[0015] Finally, assign values ​​to the data attributes of the event entity to form a set of event entity-attribute-attribute value; assign values ​​to the data attributes of the relationship to form a set of relationship-attribute-attribute value, and construct a compressor fault diagnosis causal graph.

[0016] Step 3: normalizing the question sentence, including: removing extra punctuation marks and redundant words; constructing a standardized question sentence template, including: the answering robot receives the user input of the compressor failure problem from the client, and performs question parsing on the question asked by the user; question parsing includes: question normalization, semantic vector recall, semantic slot filling and intent recognition; converting the question sentence input by the user into a structured query sentence;

[0017] Furthermore, the semantic vector recall is established as follows: all questions are vector-encoded, then the sentences are vectorized, and finally the relevant question corpora are recalled;

[0018] The semantic slot filling is established as follows: the semantic slot filling model is to perform a named entity recognition task on the questions asked by the user, identify all entities in the questions, and then recall the required event entities through vector recall;

[0019] The intention recognition is established as follows: collect a question set related to compressor failure problems. The questions queried by the user may have one or more intentions. Perform one-hot vector labeling on the question set and label the intentions; use the Bert algorithm and the TextCNN algorithm for multi-label text classification, and train the Bert algorithm and the TextCNN algorithm with the labeled question set; finally, perform intention recognition on the detailed compressor-related questions queried by the user, so that the answering robot can obtain the intention of the user's query;

[0020] Step 4: Generate a standardized answer, including: recall relevant question corpora through semantic vector recall; extract the event entities in the question through the slot filling task; then obtain the intention of the user through the intention recognition answering robot, read the fault ontology model category, relationship and constraints of the intention in the compressor fault diagnosis event graph, and generate a natural language answer through the database query statement.

[0021] Furthermore, it specifically includes:

[0022] Read triples in the compressor fault diagnosis event graph, identify one or more intentions of the user, and specific compressor fault events; read the attribute values of the compressor fault event entity attributes, including data attributes and object attributes, and generate one or more answers and convert them into natural language answers to return to the user in combination with the constraints.

[0023] The beneficial effects of the present invention:

[0024] 1. By collecting corpora in the petrochemical safety field and compressor fault corpora, sorting out the relatively complex and detailed questions about compressor faults asked by users, organizing them into a training set in the form of manual annotation and training an entity category recognition model, a relationship category recognition model and a question category recognition model, further standardizing the questions and matching them with the question templates to convert them into structured query statements, querying in the event graph, and converting the results into natural language answers.

[0025] 2. Storing and presenting compressor fault knowledge in the form of an event logic graph can answer sentences for many complex and delicate problems, parse and match user questions into question templates, and the generated answers are targeted, without blindness and mismatch with the questions raised by users; it can effectively reduce human intervention, strengthen the connection between the business needs of technicians and the development and production data, and conduct human-computer interaction and question answering in the form of natural language, providing convenience for users to obtain compressor fault knowledge, improving the user experience, helping to improve the work efficiency of users, reducing economic losses, and having good scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flowchart of the compressor fault diagnosis method based on the event logic graph of the present invention;

[0027] Figure 2 is a schematic structural diagram of the fault ontology model in the petrochemical field of the present invention;

[0028] Figure 3 is a schematic structural diagram of the entity event logic graph network of the instance of the present invention for the entity "compressor fault mode event" class of the compressor event entity;

[0029] Figure 4 is a schematic diagram of the annotation of the compressor fault relationship category of the present invention;

[0030] Figure 5 is a schematic diagram of the annotation of the compressor fault question intention of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The present invention will be further described below with reference to the drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, so it only shows the components related to the present invention.

[0032] As Figure 1 shown, a compressor fault diagnosis method based on the event logic graph includes the following steps:

[0033] Step 1: Construct a compressor fault diagnosis ontology knowledge model, including: constructing the categories and attributes of the compressor fault diagnosis ontology model; wherein, the categories of the compressor fault diagnosis ontology model include compressor fault modes, fault causes, equipment structures, fault impacts or consequences, and maintenance measures; the attributes of the compressor fault ontology model include the data attributes of each category itself and the object attributes between each category.

[0034] For example, the form of human-computer interaction Q&A is as follows: The user inputs a detailed question about the compressor failure, such as "What is the specific reason for the deviation failure that the filter pressure difference in the compressor air system is less than the standard pressure difference?" or "What are the treatment measures for the deviation failure that the filter pressure difference in the compressor air system is less than the standard pressure difference?" The response output is "The specific reason for the deviation failure that the filter pressure difference in the compressor air system is less than the standard pressure difference is that the filter element is blocked" or "The treatment measure for the deviation failure that the filter pressure difference in the compressor air system is less than the standard pressure difference is to clean and replace the filter element."

