An intelligent fault diagnosis collaboration method enhanced by fusion of large models and fault trees

Abstract

The application provides an intelligent troubleshooting collaborative method fusing a large model and a fault tree enhancement, and belongs to the technical field of aero-engine fault diagnosis and maintenance. The method comprises the following steps: receiving a fault picture, language description and time sequence data, and converting the fault picture, language description and time sequence data into a formatted fault description by a large model; constructing a domain knowledge base, combining aero-engine domain knowledge and fault case knowledge, and constructing an engine component, a fault phenomenon, a cause and a solution scheme into a knowledge graph; a fault tree dynamic updating module locates a fault tree position according to the formatted fault description and a BOM tree, and updates an edge weight of the fault tree; and a multi-troubleshooting path synchronization module generates a report according to the determined fault part and fault condition, and fuses multi-source evidence by using a Bayesian network. The method effectively processes multi-modal information, dynamically optimizes the fault tree, significantly improves the accuracy and efficiency of troubleshooting of complex equipment, has the ability of automatically generating a troubleshooting path by using the fault tree to enhance the large model, and realizes intelligent troubleshooting collaboration.

Description

Technical Field

[0001] This invention relates to the field of aero-engine fault diagnosis and maintenance technology, specifically to an intelligent troubleshooting collaborative method that integrates large models and fault tree enhancement. Background Technology

[0002] In the field of equipment troubleshooting, traditional methods often struggle to efficiently handle multimodal fault information. They lack effective integration and identification mechanisms for various input formats, such as fault images, verbal descriptions, and time-series data, resulting in inaccurate and incomplete fault descriptions. Furthermore, existing fault trees are typically static, failing to cover all possible fault scenarios and unable to dynamically adjust based on real-time fault information and equipment status, thus limiting the accuracy and efficiency of troubleshooting. In addition, when dealing with multi-component faults, traditional methods struggle to achieve parallel reasoning between different components, leading to low troubleshooting efficiency.

[0003] Existing collaborative troubleshooting methods require manual description of the fault phenomenon, breakdown and preliminary diagnosis based on the BOM tree, localization based on the fault tree, and finally, full-process traceability and investigation of related parts, including part process investigation and batch processing information. Based on relevant experience and historical data, the cause of the fault and feasible solutions are then proposed, culminating in a zero-defect report. This entire process is time-consuming and labor-intensive. Therefore, there is an urgent need for an intelligent troubleshooting method that can effectively handle multimodal input, dynamically update the fault tree, achieve multi-fault tree collaborative reasoning, automatically troubleshoot faults, and generate zero-defect reports. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent troubleshooting collaborative method that integrates large-scale models and fault tree enhancement, comprising the following steps:

[0005] Step S1: Receive fault images, language descriptions, and time series data through the multimodal input module. Extract features through image recognition, natural language processing, and signal processing, and transcribe the large model into a formatted fault description.

[0006] Step S2: Domain knowledge base construction. Combining knowledge of the aero-engine domain and fault case knowledge, a knowledge graph is constructed to represent engine components, fault phenomena, causes, and solutions, supporting complex relationship queries.

[0007] Step S3: Based on the formatted fault description and BOM tree, locate the fault tree position using the knowledge graph, and update the fault tree edge weights using the gradient descent algorithm by combining historical and real-time data, thereby dynamically updating the knowledge graph.

[0008] Step S4: Based on the identified faulty parts and fault conditions, trace the fault tree along multiple directions such as process design, product design, procurement and supply, production and manufacturing, service support, and simulation experiments. Parallel processing and result integration of different traceability paths are achieved through distributed computing.

[0009] Step S5: When generating a report from a large model, use a Bayesian network to fuse multi-source evidence to ensure that the multiple investigation paths and generated solutions in step S4 conform to the fault tree constraints.

[0010] The present invention has the following beneficial effects:

[0011] 1. The method of this invention integrates a large model into the troubleshooting collaboration process, and utilizes its powerful learning ability and rich knowledge base to achieve information extraction of fault-related multimodal data and fusion of multi-source data without consuming resources for additional training.

