Intelligent question and answer method and system for equipment operation and maintenance based on context fusion on knowledge graph

Abstract

The application discloses a device operation and maintenance intelligent question and answer method and system based on knowledge graph context fusion, which firstly constructs a knowledge text dataset in the field of industrial device operation and maintenance, obtains an industrial device operation and maintenance knowledge graph, and divides the knowledge graph into communities to obtain knowledge communities.Secondly, the user input question and the knowledge communities are represented by vectors respectively. Then, the diversity correlation between the question vector and the community vector is calculated, and a similarity confidence distribution reflecting different correlation degrees is generated, and adaptive adjustment is performed through a community confidence discount strategy. Finally, the similarity confidence distribution after adaptive adjustment is fused, the knowledge graph context is input into an industrial device operation and maintenance large language model, and the answer to the industrial device operation and maintenance question is generated. The application improves the reliability of the constructed knowledge graph context, and provides an efficient and decision-making reliable device operation and maintenance intelligent question and answer method for industrial device operation and maintenance.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, specifically to an intelligent question-answering method and system for equipment operation and maintenance based on knowledge graph context fusion. Background Technology

[0002] Large industrial equipment is widely used in key industrial sectors such as chemical, energy, power, and transportation. Examples include wind turbines, gas turbines, and large compressors. These critical devices typically operate in complex environments with fluctuating operating conditions, and their operational status directly impacts production safety and economic efficiency. When equipment malfunctions or malfunctions, maintenance personnel often need to quickly locate the fault, analyze the cause, and develop a solution, placing high demands on efficiency and expertise. Currently, in industrial equipment maintenance, relevant knowledge is primarily stored in unstructured or weakly structured texts such as technical manuals, maintenance procedures, and historical repair records. Maintenance personnel typically rely on manual retrieval or expert experience for problem-solving, making it difficult to respond quickly and accurately to complex issues. In recent years, large language models, due to their outstanding performance in natural language understanding and knowledge-based question answering, have been introduced into industrial maintenance scenarios to assist personnel in problem consultation and solution generation. However, existing operation and maintenance question-answering methods based on large language models generally suffer from problems such as insufficient reliance on professional domain knowledge, difficulty in assessing the credibility of knowledge sources, and insufficient stability of reasoning results, or even hallucinations, which restrict their application effectiveness in actual industrial equipment operation and maintenance scenarios. Summary of the Invention

[0003] To address the problems of fragmented equipment operation and maintenance knowledge, limited question matching, and insufficient reliability of inference results in existing technologies, this invention proposes an intelligent question-answering method and system for equipment operation and maintenance based on knowledge graph context fusion. Compared with existing technologies, this invention achieves multi-dimensional matching between user questions and knowledge communities through differentiated similarity measurement, and combines community reliability discounting and reliability fusion strategies to regulate the credibility of search results, thereby improving the reliability of the constructed knowledge graph context and providing support for equipment operation and maintenance decision-making and intelligent question answering.

[0004] In one aspect, this invention provides an intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion, the method comprising the following steps:

[0005] S1: Construct a knowledge text dataset for the field of industrial equipment operation and maintenance.

[0006] S2: Construct an industrial equipment operation and maintenance knowledge graph based on the knowledge text dataset obtained in S1, and divide the knowledge graph into communities to obtain multiple knowledge communities that represent different operation and maintenance knowledge units.

[0007] S3: Vectorize the user-inputted question and the knowledge community obtained in S2 to obtain the question vector and the community vector, respectively.

[0008] S4: Calculate the diverse associations between the problem vector and the community vector obtained in S3 using differentiated similarity metrics, and generate similarity confidence distributions that reflect different degrees of association.

[0009] S5: Propose a community reliability discount strategy to adaptively adjust the similarity reliability distribution obtained in S4, so as to improve the reliability of the reliability distribution.

[0010] S6: The adaptively adjusted similarity reliability distribution is fused, a knowledge graph context is constructed based on the fusion result, and the obtained knowledge graph context is input into the industrial equipment operation and maintenance large language model to generate answers to industrial equipment operation and maintenance questions.

[0011] Specifically, S3 includes the following steps:

[0012] S3.1: Receive user input of industrial equipment operation and maintenance questions, encode the questions using a text embedding model, and generate question vectors to represent the semantic features of the questions. .

[0013] S3.2: For each knowledge community, extract the semantic attribute features and topological structure features within the knowledge community, and construct the initial graph feature representation of the knowledge community; the semantic attribute features include the node feature matrix generated based on the entity description within the community, and the topological structure features include adjacency information representing the connection relationship between entities within the community.

[0014] S3.3: The initial graph feature representation is processed using a graph neural network. Through message passing between nodes and aggregation of neighborhood features, the topological semantic relationships within the community are captured. Global pooling is then performed on the aggregated features to generate a community vector representing the overall characteristics of the knowledge community. Indicates the first Community vector of a knowledge community.

