Equipment fault diagnosis method, device and storage medium based on large model, causal library and knowledge graph

Abstract

The application relates to the technical field of equipment fault detection, and discloses an equipment fault diagnosis method and device based on a large model, a causal library and a knowledge graph and a storage medium, which comprises the following steps: pre-processing a plurality of abnormal channel data of an industrial equipment sensor, screening out other abnormal channel fault information, obtaining historical fault data, learning a causal library through a Bayesian network, and tracing a same fault cause through a multivariate causal relationship; correcting confidence according to the other abnormal channel fault information to form a fault list, and the like, a directed acyclic graph of a causal relationship between storage equipment and variables is constructed, a reverse depth-first traversal and a multidimensional confidence calculation and correction mechanism are combined, accurate fault root cause tracing and quantitative sorting are realized, a fault knowledge graph is constructed in a triple structure, accurate association of fault causes, fault phenomena and maintenance schemes is realized, and the limitation of single technical diagnosis is solved.

Description

Technical Field

[0001] This invention relates to the field of equipment fault detection technology, specifically to equipment fault diagnosis methods, equipment, and storage media based on large models, causal libraries, and knowledge graphs. Background Technology

[0002] As the core carrier of manufacturing production, energy supply, and infrastructure operation, the operational stability of industrial equipment directly determines production efficiency, operational safety, and economic benefits. In particular, major equipment such as steam turbines, fans, and high-speed rail bearings are prone to failure due to wear and tear of parts, circuit faults, and abnormal parameters under extreme service conditions and complex operating environments. In severe cases, this may lead to equipment shutdown, production interruption, or even safety accidents and huge economic losses.

[0003] Currently, fault diagnosis in industrial equipment mainly relies on traditional diagnostic methods and preliminary intelligent diagnostic schemes. However, these methods still have many insurmountable shortcomings in practical applications. First, traditional fault diagnosis methods are primarily based on manual experience and periodic maintenance, overly dependent on the professional skills of maintenance personnel, resulting in low diagnostic efficiency, inaccurate root cause localization, and strong subjectivity. On the one hand, for complex equipment with multi-channel and multi-variable coupling, the mapping relationship between fault phenomena and root causes is complex, making it difficult for humans to quickly sort out the correlation between multiple anomalies, easily leading to misdiagnosis and missed diagnosis. On the other hand, the periodic maintenance model is blind, with a large number of equipment being replaced before reaching the end of their service life, resulting in resource waste and economic burden, while some potential early faults are difficult to detect in time, leading to the expansion of faults.

[0004] Secondly, existing intelligent diagnostic methods are mostly based on single technologies, lacking the synergistic advantages of multi-technology integration. Some solutions rely solely on data-driven machine learning models to identify faults by analyzing data collected by sensors. However, these methods lack in-depth exploration of fault mechanisms, making it difficult to reveal the inherent laws governing fault initiation and evolution. Furthermore, they suffer from poor compatibility with multi-source heterogeneous data (such as binary streams, CSV, and real-time data streams), low data preprocessing efficiency, and susceptibility to noisy data, resulting in insufficient diagnostic accuracy. Other solutions build fault correlations based on knowledge graphs, but lack causal reasoning capabilities, failing to accurately trace the root cause of faults. They can only match surface fault phenomena with maintenance solutions, making it difficult to address multi-factor coupled faults in complex equipment.

[0005] Furthermore, in existing technologies, fault diagnosis, root cause tracing, and maintenance plan generation are mostly independent steps, lacking fully automated integration. Multi-format anomaly data collected by sensors is difficult to standardize, the confidence assessment of fault root causes lacks scientific quantitative basis, maintenance plan generation lacks systematic and standardized guidance, and standardized maintenance reports cannot be automatically generated and archived after diagnosis. This results in inefficient fault diagnosis processes that fail to meet the needs of large-scale, intelligent operation and maintenance of industrial equipment. Summary of the Invention

