Industrial equipment self-evolution predictive maintenance system based on large language model and knowledge graph

Abstract

The application discloses an industrial equipment self-evolution predictive maintenance system based on a large language model and a knowledge graph, belongs to the technical field of intelligent operation and maintenance and predictive maintenance of industrial equipment, and comprises the following steps: obtaining real-time equipment operation logs, unstructured operation and maintenance manuals and structured data of equipment, generating real-time event entities and operation and maintenance knowledge triples by using a large language model; generating a fused knowledge graph based on the structured data of equipment and the operation and maintenance knowledge triples; generating a scenario-based event subgraph based on the real-time event entities; generating a fault cause and effect chain based on reasoning of the scenario-based event subgraph; generating a self-evolution diagnosis report based on the fault cause and effect chain; and updating the fused knowledge graph based on a maintenance confirmation signal. Through semantic analysis, knowledge graph reasoning and maintenance feedback closed-loop updating, the application realizes fault cause and effect chain tracing, diagnosis report generation and dynamic correction of a knowledge path.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and predictive maintenance technology for industrial equipment, and in particular to a self-evolving predictive maintenance method and system for industrial equipment based on large language models and knowledge graphs. Background Technology

[0002] Industrial equipment is widely used in power, petrochemical, metallurgical, rail transportation, and intelligent manufacturing industries. The operational status of this equipment directly impacts the continuity and safety of production systems. With the development of the Industrial Internet, sensor networks, and digital operation and maintenance systems, a large amount of real-time equipment operation logs, alarm records, maintenance records, operation tickets, fault reports, and maintenance manuals are generated during equipment operation. This data includes both structured information such as sensor values, equipment numbers, and component levels, as well as unstructured semantic information such as fault descriptions, maintenance experience, cause analysis, and handling procedures.

[0003] Existing predictive maintenance solutions largely rely on sensor time-series data, statistical models, or machine learning models to detect anomalies and predict faults in equipment operation. While these solutions can identify abnormal changes in parameters such as temperature, vibration, current, and pressure, processing equipment operation logs, maintenance manuals, and expert experience texts typically requires manual rule compilation or offline knowledge base construction, making it difficult to fully utilize the contextual semantic information in unstructured text. Furthermore, traditional fault diagnosis models often rely primarily on correlations in historical samples, making it difficult to combine equipment component structure, fault mechanisms, maintenance procedures, and expert experience for causal chain tracing.

[0004] Knowledge graphs can represent the structured relationships between equipment, components, faults, causes, phenomena, and maintenance measures in the form of nodes and edges. However, most existing knowledge graphs are statically constructed, and updates rely on manual editing, making it impossible to continuously adjust the reliability of diagnostic paths based on actual maintenance results. When the long-term operating environment of equipment changes, fault modes evolve, or new hidden faults appear, static knowledge bases are prone to problems such as lagging diagnostic logic, inaccurate path weights, and insufficient report interpretation. Therefore, existing predictive maintenance technologies suffer from insufficient utilization of unstructured semantic information, weak fault root cause tracing capabilities, and difficulty in adaptive evolution of the knowledge base. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a self-evolving predictive maintenance system for industrial equipment based on a large language model and knowledge graph. Through semantic parsing, knowledge graph reasoning, and closed-loop maintenance feedback updates, it enables fault causal chain tracing, diagnostic report generation, and dynamic correction of knowledge paths.

[0006] The above objectives can be achieved through the following approach: A self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs includes: acquiring real-time equipment operation logs, unstructured maintenance manuals, and structured equipment data; performing semantic parsing of the real-time equipment operation logs and unstructured maintenance manuals using large language models to generate real-time event entities and maintenance knowledge triples; performing graph initialization processing based on the structured equipment data to generate a basic equipment knowledge graph; and performing knowledge fusion processing on the basic equipment knowledge graph and maintenance knowledge triples to generate a fused knowledge graph; performing semantic mapping processing based on the real-time event entities and the fused knowledge graph to generate a contextualized event subgraph reflecting the current equipment status; performing causal path reasoning processing on the fused knowledge graph based on the contextualized event subgraphs to generate a fault causal chain pointing to potential fault nodes; performing large language model report generation processing based on the fault causal chain to generate a self-evolving diagnostic report containing fault causes, related evidence, and maintenance suggestions; and performing path weight adjustment processing on the fused knowledge graph based on maintenance confirmation signals and fault causal chains to generate an updated fused knowledge graph.

[0007] Optionally, the steps of extracting real-time event entities from device operation logs using a large language model and extracting knowledge from unstructured operation and maintenance manuals to obtain operation and maintenance knowledge triples include: parsing the device operation logs for time sequence and content, using the named entity recognition function of the large language model to identify key information such as time, components, parameters, and status descriptions to form structured real-time event entities; and processing the unstructured operation and maintenance manuals into chapters and paragraphs, using the relation extraction function of the large language model to extract entity relation pairs and construct operation and maintenance knowledge triples.

[0008] Optionally, the steps of fusing the operation and maintenance knowledge triples with the device basic knowledge graph to generate a fused knowledge graph include: standardizing the entities in the operation and maintenance knowledge triples to obtain standard entities; aligning the standard entities with the nodes in the device basic knowledge graph, establishing a link if a matching node exists, and creating a new node if no matching node exists; and using the relationships in the operation and maintenance knowledge triples as edges to connect the aligned or created nodes, thereby integrating operation and maintenance knowledge into the device basic knowledge graph and completing the construction of the fused knowledge graph.

[0009] Optionally, in the fused knowledge graph, the step of performing graph path search starting from nodes in the contextualized event subgraph to identify and extract fault causal chains pointing to potential fault nodes includes: setting nodes in the contextualized event subgraph as active nodes and assigning them initial activation strength; starting from the active node, traversing along the directed edges in the fused knowledge graph, propagating the activation strength according to the edge weights to form an activation path; when the cumulative activation strength of the activation path exceeds a preset fault activation threshold, determining that the path is a fault causal chain.

[0010] Optionally, the steps of driving a large language model to generate a self-evolving diagnostic report based on the fault causal chain include: parsing the fault causal chain into a series of logically clear diagnostic steps; matching the original text source of the node in the knowledge graph with each diagnostic step as diagnostic evidence; organizing the diagnostic steps and diagnostic evidence into natural language prompts, inputting them into the large language model, and generating a self-evolving diagnostic report.

[0011] Optionally, the step of dynamically adjusting the path weights constituting the fault causal chain in the fused knowledge graph based on the maintenance confirmation signal includes: receiving and parsing the maintenance confirmation signal, classifying it as a correct diagnosis signal or an incorrect diagnosis signal; if it is a correct diagnosis signal, increasing the weights of all edges in the fault causal chain; if it is an incorrect diagnosis signal, decreasing the weights of all edges in the fault causal chain, thereby achieving dynamic correction of knowledge.