[0015] S4 includes the following steps:

[0016] S4.1: Define a differentiated set of similarity metrics ,in Let K represent the k-th similarity measure, where K represents the number of similarity measures used in this method. Among them, cosine similarity is used to measure semantic direction consistency, Euclidean similarity is used to characterize the overall semantic position proximity, Manhattan similarity is used to reflect the accumulation of differences in multidimensional semantic features, and dot product similarity is used to represent the activation strength of the question semantics on the community semantic features.

[0017] S4.2: For the k-th similarity measure Calculate the problem vectors respectively The similarity between the vectors and the knowledge community vectors, using Represents the problem vector and the first The similarity results of each knowledge community are used to obtain the set of question-community similarity results under the k-th measure. ,in This indicates that the problem vector is computed using the k-th measure. With the Community vector of a knowledge community The similarity between them.

[0018] S4.3: Set of similarity results for the k-th measure Based on similarity, select the top-ranked similarity values ​​under the k-th measure, from largest to smallest. The community is used as the candidate community set under the k-th measure, where are positive integers and After obtaining the candidate community sets for all measures in sequence, the union of all candidate community sets is taken to form a joint candidate community set. That is, the identification framework; generated through combination enumeration. The complete set of mathematical subsets constitutes power set , For set The number of elements in the array.

[0019] S4.4: Normalize the similarity results within the candidate community set corresponding to the k-th measure to obtain the similarity reliability distribution of the k-th measure defined on the joint candidate community set. For knowledge communities belonging to the candidate community set of the kth measure, their reliability value is determined by the normalized similarity result. For knowledge communities belonging to the joint candidate community set but not to the candidate community set of the kth measure, their reliability value is set to 0. The similarity reliability distribution of all measures is obtained in turn.

[0020] S5 includes the following steps:

[0021] S5.1: For the k-th measure, based on the distribution pattern index and combined with the reliability value characteristics of each joint candidate community corresponding to the k-th measure, a discount factor is determined for each knowledge community in the joint candidate community set, forming a discount parameter vector under the k-th measure. .

[0022] S5.2: For the k-th measure, based on the discount parameter vector For the similarity reliability distribution The reliability values ​​of each knowledge community in the text are discounted item by item to obtain the reliability distribution after community reliability discounting. Output the distribution of K similarity reliability scores after community reliability discount.

[0023] S6 includes the following steps:

[0024] S6.1: Using the combination rule in Dempster-Shafer evidence theory, fuse the K similarity reliability distributions output from S5 after community reliability discounting to obtain the comprehensive reliability distribution. .

[0025] S6.2: The comprehensive reliability distribution is quantized using an approximate probability function to obtain the quantized comprehensive reliability distribution. The reliability results of single-point subsets are sorted, and the top-ranked ones are selected. This serves as a target knowledge community with high relevance to current industrial equipment operation and maintenance issues. Based on this target knowledge community, knowledge content related to the issues is extracted from the corresponding industrial equipment operation and maintenance knowledge graph to construct a knowledge graph context for input to the large language model of industrial equipment operation and maintenance.

[0026] S6.3: Using the industrial equipment operation and maintenance question-and-answer samples obtained in S1, fine-tune the parameters of the large language model to obtain the industrial equipment operation and maintenance large language model; input the knowledge graph context and the user input question into the industrial equipment operation and maintenance large language model, and generate natural language answers to the corresponding industrial equipment operation and maintenance questions after the large model inferences.

[0027] In another aspect, the present invention provides an intelligent question-and-answer system for equipment operation and maintenance based on knowledge graph context fusion, comprising the following modules:

[0028] The knowledge text module is used to build knowledge text datasets in the field of industrial equipment operation and maintenance.

[0029] The knowledge community module constructs an industrial equipment operation and maintenance knowledge graph based on a knowledge text dataset, and divides the knowledge graph into communities to obtain multiple knowledge communities that represent different operation and maintenance knowledge units.

[0030] The question vector and community vector modules are used to vectorize the user-input questions and knowledge communities, respectively, to obtain question vectors and community vectors.

[0031] The similarity reliability distribution module is used to calculate the diverse associations between question vectors and community vectors through differentiated similarity metrics, and to generate similarity reliability distributions that reflect different degrees of association.

[0032] The adaptive adjustment module is used to adaptively adjust the similarity reliability distribution through a community reliability discounting strategy.

[0033] The industrial equipment operation and maintenance question answering module is used to fuse adaptively adjusted similarity reliability distributions, construct a knowledge graph context input to the industrial equipment operation and maintenance large language model, and generate answers to industrial equipment operation and maintenance questions.

[0034] The beneficial effects of this invention are:

[0035] This invention proposes an intelligent question-answering method and system for equipment operation and maintenance based on knowledge graph context fusion. Compared with existing technologies, this invention achieves multi-angle matching between user questions and knowledge communities through differentiated similarity metrics, and combines community reliability discounting and reliability fusion strategies to constrain the consistency of search results, thereby improving the reliability of the constructed knowledge graph context. Overall, this invention provides an efficient and reliable intelligent question-answering method for industrial equipment operation and maintenance. Attached Figure Description

[0036] Figure 1 This is an overall flowchart of the present invention;

[0037] Figure 2 This is a flowchart of multi-measure similarity calculation and reliability distribution generation. Detailed Implementation

[0038] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, the invention will now be described in more detail with reference to the accompanying drawings and embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. The components shown in the drawings and embodiments can be implemented in various configurations and designs, and the specifically described implementations are intended to illustrate selected implementation methods and are not intended to limit the scope of the invention.