[0035] Step 2: Construct a compressor failure diagnosis event logic graph, including: collecting and processing relevant corpora for compressor failure diagnosis, performing named entity recognition on it, extracting system entities, component entities, and part entities in the compressor required in the compressor event logic graph to expand the event logic graph and automatically update the entity part in the event logic graph; performing relation extraction on the collected compressor failure diagnosis corpora, and storing the event logic graph in the form of event entity-relation-event entity and event entity-attribute-attribute value triples in a graph database;

[0036] Further, it specifically includes:

[0037] First, perform a named entity recognition task on the collected compressor failure-related corpora;

[0038] Secondly, use the BIO sequence annotation framework to annotate event entities, use the BILSTM algorithm for training, and then use the CRF algorithm for decoding. Using the CRF algorithm to decode statements will not cause disorder;

[0039] Thirdly, annotate the relation categories for the compressor failure diagnosis-related corpora to obtain a relation category training set; use the CasRelation algorithm to extract relations to extract the triples of the event logic graph, and use the CasRelation algorithm to extract and form a set of event entity-relation-event entity triples;

[0040] Finally, assign values to the data attributes of event entities to form a set of event entity-attribute-attribute values; assign values to the data attributes of relations to form a set of relation-attribute-attribute values, and construct a compressor failure diagnosis event logic graph.

[0041] The event graph of compressor faults is a symbolic representation of the real physical world. It organizes the compressor fault information accumulated from a large number of information sources on the Internet into knowledge about compressor faults that can be utilized. A search engine based on the event graph can visually feedback graph-structured compressor fault knowledge to users in a graphical way. Users no longer need to browse a large amount of information and use traditional search engines to search for knowledge, and can accurately obtain and handle the knowledge for dealing with compressor faults. It should be noted here that the event graph is composed of entity relations with attributes linked together. In the event entity-relation-event entity, the nodes represent event entities, and the edges represent the relations between event entities. For example, in "compressor fault cause", "filter clogging is included", where "included" is the relation. This kind of knowledge is usually defined by business experts in the field of compressor faults based on professional knowledge. In the event entity-attribute-attribute value, the nodes represent event entities, and the attribute is the data attribute of the entity. For example, "the pressure difference of the compressor filter is less than the standard pressure difference", the event entity is the pressure difference of the compressor filter, the attribute is pressure, and the attribute value is -1 kPa;

[0042] Step 3. The normalization process of the question includes: removing redundant punctuation marks and redundant words; constructing a standardized question template, including: the response robot receives the user's input about compressor faults from the client and parses the question asked by the user; the question parsing includes: question normalization, semantic vector recall, semantic slot filling, and intent recognition; converting the question input by the user into a structured query statement;

[0043] The operation method for generating a structured query statement is as follows: The machine receives the question input by the user from the client, normalizes the question, and deletes redundant punctuation marks (such as the question mark at the end of each question); performs BIO sequence annotation on the event entities in the question. The named entity recognition algorithm can adopt existing models (based on models such as BILSTM, without limitation). Subsequently, the annotated BIO sequence is decoded, and the decoding algorithm can adopt existing models (based on models such as CRF, without limitation). Then, the machine performs an intent recognition task on the question, sets a label for the intent of the question. The user may have one or more intents, and the question is marked with a one-hot vector. It is marked as many times as there are intents;

[0044] The text classification algorithm can adopt existing models (such as models based on Bert + TextCNN, without limitation); cut the question sentence, extract the events, event attributes, and interrogative words in the question sentence, then perform synonym replacement on the words, extract event entities, attributes, relationships, and limiting conditions. Limiting conditions such as the types of compressors are extracted if there are any, and ignored if not. Event entities of compressor failures such as the differential pressure of the compressor filter being less than the standard differential pressure, and the inlet temperature of the compressor air system being less than the standard temperature, attributes such as pressure, temperature, etc., relationships such as inclusion, composition, generation, cause, etc. Different types of compressors have different corresponding failure occurrences, failure causes, and failure handling methods. Limiting conditions refer to the restrictions imposed on the main event entity in the question sentence. There is a one-to-one correspondence between the event entities of compressor failures and the entities. Mark the part-of-speech of the event entity and match the answer template, that is, match the normalized user input with the question templates in the question template library to obtain the question template corresponding to the user input question. According to the question template, the entity-attribute-relationship and limiting conditions included in the question sentence.

