Aeronautical manufacturing field fault knowledge graph modeling method based on OPC UA protocol and improved CASREL model

Abstract

The invention discloses an aeronautical manufacturing field fault knowledge graph modeling method based on an OPC UA protocol and an improved CASREL model. The aeronautical manufacturing field fault knowledge graph modeling method specifically comprises the following steps that a sensor monitors and collects maintenance data; multi-source heterogeneous data are collected through an OPC UA data transmission protocol and are refined and stored according to types; preprocessing the data; carrying out data set division and model pre-training based on RoBERTa; performing triple extraction based on an improved CasRel model; constructing a knowledge graph; key entities are extracted through a BiLSTM-CRF model, and a second-step sub-graph of the knowledge graph is generated; performing knowledge question answering and reasoning based on a large language model, analyzing fault causes and generating a maintenance strategy; an online crawling mechanism is integrated, and improvement, optimization and self-learning of the knowledge base are achieved; the method can be widely applied to the fields of aviation equipment maintenance, fault processing and related decision support, the intelligent level and efficiency of fault management can be remarkably improved, and accurate and intelligent support is provided for fault diagnosis and maintenance decision of aviation equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of aviation manufacturing, and relates to the problem of fault diagnosis of aviation manufacturing equipment, and in particular to a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and an improved CASREL model. Background Art

[0002] As the fault forms in the field of aviation manufacturing become increasingly complex, the difficulty of fault diagnosis is also increasing. In the traditional aviation manufacturing fault handling process, this task is usually completed manually by professionals querying the system inspection manual. Due to the large number of mechanical and electrical components in aviation manufacturing, and with the increase of high-intensity operations and equipment service life, faults occur frequently and the causes of faults are diverse. The fault handling of aviation manufacturing equipment requires staff to have rich experience and high comprehensive business capabilities. As a result, many equipment failures are difficult to accurately find the real cause of the fault due to the lack of professional help. When the equipment fails, the staff needs to quickly analyze the operating status of the equipment, accurately determine the cause of the fault, and take appropriate treatment measures. However, a large number of fault cause records and disposal measures often exist in text form, and the staff needs to repeatedly memorize and consult relevant materials, resulting in low fault handling efficiency. If the fault cannot be handled in a timely and correct manner, it may lead to stagnation of industrial production at the least, and may endanger the lives of operators and maintenance personnel at the worst. Therefore, it is particularly important to ensure the safe and efficient operation of the equipment and develop intelligent fault diagnosis methods for aviation manufacturing equipment.

[0003] Existing fault diagnosis technologies can be roughly divided into three categories: (1) fault diagnosis methods based on analytical models; (2) fault diagnosis methods based on data-driven; and (3) fault diagnosis methods based on knowledge. Fault diagnosis methods based on analytical models usually abstract the real process into an accurate analytical model through mathematical formulas, and use the residual between the model output value and the actual measurement value to perform fault diagnosis. This method relies on a deep understanding of the mechanism of the fault process. However, when the analytical model does not match the actual situation, it will produce large diagnostic errors, especially for rotating machinery equipment with complex mechanisms and changeable working conditions. The applicability is poor. Fault diagnosis methods based on data-driven have difficulties in dealing with data noise, and the model has poor interpretability, making it difficult to cope with the diversity and complexity of CNC equipment failures.

[0004] In contrast, knowledge-based fault diagnosis methods have good interpretability, do not need to build complex mathematical mechanism models, and can continuously update fault knowledge. In recent years, knowledge graphs have been widely used as a knowledge-based fault diagnosis method tool. Knowledge graphs can integrate and filter data from fault diagnosis reports and fault case libraries by expanding the description of fault events and fault components, thereby extracting valuable information as the basis for knowledge construction, and then using the knowledge base to more accurately locate the parts of system failures and their causes. The construction of knowledge graphs usually includes entity recognition and relationship extraction technology. Traditional entity and relationship extraction methods are usually performed independently, lacking efficiency and accuracy, so it is necessary to focus on the joint extraction of entities and relationships.

