Equipment operation and maintenance knowledge large model construction method and intelligent question answering system and equipment

Abstract

The invention relates to the technical field of artificial intelligence and new-generation information, in particular to an equipment operation and maintenance knowledge large model construction method and an intelligent question answering system and equipment. The method comprises the following steps: S1, obtaining a question description text input by a user; s2, performing semantic analysis and key information extraction on the question description text input by the user to obtain key information; s3, matching corresponding knowledge nodes and triads containing the knowledge nodes from the equipment maintenance knowledge graph based on the key information, and adding the triads into a candidate triad set; s4, calculating a semantic score of each triad in the candidate triad set through a deep semantic perception embedding model, and sorting the triads in the candidate triad set based on the semantic scores; and S5, generating an answer to the question based on the candidate triple set by using a language large model. According to the invention, the accuracy and reliability of equipment operation and maintenance knowledge answering are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and new generation information technology, in particular to a device operation and maintenance knowledge large model construction method, an intelligent question and answer system and a device. BACKGROUND

[0002] The smooth operation of modern air transportation industry highly depends on the huge and complex airport baggage handling system (BHS). As one of the core infrastructures of airport ground services, BHS undertakes the key task of efficient and accurate sorting and transfer of passenger baggage. However, BHS is composed of tens of thousands of mechanical and electrical equipment (such as conveyor belts, sorting machines, scanners, turntables, etc.), which are in long-term high-intensity and high-load operation state, and equipment failure is inevitable. Once a failure occurs, it may not only cause baggage handling delay and misdelivery, but also may cause large-scale flight delays, resulting in huge economic losses and reputation risks. Therefore, to realize efficient, accurate and proactive maintenance of BHS equipment, to minimize the failure rate and shorten the maintenance response time, has become one of the core challenges to ensure the efficient operation of the airport.

[0003] Currently, the maintenance management of BHS equipment mainly relies on preventive maintenance plans, sensor data-based condition monitoring and maintenance engineers' accumulated domain experience knowledge. Although these methods have played an important role in ensuring system operation, they still face significant challenges: (1) Fragmentation and implicitness of knowledge: equipment failure modes, maintenance experience, best practices and other knowledge are scattered in different documents, historical work order records and individual engineers' experience, lacking systematic and structured integration and sedimentation, making it difficult to effectively share, reuse and inherit. (2) Insufficient complex correlation mining: BHS equipment failure is not an isolated event, but may be triggered by multiple factors (such as aging of specific components, operation under specific working conditions, linkage failure with other equipment, etc.). Existing maintenance knowledge management methods are difficult to effectively capture the deep and multi-dimensional logical correlations between equipment, failure, maintenance operation and working conditions. (3) Insufficient human-machine collaboration adaptability: existing maintenance support systems rely too much on professional engineers' domain knowledge, lack of interactive guidance mechanism for non-professional technical personnel, resulting in low knowledge transfer efficiency and delayed emergency response.

[0004] With the breakthrough of artificial intelligence technology, the intelligent question and answer system based on knowledge graph has become the core solution to improve the operation and maintenance efficiency. Through structured knowledge modeling and semantic reasoning, it can quickly respond to complex problems such as equipment fault diagnosis and maintenance process query. However, the adaptability of existing knowledge graph construction methods in the field of equipment operation and maintenance is insufficient, resulting in significant defects in the accuracy and reliability of the question and answer system. First, equipment operation and maintenance events usually involve multi-dimensional associated information, but existing knowledge graph construction methods lack fine-grained splitting of event arguments, resulting in fragmentation of semantic association in the knowledge graph, and the question and answer system is prone to information omission or incorrect association when matching complex problems. Second, there are a large number of polysemy phenomena in the field of equipment operation and maintenance, and traditional knowledge graph construction relies on manual annotation or simple rule matching, making it difficult to dynamically identify context semantics. Finally, there are a large number of implicit associations in equipment operation and maintenance knowledge, but existing methods are mostly based on explicit relationship extraction, ignoring the depth aggregation of neighbor node semantics.

[0005] The above problems restrict the landing effect of existing question and answer systems in the equipment operation and maintenance scenario, and there is an urgent need for a more fine-grained, dynamic and semantic-aware knowledge graph construction method to improve the accuracy and reliability of operation and maintenance knowledge question and answer. SUMMARY

[0006] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0007] In order to solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0008] The device operation and maintenance knowledge large model construction method comprises:

[0009] S1: obtaining a problem description text input by a user;

[0010] S2: performing semantic analysis and key information extraction on the problem description text input by the user to obtain key information;

[0011] S3: matching corresponding knowledge nodes and triples containing the knowledge nodes from the equipment maintenance knowledge graph based on the key information, and adding the triples to a candidate triple set; wherein the equipment maintenance knowledge graph is constructed based on a five-tuple model, and the five-tuple model splits event text information into five types of event arguments, including event subject, subject action, subject attribute, subject state description and maintenance behavior;

[0012] S4: calculating a semantic score of each triple in the candidate triple set through a deep semantic-aware embedding model, and ranking the triples in the candidate triple set based on the semantic score; wherein the deep semantic-aware embedding model aggregates neighbor node information in multiple semantic subspaces through a double attention mechanism, capturing potential semantic dependencies and polysemy representation in the device-maintained knowledge graph;

[0013] S5: generating the question answer based on the candidate triple set by using a language large model.

[0014] Preferably, in step S3, the processing steps for constructing the device-maintained knowledge graph are as follows:

[0015] S301: obtaining an event data set containing a plurality of event text information;

[0016] S302: splitting the event text information into event arguments by a five-tuple model, establishing event relationships between each two event arguments, and obtaining a plurality of event relationship triples; the five-tuple model splits the event text information into five types of event arguments, i.e., event subject, subject action, subject attribute, subject state description, and maintenance behavior;

[0017] S303: generating logical relationships between each two event arguments by a device-maintenance ontology, and obtaining a plurality of logical relationship triples;

[0018] S304: performing steps S302 and S303 for each event text information in the event data set, and inputting all the obtained event relationship triples and logical relationship triples into a graph database to generate the device-maintenance knowledge graph.