[0006] To address the aforementioned shortcomings and deficiencies in existing technologies, this invention provides a method, device, and storage medium for equipment fault diagnosis based on a large model, causal database, and knowledge graph, comprising the following steps: Step 1: Preprocess the abnormal channel data collected by the sensors of industrial equipment, convert the abnormal channel data of different formats into a unified structured time-series data format, and output the fault information of each abnormal channel. Step 2: Filter out fault information related to the channel under test Related information on other abnormal channel faults; obtain historical fault data based on existing equipment maintenance manuals and historical fault maintenance records, and learn a causal library from the historical fault data through Bayesian network. The causal library stores a directed acyclic graph (DAG) of causal relationships between and within each device. The same fault cause is traced through multivariate causal relationships. Starting from the node corresponding to the fault information of the channel under test, a reverse depth-first traversal (DFS) is performed in the DAG of the causal library to find the initial confidence of each parent node as the root cause of the fault. The confidence is then corrected based on the fault information of other abnormal channels. The corrected confidence is sorted from high to low, and the top 5 most likely fault causes are selected to form a fault list. Step 3: Construct a fault graph using a triplet structure based on existing equipment maintenance manuals, historical fault maintenance records, and equipment manufacturer technical documents; map each fault cause node in the fault list output in Step 2 to a fault cause entity in the knowledge graph, query the fault phenomenon entity and maintenance index entity corresponding to the fault cause entity, generate a structured solution, and output a set of maintenance solutions sorted by confidence. Step 4: Preset the repair report template and repair report output format, extract and integrate the data sources output from Step 1, Step 2 and Step 3, fill in and format the report content according to the repair report template, and complete the conversion and output of the single event repair report of the repair plan according to the preset repair report output format requirements and archive it.

[0007] In step 1, the abnormal channel data exists in the form of unstructured binary stream, comma-separated value CSV, or real-time data stream, containing channel number, channel name, timestamp, sample value, and channel image information; Preprocessing of multiple abnormal channel data collected by industrial equipment sensors includes data cleaning, format standardization, and missing value imputation.

[0008] Other abnormal fault channel information in step 2 Where A is the set of indices of all abnormal channels in this event, and j is the index of the abnormal channel in set A other than i. This represents the fault information for the j-th abnormal channel; the causal database is G=(V,E), where: It is a set of nodes, which includes device nodes and variable nodes. Device nodes include complete equipment and equipment components; variable nodes include temperature, pressure, current, etc. Let be a set of directed edges, each edge Indicates from node To the node The causal relationship.

[0009] In step 2, the node corresponding to the fault information of the channel under test Starting from the causal database, perform a reverse depth-first traversal (DFS) in the DAG to find all possible causes. abnormal parent node set And record each parent node to Path length For each parent node Calculate its initial confidence level as a root cause of the failure. : ; in: For the edge The causal strength; parent node To the node Path length; parent node The frequency of root causes in historical fault data is obtained from statistics in the historical fault database; All are weight coefficients, satisfying α+β+γ=1, and are obtained through grid search or reinforcement learning optimization; The confidence level is corrected by analyzing fault information from multiple other abnormal channels. If the parent nodes of these other abnormal channels all point to the parent node... Then for Perform weighted boosting: where k points to the parent node. The number of abnormal channels, where λ is the boosting factor (0 < λ < 1), used to strengthen the confidence that multiple abnormal channels point to the same root cause; Based on the corrected confidence level Sort the nodes from highest to lowest, select the top 5 as the most likely fault phenomena and causes, and form a fault list. : Each entry contains a root cause node. Fault symptoms and the corresponding corrected confidence level. .

[0010] The fault map in step 3 adopts a triplet structure, which mainly includes: Fault Cause Entity This includes sensor damage and poor wiring contact; Fault phenomenon entity This includes abnormal channel values ​​and image distortion; Repair Index Entity This includes replacing sensors and rewiring; There are three types of logical relationships between entities: cause, correspondence, and inclusion; Among them, it leads to the causal relationship between the entity representing the cause of the failure and the entity representing the phenomenon of the failure; This corresponds to the relationship between the fault phenomenon entity and the maintenance index entity in terms of the solution association; It contains hierarchical relationships between entities.