[0012] 2. This invention constructs a knowledge graph in the field of aero-engine maintenance, which stores components, fault phenomena, causes and solutions in a structured manner in the form of triples. It supports complex relationship retrieval such as "cross-component fault propagation path query" and "historical case similarity matching", effectively solving the problem that industry experience is difficult to quantify and utilize.

[0013] 3. This invention enables dynamic updating of the fault tree and, combined with BOM tree location of faulty components, allows the fault tree to integrate equipment operation data (such as sensor timing signals and process parameters) and historical fault probability distribution in real time. This avoids the troubleshooting bias caused by fixed parameters in static fault trees and achieves an upgrade from experience-driven to data-experience dual-driven.

[0014] 4. This invention constructs a multi-fault tree collaborative network, which realizes parallel reasoning of fault trees across components through a distributed computing framework. This mechanism can significantly improve the efficiency of fault diagnosis.

[0015] 5. This invention constructs a multi-source evidence fusion mechanism based on Bayesian networks and forces the investigation path to conform to the fault tree logic constraints, ensuring the feasibility and optimality of the solution. The generated zeroing report not only includes the cause of the fault and the repair plan, but also provides probabilistic risk assessment and evidence chain tracing, providing quantifiable scientific basis for equipment maintenance decisions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the intelligent troubleshooting collaborative method that integrates large models and fault tree enhancement according to the present invention.

[0017] Figure 2 This is a schematic diagram of the framework structure of the method of the present invention;

[0018] Figure 3 This is a structural block diagram of the intelligent troubleshooting collaborative system that integrates large model and fault tree enhancement according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0020] Example 1

[0021] like Figure 1 As shown in the figure, an intelligent troubleshooting collaborative method integrating large model and fault tree enhancement provided by an embodiment of the present invention includes the following steps:

[0022] Step S1: Receive fault images, language descriptions, and time series data through the multimodal input module. Extract features through image recognition, natural language processing, and signal processing, and transcribe the large model into a formatted fault description.

[0023] Step S2: Domain knowledge base construction. Combining knowledge of the aero-engine domain and fault case knowledge, a knowledge graph is constructed to represent engine components, fault phenomena, causes, and solutions, supporting complex relationship queries.

[0024] Step S3: The fault tree dynamic update module locates the fault tree position based on the formatted fault description and BOM tree, combined with the knowledge graph, and updates the edge weights of the fault tree using the gradient descent algorithm by combining historical and real-time data, thereby dynamically updating the knowledge graph.

[0025] Step S4: The multi-path troubleshooting synchronization module traces along multiple directions such as process design, product design, procurement and supply, production and manufacturing, service support, and simulation experiments based on the identified faulty parts and fault conditions. It achieves parallel processing and result integration of different troubleshooting paths through distributed computing.

[0026] Step S5: When generating a report from a large model, the zeroing report generation module uses a Bayesian network to fuse multi-source evidence to ensure that the multiple investigation paths and generated solutions in S4 conform to the fault tree constraints.

[0027] In one embodiment, step S1 above: receiving fault images, language descriptions, and time-series data through a multimodal input module, extracting features through image recognition, natural language processing, and signal processing, and transcribing the large model into a formatted fault description, specifically includes:

[0028] Step S11: Image recognition processing, using convolutional neural networks and ResNet models to identify faulty parts in fault images and extract features such as shape and color changes of the faulty parts;

[0029] Step S12: Natural Language Processing. Using pre-trained models such as BERT, fine-tuning is performed by combining domain knowledge and historical troubleshooting collaboration data to identify fault intent, analyze information such as the time, phenomenon, and location of the fault, and quickly locate similar cases by matching historical fault descriptions through semantic search.

[0030] Step S13: Time series data processing, extracting characteristic parameters such as vibration frequency and amplitude from the time series data through Fourier transform or wavelet analysis;

[0031] Step S14: Large model transcribing, generating a formatted fault description containing fault location, phenomenon, and characteristic parameters according to a preset format, and integrating multimodal features.