[0039] In one aspect, this invention provides an intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion. For specific steps, please refer to [link / reference needed]. Figure 1 This includes the following steps:

[0040] S1: Construct a knowledge text dataset for the field of industrial equipment operation and maintenance.

[0041] S1.1: Unified collection of unstructured or semi-structured operation and maintenance data from different sources is carried out through document parsing and text extraction technology: raw text data is obtained from the maintenance record database, fault report document library and equipment operation and maintenance standard documents related to industrial equipment operation and maintenance.

[0042] S1.2: The original text data is cleaned, segmented, and semantically standardized to eliminate noise and unify the mapping of professional terms, resulting in a standardized operation and maintenance knowledge text dataset.

[0043] S1.3: Extract knowledge entries from the operation and maintenance knowledge text dataset obtained in S1.2, and construct an industrial equipment operation and maintenance question and answer sample set containing multiple question and answer pairs.

[0044] S2: Construct an industrial equipment operation and maintenance knowledge graph based on the knowledge text dataset obtained in S1, and divide the knowledge graph into communities to obtain multiple knowledge communities that represent different operation and maintenance knowledge units.

[0045] S2.1: The industrial equipment operation and maintenance knowledge text dataset obtained in S1 is divided into text units. Based on named entity recognition and relation extraction technology, operation and maintenance elements such as equipment objects, fault phenomena, fault locations, and handling measures are identified from the operation and maintenance knowledge text dataset as entity nodes. At the same time, the causal relationship and handling dependency relationship between entity nodes are identified to construct structured graph data and obtain the industrial equipment operation and maintenance knowledge graph.

[0046] S2.2: Based on the industrial equipment operation and maintenance knowledge graph obtained in S2.1, a graph community discovery algorithm is used to divide the knowledge graph into communities according to the relationships between entity nodes, thereby obtaining multiple distinct knowledge communities. This graph community discovery algorithm is the Leiden algorithm. Indicates the first A knowledge community This represents the total number of knowledge communities, hence .

[0047] S3: Vectorize the user-inputted question and the knowledge community obtained in S2 to obtain the question vector and the community vector, respectively.

[0048] Specifically, the input question is encoded using the nomic-embed-text model, outputting a 768-dimensional question vector. Then, for each knowledge community, semantic attribute features and topological structure features are extracted to construct an initial graph feature representation of the knowledge community. The semantic attribute features include a node feature matrix generated based on entity descriptions within the community, and the topological structure features include adjacency information representing the connections between entities within the community. The initial graph feature representation is processed using a graph neural network. Through message passing between nodes and neighborhood feature aggregation, the topological semantic associations within the community are captured. Global pooling is then performed on the aggregated features to generate a 768-dimensional community vector representing the overall features of the knowledge community. Indicates the first Community vector of a knowledge community.

[0049] S4: Calculate the diverse associations between the problem vector and the community vector obtained in S3 using differentiated similarity metrics, and generate similarity confidence distributions that reflect different degrees of association.

[0050] S4.1: Define a differentiated set of similarity metrics ,in Let K represent the k-th similarity measure, where K represents the number of similarity measures used. These measures include cosine similarity, Euclidean similarity, Manhattan similarity, and dot product similarity. Cosine similarity... Used to measure semantic direction consistency Represents the dot product of vectors. Representing the 2-norm; Euclidean similarity Used to characterize the overall semantic proximity; Manhattan similarity Used to reflect the accumulation of differences in multidimensional semantic features. Representing the norm; dot product similarity The Sigmoid function is used to represent the activation strength of question semantics to community semantic features. Represented as This function is used to map the dot product result to... The interval is used to ensure the comparability of outputs from different measures; different similarity measures characterize the relationship between the problem vector and the knowledge community vector from different perspectives, thus forming a multi-angle and multi-level relationship result.

[0051] S4.2: For the k-th similarity measure Calculate the problem vectors respectively The similarity between the vectors and the knowledge community vectors, using Represents the problem vector and the first The similarity results of each knowledge community are used to obtain the set of question-community similarity results under the k-th measure. ,in Denotes the computation problem vector under the k-th measure. With the Community vector of a knowledge community The similarity between them.

[0052] S4.3: Set of similarity results for the k-th measure Based on similarity, select the top-ranked similarity values ​​under the k-th measure, from largest to smallest. The community is used as the candidate community set under the k-th measure, where are positive integers and After obtaining the candidate community sets for all measures in sequence, the union of all candidate community sets is taken to form a joint candidate community set. That is, the identification framework; generated through combination enumeration. The complete set of mathematical subsets constitutes power set , For set The number of elements in;

[0053] S4.4: Normalize the similarity results within the candidate community set corresponding to the k-th measure to obtain the similarity reliability distribution of the k-th measure defined on the joint candidate community set. For knowledge communities belonging to the candidate community set of the kth measure, their reliability value is determined by the normalized similarity result. For knowledge communities belonging to the joint candidate community set but not to the candidate community set of the kth measure, their reliability value is set to 0. The similarity reliability distribution of all measures is obtained in turn.