[0012] Optionally, after identifying and extracting the fault causal chain pointing to the potential fault node, the method further includes: simulating possible future changes in equipment operation instructions or environmental parameters based on the fused knowledge graph to generate simulated event entities; mapping the simulated event entities to the fused knowledge graph to perform a forward-looking graph path search to predict the potential fault causal chain that may be triggered; and generating risk warning information based on the potential fault causal chain.

[0013] Optionally, after obtaining the updated fused knowledge graph, the process further includes: extracting fault causal chains from the updated fused knowledge graph that have undergone multiple positive verifications and have significantly increased weights to form a general fault mode template; and applying the general fault mode template to the knowledge graph initialization or update process of other industrial equipment of the same type to achieve cross-device knowledge transfer and sharing.

[0014] Based on the same inventive concept, this invention also provides an industrial equipment self-evolving predictive maintenance system based on a large language model and knowledge graph. The system includes: a data acquisition and semantic parsing module, used to acquire real-time equipment operation logs, unstructured operation and maintenance manuals, and structured equipment data of industrial equipment; performing semantic parsing processing on the real-time equipment operation logs and unstructured operation and maintenance manuals using a large language model to generate real-time event entities and operation and maintenance knowledge triples; and a knowledge graph construction and fusion module, used to perform graph initialization processing based on the structured equipment data to generate a basic equipment knowledge graph; and performing knowledge fusion processing on the basic equipment knowledge graph and operation and maintenance knowledge triples to generate a fused knowledge graph. The state mapping and contextualization module performs semantic mapping based on real-time event entities and the fused knowledge graph to generate a contextualized event subgraph reflecting the current device state. The causal path reasoning module performs causal path reasoning based on the contextualized event subgraph in the fused knowledge graph to generate a fault causal chain pointing to potential fault nodes. The diagnostic report generation module performs large language model report generation based on the fault causal chain to generate a self-evolving diagnostic report containing fault causes, related evidence, and maintenance suggestions. The knowledge self-evolution update module adjusts the path weights of the fused knowledge graph based on maintenance confirmation signals and the fault causal chain to generate an updated fused knowledge graph.

[0015] Compared with the prior art, the present invention has the following advantages: This invention uses a large language model to perform semantic parsing on real-time device operation logs and unstructured operation and maintenance manuals, enabling device operation events and operation and maintenance experience texts to be transformed into computable and inferable real-time event entities and operation and maintenance knowledge triples.

[0016] This invention integrates structured equipment data with operation and maintenance knowledge triples into a fused knowledge graph, enabling the expression of equipment component relationships, fault modes, fault causes, maintenance processes, and textual evidence within a unified graph structure.

[0017] This invention uses contextualized event subgraphs for causal path reasoning, which can form a fault causal chain from the current equipment event to the potential fault node, enabling the diagnostic report to have path tracing and evidence correlation capabilities.

[0018] This invention dynamically adjusts the path weights in the fused knowledge graph based on maintenance confirmation signals, enabling the knowledge base to be continuously updated in actual operation and maintenance feedback, forming a closed-loop self-evolutionary mechanism of prediction, diagnosis, verification and feedback. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the self-evolutionary predictive maintenance method for industrial equipment based on a large language model and knowledge graph, according to an embodiment of the present invention.

[0021] Figure 2 This is a framework diagram of the self-evolving predictive maintenance system for industrial equipment based on a large language model and knowledge graph, according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.

[0023] Reference Figure 1 One embodiment of the present invention provides an industrial equipment self-evolving predictive maintenance system based on a large language model and knowledge graph. Through semantic parsing, knowledge graph reasoning and maintenance feedback closed-loop updates, it realizes fault causal chain tracing, diagnostic report generation and dynamic correction of knowledge paths.

[0024] S1. Obtain real-time equipment operation logs and unstructured operation and maintenance manuals of industrial equipment. Use a large language model to extract events from the equipment operation logs to obtain real-time event entities, and extract knowledge from the unstructured operation and maintenance manuals to obtain operation and maintenance knowledge triples. In one embodiment of the present invention, step S1 includes the following steps: The device operation logs are analyzed in terms of time sequence and content. The named entity recognition function of the large language model is used to identify key information such as time, components, parameters and status descriptions, forming structured real-time event entities. The unstructured operation and maintenance manual is divided into chapters and paragraphs. The relation extraction function of the large language model is used to extract entity relation pairs and construct operation and maintenance knowledge triples.

[0025] Specifically, real-time equipment operation logs are provided by the industrial equipment field data acquisition system and the operation and maintenance management system, while unstructured operation and maintenance manuals are provided by equipment repair manuals, maintenance procedures, and historical maintenance documents. For the first... There are 10 real-time device operation log entries:

[0026] In the formula, Indicates the first Real-time device operation log; , , , , , , , , The sequence represents the device number, data acquisition time, component name, parameter name, parameter value, alarm code, alarm level, operating status description, and operation record.

[0027] When parsing the time sequence and content of real-time device operation logs, the timestamps of real-time device operation logs from different sources are first normalized according to a unified time base to generate a unified time log sequence. Then, the unified time log sequence undergoes field segmentation, missing field completion, and abnormal character cleaning to generate standardized log text. Timestamp normalization converts local time, millisecond-level timestamps, or device controller timing values ​​from different acquisition systems to a unified system time; field segmentation separates log fields from the real-time device operation logs; missing field completion fills in missing content based on adjacent log records, device ledger fields, or alarm code mapping tables; and abnormal character cleaning removes garbled characters, duplicate separators, and invalid control characters.

[0028] A unified time log sequence can be represented as:

[0029] In the formula, Represents a unified time log sequence; , , , This represents the real-time device operation logs sorted according to a unified time base. This indicates the number of logs in a unified time log sequence.

[0030] After performing field segmentation, missing field completion, and abnormal character cleaning on the unified time log sequence, a normalized log text set is generated, denoted as:

[0031] In the formula, Represents a set of normalized log texts; Indicates the first The standardized log text corresponding to each real-time device operation log.

[0032] The large language model is a pre-trained language model fine-tuned using industrial operation and maintenance corpus instructions. The industrial operation and maintenance corpus includes historical equipment operation logs, manually annotated fault work orders, equipment maintenance records, and excerpts from the operation and maintenance manual. The input to the large language model is normalized log text, and the output is a structured event result containing time, component, parameter, parameter value, status description, and event type. For the first... A normalized log text, a real-time event entity can be represented as:

[0033] In the formula, Indicates the first One real-time event entity; , , , , , The sequence represents the event time, component name, parameter name, parameter value, state description, and event type, respectively.