[0011] Step 3 specifically involves: adding each root cause node in the fault list... Mapped to the fault cause entity in the knowledge graph ,like Description and In a knowledge graph, if the cosine similarity of the standardized name or semantic tag of an entity is greater than the threshold θ, then the entity is determined to be the same entity. Fault symptom query: Search for related information in the knowledge graph. There exist entities that cause failures in the relationship. That is, to find all that satisfy ⟨ ,lead to, > triples; Repair Index Query: Querying within the knowledge graph Maintenance index entities with corresponding relationships exist That is, to find all that satisfy ⟨ ,correspond, > triples; For each entity causing the failure Integrate the corresponding fault phenomenon entities With maintenance index entity Generate a detailed repair plan, including: Repair steps: Extract relevant information from the knowledge graph. The associated standardized operating procedures are arranged in the order of operation; Required tools: querying from knowledge graphs and Related tool models and specifications; Spare part model: Searched from the knowledge graph Related spare parts models and specifications; Estimated working hours: The average working hours for similar maintenance tasks, obtained from historical fault repair records. Safety Precautions: Safety guidelines relevant to the current maintenance operation extracted from the equipment maintenance manual; The structured representation of each scheme is as follows: ; in The cause of the malfunction, This is a fault phenomenon. For repair procedures, For the necessary tools, For spare parts model, To estimate working hours, Safety precautions; The solutions are then sorted by confidence level from highest to lowest, with the repair solution with the highest confidence level being recommended first.

[0012] The maintenance report template includes: summary unit, anomaly detection unit, image analysis unit, fault tracing unit, maintenance plan unit, and appendix unit; the maintenance report output formats include PDF, HTML, and JSON.

[0013] The equipment fault diagnosis device based on a large model, causal database, and knowledge graph includes a fault event determination module, a causal tracing module, a maintenance plan generation module, and a maintenance report generation module. The fault event determination module identifies possible fault phenomena and causes in the tested channel. The causal tracing module, combining causal relationships between variables, traces the source of the same fault cause and phenomenon, identifying the five most likely causes. The maintenance plan generation module, using the knowledge graph, searches for corresponding maintenance indexes and generates maintenance plans. The maintenance report generation module generates a single-event maintenance report, including anomaly detection results, an image of the abnormal tested channel, equipment information, physical meaning description, function description, possible fault phenomena, causes, and corresponding maintenance plans.

[0014] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements an equipment fault diagnosis method based on a large model, a causal library, and a knowledge graph.

[0015] An electronic device includes at least one processor and at least one memory connected to the processor, the processor being configured to invoke program instructions in the memory to execute an equipment fault diagnosis method based on a large model, causal library, and knowledge graph.

[0016] The present invention has the following beneficial effects: 1. This invention achieves standardized conversion of heterogeneous data such as unstructured binary streams, CSV files, and real-time data streams through multi-source anomaly channel data preprocessing. Combined with data cleaning and missing value imputation, it ensures the integrity and accuracy of fault information. By building a causal database and using reverse DFS traversal, confidence calculation, and correction mechanisms, it achieves accurate tracing and quantitative ranking of fault root causes, effectively avoiding the subjectivity, misjudgment, and omissions of traditional manual diagnosis, significantly improving the accuracy and efficiency of fault diagnosis, and providing reliable preliminary support for subsequent maintenance work.

[0017] 2. This invention integrates a large model, a causal database, and a knowledge graph. By mining the inherent causal relationships between equipment and variables through the causal database, it addresses the deficiency of data-driven diagnostic methods in lacking support for fault mechanisms. Through the triple structure of the knowledge graph, it achieves accurate correlation between fault causes, fault phenomena, and maintenance solutions, making up for the lack of causal reasoning capabilities in existing knowledge graphs. By optimizing data preprocessing and semantic matching efficiency through the large model, it achieves synergistic effects of multiple technologies, which can efficiently cope with complex equipment faults involving multiple channels and multiple variables, thus broadening the scope of application for fault diagnosis.

[0018] 3. This invention constructs a fully automated system encompassing data preprocessing, root cause tracing, maintenance plan generation, and maintenance report archiving. Through a confidence correction mechanism based on a causal database, it provides a scientific quantitative basis for identifying root causes of faults, ensuring the rationality of maintenance plan priorities. By pre-setting standardized maintenance report templates and multi-format output requirements, it achieves standardized integration and archiving of diagnostic data, fault information, and maintenance plans, solving the problems of disconnected diagnostic processes and non-standardized procedures in existing technologies, and meeting the management needs of large-scale, intelligent operation and maintenance of industrial equipment. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the unit composition and process of the equipment fault diagnosis method based on a large model, causal database and knowledge graph of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail, as follows: Figure 1As shown, this embodiment takes the fault diagnosis of a large industrial steam turbine equipment as an example, and elaborates in detail the specific implementation process of the equipment fault diagnosis method, device, computer-readable storage medium and electronic equipment based on a large model, causal database and knowledge graph. The steam turbine equipment includes 30 monitoring channels such as temperature sensor, pressure sensor, and current sensor, which can collect abnormal data in multiple formats. The method described in this invention is applicable to realize the whole process fault diagnosis, including the following steps: Step 1: Preprocess the abnormal channel data collected by the industrial equipment sensors, convert the abnormal channel data of different formats into a structured time-series data format, and output the fault information of each abnormal channel. In this embodiment, the abnormal channel data collected by the industrial equipment sensors exists in three formats: unstructured binary stream (temperature sensor channels 1~10), comma-separated value CSV (pressure sensor channels 11~20), and real-time data stream (current sensor channels 21~30). All abnormal channel data includes channel number, channel name, timestamp, sampled value, and channel image information.