[0054] S5: Propose a community reliability discount strategy to adaptively adjust the similarity reliability distribution obtained in S4, so as to improve the reliability of the reliability distribution.

[0055] S5.1: Similarity reliability distribution for the k-th definition on the joint candidate community set The distribution morphology index is defined to characterize the structural features of the confidence distribution, including the dispersion index and the discrimination strength index. The dispersion index is obtained by normalizing the information entropy of the confidence distribution with maximum entropy, where the maximum entropy is the entropy value when the confidence value is uniformly distributed. The discrimination strength index is the difference between the maximum confidence and the second largest confidence.

[0056] For the k-th measure, a similarity reliability distribution is defined on the joint candidate community set based on the k-th measure. Determine the discount parameter vector using the following steps:

[0057] (1) Knowledge communities with a confidence value greater than 0 in the similarity confidence distribution are selected as effective communities. The information entropy of the confidence distribution is calculated based on the proportion of the confidence value of the effective community in the total confidence of the effective community. The dispersion index is obtained by normalizing the maximum entropy. Simultaneously, the effective community reliability scores are sorted from highest to lowest, and the difference between the highest and second-highest reliability scores is calculated to obtain the discrimination strength index. ;

[0058] (2) Based on the dispersion index Differentiating Intensity Indicators ,Will and Perform a weighted summation to construct the interval adjustment coefficient corresponding to the k-th measure. Among them, the preset discount range This is used to define the acceptable range of the discount factor to ensure the stability of discount adjustments under different measures; within the preset discount range The discount interval for the k-th measure is determined based on the interval adjustment coefficient. ,in equal , equal and The summation result:

[0059]

[0060]

[0061] in, The weight is set to a preset parameter and can be set based on experience. and Let these represent the lower and upper bounds of the preset discount range, respectively, and satisfy the following conditions: .

[0062] (3) Discount interval for the k-th measure Within this framework, a linear mapping is performed based on the reliability values ​​of each knowledge community under the k-th measure to obtain the discount factor corresponding to each knowledge community. Corresponding discount factor equal and The summation result:

[0063]

[0064] in, This represents the knowledge community in the joint candidate community set under the k-th measure. The corresponding reliability value, Describing a knowledge community under the k-th measure The corresponding discount factor, This is a positive integer index variable used to represent the knowledge community ID in the joint candidate community set;

[0065] (4) Obtain the discount parameter vector under the kth measure. , Indicates the k-th measure. Each knowledge community corresponds to a discount factor for its reliability. The discount factor represents the unreliability of the reliability; the smaller the discount factor, the more reliable the reliability, and vice versa. The discount parameter vectors for all measures are obtained sequentially.

[0066] S5.2: For the k-th measure, based on the discount parameter vector obtained in S5.1 For the similarity reliability distribution The reliability values ​​of each knowledge community in the text are discounted item by item to obtain the reliability distribution after community reliability discounting. The discount process is implemented using set transformation operators: for any subset The community reliability discounted reliability value is composed of all subsets. The original reliability values ​​are obtained by weighted summation, where right The original reliability values ​​are weighted by the discount kernel function. Control; discount kernel function based on and The set inclusion relationship between the knowledge communities is used to multiply and combine the discount factor and the retention coefficient of the corresponding knowledge community reliability, thereby realizing the weighted propagation of community-level reliability in the set space:

[0067]

[0068]

[0069] in, This represents a set of joint candidate communities, serving as an identification framework for the current problem, whose elements... The first in the corresponding identification framework A knowledge community; and for Any subset of represents a set hypothesis consisting of multiple candidate knowledge communities; Representing a subset The corresponding original reliability value, Represents a subset after community reliability discount The corresponding reliability value; for The complement; Indicates the k-th measure. The retention coefficient of each knowledge community corresponds to a certain level of reliability, which is used to characterize the reliable portion retained during the discounting process, and satisfies the following conditions: ; This indicates a series of multiplication operations.

[0070] S5.3: Output the K similarity reliability distributions after community reliability discount, as input for the subsequent fusion of multiple similarity reliability distributions.

[0071] S6: The adaptively adjusted similarity reliability distribution is fused, a knowledge graph context is constructed based on the fusion result, and the obtained knowledge graph context is input into the industrial equipment operation and maintenance large language model to generate answers to industrial equipment operation and maintenance questions.

[0072] S6.1: Using the combination rule in Dempster-Shafer evidence theory, fuse the K similarity reliability distributions after the community reliability discount output in S5 to obtain the comprehensive reliability distribution.

[0073] S6.2: The comprehensive confidence distribution obtained in S6.1 is quantified using an approximate probability function, which is the Pignistic probability function:

[0074]

[0075] in, for A subset of , used to represent the set variable traversed during the summation process; Represents a set The number of elements in; for A subset of is used to represent the set of candidate knowledge communities for decision evaluation.