[0005] In addition, artificial intelligence and big data-driven technologies have been widely used in the field of intelligent fault diagnosis. With the development and advancement of large language model technology, the system's natural language understanding and knowledge reasoning capabilities have been brought to a new level, such as Tsinghua University's ChatGML, Baidu's ERNIE, OpenAI's ChatGPT, etc. Among them, ERNIE has shown great value potential in various application fields due to its powerful Chinese natural language processing and deep learning capabilities, and has been pre-trained on a large corpus. Summary of the invention

[0006] The present invention provides a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model. The knowledge extraction process is optimized by improving the CasRel model, and knowledge reasoning and question answering are performed in combination with a large language model to solve the problems of low fault processing efficiency and insufficient accuracy in existing aviation equipment fault diagnosis methods, and to achieve efficient and accurate fault diagnosis and knowledge question answering.

[0007] In order to achieve the above technical objectives, the present invention provides the following technical solutions:

[0008] A fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model specifically includes the following steps:

[0009] S1. Set sensors to monitor the processing status of workpieces and collect maintenance data. When the sensors detect abnormal equipment data and fault status, the system triggers an alarm according to the set safety threshold and generates maintenance records and fault logs, i.e., multi-source heterogeneous maintenance data.

[0010] S2. Collect multi-source heterogeneous maintenance data through the OPC UA data transmission protocol, and store them in detail according to the data type to provide a data basis for the subsequent knowledge graph construction; the multi-source heterogeneous maintenance data is divided into structured data, i.e., manufacturer, equipment model, maintenance time, maintenance personnel, and unstructured data, i.e., maintenance report, maintenance manual; the data type includes equipment information, fault record, operation steps, alarm number, and maintenance category;

[0011] S3, preprocessing the collected unstructured data;

[0012] S4, data set division and model pre-training: divide the fault description text obtained after pre-processing into training set, validation set and test set, and convert the fault description text into word embedding, position embedding and fragment embedding vectors; encode it through the RoBERTa pre-training model, extract the context feature representation of the encoded fault description text, and convert it into a tensor containing context semantics;

[0013] S5, taking the tensor obtained in step S4 as input, extracting triples based on the improved CasRel model;

[0014] S6. Import the triples extracted in step S5 into the Neo4j graph database to construct a knowledge graph;

[0015] S7, input the fault description text into the BiLSTM-CRF model, extract the key entities, search in the knowledge graph, and generate a two-step subgraph of the knowledge graph related to the key entities of the fault description;

[0016] S8. Conduct knowledge question answering and reasoning based on a large language model; perform fault cause analysis and generate maintenance strategies based on fault descriptions; and integrate an online crawling mechanism to improve, optimize, and self-learn the knowledge base.

[0017] Further, in step S1:

[0018] The sensors include: a three-axis force sensor, a three-axis vibration sensor, an acoustic emission sensor, a temperature sensor, and a pressure sensor;

[0019] The processing status includes: the vibration status of the machine tool, temperature changes, tool processing and workbench status information, pressure of key parts of the hydraulic system and pneumatic system, and electrical status.

[0020] Furthermore, step S3 specifically includes: cleaning and deduplicating the unstructured data, and performing word segmentation, part-of-speech tagging, and stop word removal to ensure the readability and consistency of the data.

[0021] Furthermore, step S5 specifically includes:

[0022] S51, setting the entity categories and inter-entity relationship categories to be extracted;

[0023] S52, input the tensor output by the RoBERTa pre-trained model in step S4 into the head entity tagger, and use two binary classifiers to assign values ​​to the positions of the entities. The formula is expressed as:

[0024]

[0025] Among them, x i W represents the encoding vector of the i-th character in the input fault description text; start , W end are the weight coefficients of the start and end positions of the head entity respectively; b start 、b end are the biases of the start and end positions of the head entity respectively; σ is the sigmoid activation function; and Respectively represent the probability of each vector being determined as the start position and end position of the head entity, set a threshold, if the probability is greater than the threshold, assign the position a value of 1, otherwise it is 0;

[0026] S53, identifying the boundary of the header entity by using the maximum likelihood function, and identifying the header entity in the fault description text;