[0019] Preferably, in step S302, the event relationships include:

[0020] 1) the relationship between the fault cause event and the fault phenomenon event: Cause;

[0021] 2) the relationship between the maintenance method event and the fault phenomenon event: Solve;

[0022] 3) the relationship between the preventive measure event and the fault phenomenon event: Prevent;

[0023] In step S303, the logical relationships defined by the device-maintenance ontology include:

[0024] 1) the device hierarchical structure relationship: Has;

[0025] 2) the fault evolution logical relationship: Cause, Solve, Prevent;

[0026] 3) the maintenance decision rule relationship: Occur, Take;

[0027] 4) Similar relationship: Similar to.

[0028] The device operation and maintenance intelligent question and answer system is implemented based on a device operation and maintenance knowledge large model construction method, and includes the following steps:

[0029] A question input module is configured to obtain a question description text input by a user;

[0030] A question analysis module is configured to perform semantic analysis and key information extraction on the question description text input by the user to obtain key information;

[0031] An entity matching module is configured to match corresponding knowledge nodes and triples containing the knowledge nodes from a device maintenance knowledge graph based on the key information, and add the triples to a candidate triple set;

[0032] A semantic score calculation module is configured to calculate a semantic score of each triple in the candidate triple set by using a deep semantic perception embedding model, and sort the triples in the candidate triple set based on the semantic score;

[0033] An answer generation module is configured to generate a question answer based on the candidate triple set by using a language large model after expert-assisted decision-making on the sorted candidate triple set.

[0034] A computer device includes one or more processors.

[0035] The processor is configured to store one or more programs.

[0036] When the one or more programs are executed by the one or more processors, the device operation and maintenance knowledge large model construction method is implemented.

[0037] Compared with the prior art, the device operation and maintenance knowledge large model construction method and the intelligent question and answer system have the following beneficial effects:

[0038] The device maintenance knowledge graph constructed by the five-tuple model (event subject, action, attribute, state, maintenance behavior) can systematically disassemble the complex semantic relationship in the device operation and maintenance event. When matching the key information, the five-tuple structure can accurately locate the corresponding event node and its associated triple in the knowledge graph, avoiding the matching omission or error caused by semantic ambiguity in the traditional method, and the construction of the candidate triple set covers multi-dimensional knowledge related to the question, providing a more complete information basis for answer generation. At the same time, the five-tuple model expands the traditional triple (fault phenomenon, relationship, solution measure) to a fine-grained five-tuple, which not only improves the expression ability of knowledge granularity, but also enhances the multi-hop reasoning ability and context modeling ability of the knowledge graph, providing a high-quality structured semantic foundation for subsequent knowledge fusion and semantic calculation, which can improve the modeling ability and knowledge expression precision of the intelligent question and answer model for the fine-grained logical structure contained in the complex maintenance event, support accurate maintenance guidance, and improve the accuracy and reliability of the device operation and maintenance knowledge answer.

[0039] The application designs a deep semantic perception based embedding model (DSPE), which effectively fuses text semantic information, graph structure information and attribute features, solves the problems of inconsistent representation and low information integration efficiency of traditional models in processing heterogeneous relationships and multi-source features. The model aggregates neighbor node information in multiple semantic subspaces through a double attention mechanism, which can capture the potential semantic dependence between triples in the knowledge graph. By calculating the semantic score and sorting, the device operation and maintenance knowledge question and answer system can preferentially output the triple (such as the maintenance behavior record with high correlation) most matched with the problem semantics, avoid interference of low value information, and further improve the accuracy and reliability of the device operation and maintenance knowledge answer. At the same time, the model realizes high-quality initialization of entity attribute embedding by introducing a pre-trained language model and a contrast learning mechanism, which can more fully excavate the semantic features of nodes; in terms of local semantic modeling, a cross-attention mechanism encoder based on Transformer is constructed to jointly model the head entity and relationship in the triple, focusing on the tail node prediction task; the multi-head cross-attention enhances the model's ability to distinguish between heterogeneous relationship types, improves the accuracy of relationship reasoning in different semantic environments, and significantly improves the comprehensive performance of the model in knowledge graph completion, fault cause analysis and intelligent recommendation tasks.

[0040] The application generates an answer to a question based on a candidate triple set through a large language model. The large language model can convert discrete triple information into natural language text through pre-training language knowledge, and the answer generated by the language model is more in line with human expression habits than directly splicing triples, thereby improving the effectiveness of device operation and maintenance knowledge answers. At the same time, the large language model can combine the multi-hop relationship in the candidate triple to perform logical reasoning and generate a complete answer containing step-by-step instructions or suggestions. In addition, when the recommended answer generated by the deep semantic perception embedding model (DSPE) is not in the current knowledge graph, the new triple can be dynamically completed into the knowledge graph after the expert auxiliary judgment is reasonable, and the large language model does not need to be retrained to adjust the answer content based on the latest triple information. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to make the objectives, technical solutions and advantages of the present application clearer, the following will further describe the present application with reference to the accompanying drawings, in which:

[0042] Figure 1 A logical block diagram of a device operation and maintenance knowledge model construction system.

[0043] Figure 2 An example diagram of a device maintenance knowledge graph.

[0044] Figure 3 A logical block diagram of a deep semantic perception embedding model. DETAILED DESCRIPTION

[0045] In order to make the objectives, technical solutions and advantages of the present application clearer, the following will further describe the present application with reference to the accompanying drawings, in which:

[0046] It should be noted that similar reference numerals and letters refer to like items in the accompanying drawings, and once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings. In the description of the application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", and the like, indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the application is usually placed during use, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In addition, the terms "first", "second", "third", and the like are only used to distinguish the description and cannot be understood as indicating or implying relative importance. In addition, the terms "horizontal", "vertical", and the like do not mean that the components must be absolutely horizontal or vertical, but can be slightly inclined. For example, "horizontal" only means that its direction is more horizontal relative to "vertical", and does not mean that the structure must be completely horizontal, but can be slightly inclined. In the description of the application, it should also be noted that unless otherwise specifically defined and limited, the terms "provided", "mounted", "connected", "linked", and the like should be broadly understood, for example, they can be fixedly connected, or detachably connected, or integrally connected; can be mechanically connected, or electrically connected; can be directly connected, or indirectly connected through an intermediate medium; can be internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances.