[0021] Furthermore, preprocessing is performed on the abnormal channel data collected by the sensors of industrial equipment, including data cleaning, format standardization, and missing value imputation; Data cleaning: Remove outliers that are significantly outside the measurement range, and use the 3σ criterion to identify and remove abnormal sampling points: If the sampling value of the i-th channel at time t is... satisfy If it is an outlier, it will be removed. This represents the average of historical normal data for this channel. Standard deviation; Format standardization: Converting abnormal channel data of different formats into a unified structured time-series data format. : ; in For the i-th channel at time... The sampled values, where m is the time series length of a single event; Missing value imputation: For missing sampling points caused by sensor disconnection or communication interruption, linear interpolation is used to imput them. in ,and , The adjacent sample values ​​are known.

[0022] Step 2: Filter out fault information related to the channel under test Related information on other abnormal channel faults; obtain historical fault data based on existing equipment maintenance manuals and historical fault maintenance records, and learn a causal library from the historical fault data through Bayesian network. The causal library stores a directed acyclic graph (DAG) of causal relationships between and within each device. The same fault cause is traced through multivariate causal relationships. Starting from the node corresponding to the fault information of the channel under test, a reverse depth-first traversal (DFS) is performed in the DAG of the causal library to find the initial confidence of each parent node as the root cause of the fault. The confidence is then corrected based on the fault information of other abnormal channels. The corrected confidence is sorted from high to low, and the top 5 most likely fault causes are selected to form a fault list. Furthermore, information on other abnormal fault channels. Where A is the set of indices of all abnormal channels in this event, and j is the index of the abnormal channel in set A other than i. For the fault information of the j-th abnormal channel; collect the equipment maintenance manual and historical fault maintenance records of the turbine for the past 5 years as historical fault data, and construct a causal library G=(V,E) through Bayesian network learning (using the HillClimbSearch algorithm of the pgmpy library). The causal library stores the directed acyclic graph (DAG) of causal relationships between and within each device, where: It is a set of nodes, which includes device nodes and variable nodes. Device nodes include complete equipment and equipment components; variable nodes include temperature, pressure, current, etc. Let be a set of directed edges, each edge Indicates from node To the node The causal relationship. For example (Unusually low cooling water pressure → abnormally high high-pressure cylinder temperature) (Cooling water pump current is abnormally low → Cooling water circuit pressure is abnormally low) (Abnormally high temperature in the intermediate-pressure cylinder → Abnormally high temperature in the high-pressure cylinder) etc.

[0023] Furthermore, the nodes corresponding to the fault information of the channel under test... Starting from the causal database, perform a reverse depth-first traversal (DFS) in the DAG to find all possible causes. abnormal parent node set And record each parent node to Path length For each parent node Calculate its initial confidence level as a root cause of the failure. : ; in: For the edge The causal strength; For nodes To the node Path length; For nodes The frequency of root causes in historical fault data is obtained statistically from the historical fault database; the weight coefficients α=0.4, β=0.3, γ=0.3, satisfying α+β+γ=1, are obtained through grid search (search range 0~1, step size 0.1) to ensure the rationality of confidence calculation; and are optimized through grid search or reinforcement learning (e.g., standard grid search, random grid search, Bayesian optimized grid search, Q-Learning, DQN, proximal policy optimization, multi-agent reinforcement learning). For example freq=0.25 freq=0.18 freq=0.15 freq=0.08 freq=0.06.