[0027] S54. For the identified head entity, combined with the self-attention mechanism, the head entity encoding vector representation obtained by weighting each head entity with the attention score is input into the tail entity annotator. The formula is expressed as:

[0028]

[0029] in, represents the encoding vector output by the head entity tagger; r is the relationship between entities; are the weight coefficients of the start and end positions of the tail entity respectively; are the bias items of the start and end positions of the tail entity respectively; σ is the sigmoid activation function; and Set thresholds for the probability of each vector being determined as the start position and end position of the tail entity. If the probability is greater than the threshold, the position is assigned a value of 1, otherwise it is 0;

[0030] S55. Obtain the tail entity boundary under the specific relationship r by maximizing the likelihood function, identify the tail entity, and simultaneously identify the relationship category between the head entity and the corresponding tail entity; finally, obtain the entity-relationship-entity triple.

[0031] More specifically, in step S5, the objective function of the overall model training is to maximize the likelihood function, and the formula of the likelihood function is expressed as:

[0032]

[0033] That is:

[0034]

[0035] Where h∈T j Represents the triple T j The head entity that appears in T j |h represents T j The triple with h as the head entity; R is all possible relations, R\T j |h represents all relations in the triple except the one with h as the head entity; Represents the head entity h∈T j There is no tail entity between them, D represents the entire training set; x j Represents the given fault description text.

[0036] More specifically, the Adam gradient descent method is used to optimize the objective function, and the training set data of each batch is randomly shuffled to avoid overfitting.

[0037] Furthermore, in the knowledge graph constructed in step S6:

[0038] Nodes represent entities, including: manufacturer, aviation equipment, faulty parts, and fault types;

[0039] Edges represent the relationship between entities, including: fault relationship, maintenance measures.

[0040] Furthermore, the large language model used is the ERNIE model based on the transformer architecture;

[0041] The online crawling mechanism is specifically as follows: by analyzing the feedback data of users on fault diagnosis suggestions, when the existing knowledge cannot solve the new fault, online analysis is performed to crawl solutions; after confirming that the results are reliable through user feedback and corpus comparison and analysis procedures, the current fault description is analyzed, new knowledge is supplemented, and existing knowledge is further improved and optimized, ultimately realizing self-learning of the knowledge base.

[0042] Based on the above technical solution, the present invention has the following beneficial effects:

[0043] The present invention can effectively collect and process fault data in the field of aviation manufacturing and construct a knowledge graph; it effectively improves the automation, intelligence and visualization level of aviation equipment fault diagnosis, reduces manual intervention, improves the accuracy and efficiency of diagnosis, and provides strong technical support for the maintenance of aviation equipment.

[0044] Aiming at the difficulty of fault diagnosis in aviation manufacturing equipment, the improved CasRel model was used to construct a knowledge graph, and the fault knowledge extraction process in the aviation field was optimized. The pre-training model and attention mechanism were introduced to enable it to better focus on key words in the aviation field and process complex fault descriptions to improve the accuracy of the knowledge graph, and solve the problems of error accumulation and overlapping triples in the traditional model. At the same time, the constructed aviation manufacturing knowledge graph was combined with the ERNIE large language model for knowledge reasoning to carry out intelligent fault diagnosis tasks in the aviation manufacturing field.

[0045] The overall model demonstrates excellent reasoning, question-answering, text generation, and knowledge expression capabilities in the process of fault diagnosis in the aviation manufacturing field, provides more comprehensive and effective knowledge reasoning and decision support, can better assist professionals in troubleshooting, and improve the efficiency of fault diagnosis. It reduces equipment downtime caused by faults and improves production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be described in detail and easily understood in conjunction with the following drawings, in which:

[0047] Figure 1 It is the overall flow chart of the present invention;

[0048] Figure 2 This is a partial visualization of the knowledge graph in the field of aviation manufacturing;

[0049] Figure 3 Fault retrieval graph for fault diagnosis question answering system;

[0050] Figure 4 This is the fault diagnosis result diagram of the fault diagnosis question and answer system. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the following Figure 1-4 It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0052] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0053] like Figure 1 As shown, a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model proposed by the present invention is shown, which specifically includes the following steps:

[0054] S1. Set sensors to monitor the processing status of workpieces and collect maintenance data. When the sensors detect abnormal equipment data and fault status, the system triggers an alarm according to the set safety threshold and generates maintenance records and fault logs.