[0047] The specific embodiments are further described in detail below:

[0048] Embodiment I:

[0049] The embodiment discloses a device operation and maintenance knowledge large model construction method.

[0050] As shown in the figure, the device operation and maintenance knowledge large model construction method comprises: Figure 1

[0051] S1: obtaining a problem description text input by a user;

[0052] S2: performing semantic analysis and key information extraction on the problem description text input by the user to obtain key information;

[0053] The key information includes key entities (device name, fault phenomenon, component type, etc.), relationships (reason, solution, composition, etc.), and user intent (diagnosis, query, guidance, etc.);

[0054] ​S3: match the corresponding knowledge nodes and triples containing the knowledge nodes from the device maintenance knowledge graph based on the key information, and add the triples to the candidate triple set;

[0055] S4: calculate the semantic score of each triple in the candidate triple set through the deep semantic perception embedding model, and sort the triples in the candidate triple set based on the semantic score;

[0056] S5: after expert-assisted decision-making on the sorted candidate triple set, use a large language model to generate an answer to the question based on the candidate triple set.

[0057] In this embodiment, the process of generating an answer to the question based on the candidate triple set using a large language model (such as GPT, LLaMA, etc.) is implemented through the following steps: first, convert the structured knowledge in the candidate triple set into intermediate text described in natural language (which can be combined with template filling); then concatenate the intermediate text with the user's original question as the input sequence of the language model; finally, based on the pre-trained semantic understanding and generation ability of the language model, automatically complete the logical connection words, adjust the sentence structure, and output a complete answer that conforms to human expression habits.

[0058] The device maintenance knowledge graph constructed by the five-tuple model (event subject, action, attribute, state, and maintenance behavior) can systematically disassemble the complex semantic relationships in device operation and maintenance events. When matching key information, the five-tuple structure can accurately locate the corresponding event nodes and their associated triples in the knowledge graph, avoiding the matching omissions or errors caused by semantic ambiguity in traditional methods, and the construction of the candidate triple set covers multi-dimensional knowledge related to the question, providing a more complete information foundation for answer generation. At the same time, the five-tuple model expands the traditional triple (fault phenomenon, relationship, and solution measure) to a fine-grained five-tuple, not only improving the expression ability of knowledge granularity, but also enhancing the multi-hop reasoning ability and context modeling ability of the knowledge graph, providing a high-quality structured semantic foundation for subsequent knowledge fusion and semantic calculation, which can improve the modeling ability and knowledge expression precision of the intelligent question and answer model for the fine-grained logical structure contained in complex maintenance events, support accurate maintenance guidance, and thus improve the accuracy and reliability of the device operation and maintenance knowledge answer.

[0059] The application designs a deep semantic perception based embedding model (DSPE), effectively fusing text semantic information, graph structure information and attribute characteristics, solving the problems of inconsistent representation and low information integration efficiency of traditional models in processing heterogeneous relationships and multi-source characteristics. The model aggregates neighbor node information in multiple semantic subspaces through a double attention mechanism, which can capture the potential semantic dependence between triples in the knowledge graph. By calculating the semantic score and sorting, the device operation and maintenance knowledge question and answer system can preferentially output the triple (such as the maintenance behavior record with high correlation) most matched with the problem semantics, avoiding interference from low-value information, and further improving the accuracy and reliability of the device operation and maintenance knowledge answer. At the same time, the model realizes high-quality initialization of entity attribute embedding by introducing a pre-trained language model and a contrast learning mechanism, which can more fully excavate the semantic characteristics of nodes; in terms of local semantic modeling, a cross-attention mechanism encoder based on Transformer is constructed to jointly model the head entity and the relationship in the triple, focusing on the tail node prediction task; the multi-head cross-attention enhances the model's ability to distinguish between heterogeneous relationship types, improves the accuracy of relationship reasoning in different semantic environments, and significantly improves the comprehensive performance of the model in knowledge graph completion, fault cause analysis and intelligent recommendation tasks.

[0060] The language large model can convert discrete triple information into natural language text through pre-trained language knowledge, and the answer generated by the language model is more in line with human expression habits than directly splicing triples, thereby improving the effectiveness of the device operation and maintenance knowledge answer. At the same time, the language large model can combine the multi-hop relationship in the candidate triple to perform logical reasoning and generate a complete answer containing step-by-step instructions or suggestions. In addition, when the recommended answer generated by the deep semantic perception based embedding model (DSPE) is not in the current knowledge graph, after expert assistance determines that it is reasonable, the new triple can be dynamically completed to the knowledge graph, and the large language model can adjust the answer content based on the latest triple information without retraining.

[0061] In order to better introduce the technical scheme of the application, the embodiments are described through the following several parts.

[0062] I. Device maintenance knowledge graph

[0063] In this embodiment, the processing steps for constructing the device maintenance knowledge graph are as follows:

[0064] S301: Obtain an event data set containing a plurality of event text information;

[0065] S302: Split the event text information into event arguments by a five-tuple model, establish the event relationship between each two event arguments, and obtain a plurality of groups of event relationship triples;

[0066] S303: generating logical relations between each two event arguments by the device maintenance ontology, obtaining several groups of logical relation triples;

[0067] S304: executing steps S302 and S303 for each event text information in the event data set, inputting all the event relation triples and logical relation triples obtained into a Neo4j graph database to generate a device maintenance knowledge graph.

[0068] 1. Five-tuple model

[0069] The five-tuple model splits the event text information into five types of event arguments, i.e., event subject (such as device, component), subject action (such as occurrence, taking), subject attribute (such as rated power, material), subject state description (such as temperature being too high), and maintenance behavior (such as replacement, inspection);

[0070] The event relations include:

[0071] 1) the relation between the fault cause event and the fault phenomenon event: Cause;

[0072] 2) the relation between the repair method event and the fault phenomenon event: Solve;

[0073] 3) the relation between the preventive measure event and the fault phenomenon event: Prevent.

[0074] In this embodiment, the five-tuple model can be obtained by training a neural network, and the trained model can automatically split the input event text information into five types of event arguments and generate corresponding event relations. Based on the splitting logic of the five-tuple model, the event text information can also be manually split into five types of event arguments and set corresponding event relations.