[0045] Step 4: Generate a standardized answer, including: recall relevant question sentence corpora through semantic vector recall; extract the event entities in the question sentence through slot filling tasks; then obtain the user's intention through intention recognition of the answering robot, read the fault ontology model category, relationship, and limiting conditions of the intention in the compressor fault diagnosis event logic graph, and generate a natural language answer through database query statements.

[0046] After the machine performs semantic vector recall, semantic slot filling, and intention recognition tasks on the user input question sentence, for example, when the user inputs "What is the reason for the deviation of the filter differential pressure parameter of the compressor and how to handle it?", the corresponding question sentence corpora obtained through semantic vector recall are "What is the reason for the failure that the differential pressure of the compressor filter is less than the standard differential pressure and how to handle it?" or "What is the reason for the failure that the differential pressure of the compressor filter is greater than the standard differential pressure and how to handle it?"; the failure here obtained through the semantic slot filling task is "the differential pressure of the filter is less than the standard differential pressure" or "the differential pressure of the filter is greater than the standard differential pressure"; through intention recognition, it is clear that the user's intention is to know the reason and handling of the failure. Through named entity recognition and relationship extraction, a compressor fault diagnosis event logic graph is constructed, and the corresponding failure reasons read from it are "filter element blockage" or "filter element looseness", and the corresponding failure handling measures read are "replace the filter element"; then the answer output by the answering robot is "The reason for the deviation failure that the differential pressure of the compressor filter is less than the standard differential pressure is filter element blockage, and the handling method is to replace the filter element", and "The reason for the deviation failure that the differential pressure of the compressor filter is greater than the standard differential pressure is filter element looseness, and the handling method is to replace the filter element";

[0047] like Figure 2 As shown in the figure, before building the compressor fault diagnosis question-answering system, it is necessary to define the compressor fault diagnosis ontology knowledge model; the parent categories are fault mode, fault cause, fault impact, maintenance measures, fault diagnosis and equipment structure. Among them, the subcategories of fault cause are equipment itself, environmental factors and human factors; the equipment structure subcategory includes equipment, the subcategory of equipment is system, the subcategory of system is component, and the subcategory of component is parts; the subcategories of fault impact are upper impact, local impact and final impact; the subcategories of maintenance measures are improvement, cleaning, inspection, adjustment, repair and care;

[0048] like Figure 4 ,5 shows the training corpus for constructing the compressor fault diagnosis event graph, Figure 4 Annotate the corpus for entity extraction; Figure 5 To extract triples of annotated corpus mainly based on extraction relations; Figure 4 , 5 The operation shown in the figure stores the extracted compressor fault diagnosis triples into the event graph, such as Figure 3 is an example diagram of the extracted compressor failure mode event category;

[0049] Take a user inputting an example question sentence "Hello, what is the cause of the fault of filter pressure difference deviation in the air-conditioning compressor, and how to deal with it?" to illustrate the knowledge question answering process;

[0050] 1. Standardize the question. If necessary, please process it. If not, please ignore it. Remove "Hello" from the above question, that is, "What is the cause of the filter pressure difference deviation in the air-conditioning compressor? How to deal with it?";

[0051] 2. Remove punctuation marks and interjections in the question, i.e. "What is the cause of the filter pressure difference deviation in the air-conditioning compressor and how to deal with it?"

[0052] 3. Remove unnecessary words, such as "Excuse me" and "What", that is, "What is the cause of the fault of filter pressure difference deviation in the air-conditioning compressor and how to deal with it";

[0053] 4. Parsing the questions, including: recalling relevant question corpus: "What is the cause of the fault that the filter pressure difference of the air-conditioning compressor is less than the standard pressure difference? How to deal with it?" or "What is the cause of the fault that the filter pressure difference of the air-conditioning compressor is greater than the standard pressure difference? How to deal with it?";

[0054] Extract event entity: "Filter pressure difference is less than standard pressure difference" or "Filter pressure difference is greater than standard pressure difference";

[0055] Extract the question intent: "cause of the fault" or "solution to the fault";

[0056] Extract the limiting condition: "air conditioner compressor";

[0057] 5. Generate a database query statement to obtain the following Cypher query statements: MATCH p=(x:fault{name: "Filter differential pressure is less than the standard differential pressure"})-[:cause]-(y:cause{}) return count(y); MATCH p=(x:fault{name: "Filter differential pressure is less than the standard differential pressure"})-[:solution]-(y:Need{}) return count(y) MATCH p=(x:fault{name: "Filter differential pressure is greater than the standard differential pressure"})-[:cause]-(y:cause{}) return count(y); MATCH p=(x:fault{name: "Filter differential pressure is greater than the standard differential pressure"})-[:solution]-(y:Need{}) return count(y);