[0055] As a preferred implementation of step S1, in step S1:

[0056] The sensors include: a three-axis force sensor, a three-axis vibration sensor, an acoustic emission sensor, a temperature sensor, and a pressure sensor;

[0057] The processing status includes: the vibration status of the machine tool, temperature changes, tool processing and workbench status information, pressure of key parts of the hydraulic system and pneumatic system, and electrical status.

[0058] In this embodiment, the acoustic emission sensor needs to be close to the side of the workpiece, the three-axis force sensor and the three-axis acceleration sensor need to be placed on the workpiece or the fixture; the temperature sensor is installed on the tool, the spindle and the motor, the pressure sensor is installed in the pipeline of the hydraulic system or the pneumatic system, and the three-axis acceleration sensor detects the acceleration of the processed workpiece in all directions and converts it into electrical signals. If the sensor detects an abnormality such as excessive stress, high temperature, abnormal vibration or insufficient pressure, the system will trigger an alarm mechanism according to the set safety threshold and generate maintenance records and fault logs.

[0059] S2. Collect multi-source heterogeneous maintenance data through the OPC UA data transmission protocol, and store them in detail according to the data type to provide a data basis for the subsequent knowledge graph construction; the multi-source heterogeneous maintenance data is divided into structured data, i.e., manufacturer, equipment model, maintenance time, maintenance personnel, and unstructured data, i.e., maintenance report, maintenance manual; the data type includes equipment information, fault record, operation steps, alarm number, and maintenance category;

[0060] In this embodiment, by modeling data entities into specific types such as equipment information, fault records, and operation steps, it is ensured that the system can handle diverse data inputs.

[0061] S3, preprocessing the collected unstructured data;

[0062] As a preferred implementation of step S3, the data preprocessing in S3 is specifically as follows:

[0063] Clean and deduplicate the unstructured data, perform word segmentation, part-of-speech tagging, and remove stop words to ensure the readability and consistency of the data. The preprocessed data is verified through multiple audits to ensure that it can meet the needs of subsequent modeling and knowledge graph construction. Unstructured data includes several fault description texts, operation steps, and fault causes. In particular, it is necessary to extract the fault description text from it for subsequent retrieval input information for knowledge questions and answers.

[0064] In this embodiment, the fault description text is segmented using Jieba word segmentation, and the words in the text are marked using part-of-speech tagging, stop words are removed, and the text is annotated using the labelbox online tagging tool to improve the tagging efficiency and the accuracy of the tagging results. Five entity categories are set, namely, faulty equipment, alarm number, execution operation, fault phenomenon, and fault cause, and five relationship categories are set. Finally, more than 5,000 clauses and more than 12,000 triples are obtained.

[0065] S4, dataset partitioning and model pre-training;

[0066] In this embodiment, the preprocessed fault description text is divided into a training set, a validation set, and a test set in a ratio of 8:1:1, and the fault description text is converted into a word embedding, a position embedding, and a fragment embedding vector through a dictionary; it is encoded through the RoBERTa pre-trained model, and the context feature representation of the encoded fault description text is extracted and converted into a tensor containing context semantics;

[0067] S5, taking the tensor obtained in step S4 as input, extracting triples based on the improved CasRel model;

[0068] As a preferred implementation of step S5, step S5 specifically includes:

[0069] S51, setting the entity categories and inter-entity relationship categories to be extracted;

[0070] S52, input the tensor output by the RoBERTa pre-trained model in step S4 into the head entity tagger, and use two binary classifiers to assign values ​​to the positions of the entities. The formula is expressed as:

[0071]

[0072] Among them, x iW represents the encoding vector of the i-th character in the input fault description text; start , W end are the weight coefficients of the start and end positions of the head entity respectively; b start 、b end are the biases of the start and end positions of the head entity respectively; σ is the sigmoid activation function; and Respectively represent the probability of each vector being determined as the start position and end position of the head entity, set a threshold, if the probability is greater than the threshold, assign the position a value of 1, otherwise it is 0;