[0075] Figure 2 For a detailed example, each event constitutes a maintenance evolution and disposal logic chain through the following semantic relations. The fault cause event and the fault phenomenon event are causally related, with a relation of "Cause". The handling method event and the fault phenomenon event are disposal-related, with a relation of "Solve". The preventive measure event and the fault phenomenon event are prevention-related, with a relation of "Prevent". Meanwhile, the fault event has a dynamic conversion characteristic, and the same event can have both fault cause and fault phenomenon attributes in different scenarios. For example, in event case 1 "the temperature of the AC motor is too high, causing the belt to slip", the temperature of the AC motor is a fault cause event, and the belt slipping is a fault phenomenon event. In event case 2 "belt slipping is easy to cause belt breakage", the belt slipping is a fault cause event. Using event splitting can deeply mine such causal relations.

[0076] 2. Equipment maintenance unit

[0077] The logical relationships defined in the equipment maintenance entity definition include:

[0078] 1) Equipment hierarchical structure: Has (inclusive); The equipment hierarchical structure includes five levels of nodes: system, equipment category, equipment, component, and part;

[0079] 2) Fault evolution logic relationship: Cause, Solve, Prevent; the fault evolution logic relationship is used to describe the generation and propagation process of equipment faults;

[0080] 3) Maintain decision rule relationships: Occur (occurs), Take; Maintain decision rule relationships are established by combining entity attributes (such as temperature, material), state (such as excessive temperature) and maintenance behavior (such as periodic inspection).

[0081] 4) Similarity relation: Similar to; To enhance the semantic reasoning ability of knowledge graphs, the "Similar to" predicate relation is introduced in the device maintenance manual to connect nodes that are semantically similar but not completely equivalent, thereby strengthening the connection strength between semantically similar entities in the graph structure and supporting semantic similarity reasoning and knowledge completion in the graph structure.

[0082] Combination Figure 2 As shown, after defining event arguments and event associations, it is necessary to determine the parameters of the event arguments, including category classification and relationship classification. Four categories are defined in the equipment maintenance ontology layer: maintenance equipment hierarchy structure class, attribute class, state class, and behavior class. The equipment hierarchy structure includes system, equipment category, equipment, component, and part, with the hierarchical relationship being "Has". Next, the relationships between each conceptual class need to be defined: the relationship between the equipment hierarchy structure and the attribute class is "Has", the relationship with the state class is "Occur", the relationship with the behavior class is "Take", and the relationship between the attribute class and the state class is also "Occur". The equipment maintenance ontology model was constructed using the Protege tool.

[0083] II. Deep Semantic Aware Embedding Model (DSPE)

[0084] like Figure 3 As shown, the processing steps of the deep semantic-aware embedding model include:

[0085] S401: Input the triples from the candidate triples set;

[0086] S402: Using the BERT model, attribute embedding is performed on all entities and relations in the equipment maintenance knowledge graph to obtain the entity embedding matrix E0∈R. N×d And relation embedding matrix R0∈Rr×d , N is the number of all entities, d is the dimension of the initial embedding, and r is the number of all relationship categories; wherein the knowledge nodes in the device maintenance knowledge graph are entities, and the edges between the knowledge nodes are relationships;

[0087] S403: Multi-dimensional latent semantic mapping: a mapping function f is constructed to map each entity in the entity embedding matrix E0into multiple different semantic spaces, to obtain the entity embedding matrix E map(y) of each semantic subspace;

[0088] The formula is:

[0089] E map(y) =f(E0);

[0090] f(E0)=σ(M y ·E0+b y );

[0091] In the formula: E map(y) ∈R N×d represents the entity embedding matrix of the yth semantic subspace; M y ∈R d×d represents the mapping matrix of the yth semantic subspace; b y represents the bias term; σ(·) represents the activation function; in order to maximize the integrity of the mapping semantic subspace, E map(y) and E0have the same embedding dimension;

[0092] S404: First-level attention mechanism: the entity embedding matrix E map(y) of each semantic subspace is input into the stacked multi-layer graph attention network, and by focusing on the importance of different neighbor nodes to the center node, neighbor information more representative or influential to the current node is identified and strengthened, to obtain the enhanced entity embedding matrix

[0093] S405: Second-level attention mechanism: based on the attention mechanism of multiple semantic subspaces, the enhanced entity embedding matrix of each semantic subspace in the lth layer graph attention network is calculated, to obtain the updated entity embedding matrix E l+1 of the (l+1)th layer graph attention network;

[0094] S406: Local semantic perception mechanism: for a triple (h, r, t) in the candidate triple set, the entity embedding vector and the relationship embedding vector are obtained from the entity embedding matrix E l+1 and the relationship embedding matrix R0, and the corresponding output representation matrix H out is calculated through the cross-attention mechanism and the feedforward layer;

[0095] S407: The output of the triplet (h, r, t) is represented by the matrix H through the decoder out is decoded to obtain the semantic score of the triplet (h, r, t);

[0096] The formula is expressed as:

[0097] score=H out ·e t ;

[0098] In the formula: e t ∈R d is the embedding representation of the tail node of the triplet (h, r, t).

[0099] 1. Attribute semantic embedding initialization

[0100] Specifically, the following method is used to generate attribute semantic embeddings of device maintenance knowledge graph nodes:

[0101] S2011: For each node, two forward calculations are performed on the same node through the dropout mechanism of the BERT model to generate positive sample vectors with semantic consistency, and the representations of other nodes in the same batch are taken as negative samples. In the training process with a batch size of N, each node corresponds to 1 positive sample and N-1 negative sample combinations.

[0102] The specific steps are as follows:

[0103] 1) Use BERT to calculate node embedding:

[0104] H i =BERT(T i );

[0105] In the formula: H i ∈R L×768 (L is the length of the text sequence), T i is the entity text description;

[0106] 2) Extract the [CLS] token as the global semantic representation vector:

[0107]

[0108] In the formula: is the embedding representation of the node, H i [0] represents the 0th hidden layer output taken from the BERT model output;

[0109] 3) The method for obtaining the initial sample data set is to input the text information of the node into the BERT model twice to obtain the original node and the positive sample Input the remaining N-1 nodes into the BERT model to obtain negative samples

[0110] S2012: Calculate the semantic distance sim of the positive sample and the original node, and the negative sample and the original node, using the cosine similarity based on L2 norm, and the calculation method is:

[0111]

[0112] In the formula, u and v represent two different node embedding vectors, and d is the dimension of the vector;

[0113] S2013: The model training adopts a batch contrast learning framework, and the contrast loss function L a is used to optimize the parameters.