[0024] Substitute into the formula to calculate the initial confidence level: =0.4×0.85 + 0.3×(1 / 1) + 0.3×0.25 = 0.34 + 0.3 + 0.075 =0.715; =0.4×0.72 + 0.3×(1 / 1) + 0.3×0.18 = 0.288 + 0.3 + 0.054 =0.642; =0.4×0.68 + 0.3×(1 / 1) + 0.3×0.15 = 0.272 + 0.3 + 0.045 =0.617; =0.4×0.55 + 0.3×(1 / 2) + 0.3×0.08 = 0.22 + 0.15 + 0.024 =0.394; =0.4×0.50 + 0.3×(1 / 2) + 0.3×0.06 = 0.2 + 0.15 + 0.018 =0.368.

[0025] The confidence level is corrected by analyzing fault information from other abnormal channels. If the parent nodes of multiple abnormal channels all point to... Then for Perform weighted boosting: , where k is a pointer The number of abnormal channels, λ is the boosting factor = 0.1 (0 < λ < 1), used to strengthen the confidence that multiple abnormal channels point to the same root cause; Analysis revealed other abnormal channels. , , The parent nodes all point to That is, k=3, therefore only for The confidence level is adjusted accordingly: =0.715×(1+0.1×3)=0.715×1.3=0.9295; Other parent nodes do not have multiple abnormal channels pointing to them, and the confidence level remains unchanged.

[0026] Based on the corrected confidence level Sort the nodes from highest to lowest, select the top 5 as the most likely fault phenomena and causes, and form a fault list. : T={( The cooling water pressure is abnormally low, at 0.9295. The intermediate pressure cylinder temperature is abnormally high, 0.642°C. The high-pressure cylinder temperature sensor itself is faulty, 0.617. Turbine cylinder block cooling module failure, 0.394), ( Temperature acquisition line fault, 0.368); Each entry contains the root cause node of the fault, the fault phenomenon, and the corresponding corrected confidence value. .

[0027] Step 3: Construct a fault graph using a triplet structure based on existing equipment maintenance manuals, historical fault maintenance records, and equipment manufacturer technical documents; map each fault cause node in the fault list output in Step 2 to a fault cause entity in the knowledge graph, query the fault phenomenon entity and maintenance index entity corresponding to the fault cause entity, generate a structured solution, and output a set of maintenance solutions sorted by confidence. Knowledge Graph: A fault graph constructed based on the turbine's equipment maintenance manual, historical fault maintenance records, and equipment manufacturer's technical documents. It uses a triplet structure ⟨Subject-Relationship-Object> and mainly includes: Fault Cause Entity (e.g., "sensor damage" or "poor circuit contact"); Fault phenomenon entity (e.g., "abnormal channel values", "image distortion"); Repair Index Entity (e.g., "replace sensor" or "rewire"); the logical relationships between entities are of three types: cause, correspondence, and inclusion; Among them, it leads to the causal relationship between the entity representing the cause of the failure and the entity representing the phenomenon of the failure; This corresponds to the relationship between the fault phenomenon entity and the maintenance index entity in terms of the solution association; It contains hierarchical relationships between entities.

[0028] Each root cause node in the fault list Mapped to the fault cause entity in the knowledge graph ,like Description and In a knowledge graph, if the cosine similarity of the standardized name or semantic tag of an entity is greater than the threshold θ, then the entity is determined to be the same entity. Fault symptom query: Search for related information in the knowledge graph. There exist entities that cause failures in the relationship. That is, to find all that satisfy ⟨ ,lead to, > triples; Repair Index Query: Querying within the knowledge graph Maintenance index entities with corresponding relationships exist That is, to find all that satisfy ⟨ ,correspond, > triples; For each entity causing the failure Integrate the corresponding fault phenomenon entities With maintenance index entity Generate a detailed repair plan, including: Repair steps: Extract relevant information from the knowledge graph. The associated standardized operating procedures are arranged in the order of operation; Required tools: querying and interpreting knowledge graphs Related tool models and specifications; Spare part model: Searched from the knowledge graph Related spare parts models and specifications; Estimated working hours: The average working hours for similar maintenance tasks, obtained from historical fault repair records. Safety Precautions: Safety guidelines relevant to the current maintenance operation extracted from the equipment maintenance manual; The structured representation of each scheme is as follows: The cooling water pressure is abnormally low. Temperature is too high and pressure is too low 1. Stop the turbine and disconnect the power supply to the cooling water system; 2. Disconnect the cooling water pipe connections and check for blockages; 3. Use a high-pressure water jet to clear any blockages from the pipes; 4. Reinstall the pipe connections, reconnect the power supply, and start the cooling water pump; 5. Monitor the cooling water pressure and turbine temperature to confirm that the fault has been resolved. High-pressure water gun, wrench, pressure gauge Sealing gasket (Model: GB / T 12350-2017) 1.5 hours 1. Power must be disconnected before operation to prevent electric shock; 2. Protective gloves must be worn when disassembling pipes to prevent burns; 3. When clearing blockages, avoid debris entering the pipes, as this may affect the cooling effect.