[0073] S53, identifying the boundary of the header entity by using the maximum likelihood function, and identifying the header entity in the fault description text;

[0074] S54. For the identified head entity, combined with the self-attention mechanism, the formula of the self-attention mechanism is expressed as:

[0075]

[0076] Where Q, K, V are the vectors of input query, key, and value vectors; d k is the dimension of the key vector; softmax is the normalized exponential function;

[0077] The head entity encoding vector representation obtained by weighting each head entity with the attention score is input into the tail entity tagger, and the formula is expressed as:

[0078]

[0079] in, represents the encoding vector output by the head entity tagger; r is the relationship between entities; are the weight coefficients of the start and end positions of the tail entity respectively; are the bias items of the start and end positions of the tail entity respectively; σ is the sigmoid activation function; and Set thresholds for the probability of each vector being determined as the start position and end position of the tail entity. If the probability is greater than the threshold, the position is assigned a value of 1, otherwise it is 0;

[0080] S55, obtaining the tail entity boundary under the specific relationship r by maximizing the likelihood function, identifying the tail entity, and identifying the relationship category between the head entity and the corresponding tail entity; finally obtaining an entity-relationship-entity triple;

[0081] More specifically, in this embodiment, the objective function of the overall model training is to maximize the likelihood function, and the formula of the likelihood function is expressed as:

[0082]

[0083] That is:

[0084]

[0085] Where h∈T j Represents the triple T j The head entity that appears in T j |h represents T j The triple with h as the head entity; R is all possible relations, R\T j |h represents all relations in the triple except the one with h as the head entity; Represents the head entity h∈T j There is no tail entity between them, D represents the entire training set; x j Represents the given fault description text.

[0086] At the same time, the Adam optimizer is used to optimize the objective function. The learning rate of the Adam algorithm is set to 0.01, and the Dropout rate is set to 0.2. The training set, validation set, and test set are input respectively; finally, the entity-relationship-entity triples are generated.

[0087] S6. Import the triples extracted in step S5 into the Neo4j graph database to construct a knowledge graph;

[0088] In this embodiment, the development environment adopted is Windows 11, 64-b operating system, Intel Core i9-11950H CPU, 32 GB memory, RTX 3090 64G GPU. Python 3.10 is used for development, and Neo4j graph database version 5.8.0 is used. In the obtained knowledge graph, nodes represent entities such as aviation equipment, faulty parts, and fault types, and edges represent associations such as fault relationships and maintenance measures between entities. Through the graphical visualization function of Neo4j, aviation equipment, faulty parts and their associated information can be intuitively viewed, which is convenient for complex fault analysis.

[0089] S7. Input the fault description text into the BiLSTM-CRF model, extract the key entities, search in the knowledge graph, and generate a two-step subgraph of the knowledge graph related to the key entities of the fault description, such as Figure 2 As shown;

[0090] In this embodiment, the BiLSTM-CRF model includes an LSTM layer and a CRF model; the LSTM layer is used for feature extraction, and the CRF model is used for decoding; the LSTM layer is set to have 128 hidden state layers, the word vector dimension is 128, and the training rounds are 30. The batch size is set to 64, and the learning rate is set to 0.001. The model is trained using the stochastic gradient descent algorithm. The process of extracting features in the LSTM layer is:

[0091] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0092] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0093]

[0094] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0095] h t =o t *tanh(C t );

[0096] Among them, x t is the input at time t, f t is the output of the forget gate, i t and is the output of the input gate, C t is the output of the state update, o t and h t is the output of the output gate, σ represents the sigmoid function, W f ,W i ,W C and W o Weight matrix, b f ,b i ,b C and b o is the bias vector.

[0097] The decoding process of the CRF model is:

[0098]

[0099]

[0100] Among them, P(y|x) is the probability of label sequence y appearing, f k (y i-1 ,y i ,x,i) is the kth characteristic function of the model; w k is the characteristic function f k The weight coefficient of ; Z(x) is the normalization factor.

[0101] S8. Conduct knowledge question answering and reasoning based on a large language model; perform fault cause analysis and generate maintenance strategies based on fault descriptions; and integrate an online crawling mechanism to improve, optimize, and self-learn the knowledge base.