[0114]

[0115] In the formula, and is the feature vector of the positive sample pair generated by randomly discarding the mask generated by inputting the same node i twice into the BERT encoder, is the negative sample feature vector generated by other nodes, N is the number of all nodes in a batch, τ is the temperature coefficient, and the default value is 0.05, and the sim function is used to measure the semantic distance between different node feature vectors.

[0116] S2014: The joint classification optimization scheme of fusing BERT deep semantic encoding and Focal Loss dynamic adjustment:

[0117] The specific steps are as follows:

[0118] 1) Input the obtained node embedding vector into the classifier layer in turn:

[0119]

[0120] y i = softmax(W2z i +b2);

[0121] W1∈R 256×768 ,b1∈R 256 ;

[0122] W2∈R 256×768 ,b2∈R C (C is the number of categories, and the classification task this time is 8 categories);

[0123] 2) Gradient back propagation, using the chain rule to calculate the parameter gradient:

[0124]

[0125] where the BERT parameters adopt a hierarchical learning rate:

[0126] η BERT = α · η base · 2 -(12-l) ;

[0127] l is the Transformer layer number, α = 0.8 is the attenuation coefficient, η BERT , η base indicates that the learning rate of BERT-base-Chinese is used;

[0128] 3) Focal Loss optimization module, using multi-class Focal Loss to assist in completing the BERT fine-tuning task

[0129] Define the class balance coefficient w c :

[0130]

[0131] N c is the number of classes of sample c, N max is the number of samples of the largest class, δ = 10 is the smoothing factor;

[0132] Multi-class Focal Loss formula:

[0133]

[0134] p t = y i t is the probability of the model predicting the class, and γ is the focusing factor (the default value is 2);

[0135] S2015: During training, the parameters of the encoder are updated in reverse based on the contrastive loss and the Focal Loss loss.

[0136] Finally, the attribute embedding of the knowledge node is obtained for all entities and relations, and the entity embedding matrix E0∈R N×d and the relation embedding matrix R0∈R r×d .

[0137] 2. Multi-dimensional latent semantic mapping

[0138] Specifically, the multi-dimensional latent semantic mapping is realized by the following steps:

[0139] S2021: Introduce a mapping function f to map each entity to multiple different semantic spaces, and the mapping formula f is defined as follows:

[0140] f(E0) = σ(M y · E0 + b y );

[0141] E map(y) = f(E0).

[0142] S2022: Use Hilbert-Schmidt independence criterion (HSIC) to strengthen the independence between different semantic subspaces;

[0143] The specific steps are as follows:

[0144] 1) Define the RBF kernel function

[0145]

[0146] The a row and b row vectors of the selected subspace represent the kernel function matrix K (i) ab items. and respectively represent the embedding vectors of two different nodes of the i-th semantic subspace, and σ is the standard deviation;

[0147] 2) Calculate the centering matrix:

[0148]

[0149] I N is an N×N identity matrix, 1 N is an N-dimensional all-1 vector, and H is used to center the kernel matrix to remove the mean effect;

[0150] 3) When the number of mapped subspaces exceeds two, the formula for uniformly calculating the HSIC loss between multiple subspaces is:

[0151]

[0152] K (i) = k(E map(i) , E map(i) ), K (j) = k(E map(j) , E map(j) );

[0153] k represents the RBF kernel function, Tr(·) represents the trace operation of the matrix, that is, the sum of the diagonal elements, and H is the centering matrix, which measures the statistical independence between two random variables. This is because the semantic subspaces are not necessarily a single linear relationship, and may contain multiple nonlinear relationships.

[0154] 3, Relationship enhancement of graph structure perception

[0155] 3.1, First-level attention mechanism

[0156] The first-level attention mechanism introduces a relationship-enhanced neighbor aggregation strategy, which aggregates the neighbor nodes themselves and also takes the relationship types of the edges connected to them into account. The specific processing steps are as follows:

[0157] S4041: Calculate the entity embedding matrix E of the yth semantic subspace map(y) The entity embedding vector of any neighbor node t of the center node p in the lth layer graph attention network is multiplied by the relationship of the neighbor node t, and the dot product of the resulting result and the entity embedding vector of the center node p in the lth layer graph attention network is calculated to obtain the attention A in exponential form t ;

[0158] The formula is as follows:

[0159]

[0160] Where, for the node q in the semantic subspace y, the multiplication calculation is completed on the lth layer embedding vector of the node q and the relationship :

[0161]

[0162] S4042: Calculate the attention coefficients of the center node for each neighbor node in the yth semantic subspace in the lth layer graph attention network Get the attention coefficient sequence {α1,α2...α t-1 ,α t} of all neighbor nodes;

[0163] The formula is as follows:

[0164]

[0165] In the formula, N t represents the set of all neighbors and relationships of the center node; A q is the attention of a certain neighbor node;

[0166] S4043: Arrange the attention coefficient sequence in order of size, and select the top 15 neighbor node attention coefficients {α1,α2...α 14 ,α 15}; if the number of neighbor nodes of a certain node is less than 15, select all the neighbor nodes;

[0167] S4044: Aggregating the attention coefficient sequence of the neighbor nodes selected by the lth layer graph attention network to the center node to obtain the entity embedding vector of the center node of the (l+1)th layer graph attention network

[0168] The formula is expressed as:

[0169]

[0170] In the formula: N (t,15) represents the node and relationship set ranked in the top 15 of the neighbor subset attention coefficient; (W map(y) ) l is a trainable parameter; and σ represents a linear activation function.