[0034] When extracting knowledge from unstructured operations and maintenance manuals, the manuals are first segmented according to document titles, chapter numbers, paragraph boundaries, and table titles, generating a set of manual text paragraphs. For the first... The manual text paragraphs include:

[0035] In the formula, Indicates the first A paragraph of manual text; , , , These represent the document title, chapter number, paragraph number, and paragraph text content in that order.

[0036] The set of manual text paragraphs is input into a large language model, which extracts component entities, fault phenomenon entities, cause entities, troubleshooting step entities, maintenance action entities, and spare parts entities, and identifies the relationship types between different entities. Relationship types include belonging, causing, accompanying, troubleshooting, maintenance, and replacement relationships. The maintenance knowledge triplet is expressed using a head entity, relation, and tail entity structure, denoted as:

[0037] In the formula, Indicates the first A triplet of operation and maintenance knowledge; , , They represent the head entity, entity relationship, and tail entity in that order.

[0038] The training samples for the large language model consist of industrial equipment log text, maintenance manual paragraphs, and historical fault work orders. Supervised labels include event entity boundaries, entity categories, entity relation categories, and triplet extraction results. During training, entity recognition cross-entropy loss and relation classification cross-entropy loss are used as optimization objectives to obtain a large language model capable of performing named entity recognition and relation extraction on industrial equipment maintenance text. The structure of the large language model includes a text embedding layer, a multi-layer self-attention encoding layer, an entity recognition output layer, and a relation classification output layer. The text embedding layer receives segmented, normalized log text or manual text paragraphs and outputs a sequence of word vectors. The multi-layer self-attention encoding layer receives the word vector sequence and outputs contextual semantic features. The entity recognition output layer outputs a sequence of entity labels based on the contextual semantic features. The relation classification output layer outputs entity relation categories based on entity pair semantic features.

[0039] For example, a compressor generates a real-time equipment operation log at 10:15:30. In this log, the equipment number is C-203, the acquisition time is 10:15:30, the component name is "Level 1 Bearing," the parameter name is "Temperature," the parameter value is 92℃, the alarm code is TEMP-HIGH-01, the alarm level is Level 2 alarm, the operating status is described as "High Temperature Alarm," and the operation record is empty. Then:

[0040] After performing timestamp normalization, field segmentation, missing field completion, and abnormal character cleaning on the real-time equipment operation log, the standardized log text is obtained: "10:15:30, C-203 compressor first-stage bearing temperature is 92℃, alarm code is TEMP-HIGH-01, alarm level is level two alarm, status is high temperature alarm". The standardized log text is input into a large language model, which identifies the event time as 10:15:30, component name as first-stage bearing, parameter name as temperature, parameter value as 92℃, status description as high temperature alarm, and event type as bearing anomaly, generating a real-time event entity:

[0041] Regarding the section in the maintenance manual that states, "High temperature in the primary bearing is usually caused by insufficient lubrication or insufficient cooling water flow. The lubricating oil pressure, cooling water inlet valve, and bearing wear should be checked," assuming the document title is "C-203 Compressor Repair Manual," the chapter number is Chapter 4, Section 2, and the paragraph number is Paragraph 3, then: ; Input the manual text paragraphs into the large language model. The large language model extracts the component entity as the first-level bearing, the fault phenomenon entity as high temperature, the cause entities as insufficient lubrication and insufficient cooling water flow, and the troubleshooting steps entities as checking the lubricating oil pressure and checking the cooling water inlet valve. It then generates a maintenance knowledge triplet: ; ; ; ; Therefore, real-time event entities are generated based on real-time device operation logs, and operation and maintenance knowledge triples are generated based on unstructured operation and maintenance manuals.

[0042] S2. Initialize and construct a basic knowledge graph of the equipment based on the pre-set structured equipment data, and integrate the operation and maintenance knowledge triples with the basic knowledge graph of the equipment to generate a fused knowledge graph that includes the relationship between equipment components, fault modes and maintenance processes; In one embodiment of the present invention, step S2 includes the following steps: Standardize the entities in the operation and maintenance knowledge triples to obtain standard entities; Align standard entities with nodes in the device basic knowledge graph; if a matching node exists, establish a link; otherwise, create a new node. By using the relationships in the O&M knowledge triples as edges to connect aligned or created nodes, O&M knowledge is integrated into the device basic knowledge graph, thus completing the construction of the fused knowledge graph.

[0043] Specifically, the structured equipment data is provided by equipment ledgers, sensor point tables, and a maintenance management database. This structured data is used to initialize and construct a basic equipment knowledge graph. The basic equipment knowledge graph uses equipment, components, measurement points, failure modes, maintenance processes, and spare parts as nodes, and inclusion relationships, connection relationships, monitoring relationships, failure relationships, and maintenance relationships as edges. For the basic equipment knowledge graph, we have:

[0044] In the formula, Represents a basic knowledge graph of the equipment; Represents a set of nodes; Represents the set of edges; This represents the set of node attributes and edge attributes.

[0045] When standardizing entities in the maintenance knowledge triplet, the first step is to perform name cleaning, synonym replacement, number mapping, and unit unification on the head and tail entities to generate standard entities. Name cleaning removes invalid symbols and redundant numbers from entity names; synonym replacement maps different names for the same component or fault phenomenon to a unified name; number mapping maps equipment codes, component codes, and measurement point codes to standard codes in the equipment ledger; and unit unification converts different dimensional expressions into preset standard units.

[0046] For the A triplet of operation and maintenance knowledge After standardization, we get:

[0047] In the formula, Indicates the first A standardized triplet of operation and maintenance knowledge; Represents the standardized head entity; Represents the standardized tail entity; Indicates entity relationships.

[0048] When aligning a standard entity with a node in the device basic knowledge graph, the standard entity is first used as the entity to be aligned. Then, the node matching score between the entity to be aligned and the candidate node in the device basic knowledge graph is calculated based on the entity name, entity code, entity type, and entity description. When the node matching score is greater than or equal to the preset node alignment threshold, the entity to be aligned is linked to the corresponding candidate node. When the node matching score is less than the preset node alignment threshold, a new node is created in the device basic knowledge graph.

[0049] For the node matching score between the entity to be aligned and the candidate nodes, we have:

[0050] In the formula, Indicates the node matching score; Indicates the score for name matching; Indicates the encoded matching score; Indicates the type matching score; , , These represent the weights of the corresponding matching scores, and .