[0102] As a preferred implementation of step S8, step S8 is specifically:

[0103] Use the ERNIE large language model based on the transformer architecture;

[0104] The knowledge subgraph is embedded into the ERNIE model, and the knowledge information triples of the input knowledge graph subgraph are generated. The optimal answer y is predicted through LLM. The implementation formula is:

[0105]

[0106] Among them, LM is the ERNIE model based on the transformer architecture. is the parameter of the large model, h is the hidden layer output of the model, and t represents the set of retrieved triples.

[0107] The ERNIE model is used as a knowledge reasoning tool for fault detection and troubleshooting. The online crawling mechanism is as follows: by analyzing the feedback data of users on fault diagnosis suggestions, when the existing knowledge cannot solve the new fault, online analysis is performed to crawl solutions; after confirming that the results are reliable through user feedback and corpus comparison and analysis procedures, the current fault description is analyzed, new knowledge is supplemented, and existing knowledge is further improved and optimized, ultimately realizing self-learning of the knowledge base.

[0108] Finally, through the method proposed by the present invention, it is possible to construct Figure 3 , Figure 4 The fault diagnosis question-answering system shown in Figure 3 The fault search interface is displayed, and you can enter the fault description text to search; Figure 4 It shows the knowledge question-answering process: for the fault phenomenon described by the user, related phenomena are speculated and possible causes are given as reference; the entire fault reasoning process is visualized, the diagnosis results are detailed, and the fault cause and maintenance strategy are given.

[0109] In summary, the method proposed in the present invention can effectively improve the automation, visualization and intelligence level of aviation equipment fault diagnosis, reduce manual intervention, improve the accuracy and efficiency of diagnosis, and provide strong technical support for the maintenance of aviation equipment.

[0110] In this specification, the terms "one embodiment", "some embodiments", "examples", "specific examples" or "some examples" refer to at least one embodiment or example described in combination with specific features, structures, materials or characteristics. These described specific features, structures, materials or characteristics can be combined in an appropriate manner in one or more embodiments or examples. In addition, the technician can combine and combine different embodiments or examples and their features described in this specification without contradiction.

[0111] The logic and / or steps shown in the flowchart or described in other ways can be regarded as a sequence of executable instructions for implementing logical functions. These instructions can be implemented in any computer-readable medium for use by instruction execution systems, devices or equipment. These systems, devices or equipment include processor systems or other systems capable of receiving and executing instructions.

[0112] The above embodiments have described the principles and implementations of the present invention in detail, and have used specific examples to illustrate the working principles thereof. These examples are only used to help understand the method of the present invention and its core concept. Meanwhile, according to the concept of the present invention, the actual implementation methods and application scope may vary. Therefore, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model, characterized in that: The specific steps include: S1. Set sensors to monitor the processing status of workpieces and collect maintenance data. When the sensors detect abnormal equipment data and fault status, the system triggers an alarm according to the set safety threshold and generates maintenance records and fault logs, i.e., multi-source heterogeneous maintenance data. S2. Collect multi-source heterogeneous maintenance data through the OPC UA data transmission protocol, and store them in detail according to the data type to provide a data basis for the subsequent knowledge graph construction; the multi-source heterogeneous maintenance data is divided into structured data, i.e., manufacturer, equipment model, maintenance time, maintenance personnel, and unstructured data, i.e., maintenance report, maintenance manual; the data type includes equipment information, fault record, operation steps, alarm number, and maintenance category; S3, preprocessing the collected unstructured data; S4, data set division and model pre-training: divide the fault description text obtained after pre-processing into training set, validation set and test set, and convert the fault description text into word embedding, position embedding and fragment embedding vectors; encode it through the RoBERTa pre-training model, extract the context feature representation of the encoded fault description text, and convert it into a tensor containing context semantics; S5, taking the tensor obtained in step S4 as input, extracting triples based on the improved CasRel model; S6. Import the triples extracted in step S5 into the Neo4j graph database to construct a knowledge graph; S7, input the fault description text into the BiLSTM-CRF model, extract the key entities, search in the knowledge graph, and generate a two-step subgraph of the knowledge graph related to the key entities of the fault description; S8, knowledge question answering and reasoning based on large language models; Perform fault cause analysis and generate maintenance strategies based on the fault description; integrate online crawling mechanism to achieve improvement, optimization and self-learning of the knowledge base.