[0171] S4045: Performing steps S4041 to S4044 on all center nodes in the entity embedding matrix E map(y) of the yth semantic subspace to obtain the enhanced entity embedding matrix E l of the yth semantic subspace in the lth layer graph attention network

[0172] 3.2, second-level attention mechanism

[0173] The second-level attention mechanism aggregates the information of multiple semantic subspaces for node information updating. The specific processing steps are as follows:

[0174] S4051: Inputting the enhanced entity embedding matrix E l of the yth semantic subspace in the lth layer graph attention network and the entity embedding vector set E l aggregated by the lth layer graph attention network into an MLP (Multi-Layer Perceptron) to obtain MLP(E l ) and

[0175] The calculation process of the MLP is as follows:

[0176] 1) The output of the lth layer MLP is:

[0177] h (l) = σ(W (l) h (l-1) + b (l) );

[0178] In the formula: h (l-1) is the output of the (l-1)th layer MLP; W (l) and b (l) are the weight matrix and bias vector of the lth layer MLP, respectively; and σ(·) is a nonlinear activation function.

[0179] 2) The final output is: y = h (l)where L is the number of layers of the MLP, and y is the final output of the MLP;

[0180] S4052: Calculate the attention coefficient β of the y-th semantic subspace y ;

[0181] The formula is:

[0182]

[0183] Where: exp(·) represents the exponential;

[0184] S4053: Update the entity embedding matrix E of the (l+1)-th layer graph attention network by the following formula l+1 ;

[0185] For each layer of the graph attention network, a single node e p needs to aggregate information from multiple subspaces to update:

[0186]

[0187] 4. Local semantic perception mechanism

[0188] The specific processing steps of local semantic perception include:

[0189] S4061: For a triple (h, r, t), obtain entity embedding vectors e l+1 , e h ∈R t and relation embedding vector r∈R d from the entity embedding matrix E d and the relation embedding vector R0; concatenate e h and r to obtain e hr ∈R 2d , and take e hr and r as input;

[0190] S4062: Calculate the enhanced feature representation by cross-attention mechanism; cross-attention mechanism dynamically allocates attention weights by calculating the interaction between units in the input sequence.

[0191] 1) Calculate the query matrix Q, the key matrix K, and the value matrix V by e hr and r:

[0192] The formula is:

[0193] Q=rW Q ;

[0194] K=e hr W K ;

[0195] V = e hr W V ;

[0196] wherein: is a learnable weight matrix;

[0197] 2) Calculate the attention score matrix A:

[0198]

[0199] wherein: softmax(·) is used to normalize the attention score;

[0200] 3) Calculate the weighted enhanced feature representation H:

[0201] H = AV;

[0202] wherein: is the enhanced feature representation after cross-attention mechanism;

[0203] S4063: On the basis of the enhanced feature representation H, capture different aspects of information through multi-head attention mechanism, and obtain a comprehensive feature representation;

[0204] The formula is:

[0205] M = Concat[H1, H1,..., H h ]W O ;

[0206] wherein: M ∈ R 2×d represents the comprehensive feature representation; h represents the number of attention heads; Concat is a concatenation operation; is a learnable weight coefficient;

[0207] S4064: Enhance the knowledge expression ability of the comprehensive feature representation M through the feedforward layer FFN(·), and obtain an output representation matrix H out ;

[0208] The formula is:

[0209] H out = FFN(M);

[0210] wherein, the calculation formula of FFN(M) is:

[0211] FFN(M) = max(0, M W1 + b1) W2 + b2;

[0212] wherein: W1 ∈ R d×d , W2 ∈ R d×d , b1 ∈ R d , b2 ∈ R d represent a learnable weight matrix.denotes a learnable bias term, H out ∈ R d .

[0213] 5. Decoder decodes

[0214] Specifically, the decoding of the embedding vectors is implemented by the following steps:

[0215] S2051: Calculate the score of each candidate triple by using the bidirectional inner product as the basic decoding paradigm:

[0216] score = H out · e t ;

[0217] where e t ∈ R d is the embedding representation of the tail node of the triple;

[0218] S2052: Convert the score into a probability output:

[0219] p T = σ(score);

[0220] where σ(·) represents an activation function;

[0221] S2053: Create negative samples, for each positive sample (h, t, t), generate k = 5 negative samples, where h and t are randomly negative sampled, obtaining negative samples h ′ and t ′ :

[0222] N hr = {(h ′ , r, t) | h ′ ∈ ε \ h ∪ (h, r, t ′ ) | t ′ ∈ ε \ t};

[0223] where N hrt is the set of all negative samples;

[0224] S2054: Calculate the binary cross-entropy loss L t :

[0225]

[0226] where y T represents the label information of the node, if the predicted node is correct, it is 1, otherwise it is 0;

[0227] S2055: Calculate the regularization term loss L ω , defined as:

[0228]

[0229] where W τ represents all trainable parameters, and ω is a regularization coefficient;

[0230] S2056: Multi-task loss calculation. All losses of the model can be summarized as a multi-task learning method, and the total loss of the model task is:

[0231] L = L t + Loss HSIC + L ω ;

[0232] S2057: The model finally uses the Adam optimizer to train and update all parameters.

[0233] To improve the generalization ability and stability of the model, the application adopts multi-task contrast learning and K-fold cross-validation strategy in the training process, which alleviates the problems of unbalanced training samples and overfitting, and improves the adaptability of the model in different maintenance scenarios. Experimental results show that this method is superior to the mainstream embedding model in multiple indicators, effectively improving the knowledge reasoning ability and recommendation accuracy in device maintenance tasks.

[0234] III. Expert assisted decision making

[0235] Specifically, the expert assisted decision making is achieved by the following steps:

[0236] S501: If the domain expert manually judges the ranked candidate triplets (especially the highest ranked result) and determines that they meet the problem solving requirements and are reasonable and effective, the system:

[0237] Organizes the input information in combination with the pre-defined language large model prompt word template.

[0238] The template usually contains three layers of structure:

[0239] Prompt layer: clearly define the task requirements (such as "generate maintenance suggestions");

[0240] Knowledge layer: inject structured knowledge of the top-ranked candidate entities and associated knowledge subgraphs;

[0241] Constraint layer: set output specifications (such as language, format, and avoidance of professional terms). Then, call the pre-selected LLM. The LLM generates a final high-quality, readable natural language answer text for the user based on the injected structured knowledge and prompt template;

[0242] S502: If the domain expert determines that the candidate triplet set does not meet the requirements or is not satisfied with the answer text generated by the LLM, the system triggers a feedback mechanism: the expert submits rejection information (which can mark specific errors or deficiencies) through the user interface.