[0029] Sort the solutions by confidence level from highest to lowest, generate a set of repair solutions, and prioritize the repair solution corresponding to "abnormally low cooling water pressure" with the highest confidence level.

[0030] Step 4: Preset the repair report template and repair report output format, extract and integrate the data sources output from Step 1, Step 2 and Step 3, fill in and format the report content according to the repair report template, and complete the conversion and output of the single event repair report of the repair plan according to the preset repair report output format requirements and archive it.

[0031] Furthermore, the data sources include: Anomaly detection results include: number of abnormal channels, channel number, channel name, anomaly type, and confidence level; Abnormal channel images: including preprocessed waveforms, magnified detail images, and spectrograms; Equipment information: including equipment model, installation location, operating time, and operating parameters; Physical meaning description: including the object being measured, the unit of measurement, and the description of its function; Fault information: including the top 5 most likely causes of the fault and their confidence levels; Repair plan: including detailed repair steps, repair tools, spare parts, repair time, and safety precautions.

[0032] Repair report templates include: Abstract Unit: This unit summarizes the core information of this incident, including the number of abnormal channels, the root cause of the fault with the highest confidence level obtained from the fault list after tracing the cause and effect database and adjusting the confidence level, and the recommended repair plan. Anomaly Detection Unit: Lists the number, name, anomaly type, confidence level, and targeted handling suggestions generated based on the fault map matching results of all anomaly channels in tabular form; Image analysis unit: Inserts abnormal images of the channel under test and marks the abnormal areas; Fault tracing unit: The cause-and-effect tracing path is displayed in the form of a flowchart, highlighting the confidence ranking of the 5 root causes of the fault in the fault list; Repair Solution Unit: Lists repair solutions for each fault according to priority, highlighting the detailed steps of the recommended solution; Appendix: Includes data fragments from the abnormal channels, statistical analysis results, and knowledge graph query records; Repair report output formats include PDF, HTML, and JSON. PDF format: Read the template, fill in the data for each unit, and generate a standardized PDF report. The naming format is "Equipment Name-Fault Diagnosis Date-Report Type.pdf" for easy printing, archiving, and on-site review. HTML format: Render template, populate data to generate HTML page, supports online viewing in browser, easy to display and interact with on the web; JSON format: Data sources (preprocessed data, fault lists, maintenance plans, etc.) are converted into JSON strings for easier subsequent system calls and data storage. This facilitates integration with the Enterprise Asset Management (EAM) system, enabling automatic data synchronization. The final output is a single-event maintenance report, which is then archived in the fault diagnosis database.

[0033] This invention also provides an equipment fault diagnosis device based on a large model, causal database, and knowledge graph, including a fault event determination module, a causal tracing module, a maintenance plan generation module, and a maintenance report generation module. The hardware configuration, software implementation, and operation process of each module are as follows: The module for determining the fault event to be tested: The hardware uses a data acquisition card, and the software uses Python to write data acquisition and preprocessing programs. It runs in the aforementioned hardware environment and its functions are to acquire abnormal sensor data, perform the preprocessing operation in step 1, output the fault information of each abnormal channel, and interact with other modules through the TCP / IP protocol. Causal tracing module: The hardware uses CPU (Intel Core i9-12900K) + memory (64GB), and the software uses the networkx library to implement DFS traversal and the pgmpy library to implement Bayesian network learning. Its functions are to build a causal library, perform root cause tracing and confidence calculation in step 2, make corrections, and output a fault list. Repair solution generation module: The hardware uses Neo4j graph database server (CPU: Intel Core i7-12700K, memory 32GB), and the software uses Neo4j Python driver (neo4j library). Its functions are to build a fault knowledge graph, perform entity mapping and association query in step 3, and generate a set of structured repair solutions. Repair report generation module: The hardware uses a local server + cloud server, and the software uses ReportLab, Django, and json libraries. Its functions include filling in the report template, format conversion, output and archiving in step 4, and supporting simultaneous output of multiple formats.