[0076] The quantified comprehensive reliability distribution is obtained, and then... The reliability results of single-point subsets are sorted, and the top-ranked ones are selected. This serves as a target knowledge community with high relevance to current industrial equipment operation and maintenance issues. Based on this target knowledge community, knowledge content related to the issues is extracted from the corresponding industrial equipment operation and maintenance knowledge graph to construct a knowledge graph context for input to the large language model of industrial equipment operation and maintenance.

[0077] S6.3: Using the industrial equipment operation and maintenance question-and-answer samples obtained in S1, fine-tune the parameters of the large language model to obtain the industrial equipment operation and maintenance large language model; the baseline large model used is the Qwen2-7B model, and LoRa is used to fine-tune the parameters of the model; then, the knowledge graph context obtained above and the user input questions are input into the industrial equipment operation and maintenance large language model, and the large model generates natural language answers to the corresponding industrial equipment operation and maintenance questions after inference.

[0078] In another aspect, the present invention provides an intelligent question-and-answer system for equipment operation and maintenance based on knowledge graph context fusion, such as... Figure 2 As shown, it includes the following modules:

[0079] The knowledge text module is used to build knowledge text datasets in the field of industrial equipment operation and maintenance.

[0080] The knowledge community module constructs an industrial equipment operation and maintenance knowledge graph based on a knowledge text dataset, and divides the knowledge graph into communities to obtain multiple knowledge communities that represent different operation and maintenance knowledge units.

[0081] The question vector and community vector modules are used to vectorize the user-input questions and knowledge communities, respectively, to obtain question vectors and community vectors.

[0082] The similarity reliability distribution module is used to calculate the diverse associations between question vectors and community vectors through differentiated similarity metrics, and to generate similarity reliability distributions that reflect different degrees of association.

[0083] The adaptive adjustment module is used to adaptively adjust the similarity reliability distribution through a community reliability discounting strategy.

[0084] The industrial equipment operation and maintenance question answering module is used to fuse adaptively adjusted similarity reliability distributions, construct a knowledge graph context input to the industrial equipment operation and maintenance large language model, and generate answers to industrial equipment operation and maintenance questions.

[0085] Example:

[0086] The specific applications of the present invention will be further described below with reference to specific embodiments:

[0087] First, raw text data is obtained from the maintenance record database, fault report document library, and equipment operation and maintenance standard documents related to offshore wind turbine equipment operation and maintenance. Then, the collected raw text is cleaned, segmented, and semantically standardized to eliminate noise content and unify the mapping of professional terms, resulting in a standardized wind turbine operation and maintenance knowledge text dataset. Finally, knowledge items are extracted from the corpus to construct a wind turbine equipment operation and maintenance question and answer sample set containing multiple question-answer pairs. In this embodiment, a question and answer sample set containing 1451 question and answer samples is constructed.

[0088] This embodiment uses Microsoft's Graphrag architecture to unitize text content, forming multiple text units. Based on named entity recognition and relation extraction technology, it identifies maintenance elements such as equipment objects, fault phenomena, fault locations, and handling measures as entity nodes from the text dataset. Simultaneously, it identifies causal relationships and handling dependencies between entity nodes, obtaining a wind turbine equipment maintenance knowledge graph. Then, the Leiden community discovery algorithm is used to divide the obtained knowledge graph into communities, obtaining multiple distinct knowledge communities. Indicates the first A knowledge community This represents the total number of knowledge communities, hence In this embodiment, the community detection algorithm identified 204 knowledge communities, namely... .

[0089] When a user inputs the question "When a wind turbine experiences a low yaw motor speed fault, suspected to be caused by communication line interference, what are the corresponding handling measures?", the question is first encoded using the nomic-embed-text model, outputting a 768-dimensional question vector. [0.017289, 0.025108, -0.167471, 0.041056, ... , -0.028673]; then, semantic attribute features and topological structure features are extracted from each knowledge community to construct an initial graph feature representation of the knowledge community; the initial graph feature representation is processed using a graph neural network, and the topological semantic associations within the community are captured through message passing between nodes and neighborhood feature aggregation, and global pooling is performed on the aggregated features to generate a 768-dimensional community vector representing the overall features of the knowledge community. Indicates the first Community vector of a knowledge community.

[0090] This embodiment sets a set of four differentiated similarity measures. ,in This indicates the k-th similarity measure, where K = 4 represents the number of similarity measures used. The measures described in this embodiment include cosine similarity, Euclidean similarity, Manhattan similarity, and dot product similarity. Different similarity measures characterize the relationship between the question vector and the knowledge community vector from different perspectives, thereby forming a multi-angle and multi-level relationship result.

[0091] For the k-th similarity measure Calculate the problem vectors respectively Regarding the similarity between the question vector and the knowledge community vector, the results of the similarity calculation of the question vector and the knowledge community vector under different similarity measures in this embodiment are shown in Table 1.