[0243] The system transmits the rejection information and context (including the original question, the rejected candidate sequence / answer) to the knowledge graph management and model training module through a preset backhaul path (such as an API interface). The feedback information is used for: (a) refining the retrieval algorithm: adjusting the semantic understanding module or graph matching strategy; (b) optimizing the embedding model: as negative samples for subsequent DSPE model fine-tuning training, enhancing the semantic understanding accuracy of the model and the adaptability of the retrieval result.

[0244] IV. Dynamic completion and intelligent evolution of the knowledge graph

[0245] In this embodiment, the device maintaining the knowledge graph calculates the cosine similarity between the knowledge nodes in the knowledge graph and the problem description text input by the user through the deep semantic perception embedding model, and takes the knowledge nodes with a cosine similarity exceeding a threshold value as candidate knowledge nodes; generates new relationships and constructs new triples based on the candidate knowledge nodes and the problem description text input by the user; and updates the new triples to the device maintaining the knowledge graph.

[0246] Specifically, the dynamic completion and intelligent evolution of the knowledge graph are implemented through the following steps:

[0247] S3051: In analyzing the user query question statement, the DSPE model, in addition to returning the answers existing in the graph, also identifies the knowledge nodes highly relevant to the question based on semantic similarity (cosine similarity) calculation;

[0248] S3052: Judgment and completion: the system can automatically or semi-automatically determine this as a valid new triple based on a preset confidence threshold and / or the final confirmation of a domain expert (manual review logical relevance), and supplement it to the storage of the current device maintaining the knowledge graph.

[0249] Embodiment Two

[0250] The embodiment discloses a device operation and maintenance intelligent question answering system, which is based on the device operation and maintenance knowledge large model construction method in embodiment one.

[0251] The device operation and maintenance intelligent question answering system comprises:

[0252] A question input module configured to obtain a problem description text input by a user;

[0253] A question analysis module configured to perform semantic analysis and key information extraction on the problem description text input by the user to obtain key information;

[0254] An entity matching module configured to match corresponding knowledge nodes and triples containing the knowledge nodes from a device maintenance knowledge graph based on the key information, and add the triples to a candidate triple set;

[0255] The semantic score calculation module is configured to calculate a semantic score of each triple in the candidate triple set by using a deep semantic perception embedding model, and sort the triples in the candidate triple set based on the semantic score.

[0256] The answer generation module is configured to generate an answer to the question based on the candidate triple set by using a language large model after expert-assisted decision-making on the sorted candidate triple set.

[0257] Embodiment three:

[0258] The embodiment discloses a computer device.

[0259] The computer device comprises a processor, a memory, and a computer program stored in the memory and executable on the processor. The processor implements the steps in the above various device operation and maintenance knowledge model construction method embodiments when executing the computer program. Alternatively, the processor implements the functions of each module / unit in the above various system embodiments when executing the computer program.

[0260] Illustratively, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application. The one or more modules / units can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the device operation and maintenance intelligent question and answer system.

[0261] The computer device can be a desktop computer, a notebook computer, a palm computer, and a cloud server, etc. The computer device can include, but is not limited to, a processor and a memory.

[0262] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application rather than limit the technical solutions. Those of ordinary skill in the art should understand that modifications or equivalent replacements to the technical solutions of the present application without departing from the spirit and scope of the technical solutions should be covered in the scope of the claims of the present application.

Claims

1. A method for constructing a device operation and maintenance knowledge large model, characterized in that, Comprise: S1: acquire user input question description text; S2: the user input question description text is carried out semantic analysis and key information extraction, and the key information is obtained; S3: based on the key information, the corresponding knowledge node and the triple containing the knowledge node are matched from the equipment maintenance knowledge graph, and the triple is added to the candidate triple set; Wherein the equipment maintenance knowledge graph is constructed based on the five tuple model, and the five tuple model splits the event text information into five types of event arguments, including event subject, subject action, subject attribute, subject state description and maintenance behavior; S4: the semantic score of each triple in the candidate triple set is calculated by a deep semantic perception embedding model, and the triples in the candidate triple set are sorted based on the semantic score; Wherein the deep semantic perception embedding model aggregates neighbor node information in multiple semantic subspaces through a double attention mechanism to capture the potential semantic dependency and polysemy representation in the equipment maintenance knowledge graph; S5: using a language large model to generate a question answer based on the candidate triple set.

2. The device operation and maintenance knowledge large model construction method of claim 1, wherein: In step S3, the processing steps of constructing the equipment maintenance knowledge graph are as follows: S301: acquire an event data set containing a plurality of event text information; S302: split the event text information into event arguments by the five tuple model, establish the event relationship between each two event arguments, and obtain a plurality of event relationship triples; The five tuple model splits the event text information into five types of event arguments, including event subject, subject action, subject attribute, subject state description and maintenance behavior; S303: generate the logical relationship between each two event arguments through the equipment maintenance ontology, and obtain a plurality of logical relationship triples; S304: execute steps S302 and S303 for each event text information in the event data set, input all the obtained event relationship triples and logical relationship triples into a graph database to generate the equipment maintenance knowledge graph.

3. The device operation and maintenance knowledge large model construction method of claim 2, wherein: In step S302, the event relationship includes: 1) the relationship between the fault cause event and the fault phenomenon event: Cause; 2) the relationship between the repair method event and the fault phenomenon event: Solve; 3) the relationship between the preventive measure event and the fault phenomenon event: Prevent; In step S303, the logical relationship defined by the equipment maintenance ontology includes: 1) device hierarchical structure relationship: Has; 2) fault evolution logical relationship: Cause, Solve, Prevent; 3) maintenance decision rule relationship: Occur, Take; 4) similarity relationship: Similar to.