[0034] The modules work together, and the process is as follows: the module for determining the fault event to be tested collects preprocessed data → the causal tracing module generates a fault list → the repair plan generation module generates a repair plan → the repair report generation module generates a report and archives it.

[0035] The present invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the equipment fault diagnosis method based on a large model, causal library, and knowledge graph, including code for modules such as data preprocessing, causal library construction, root cause tracing, knowledge graph query, and maintenance report generation.

[0036] This invention also provides an electronic device, including at least one processor and at least one memory connected to the processor. The processor is used to call program instructions in the memory to execute the equipment fault diagnosis method based on a large model, causal library, and knowledge graph. Specifically, the processor receives abnormal data collected by sensors through an RS485 interface and calls a preprocessing program for processing; it calls a causal tracing program to construct a causal library and complete root cause tracing; it calls a knowledge graph query program to generate a maintenance plan; and it calls a report generation program to generate and archive multi-format maintenance reports. Simultaneously, this electronic device supports manual intervention (such as manually adjusting confidence parameters and modifying maintenance plans) and has additional functions such as data query and report statistics.

[0037] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any modifications, alterations, substitutions, and variations made by those skilled in the art to the above embodiments are within the scope of the present invention.

Claims

1. An equipment fault diagnosis method based on large models, causal bases, and knowledge graphs, characterized in that: Includes the following steps: Step 1: Preprocess the abnormal channel data collected by the sensors of industrial equipment, convert the abnormal channel data of different formats into a unified structured time-series data format, and output the fault information of each abnormal channel. Step 2: Filter out fault information related to the channel under test Related information on other abnormal channel faults; obtain historical fault data based on existing equipment maintenance manuals and historical fault maintenance records, and learn a causal library from the historical fault data through Bayesian network. The causal library stores a directed acyclic graph (DAG) of causal relationships between and within each device. The same fault cause is traced through multivariate causal relationships. Starting from the node corresponding to the fault information of the channel under test, a reverse depth-first traversal (DFS) is performed in the DAG of the causal library to find the initial confidence of each parent node as the root cause of the fault. The confidence is corrected based on the fault information of other abnormal channels. The corrected confidence is sorted from high to low, and the top 5 fault causes are selected to form a fault list. Specifically, in step 2, the node corresponding to the fault information of the channel under test... Starting from the causal database, perform a reverse depth-first traversal (DFS) in the DAG to find all causes abnormal parent node set And record each parent node to Path length For each parent node Calculate its initial confidence level as a root cause of the failure. : ;in: For the edge The causal strength; parent node To the node Path length; parent node The frequency of root causes in historical fault data is obtained from statistics in the historical fault database; All are weight coefficients, satisfying α+β+γ=1, and are obtained through grid search or reinforcement learning optimization; The confidence level is corrected by analyzing fault information from multiple other abnormal channels. If the parent nodes of these other abnormal channels all point to the parent node... Then for Perform weighted boosting: where k points to the parent node. The number of abnormal channels, where λ is the boosting factor, 0 < λ < 1, used to strengthen the confidence that multiple abnormal channels point to the same root cause; Based on the corrected confidence level Sort the faults from highest to lowest, select the top 5 nodes as the fault symptoms and causes, and form a fault list. : Each entry contains a root cause node. Fault symptoms and the corresponding corrected confidence level. ; Step 3: Construct a fault graph using a triplet structure based on existing equipment maintenance manuals, historical fault maintenance records, and equipment manufacturer technical documents; map each fault cause node in the fault list output in Step 2 to a fault cause entity in the knowledge graph, query the fault phenomenon entity and maintenance index entity corresponding to the fault cause entity, generate a structured solution, and output a set of maintenance solutions sorted by confidence. The fault map uses a triplet structure and includes: the entity representing the fault cause. This includes sensor damage, poor wiring contact; physical symptoms of the fault. This includes abnormal channel values ​​and image distortion; repair index entities. This includes replacing sensors and rewiring; There are three types of logical relationships between entities: cause, correspondence, and inclusion. Among them, cause represents the causal relationship between the entity causing the fault and the entity exhibiting the fault phenomenon; correspondence represents the solution association between the entity exhibiting the fault phenomenon and the entity exhibiting the maintenance index; and inclusion represents the hierarchical relationship between entities. Each root cause node in the fault list Mapped to the fault cause entity in the knowledge graph ,like Description and In a knowledge graph, if the cosine similarity of the standardized name or semantic tag of an entity is greater than the threshold θ, then the entity is determined to be the same entity. Fault symptom query: Search for related information in the knowledge graph. There exist entities that cause failures in the relationship. That is, to find all that satisfy < ,lead to, The triplet of >; Repair Index Query: Querying within the knowledge graph Maintenance index entities with corresponding relationships exist That is, to find all that satisfy < ,correspond, The triplet of >; For each entity causing the failure Integrate the corresponding fault phenomenon entities With maintenance index entity Generate a detailed repair plan, including: Repair steps: Extract relevant information from the knowledge graph. The associated standardized operating procedures are arranged in the order of operation; Required tools: querying from knowledge graphs and Related tool models and specifications; Spare part model: Searched from the knowledge graph Related spare parts models and specifications; Estimated working hours: The average working hours for similar maintenance tasks, obtained from historical fault repair records. Safety Precautions: Safety guidelines relevant to the current maintenance operation extracted from the equipment maintenance manual; The structured representation of each scheme is as follows: ; in The cause of the malfunction, This is a fault phenomenon. For repair procedures, For the necessary tools, For spare parts model, To estimate working hours, Safety precautions; The repair solutions are then sorted by confidence level from highest to lowest, with the highest confidence level being the most recommended. Step 4: Preset the repair report template and repair report output format, extract and integrate the data sources output from Step 1, Step 2 and Step 3, fill in and format the report content according to the repair report template, and complete the conversion and output of the single event repair report of the repair plan according to the preset repair report output format requirements and archive it.