[0051] After node alignment or new node creation is completed, the relations in the operation and maintenance knowledge triples are written as edges into the device basic knowledge graph. Edge attributes include relation type, source document, source paragraph, extraction confidence, and initial path weight. The initial path weight is determined based on the relation extraction confidence output by the large language model. Thus, the operation and maintenance knowledge triples are integrated into the device basic knowledge graph to generate a fused knowledge graph. For the fused knowledge graph, we have:

[0052] In the formula, This indicates the fusion of knowledge graphs; Represents the merged set of nodes; Represents the set of edges after merging; This represents the set of node attributes and edge attributes after merging. This represents the set of edge weights.

[0053] For example, the equipment structured data includes the equipment node "C-203 Compressor", the component node "Level 1 Bearing", and the measurement point node "Level 1 Bearing Temperature". (Operation and Maintenance Knowledge Triad) After input, the head entity "Segment 1 Bearing" is first standardized to obtain the standard entity "Level 1 Bearing", which generates:

[0054] Let the name matching score between the standard entity and the candidate node "Level 1 Bearing" be quantified. Encoding matching score Type matching score Weight , , ,but:

[0055] If the preset node alignment threshold is Then, the standard entity "Level 1 Bearing" is linked to the "Level 1 Bearing" node in the equipment basic knowledge graph. Since "Insufficient Lubrication" does not exist in the equipment basic knowledge graph, a new node "Insufficient Lubrication" is created, and "High Temperature Cause" is used as an edge to connect the "Level 1 Bearing" node and the "Insufficient Lubrication" node, generating a new relationship path in the fused knowledge graph.

[0056] S3. Calculate the semantic similarity between real-time event entities and nodes in the fused knowledge graph using a large language model. Based on the similarity results, map the real-time event entities to the fused knowledge graph to construct a contextualized event subgraph that reflects the current device status. In one embodiment of the present invention, step S3 includes the following steps: The textual description information in real-time event entities is encoded into event feature vectors using a large language model; The attribute information of nodes in the fused knowledge graph is encoded into node feature vectors using a large language model; The semantic similarity is obtained by calculating the cosine similarity between the event feature vector and the feature vector of each node within the same vector space.

[0057] Specifically, real-time event entities are generated in step S1, and the fused knowledge graph is generated in step S2. When encoding real-time event entities, the component names, parameter names, parameter values, state descriptions, and event types are concatenated into event semantic text, which is then input into the semantic encoding layer of the large language model to generate event feature vectors. For the first... There are 1 real-time event entity:

[0058] In the formula, Indicates the first The event feature vector corresponding to each real-time event entity; Represents the semantic encoding layer of a large language model; Indicates by the first The event semantic text obtained by converting a real-time event entity.

[0059] When encoding nodes in the fused knowledge graph, the node name, node type, node alias, node description, and source text summary are concatenated into the node semantic text, which is then input into the semantic encoding layer of the large language model to generate node feature vectors. For the first... There are nodes, including:

[0060] In the formula, Indicates the first The node feature vectors corresponding to each node; Indicates by the first The semantic text of a node is obtained by converting the attribute information of each node.

[0061] Semantic similarity is obtained by calculating the cosine similarity between event feature vectors and node feature vectors within the same vector space.

[0062] In the formula, Indicates the first Real-time event entity and the first Semantic similarity between nodes; This represents the inner product of the event feature vector and the node feature vector; and These represent the magnitudes of the corresponding vectors.

[0063] Real-time event entities are mapped to the fused knowledge graph based on semantic similarity. If the maximum semantic similarity is greater than or equal to a preset mapping threshold, the real-time event entity is mapped to the corresponding node; if the maximum semantic similarity is less than the preset mapping threshold, the real-time event entity is written into the fused knowledge graph as a temporary event node and marked as a node to be confirmed. After mapping, the associated component nodes, fault phenomenon nodes, cause nodes, maintenance process nodes, and corresponding edge relationships are extracted from the mapped nodes to generate a contextualized event subgraph. For the contextualized event subgraph, we have:

[0064] In the formula, Represents a contextualized event subgraph; This represents the set of nodes in a contextualized event subgraph. This represents the set of edges in a contextualized event subgraph. This represents the set of node attributes and edge attributes in the contextualized event subgraph.

[0065] For example, real-time event entities This is converted to event semantic text: "Level 1 bearing temperature 92℃, status is high temperature alarm, event type is bearing abnormality." The large language model encodes the event semantic text, generating an event feature vector. For ease of explanation, the component status dimension and fault semantic dimension from the event feature vector are selected as normalized example vectors, resulting in:

[0066] The knowledge graph contains nodes such as "high temperature of primary bearing," "insufficient lubrication," "insufficient cooling water flow," and "bearing wear." The large language model encodes the node name, node type, node description, and source text summary for each of these nodes, resulting in node feature vectors.

[0067]

[0068]

[0069]

[0070] Based on the cosine similarity calculation formula, real-time event entities The semantic similarity between this node and the node “high temperature of first-level bearing” is:

[0071] Real-time event entities The semantic similarity between the node and the node with "insufficient lubrication" is:

[0072] Real-time event entities The semantic similarity between this node and the node "insufficient cooling water flow" is:

[0073] Real-time event entities The semantic similarity between the node and the node "bearing wear" is:

[0074] This yields the real-time event entity. The semantic similarity results between each candidate node are , , and If the preset mapping threshold is Then only the semantic similarity corresponding to the node "high temperature of first-level bearing" is considered. The semantic similarity is greater than the preset mapping threshold, and it is the maximum value among all candidate nodes. Therefore, the real-time event entity... Mapped to the "Level 1 Bearing High Temperature" node in the fused knowledge graph.

[0075] After mapping, using the "high temperature of primary bearing" node as the center, extract the "insufficient lubrication," "insufficient cooling water flow," "bearing wear," and "check lubricating oil pressure" nodes that are directly or indirectly related to it from the fused knowledge graph. Then, extract the corresponding causal relationships, troubleshooting relationships, and maintenance relationships to generate a contextualized event subgraph. Specifically, when imported, the contextualized event subgraph is as follows:

[0076] The node set is as follows:

[0077] The edge set is:

[0078] The set of node attributes and edge attributes is as follows:

[0079] Therefore, the contextualized event subgraph can be specifically represented as:

[0080] The contextualized event subgraph is used to characterize the local fault association structure triggered by "high temperature of the first-level bearing" in the current equipment state, and serves as the input for subsequent fault causal chain search.

[0081] S4. In the integrated knowledge graph, graph path search is performed starting from the nodes in the contextualized event subgraph to identify and extract the fault causal chain pointing to the potential fault node. In one embodiment of the present invention, step S4 includes the following steps: Set the nodes in the contextualized event subgraph as active nodes and assign them an initial activation strength; Starting from the activation node, traverse along the directed edges in the fused knowledge graph, propagate the activation intensity according to the edge weights, and form an activation path; When the cumulative activation intensity of an activation path exceeds a preset fault activation threshold, the path is determined to be a fault causal chain.