2. According to claim 1, a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model is characterized in that: In step S1: The sensors include: a three-axis force sensor, a three-axis vibration sensor, an acoustic emission sensor, a temperature sensor, and a pressure sensor; The processing status includes: the vibration status of the machine tool, temperature changes, tool processing and workbench status information, hydraulic system, pneumatic system pressure and electrical status.

3. According to claim 1, a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model is characterized in that: Step S3 specifically includes: cleaning and deduplicating the unstructured data, and performing word segmentation, part-of-speech tagging, and stop word removal to ensure the readability and consistency of the data.

4. According to claim 1, a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model is characterized in that: Step S5 specifically includes: S51, setting the entity categories and inter-entity relationship categories to be extracted; S52, input the tensor output by the RoBERTa pre-trained model in step S4 into the head entity tagger, and use two binary classifiers to assign values ​​to the positions of the entities. The formula is expressed as: Among them, x i W represents the encoding vector of the i-th character in the input fault description text; start , W end are the weight coefficients of the start and end positions of the head entity respectively; b start 、b end are the biases of the start and end positions of the head entity respectively; σ is the sigmoid activation function; and Respectively represent the probability of each vector being determined as the start position and end position of the head entity, set a threshold, if the probability is greater than the threshold, assign the position a value of 1, otherwise it is 0; S53, identifying the boundary of the header entity by using the maximum likelihood function, and identifying the header entity in the fault description text; S54. For the identified head entity, combined with the self-attention mechanism, the head entity encoding vector representation obtained by weighting each head entity with the attention score is input into the tail entity annotator. The formula is expressed as: in, represents the encoding vector output by the head entity tagger; r is the relationship between entities; are the weight coefficients of the start and end positions of the tail entity respectively; are the bias items of the start and end positions of the tail entity respectively; σ is the sigmoid activation function; and Set thresholds for the probability of each vector being determined as the start position and end position of the tail entity. If the probability is greater than the threshold, the position is assigned a value of 1, otherwise it is 0; S55. Obtain the tail entity boundary under the specific relationship r by maximizing the likelihood function, identify the tail entity, and simultaneously identify the relationship category between the head entity and the corresponding tail entity; finally, obtain the entity-relationship-entity triple.

5. According to claim 4, a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model is characterized in that: In step S5, the objective function of the overall model training is to maximize the likelihood function, and the formula of the likelihood function is expressed as: That is: Where h∈T j Represents the triple T j The head entity that appears in T j |h represents T j The triple with h as the head entity; R is all possible relations, R\T j |h represents all relations in the triple except the one with h as the head entity; Represents the head entity h∈T j There is no tail entity between them, D represents the entire training set; x j Represents the given fault description text.

6. According to claim 5, a fault knowledge graph modeling method in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model is characterized in that: The Adam gradient descent method is used to optimize the objective function, and the training set data of each batch is randomly shuffled to avoid overfitting.

7. According to claim 1, a method for modeling a fault knowledge graph in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model is characterized in that: In the knowledge graph constructed in step S6: Nodes represent entities, including: aviation equipment, faulty parts, and fault types; Edges represent the relationship between entities, including: fault relationship, maintenance measures.

8. According to claim 1, a method for modeling a fault knowledge graph in the aviation manufacturing field based on the OPC UA protocol and the improved CASREL model, characterized in that: In step S8: The large language model used is the ERNIE model based on the transformer architecture; The online crawling mechanism is specifically as follows: by analyzing the feedback data of users on fault diagnosis suggestions, when the existing knowledge cannot solve the new fault, online analysis is performed to crawl solutions; after confirming that the results are reliable through user feedback and corpus comparison and analysis procedures, the current fault description is analyzed, new knowledge is supplemented, and existing knowledge is further improved and optimized, ultimately realizing self-learning of the knowledge base.

Images