[0092] Table 1 shows the set of similarity results for the k-th measure. Based on similarity, select the top-ranked similarity values ​​under the k-th measure, from largest to smallest. The communities are selected as the candidate community set under the k-th measure; the measure is obtained. The corresponding candidate community set is ,measure The corresponding candidate community set is ,measure The corresponding candidate community set is ,measure The corresponding candidate community set is Then, a joint candidate community was further obtained. .

[0093] Table 1. Partial results of similarity calculation between question vectors and knowledge community vectors under different similarity measures.

[0094]

[0095] The similarity results within the candidate community set corresponding to the k-th measure are normalized to obtain the similarity reliability distribution of the k-th measure defined on the joint candidate community set. The similarity reliability distributions of all measures are obtained in turn as shown in Table 2.

[0096] Table 2. Distribution of similarity reliability of different similarity measures on the joint candidate knowledge community set.

[0097]

[0098] For the similarity reliability distribution defined by the k-th similarity measure on the joint candidate community set, knowledge communities with a reliability greater than 0 are first selected as effective communities. Information entropy is calculated for the effective communities, and the dispersion index is obtained through maximum entropy normalization. Simultaneously, the difference between the maximum and second-highest reliability is calculated to obtain the discrimination strength index. An interval adjustment coefficient is calculated based on the dispersion index and the discrimination strength index, and a discount interval for the k-th measure is determined within a preset interval of 0.01–0.5. Then, a linear mapping is performed based on the reliability values ​​of each knowledge community under the k-th measure, generating corresponding discount factors within this discount interval. Finally, the reliability discount factors corresponding to the joint candidate knowledge communities under each similarity measure are obtained, as shown in Table 3.

[0099] Table 3. Discount Factors for Knowledge Community Reliability under Various Similarity Measures

[0100]

[0101] For the k-th measure, the reliability values ​​corresponding to each knowledge community in the similarity reliability distribution are discounted item by item according to the discount factor, and the reliability distribution results after community reliability discount are shown in Table 4.

[0102] Table 4. Distribution of similarity reliability after community reliability discount for each similarity measure.

[0103]

[0104] Using the Dempster-Shafer combination rule, the four similarity reliability distributions after discounting were fused together to obtain the comprehensive reliability distribution results, as shown in Table 5.

[0105] Table 5. Distribution and Ranking of Comprehensive Reliability of Knowledge Communities Based on Dempster-Shafer Combination Rules

[0106]

[0107] The obtained comprehensive reliability distribution is quantified using an approximate probability function and used for decision-making to obtain the quantified reliability. The reliability results are then ranked, and the top-ranked results are selected. As a target knowledge community with a high degree of relevance to current industrial equipment operation and maintenance issues, the results are shown in Table 6. Based on the target knowledge community and its quantitative reliability value, knowledge content related to the problem is extracted from the corresponding industrial equipment operation and maintenance knowledge graph to construct a knowledge graph context for input to the industrial equipment operation and maintenance large language model.

[0108] Table 6. Quantitative Reliability Values ​​and Ranking Results of the Target Knowledge Communities

[0109]