[0082] Specifically, the contextualized event subgraph is generated in step S3. Event nodes, fault phenomenon nodes, and parameter anomaly nodes in the contextualized event subgraph are set as activation nodes, and the initial activation strength is calculated based on semantic similarity and anomaly severity. For the first... There are 10 active nodes:

[0083] In the formula, Indicates the first The initial activation strength of each activated node; Indicates the first The semantic similarity of each activated node; Indicates the first The severity of the anomaly corresponding to each activated node; This represents the semantic similarity weight.

[0084] Starting from the activated node, the system traverses along the directed edges in the fused knowledge graph. These directed edges include edges pointing from the fault phenomenon to the fault cause, edges pointing from the fault cause to the fault mode, and edges pointing from the fault mode to the maintenance process. During traversal, the activation strength is propagated based on edge weights and relationship credibility. For the propagation between two adjacent nodes in an activation path, we have:

[0085] In the formula, Indicates the spread to the first Activation strength after each node; Indicates the first The activation strength of each node; Indicates the first Path weight of an edge; Indicates the first The reliability of the relationship between the edges.

[0086] For a candidate activation path, the cumulative activation intensity is:

[0087] In the formula, This represents the cumulative activation intensity of the candidate activation paths; Indicates the initial activation strength at the starting point of the path; This indicates that the propagation factors of each edge in the path are multiplied together.

[0088] When the cumulative activation intensity is greater than or equal to a preset fault activation threshold, the candidate activation path is determined to be a fault causal chain. A fault causal chain must include at least a real-time event node, a fault phenomenon node, a fault cause node, and a potential fault node. Graph path search employs a depth-limited weighted path search method, with the maximum search depth set to [value missing]. In the layer, visited nodes are marked during the search process to avoid repeated propagation of circular paths.

[0089] For example, the "high temperature of primary bearing" node in the contextualized event subgraph is set as the active node, and the semantic similarity is... abnormal severity Semantic similarity weight ,but:

[0090] The fused knowledge graph contains the path "Level 1 bearing high temperature → insufficient lubrication → bearing wear → bearing failure", with corresponding edge weights as follows: , and The relationship credibility is as follows: , and ,but:

[0091] If the preset fault activation threshold is If so, the path is determined to be a fault causal chain, and the potential fault node is "bearing failure".

[0092] S5. Based on the fault causal chain, drive the large language model to generate a self-evolving diagnostic report containing fault causes, related evidence, and maintenance suggestions. In one embodiment of the present invention, step S5 includes the following steps: The cause-and-effect chain of the failure is broken down into a series of logically clear diagnostic steps; For each diagnostic step, the original text source of the node in the fusion knowledge graph is matched as diagnostic evidence. The diagnostic steps and diagnostic evidence are organized into natural language prompts, input into a large language model, and an self-evolving diagnostic report is generated.

[0093] Specifically, the fault causal chain is generated in step S4. When parsing the fault causal chain, real-time event nodes, fault phenomenon nodes, fault cause nodes, potential fault nodes, and maintenance process nodes are read sequentially according to the path direction, and the edge relationships between adjacent nodes are converted into diagnostic steps. For the first... The diagnostic steps include:

[0094] In the formula, Indicates the first One diagnostic step; Indicates the current node; This indicates the relationship between the current node and the next node; Indicates the next node; This indicates the path confidence level corresponding to the diagnostic steps.

[0095] When matching diagnostic evidence for each diagnostic step, the current node, next node, and edge relationships in the diagnostic step are used as search keys to retrieve node attributes and edge attributes from the fused knowledge graph, yielding the original text source. The original text source includes the source document name, chapter number, paragraph number, and original text fragment. For the first... The diagnostic evidence corresponding to each diagnostic step includes:

[0096] In the formula, Indicates the first Diagnostic evidence corresponding to each diagnostic step; , , , The sequence represents the source document name, chapter number, paragraph number, and original text fragment, respectively.

[0097] The diagnostic steps and evidence are organized into natural language prompts, which include the current state of the equipment, the fault causal chain, diagnostic evidence, report generation constraints, and output format constraints. After receiving the natural language prompts, the large language model generates a self-evolving diagnostic report. For the self-evolving diagnostic report, the following applies:

[0098] In the formula, This indicates an evolutionary diagnostic report; Representing a large language model; Natural language prompts that consist of diagnostic steps and diagnostic evidence.

[0099] The self-evolving diagnostic report includes the cause of the failure, related evidence, and maintenance recommendations. The cause of the failure is generated from the cause node and potential failure node in the failure causal chain; the related evidence is generated from the original text source; and the maintenance recommendations are generated from the maintenance process node in the failure causal chain. During the generation process, the large language model organizes the report content only based on the diagnostic steps and diagnostic evidence. For content in the fused knowledge graph that does not provide an original text source, it is marked as an item requiring manual confirmation.

[0100] For example, the fault causal chain is "high temperature of primary bearing → insufficient lubrication → bearing wear → bearing failure". The path nodes in the fault causal chain are denoted as follows: , , and ,in This indicates the "high temperature of first-level bearings" node. This indicates a node with insufficient lubrication. Indicates the "bearing wear" node. The node representing "bearing failure" is denoted as . The edge relationships in the fault causal chain are denoted as follows: , and ,in This indicates the relationship caused by high temperature. Indicates that a relationship has been established. This indicates a causal relationship. The system resolves the fault causal chain into three diagnostic steps based on the path direction:

[0101]

[0102]

[0103] In this embodiment, let , , The overall path confidence of the fault causal chain is:

[0104] because Greater than the preset diagnostic confidence threshold Therefore, determining the causal chain of the fault can be used to generate a self-evolving diagnostic report.

[0105] When matching diagnostic evidence for each diagnostic step, the current node, edge relationship, and next node in the diagnostic step are used as search keys to retrieve node attributes and edge attributes from the fused knowledge graph. For the first diagnostic step... The search key "high temperature of primary bearing - cause of high temperature - insufficient lubrication" yielded the following diagnostic evidence:

[0106] For the second diagnostic step The search key "insufficient lubrication - leading to - bearing wear" yielded the following diagnostic evidence:

[0107] For the third diagnostic step The search key "bearing wear - leading to - bearing failure" yielded the following diagnostic evidence:

[0108] The system organizes diagnostic steps and diagnostic evidence into natural language prompts, which can be represented as:

[0109] In the formula, This indicates report format constraints, which include the cause of the failure, related evidence, maintenance recommendations, risk level, and items to be confirmed.