4. The device operation and maintenance knowledge large model construction method of claim 1, wherein: In step S4, the processing steps of the deep semantic perception embedding model include: S401: the triples in the candidate triple set are taken as input; S402: attribute embedding of knowledge nodes of all entities and relationships in the equipment maintenance knowledge graph is performed by a BERT model to obtain an entity embedding matrix E0e R N×d and a relationship embedding matrix R0e R r×d , N is the number of all entities, d is the dimension of initial embedding, and r is the number of all relationship categories; wherein, the knowledge nodes in the equipment maintenance knowledge graph are entities, and the edges between the knowledge nodes are relationships; S403: Multi-dimensional latent semantic mapping: construct a mapping function f to map each entity in the entity embedding matrix E0 into multiple different semantic spaces, obtaining the entity embedding matrix E of each semantic subspace aap(y) ; The formula is represented as: E map(y) = f(E0); f(E0) = σ(M y · E0 + b y ); where: E map(y) ∈ R N×d denotes the entity embedding matrix of the y-th semantic subspace; M y ∈ R d×d denotes the mapping matrix of the y-th semantic subspace; b y denotes the bias term; σ(·) denotes the activation function; S404: First-level attention mechanism: obtaining the entity embedding matrix E of each semantic subspace map(y) In the input stacked multi-layer graph attention network, the enhanced entity embedding matrix of each semantic subspace in the lth layer graph attention network is obtained S405: second-level attention mechanism: the attention mechanism based on multiple semantic subspaces enhances the entity embedding matrix of each semantic subspace in the lth layer graph attention network Attention calculation is performed to obtain the updated entity embedding matrix E of the l+1th layer graph attention network l+1 ; S406: Local semantic perception mechanism: for the triplets (h, r, t) in the candidate triplet set, the entity embedding vector and the relationship embedding vector are obtained from the entity embedding matrix E l+1 and the relationship embedding matrix R0, and the corresponding output representation matrix H out is calculated through the cross-attention mechanism and the front feedback layer; S407: The output of the triplet (h, r, t) is represented by a matrix H through a decoder out decoding to obtain the semantic score of the triplet (h, r, t); The formula is represented as: score = H out • e t ; where: e t ∈R d is the embedding representation of the tail node of the triple (h, r, t).

5. The device operation and maintenance knowledge large model construction method of claim 4, wherein: The formula is represented as: S4041: calculating the entity embedding matrix E of the yth semantic subspace map(y) Any neighbor node t of the center node p in the entity embedding vector of the lth layer graph attention network And its relationship The product of the two, and the entity embedding vector of the center node p in the lth layer graph attention network Dot product to get the attention A in exponential form t ; In step S404, the specific processing steps are as follows: S4042: The center node of the y-th semantic space calculates the attention coefficient of each neighbor node in the l-th layer graph attention network Get the attention coefficient sequence of all neighbor nodes {α1,α2...α t-1 ,α t} The formula is represented as: wherein: N t represents a set of all neighbors and relationships of the center node; A q is the attention of a certain neighbor node; S4043: arrange the attention coefficient sequence in order of size, and select the top n neighbor node attention coefficients {α1, α2...α n} S4044: aggregate the attention coefficient sequence of the neighbor node selected by the lth layer graph attention network to the center node to obtain the entity embedding vector of the center node of the (l+1)th layer graph attention network The formula is represented as: where N (t,n) represent the top-n nodes and relation set of neighbor subset attention coefficient; (W map(y) ) l is a trainable parameter; and σ represents a linear activation function. S4045: Perform steps S4041 to S4044 on all center nodes in the entity embedding matrix E map(y) of the y-th semantic subspace to obtain the enhanced entity embedding matrix 6. The device operation and maintenance knowledge large model construction method of claim 4, wherein: The formula is represented as: S4051: the y-th semantic subspace in the enhanced entity embedding matrix of the l-th layer graph attention network and the entity embedding vector set E of all semantic subspace aggregated by the l-th layer graph attention network l respectively input into the MLP to obtain MLP(E l ) and S4052: Calculate the attention coefficient β of the y-th semantic subspace y ; In step S405, the specific processing steps are as follows: The formula is represented as: S4053: Calculate the updated entity embedding matrix E of the (l+1)th layer graph attention network by the following formula l+1 ; 7. The device operation and maintenance knowledge large model construction method of claim 4, wherein: Wherein: exp(·) represents the exponential; S4061: For the triple (h,r,t), from the entity embedding matrix E l+1 Obtain the entity embedding vector e from the relation embedding vector R0 h e t And relation embedding vector r; concatenate e h And r get e hr ∈R 2d , will e hr and r as input; In step S406, the specific processing steps include: 1) by e hr and r to compute the query matrix Q, the key matrix K, and the value matrix V: S4062: calculate the enhanced feature representation through the cross attention mechanism; The formula is represented as: Q = rW Q ; K = e hr W K ; V = e hr W V ; wherein: is a learnable weight matrix; 2) Calculate the attention score matrix A: In the formula: softmax(·) is used to normalize the attention score; 3) Calculate the weighted enhanced feature representation H: H = AV; In the formula: is an enhanced feature representation through a cross-attention mechanism; S4063: Based on the enhanced feature representation H, capture different aspects of information through the multi-head attention mechanism to obtain a comprehensive feature representation; The formula is expressed as: M = Concat [H1, H2,..., H h ]W O ; In the formula, M ∈ R 2×d denotes the comprehensive feature representation; h denotes the number of attention heads; Concat is a concatenation operation; is a learnable weight coefficient; S4064: enhance the knowledge representation ability of the comprehensive feature representation M by the feedforward layer FFN(·) to obtain an output representation matrix H out .

8. An intelligent question and answer system for equipment operation and maintenance, characterized in that: Based on the device operation and maintenance knowledge large model construction method in claim 1, comprising: A question input module for obtaining a user input question description text; A question analysis module for performing semantic analysis and key information extraction on the user input question description text to obtain key information; An entity matching module for matching corresponding knowledge nodes and triples containing the knowledge nodes from the device maintenance knowledge graph based on the key information, and adding the triples to a candidate triple set; A semantic score calculation module for calculating the semantic score of each triple in the candidate triple set through a deep semantic perception embedding model, and sorting the triples in the candidate triple set based on the semantic score; An answer generation module for generating a question answer based on the candidate triple set after expert assisted decision making on the sorted candidate triple set.

9. A computer device, comprising: Comprise: One or more processors; The processor is used to store one or more programs; When the one or more programs are executed by the one or more processors, the device operation and maintenance knowledge large model construction method as claimed in any one of claims 1 to 7 is implemented.

Images