[0110] Using the obtained industrial equipment operation and maintenance question-and-answer samples, the parameters of the large language model are fine-tuned to obtain the industrial equipment operation and maintenance large language model. In this embodiment, the baseline large model used is the Qwen2-7B model, and LoRa is used to fine-tune the model parameters with a fine-tuning learning rate of [missing information]. The process involves 30 rounds. The previously obtained knowledge graph context, along with the user-input question, is then input into the industrial equipment operation and maintenance large language model. The large model infers and generates a natural language answer to the corresponding industrial equipment operation and maintenance question. In this embodiment, the large model infers and generates the answer to the user-input question as follows: When encountering a low speed of the wind turbine yaw motor and suspecting it's due to communication line interference, the following measures can be taken: 1. Inspection and Cleaning: First, ensure all connections are secure and not loose, and that the cables are not physically damaged. Sometimes, slight bending or squeezing can affect signal transmission. 2. Cable Replacement: If interference is suspected to be in the communication line itself, try replacing it with a new communication cable for testing. The new cable should meet the manufacturer's recommended standards and specifications. 3. Shielding: Ensure all cables connected to the motor have good shielding and are properly grounded during installation. This helps reduce the impact of external electromagnetic interference. 4. Adding Filters or Isolators: Adding EMI (electromagnetic interference) filters or signal isolators to the communication lines can effectively reduce the impact of external interference on system performance. 5. Check and adjust motor parameters: Sometimes, improper setting of yaw motor operating parameters can also lead to abnormal speed. Check and adjust the motor's control parameters appropriately to ensure it operates within the expected operating range. 6. Software update: If there are known software problems or compatibility issues with the control system, try updating to the latest version of the control system software, which may include optimizations to communication protocols and interference handling. 7. Professional testing and diagnosis: If the above measures cannot solve the problem, it is recommended to contact a professional maintenance team for on-site inspection. They can use specialized testing equipment to more accurately locate the problem and provide targeted solutions. 8. Recording and analysis: Record all operating steps, results, and any factors that may affect system performance in detail throughout the process. This helps in subsequent problem investigation and prevention. Compared to directly inputting the problem into the baseline large model Qwen2-7B, the answer obtained is: When dealing with a low yaw motor speed fault in a wind turbine, especially when the problem is suspected to be caused by communication line interference, the following steps can be taken for troubleshooting and resolution: 1. Confirm the fault phenomenon: First, it is necessary to clarify the actual operating status of the motor. Check whether the motor is really experiencing a low speed, which can be confirmed through the monitoring system or field instruments. 2. Inspect the communication lines: Physical inspection: Ensure all connected cables are free from physical damage and check that joints are secure. Signal testing: Use an oscilloscope or specialized tools to test the communication lines for interference, short circuits, or open circuits. Pay particular attention to the stability of data transmission. 3. Isolation method for troubleshooting: Temporarily disconnect potentially affected devices from suspected interference sources and observe if the fault disappears. Gradually reconnect potential interference sources and compare the recurrence of the fault to pinpoint the specific source of interference.4. Use shielded cables: If it is confirmed that the communication line is affected by external electromagnetic interference, try replacing or adding shielded communication cables. Ensure that all signal lines use cables with good shielding, and avoid running parallel to or crossing other high-voltage lines during wiring. 5. Software adjustments: Check if the parameter settings in the control system are reasonable, especially those related to motor control. Adjust the communication protocol and baud rate to ensure compatibility with the equipment and optimize data transmission efficiency. 6. Professional testing and maintenance: If the above methods fail to resolve the issue, further diagnosis may be necessary from a professional electrical engineer or technical support from the manufacturer. They may use more advanced testing tools to identify specific interference sources and provide targeted solutions. Comparative results show that the knowledge graph-based contextual fusion-based large-scale model reasoning method for equipment operation and maintenance can generate more practical operation and maintenance suggestions when answering questions about industrial equipment fault handling. Compared to the baseline large-scale model, its generated content not only covers fault diagnosis steps but also provides clear engineering-oriented handling measures, demonstrating stronger domain knowledge consistency and engineering feasibility, effectively enhancing the model's application value in industrial equipment operation and maintenance decision support scenarios.

Claims

1. A method for intelligent question answering of equipment operation and maintenance based on context fusion on a knowledge graph, characterized in that, Includes the following steps: S1: Construct a knowledge-text dataset for the field of industrial equipment operation and maintenance; S2: Construct an industrial equipment operation and maintenance knowledge graph based on a knowledge text dataset, and divide the knowledge graph into communities to obtain multiple knowledge communities that represent different operation and maintenance knowledge units; S3: Vectorize the user-inputted question and the knowledge community respectively to obtain the question vector and the community vector; S4: Define a differentiated set of similarity metrics , Let K be the k-th similarity measure, and K be the number of similarity measures; based on Calculate the problem vector With knowledge community vector Similarity between them; For the set of similarity results for the k-th measure, select the top similarity results under the k-th measure based on similarity ranking. The communities defined by all measures are used as the candidate community set; the union of all candidate community sets by all measures is taken to form the joint candidate community set. , The complete set of mathematical subsets constitutes power set ; Normalize the similarity results within the candidate community set corresponding to the k-th measure to obtain the similarity reliability distribution of the k-th measure defined on the joint candidate community set. ; S5: Adaptively adjust the similarity reliability distribution using a community reliability discounting strategy, as follows: S5.1: Knowledge communities with a confidence value greater than 0 in the similarity confidence distribution are considered valid communities. The dispersion index of valid communities is calculated. The difference between the maximum and second-highest reliability of the effective community is used to obtain the distinguishing strength index. ; Will and Perform a weighted summation to construct the interval adjustment coefficient corresponding to the k-th measure. Within the preset discount range The discount interval for the k-th measure is determined based on the interval adjustment coefficient. ; Within the discount interval of the k-th measure, a linear mapping is performed based on the reliability values ​​of each knowledge community under the k-th measure to obtain the knowledge community under the k-th measure. Corresponding discount factor Finally, the discount parameter vector under the k-th measure is obtained. ; S5.2: Based on the discount parameter vector For the similarity reliability distribution The reliability values ​​of each knowledge community in the text are discounted item by item to obtain the reliability distribution. ; S6: Using the combination rule in DS evidence theory, fuse the K similarity reliability distributions after community reliability discounting to obtain a comprehensive reliability distribution. After quantification using an approximate probability function, ... The reliability results of single-point subsets are sorted, and the top-ranked ones are selected. As a target knowledge community; Based on the target knowledge community, knowledge content related to the problem is extracted from the corresponding industrial equipment operation and maintenance knowledge graph to construct a knowledge graph context for input to the industrial equipment operation and maintenance large language model. The knowledge graph context and the user's input question are fed into the industrial equipment operation and maintenance language model, and after reasoning, a natural language answer to the corresponding industrial equipment operation and maintenance question is generated.