[0110] Natural language prompts After inputting the large language model, a self-evolutionary diagnostic report is generated:

[0111] The self-evolutionary diagnostic report states that the cause of the failure is "a high-confidence correlation exists between high temperature in the primary bearing and insufficient lubrication. Insufficient lubrication further leads to bearing wear and may develop into bearing failure." Supporting evidence includes "C-203 Compressor Repair Manual, Chapter 4, Section 2, Paragraph 3," "C-203 Compressor Historical Work Order No. 20240512031," and "C-203 Compressor Overhaul Procedures, Chapter 6, Section 1." Maintenance recommendations include "checking lubricating oil pressure, replenishing lubricating oil, detecting bearing clearance, and replacing the bearing according to its wear level." The risk level is "medium-high risk." The items pending confirmation are "on-site re-measurement of lubricating oil pressure and bearing vibration value."

[0112] Therefore, the system generates a self-evolving diagnostic report with evidence source constraints based on the fault causal chain and outputs the self-evolving diagnostic report to the maintenance personnel's terminal. After the maintenance personnel confirm the fault cause and maintenance recommendations, they generate a maintenance confirmation signal, which is used to dynamically adjust the edge weights in the fault causal chain.

[0113] S6. Receive the maintenance confirmation signal for the self-evolutionary diagnostic report, and dynamically adjust the path weights that constitute the fault causal chain in the fused knowledge graph according to the maintenance confirmation signal to obtain the updated fused knowledge graph. In one embodiment of the present invention, step S6 includes the following steps: Receive and parse maintenance confirmation signals, classifying them as correct diagnostic signals or incorrect diagnostic signals; If the signal is a correct diagnosis, increase the weight of all edges in the fault causal chain; If the signal is a false diagnosis, the weight of all edges in the fault causal chain is reduced, thereby achieving dynamic correction of knowledge.

[0114] Specifically, the maintenance confirmation signal is generated from feedback information from maintenance personnel, the closed-loop results of the maintenance work order, and equipment retest data. For the first... There are several maintenance confirmation signals:

[0115] In the formula, Indicates the first A maintenance confirmation signal; This indicates the corresponding self-evolutionary diagnostic report number; Indicate the actual cause of the malfunction; Indicates actual maintenance actions; This indicates the status of the equipment after maintenance.

[0116] Upon receiving a maintenance confirmation signal, the system associates the corresponding fault causal chain based on the self-evolving diagnostic report number. It then compares the fault cause in the self-evolving diagnostic report with the actual fault cause, and compares the maintenance recommendations with the actual maintenance actions. Simultaneously, it determines whether the equipment has returned to normal based on the post-maintenance equipment status. If the fault cause matches and the equipment has returned to normal, the maintenance confirmation signal is classified as a correct diagnostic signal; if the fault cause is inconsistent or the equipment has not returned to normal, the maintenance confirmation signal is classified as an incorrect diagnostic signal.

[0117] For a correct diagnostic signal, increase the weights of all edges in the fault causal chain according to the positive update rule:

[0118] In the formula, This represents the updated edge weights; This indicates the edge weights before the update; Indicates the upper limit of edge weight; This indicates a positive update step size; This indicates the confidence level of a correct diagnostic signal.

[0119] For incorrect diagnosis signals, reduce the weight of all edges in the fault causal chain according to the negative update rule:

[0120] In the formula, Indicates the lower bound of the edge weight; Indicates a negative update step size; This indicates the confidence level of the error diagnosis signal.

[0121] The confidence level of a correct diagnosis signal is obtained by weighting the consistency of fault causes, the consistency of maintenance actions, and the degree of post-maintenance state recovery; the confidence level of a wrong diagnosis signal is obtained by weighting the degree of difference in fault causes, the degree of deviation in maintenance actions, and the degree of anomaly persistence. After the weight adjustment is completed, the updated edge weights are written back to the fused knowledge graph to obtain the updated fused knowledge graph.

[0122] For example, a self-evolutionary diagnostic report number is The fault causal chain corresponding to the self-evolutionary diagnostic report is "high temperature of primary bearing → insufficient lubrication → bearing wear → bearing failure". The fault causal chain includes three edges: "high temperature of primary bearing → insufficient lubrication", "insufficient lubrication → bearing wear", and "bearing wear → bearing failure", with the following weights before the update: , and .

[0123] After completing on-site repairs, maintenance personnel send back a maintenance confirmation signal through the operation and maintenance management system. For the first... There are several maintenance confirmation signals:

[0124] In the formula, Indicates the first A maintenance confirmation signal; This indicates the corresponding self-evolutionary diagnostic report number; The value is "insufficient lubrication"; The value is "Replenish lubricating oil and check bearing wear"; The value is set to "alarm cleared and bearing temperature returned to normal".

[0125] Furthermore, maintain the temperature of the preceding bearing at [temperature value missing]. The temperature of the first-stage bearing after maintenance is The alarm status changed from "high temperature alarm" to "alarm cleared". The system will then determine the actual cause of the fault. The fault cause consistency score is obtained by comparing it with the fault cause in the self-evolutionary diagnostic report. Actual maintenance actions The maintenance action consistency score is obtained by comparing the maintenance recommendations with those in the self-evolutionary diagnostic report. Based on the equipment status after maintenance Assess the degree of equipment recovery to obtain a status recovery score. .

[0126] Set the consistency score weight for fault causes. Weighting of consistency score for maintenance actions State recovery score weight The confidence level of the correct diagnostic signal is:

[0127] Consistency score due to fault cause And the equipment status after maintenance The maintenance confirmation signal has been sent to indicate that "the alarm has been cleared and the bearing temperature has returned to normal". This is classified as a correct diagnostic signal.

[0128] Set upper limit for edge weights Positive update step size Correctly diagnose signal confidence The updated weight of the first edge "High temperature of primary bearing → Insufficient lubrication" is:

[0129] The updated weight for the second edge, "Insufficient lubrication → Bearing wear", is:

[0130] The updated weight of the third edge "Bearing Wear → Bearing Failure" is:

[0131] Therefore, the path weights of the three edges in the fault causal chain are respectively determined by... , and Updated to , and The system writes the updated edge weights back to the fused knowledge graph, which increases the activation intensity of the path "high temperature of primary bearing → insufficient lubrication → bearing wear → bearing failure" in subsequent graph path searches, thereby obtaining an updated fused knowledge graph.