[0058] 6. Retrieve and query the event graph of compressor fault diagnosis to obtain the answer;

[0059] 7. Generate the following answer template:

[0060] The queried compressor fault event entity + relationship (here are cause and solution) + the result queried in the event graph;

[0061] 8. The answering robot outputs the answer to the user: namely, "The cause of the deviation fault that the filter differential pressure of the air conditioner compressor is less than the standard differential pressure is that the filter element is blocked, and the solution is to replace the filter element" or "The cause of the deviation fault that the filter differential pressure of the air conditioner compressor is greater than the standard differential pressure is that the filter element is loose, and the solution is to replace the filter element".

[0062] Enlightened by the ideal embodiments of the present invention as described above, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of this invention. The technical scope of this invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A compressor fault diagnosis method based on an event logic graph, characterized in that It includes the following steps: Step 1: Construct a compressor fault diagnosis ontology knowledge model, including: constructing the categories and attributes of the compressor fault diagnosis ontology model; among them, the categories of the compressor fault diagnosis ontology model include compressor fault modes, fault causes, equipment structures, fault impacts, maintenance measures, and fault diagnosis; the attributes of the compressor fault ontology model include various category data attributes and object attributes between various categories; Build the relationships between categories through the object attributes between the categories of the compressor fault diagnosis ontology model; among them, the words of the relationships include: occur, include, cause, require, relationship, and generate; Step 2: Construct a compressor fault diagnosis event logic graph, including: collecting and processing compressor fault diagnosis-related corpora, and performing named entity recognition, extracting compressor system entities, component entities, and part entities in the compressor event logic graph to expand the event logic graph and the automatic update of the entity part in the event logic graph; performing relationship extraction on the collected compressor fault diagnosis corpora, and storing the event logic graph in the form of event entity-relationship-event entity and event entity-attribute-attribute value triples in the graph database; Step 2 specifically includes: First, perform named entity recognition on the collected compressor fault-related corpora, and the named entity recognition includes: system entity recognition, component and part entity recognition, and event entity recognition; Secondly, use the BIO sequence annotation framework to annotate event entities, use the BILSTM algorithm for training, and then use the CRF algorithm for decoding; Thirdly, annotate the relationship categories of the compressor fault diagnosis-related corpora to obtain a relationship category training set; use the CasRelation algorithm to extract relationships to extract the triples of the event logic graph, and use the CasRelation algorithm to extract and form a set of event entity-relationship-event entity triples; Finally, assign values to the data attributes of the event entities to form a set of event entity-attribute-attribute values; assign values to the data attributes of the relationships to form a set of relationship-attribute-attribute values, and construct a compressor fault diagnosis event logic graph; Step 3: The normalization process of the question includes: removing redundant punctuation marks and redundant words; constructing a standardized question template, including: the response robot receives the user's input of a compressor failure problem from the client, and performs question parsing on the question asked by the user; the question parsing includes: question normalization, semantic vector recall, semantic slot filling, and intent recognition; converting the question input by the user into a structured query statement; Step 4: Generate a standardized answer, including: recalling relevant question corpora through semantic vector recall; extracting the event entities in the question through slot filling; and then enabling the response robot to obtain the user's intent through intent recognition, reading the categories, relationships, and constraints of the compressor fault diagnosis ontology model to which the intent belongs in the compressor fault diagnosis event logic graph, and generating a natural language answer through a database query statement.

2. The compressor fault diagnosis method based on the event logic graph according to claim 1, characterized in that The establishment method of semantic vector recall is: encoding all questions into vectors, then vectorizing the sentences, and finally recalling relevant question corpora; The establishment method of semantic slot filling is as follows: perform named entity recognition on the question asked by the user, identify all entities in the question, and then recall the required event entities through vector retrieval; The establishment method of intention recognition is as follows: collect a question set related to compressor failure problems, and perform one-hot vector labeling on the question set; use the Bert algorithm and the TextCNN algorithm for multi-label text classification, and train the Bert algorithm and the TextCNN algorithm with the labeled question set; finally, perform intention recognition on the compressor-related questions queried by the user, so that the answering robot can obtain the intention of the user's query.

3. The compressor fault diagnosis method based on the event logic graph according to claim 1, characterized in that Step 4 specifically includes: read triples in the compressor fault diagnosis event graph, identify one or more intentions of the user and specific compressor fault events; read the attribute values of the compressor fault event entity attributes, including data attributes and object attributes, combine the limiting conditions, generate one or more answers and convert them into natural language answers to return to the user.

Images