2. The intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion according to claim 1, characterized in that, The specific implementation process of step S1 is as follows: S1.1: Unified collection of unstructured or semi-structured operation and maintenance data from different sources is carried out through document parsing and text extraction technology: raw text data is obtained from the maintenance record database, fault report document library and equipment operation and maintenance standard documents related to industrial equipment operation and maintenance. S1.2: The original text data is cleaned, segmented, and semantically standardized to eliminate noise and unify the mapping of professional terms, resulting in a standardized operation and maintenance knowledge text dataset. S1.3: Extract knowledge entries from the operation and maintenance knowledge text dataset and construct an industrial equipment operation and maintenance question and answer sample set containing multiple question-answer pairs.

3. The intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion according to claim 1, characterized in that, The specific implementation process of step S2 is as follows: S2.1: The knowledge text dataset of industrial equipment operation and maintenance obtained in S1 is divided into text units. Based on named entity recognition and relation extraction technology, the equipment objects, fault phenomena, fault locations and handling measures are identified as entity nodes from the knowledge text dataset. At the same time, the causal relationship and handling dependency relationship between entity nodes are identified to construct structured graph data and obtain the knowledge graph of industrial equipment operation and maintenance. S2.2: For the industrial equipment operation and maintenance knowledge graph, based on the relationships between entity nodes, the Leiden graph community discovery algorithm is used to divide the knowledge graph into communities, obtaining multiple distinct knowledge communities; using... Indicates the first A knowledge community This indicates the total number of knowledge communities.

4. The intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion according to claim 1, characterized in that, The specific implementation process of step S3 is as follows: S3.1: Receive user input of industrial equipment operation and maintenance questions, encode the questions using a text embedding model, and generate question vectors to represent the semantic features of the questions. ; S3.2: For each knowledge community, extract the semantic attribute features and topological structure features within the knowledge community, and construct the initial graph feature representation of the knowledge community; Semantic attribute features include a node feature matrix generated based on entity descriptions within the community, and topological features include adjacency information representing the connection relationships between entities within the community. S3.3: A graph neural network is used to process the initial graph feature representation. Through message passing between nodes and neighborhood feature aggregation, topological semantic relationships within the community are captured. Global pooling is then performed on the aggregated features to generate a community vector representing the overall characteristics of the knowledge community. Indicates the first Community vector of a knowledge community.

5. The intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion according to claim 4, characterized in that, In step S4, the similarity reliability distribution of the k-th measure defined on the joint candidate community set is obtained. Specifically, for knowledge communities belonging to the candidate community set of the kth measure, their reliability value is determined by the normalized similarity result; for knowledge communities belonging to the joint candidate community set but not to the candidate community set of the kth measure, their reliability value is set to 0; and the similarity reliability distribution of all measures is obtained in turn.

6. The intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion according to claim 5, characterized in that, In step S5.1, the dispersion index The calculation is as follows: The information entropy of the reliability distribution is calculated based on the proportion of the reliability value of the effective community in the total reliability of the effective community, and the dispersion index is obtained by normalization through maximum entropy. ; The discount range ,in equal , equal and The summation result; The knowledge community under the k-th measure Corresponding discount factor equal and The summation result, This represents the knowledge community in the joint candidate community set under the k-th measure. The corresponding reliability value, This indicates the knowledge community ID in the set of joint candidate communities.

7. The intelligent question-answering method for equipment operation and maintenance based on knowledge graph context fusion according to claim 6, characterized in that, In step S5.2, the similarity reliability distribution is... The reliability values ​​of each knowledge community in the text are discounted item by item to obtain the reliability distribution after community reliability discounting. The discount process is implemented using set transformation operators, specifically: for any subset The community reliability discounted reliability value is composed of all subsets. The original reliability values ​​are obtained by weighted summation, where right The original reliability values ​​are weighted by the discount kernel function. Control; discount kernel function based on and The set inclusion relationship between the sets is used to multiply and combine the discount factor and the retention coefficient of the corresponding reliability of the knowledge community to achieve the weighted propagation of community-level reliability in the set space.

8. A knowledge graph-based context fusion-based intelligent question-and-answer system for equipment operation and maintenance, used to implement the intelligent question-and-answer method for equipment operation and maintenance as described in any one of claims 1 to 7, characterized in that, Includes the following modules: The knowledge text module is used to build a knowledge text dataset in the field of industrial equipment operation and maintenance. The knowledge community module constructs an industrial equipment operation and maintenance knowledge graph based on a knowledge text dataset, and divides the knowledge graph into communities to obtain multiple knowledge communities that represent different operation and maintenance knowledge units. The question vector and community vector modules are used to vectorize the user-input questions and knowledge communities, respectively, to obtain question vectors and community vectors. The similarity reliability distribution module is used to calculate the diverse associations between question vectors and community vectors through differentiated similarity metrics, and to generate similarity reliability distributions that reflect different degrees of association. The adaptive adjustment module is used to adaptively adjust the similarity reliability distribution through a community reliability discounting strategy; The industrial equipment operation and maintenance question answering module is used to fuse adaptively adjusted similarity reliability distributions, construct a knowledge graph context input to the industrial equipment operation and maintenance large language model, and generate answers to industrial equipment operation and maintenance questions.

Images