2. The equipment fault diagnosis method based on a large model, causal database, and knowledge graph as described in claim 1, characterized in that, In step 1, the abnormal channel data exists in the form of unstructured binary stream, comma-separated value CSV, or real-time data stream, containing channel number, channel name, timestamp, sample value, and channel image information; Preprocessing of multiple abnormal channel data collected by industrial equipment sensors includes data cleaning, format standardization, and missing value imputation.

3. The equipment fault diagnosis method based on a large model, causal database, and knowledge graph as described in claim 1, characterized in that, Other abnormal fault channel information in step 2 Where A is the set of indices of all abnormal channels in this event, and j is the index of the abnormal channel in set A other than i. This represents the fault information for the j-th abnormal channel; the causal database is G=(V,E), where: It is a set of nodes, which includes device nodes and variable nodes. Device nodes include complete equipment and equipment components; variable nodes include temperature, pressure, and current. Let be a set of directed edges, each edge Indicates from node To the node The causal relationship.

4. The equipment fault diagnosis method based on a large model, causal database, and knowledge graph as described in claim 1, characterized in that, The maintenance report template includes: summary unit, anomaly detection unit, image analysis unit, fault tracing unit, maintenance plan unit, and appendix unit; the maintenance report output formats include PDF, HTML, and JSON.

5. An equipment fault diagnosis device based on a large model, causal database, and knowledge graph, which applies the equipment fault diagnosis method based on a large model, causal database, and knowledge graph as described in claim 1, is characterized in that... The system includes a fault event determination module, a causal tracing module, a maintenance plan generation module, and a maintenance report generation module. The fault event determination module is used to find the fault phenomenon and cause of the channel under test. The causal tracing module combines the causal relationship between variables to trace the same fault cause and phenomenon, finding 5 fault causes. The maintenance plan generation module is used to combine a knowledge graph to find the corresponding maintenance index and generate a maintenance plan. The maintenance report generation module generates a single event maintenance report, including the abnormal detection result, the abnormal channel under test image, the equipment information, the physical meaning description, the function description, the fault phenomenon, the cause, and the corresponding maintenance plan.

6. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the equipment fault diagnosis method based on a large model, causal library and knowledge graph as described in any one of claims 1 to 4.

7. An electronic device, characterized in that: It includes at least one processor and at least one memory connected to the processor, the processor being used to invoke program instructions in the memory to execute the equipment fault diagnosis method based on a large model, causal library and knowledge graph as described in any one of claims 1 to 4.

Images