[0132] S7. Based on the fused knowledge graph, simulate possible future changes in equipment operation commands or environmental parameters to generate simulated event entities; map the simulated event entities to the fused knowledge graph, perform a forward-looking graph path search, and predict potential fault causal chains that may be triggered; generate risk warning information based on potential fault causal chains. In one embodiment of the present invention, step S7 includes the following steps: Based on the fusion of knowledge graphs, simulate possible future changes in device operation commands or environmental parameters to generate simulated event entities; Simulated event entities are mapped to a fused knowledge graph to perform a forward-looking graph path search and predict potential fault causal chains that may be triggered. Risk warning information is generated based on the causal chain of potential failures.

[0133] Specifically, equipment operation instructions are provided by the production planning system and equipment control plan, while environmental parameter changes are provided by the environmental monitoring system and equipment operation prediction data. Future operating scenarios are generated based on the equipment operation instructions and environmental parameter changes, and these future operating scenarios are then converted into simulated event entities. For the first... There are several simulated event entities:

[0134] In the formula, Indicates the first One simulated event entity; , , , , , These represent, in order, simulation time, simulation components, simulation parameters, simulation parameter values, simulation state description, and simulation event type.

[0135] When mapping simulated event entities to the fused knowledge graph, the same semantic encoding and cosine similarity calculation method as in step S3 are used to determine the simulated activation nodes corresponding to the simulated event entities. Starting from the simulated activation nodes, a prospective graph path search is performed in the fused knowledge graph. Building upon the activation intensity propagation in step S4, the prospective graph path search incorporates the probability of future running scenarios and environmental impact coefficients. For simulated paths, we have:

[0136] In the formula, This represents the cumulative activation intensity of the simulated path; Indicates the probability of the future operating scenario occurring; Indicates the environmental impact coefficient; This indicates the initial activation strength of the simulated active node; Indicates the first Path weight of an edge; Indicates the first The reliability of the relationship between the edges.

[0137] When the cumulative activation intensity of the simulated path is greater than or equal to the preset risk warning threshold, the simulated path is identified as a potential fault causal chain. Risk warning information is generated based on the potential fault causal chain, including the risky device, risky component, potential fault node, triggering conditions, prediction time window, risk level, and recommended early maintenance actions.

[0138] For example, the production plan shows that the C-203 compressor will be used in the future. Within hours Load increased to The load and environmental monitoring system predict that the workshop temperature will rise to The system generates simulated event entities:

[0139] The simulated event entity is mapped to the "Level 1 Bearing High Temperature" node in the fused knowledge graph. Assume the probability of the future operating scenario occurring. Environmental impact coefficient Simulate the initial activation strength of the activated node. The result of the multiplication of the path propagation factors is ,but:

[0140] If the preset risk warning threshold is The "high temperature of primary bearing → insufficient lubrication → bearing wear → bearing failure" chain is identified as a potential fault causal chain, and a risk warning message is generated, prompting the user to check the lubricating oil pressure, bearing temperature and bearing vibration status before high-load operation.

[0141] S8. Extract fault causal chains from the updated fused knowledge graph that have undergone multiple positive verifications and have significantly increased weights to form a general fault mode template; apply the general fault mode template to the knowledge graph initialization or update process of other industrial equipment of the same type to achieve cross-device knowledge transfer and sharing.

[0142] In one embodiment of the present invention, step S8 includes the following steps: From the updated fusion knowledge graph, fault causal chains that have undergone multiple positive verifications and have significantly increased weights are extracted to form a general fault mode template. By applying the general failure mode template to the knowledge graph initialization or update process of other industrial equipment of the same type, knowledge transfer and sharing across devices can be achieved.

[0143] Specifically, the updated fused knowledge graph is obtained in step S6. The system scans the updated fused knowledge graph according to a preset cycle, and counts the number of positive verifications, the growth rate of edge weights, and the diagnosis success rate for each fault causal chain. The number of positive verifications is determined based on the number of times the maintenance confirmation signals are classified as correct diagnostic signals; the growth rate of edge weights is determined based on the difference between the current edge weight and the initial edge weight; and the diagnosis success rate is determined based on the ratio of the number of correct diagnoses to the total number of diagnoses.

[0144] For the There are several candidate fault causal chains:

[0145] In the formula, Indicates the first Candidate fault causal chain; Represents a sequence of path nodes; Indicates the number of positive validations; This indicates the average increase in edge weights; This indicates the success rate of diagnosis.

[0146] The template score is calculated based on the number of positive validations, the average increase in edge weights, and the diagnostic success rate.

[0147] In the formula, Indicates template scoring; This represents the number of positive verifications after normalization. This represents the increase in the average edge weight after normalization. , , These represent the scoring weights, and .

[0148] When the template score is greater than or equal to the preset template extraction threshold, the candidate fault causal chain is extracted into a general fault mode template. The general fault mode template includes applicable equipment type, applicable component type, typical abnormal phenomenon, typical fault cause, typical fault mode, typical maintenance action, and template path weight.

[0149] When applying a general failure mode template to the knowledge graph initialization or update process of other industrial equipment of the same type, the system first determines whether the target industrial equipment meets the template's applicability conditions based on the equipment type, component structure, and measurement point configuration. If the template's applicability conditions are met, the nodes and edges in the general failure mode template are written into the target industrial equipment's knowledge graph. If the corresponding nodes and edges already exist in the target industrial equipment's knowledge graph, the corresponding edge weights are updated according to the template path weights. If the corresponding nodes or edges do not exist, the corresponding nodes and edges are created, thereby achieving cross-device knowledge transfer and sharing.

[0150] For example, in continuous During the month-long maintenance loop, the candidate fault causal chain "bearing high temperature → insufficient lubrication → bearing wear → bearing failure" was... The secondary maintenance confirmation signal was verified as a correct diagnostic signal, and the average edge weight was adjusted accordingly. Upgraded to The diagnostic success rate is Let the normalized number of positive validations be... The growth rate of the normalized average side weight Diagnostic success rate Rating weight , , ,but:

[0151] If the preset template extraction threshold is Then, the causal chain of the candidate fault is extracted into a general fault mode template. For another compressor of the same model, if it includes bearing temperature measurement points, lubrication system and bearing vibration measurement points, the general fault mode template is written into the knowledge graph initialization process of the compressor of the same model, so that the compressor of the same model can have predictive maintenance knowledge of bearing high temperature related faults in the initial stage.

[0152] Based on the same inventive concept, such as Figure 2 As shown, this invention also provides a self-evolving predictive maintenance system for industrial equipment based on a large language model and knowledge graph. The system includes: The data acquisition and parsing module is used to acquire real-time equipment operation logs and unstructured operation and maintenance manuals of industrial equipment, and to process them using a large language model to generate real-time event entities and operation and maintenance knowledge triples. The knowledge graph construction and fusion module is used to build and maintain a fused knowledge graph based on device structured data and operation and maintenance knowledge triples. The State Mapping and Contextualization module is used to map real-time event entities to a fused knowledge graph using a large language model, and to construct a contextualized event subgraph. The causal reasoning and diagnosis module is used to perform path search based on contextualized event subgraphs in the fused knowledge graph, extract fault causal chains, and drive the large language model to generate self-evolving diagnostic reports. The knowledge self-evolution module is used to receive maintenance confirmation signals and dynamically adjust the weights of relevant paths in the fused knowledge graph based on these signals, thereby achieving closed-loop knowledge updates.

Claims

1. A self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs, characterized in that, Includes the following steps: The system obtains real-time equipment operation logs and unstructured operation and maintenance manuals of industrial equipment, extracts real-time event entities from the equipment operation logs using a large language model, and extracts operation and maintenance knowledge triples from the unstructured operation and maintenance manuals. Based on the pre-set structured equipment data, a basic knowledge graph of the equipment is initialized and constructed. The operation and maintenance knowledge triples are then fused with the basic knowledge graph of the equipment to generate a fused knowledge graph that includes the relationship between equipment components, fault modes and maintenance processes. The semantic similarity between the real-time event entity and the nodes in the fused knowledge graph is calculated using the large language model. Based on the similarity results, the real-time event entity is mapped to the fused knowledge graph to construct a contextualized event subgraph that reflects the current device status. In the fused knowledge graph, a graph path search is performed starting from the nodes in the contextualized event subgraph to identify and extract the fault causal chain pointing to the potential fault node. Based on the fault causal chain, the large language model is driven to generate a self-evolving diagnostic report containing the fault cause, related evidence, and maintenance recommendations; The system receives a maintenance confirmation signal for the self-evolutionary diagnostic report and dynamically adjusts the path weights that constitute the fault causal chain in the fused knowledge graph based on the maintenance confirmation signal to obtain an updated fused knowledge graph.

2. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to claim 1, characterized in that, The steps of extracting real-time event entities from the device operation log using a large language model and extracting operation and maintenance knowledge triples from the unstructured operation and maintenance manual include: The device operation log is parsed in terms of time sequence and content. The named entity recognition function of the large language model is used to identify key information such as time, components, parameters and status descriptions, forming structured real-time event entities. The unstructured operation and maintenance manual is divided into chapters and paragraphs. The relation extraction function of the large language model is used to extract entity relation pairs and construct the operation and maintenance knowledge triples.

3. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to claim 1, characterized in that, The step of fusing the operation and maintenance knowledge triples with the device basic knowledge graph to generate a fused knowledge graph includes: The entities in the aforementioned maintenance knowledge triple are standardized to obtain standard entities; Align the standard entity with the nodes in the device basic knowledge graph. If a matching node exists, establish a link; otherwise, create a new node. By using the relationships in the operation and maintenance knowledge triples as edges to connect the aligned or created nodes, the operation and maintenance knowledge is integrated into the device basic knowledge graph, thus completing the construction of the fused knowledge graph.

4. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to claim 1 or 3, characterized in that, The step of calculating the semantic similarity between the real-time event entity and the nodes in the fused knowledge graph using the large language model includes: The large language model is used to encode the textual description information in the real-time event entity into an event feature vector; The attribute information of nodes in the fused knowledge graph is encoded into node feature vectors using the large language model. The semantic similarity is obtained by calculating the cosine similarity between the event feature vector and each node feature vector in the same vector space.

5. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to claim 1, characterized in that, The step of performing graph path search in the fused knowledge graph, starting from nodes in the contextualized event subgraph, to identify and extract fault causal chains pointing to potential fault nodes, includes: The nodes in the contextualized event subgraph are set as active nodes and assigned an initial activation strength; Starting from the activated node, traverse along the directed edges in the fused knowledge graph, propagate the activation intensity according to the edge weights, and form an activation path; When the cumulative activation intensity of the activation path exceeds the preset fault activation threshold, the path is determined to be the fault causal chain.

6. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to claim 1, characterized in that, The step of driving the large language model to generate a self-evolving diagnostic report based on the fault causal chain includes: The fault causal chain is analyzed into a series of logically clear diagnostic steps; For each diagnostic step, the original text source of the node in the fused knowledge graph is matched as diagnostic evidence; The diagnostic steps and diagnostic evidence are organized into natural language prompts, which are then input into the large language model to generate the self-evolutionary diagnostic report.

7. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to claim 1, characterized in that, The step of dynamically adjusting the path weights constituting the fault causal chain in the fused knowledge graph based on the maintenance confirmation signal includes: Receive and parse the maintenance confirmation signal, and classify it as a correct diagnostic signal or an incorrect diagnostic signal; If the diagnosis signal is correct, then increase the weight of all edges in the fault causal chain; If the signal is a fault diagnosis, the weight of all edges in the fault causal chain is reduced, thereby achieving dynamic correction of knowledge.

8. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to any one of claims 1-7, characterized in that, After identifying and extracting the causal chain of failures pointing to potential faulty nodes, the process also includes: Based on the fused knowledge graph, simulate possible future changes in device operation instructions or environmental parameters to generate simulated event entities; The simulated event entities are mapped to the fused knowledge graph, and a forward-looking graph path search is performed to predict potential fault causal chains that may be triggered. Risk warning information is generated based on the potential fault causal chain.

9. The self-evolving predictive maintenance method for industrial equipment based on large language models and knowledge graphs according to any one of claims 1-7, characterized in that, Following the updated fusion knowledge graph, it also includes: From the updated fused knowledge graph, fault causal chains that have undergone multiple positive verifications and have significantly increased weights are extracted to form a general fault mode template. The general fault mode template can be applied to the knowledge graph initialization or update process of other industrial equipment of the same type to achieve cross-device knowledge transfer and sharing.

10. A self-evolving predictive maintenance system for industrial equipment based on large language models and knowledge graphs, characterized in that: The system includes: The data acquisition and parsing module is used to acquire real-time equipment operation logs and unstructured operation and maintenance manuals of industrial equipment, and to process them using a large language model to generate real-time event entities and operation and maintenance knowledge triples. The knowledge graph construction and fusion module is used to construct and maintain a fused knowledge graph based on the device structured data and the operation and maintenance knowledge triples. The state mapping and contextualization module is used to map the real-time event entities to the fused knowledge graph using a large language model, and to construct a contextualized event subgraph. The causal reasoning and diagnosis module is used to perform path search based on the contextualized event subgraph in the fused knowledge graph, extract the fault causal chain, and drive the large language model to generate a self-evolving diagnostic report. The knowledge self-evolution module is used to receive maintenance confirmation signals and dynamically adjust the weights of relevant paths in the fused knowledge graph according to the signals, so as to realize closed-loop update of knowledge.

Images