[0032] In one embodiment, step S2 above: domain knowledge base construction, combining aero-engine domain knowledge and fault case knowledge, constructs a knowledge graph of engine components, fault phenomena, causes, and solutions, supporting complex relationship queries, specifically including:

[0033] Step S21: Domain data collection, collecting data from areas such as aircraft engine component manuals, historical failure cases, historical zeroing reports, maintenance records, and BOM tree structures;

[0034] Step S22: Integrate unstructured data, such as maintenance reports and technical documents, with structured data, such as sensor logs and fault codes;

[0035] Step S23: Knowledge extraction and modeling. Based on named entity recognition, extract entities such as engine parts, fault phenomena, troubleshooting paths, fault causes, and solutions. Define the relationships between entities through relation extraction, including "occurred at", "caused by", "solution", etc.

[0036] Step S24: Knowledge graph construction. A knowledge graph for aero-engines is constructed using a graph database (such as Neo4j) and stored in the form of triples (entity, relation, entity).

[0037] Step S25: Establish complex relationships between entities, such as the reasoning path of "compressor → abnormal vibration → blade wear → blade replacement".

[0038] In one embodiment, step S3 above: The fault tree dynamic update module locates the fault tree position based on the formatted fault description and BOM tree, and updates the fault tree edge weights using a gradient descent algorithm by combining historical and real-time data, specifically including:

[0039] Step S31: Traverse the BOM tree, match component nodes according to the formatted fault description, and locate the initial position of the fault tree;

[0040] Step S32: Combining the fault occurrence frequency, historical maintenance data, and initial values ​​of the fault tree edge weights in the domain knowledge graph, the fault nodes in the domain knowledge graph are analyzed. To the fault phenomenon node The number of occurrences is expressed as Fault cause node The total number of times it appears in all records is expressed as Right now

[0041] (4)

[0042] Fault Cause Node To the fault phenomenon node The number of times it was repaired is expressed as , Indicate the cause of the malfunction The total number of repairs in historical maintenance records is weighted by a coefficient. and and The specific allocation can be adjusted according to the actual situation. Represents a node in the fault tree To the node The initial weights are calculated using the following formula;

[0043] (5)

[0044] Step S33: Collect equipment operation data in real time, and dynamically update the fault tree edge weights using gradient descent to reflect the current fault propagation probability. Mean Squared Error (MSE) is used as the loss function. The formula is:

[0045] (6)

[0046] in, This represents the probability of fault propagation predicted by edge weights. This represents the actual observed probability of fault propagation.

[0047] In one embodiment, step S4 above: the multi-path investigation synchronization module traces along multiple directions such as process design, product design, procurement and supply, production and manufacturing, service support, and simulation experiments based on the determined faulty parts and fault conditions, and realizes parallel processing and result integration of different investigation paths through distributed computing:

[0048] Step S41: Parallel Traceability Initialization: Based on the process design, product design, procurement and supply, production and manufacturing, service support, and simulation experiment troubleshooting paths determined in S4, computing resources are allocated to each path processing node through a distributed computing platform;

[0049] Step S42: Multi-path parallel execution, with each investigation path conducting tracing in parallel. In the process design direction, process document review, process simulation verification, and process optimization suggestions are output; in the product design direction, design drawing review, finite element analysis, and design improvement schemes are output; in the procurement and supply direction, supplier evaluation, raw material inspection, and procurement process review are performed; in the manufacturing direction, production process monitoring data review, operator operation standard inspection, and production environment assessment are performed; in the service support direction, maintenance record review, fault reporting and handling record analysis, and service support process optimization suggestions are output; in the simulation experiment direction, experimental scheme review, experimental data review, and experimental model improvement are performed.

[0050] Step S43: Real-time information interaction: Each path processing node synchronizes and checks data in real time through the message queue or shared memory of the distributed computing platform, and dynamically adjusts the traceability depth based on cross-path association rules (e.g., when a raw material problem is found in the procurement and supply direction, it triggers the production and manufacturing direction to strengthen the monitoring and verification of the processing data of that batch of parts).

[0051] Step S44: Multi-dimensional result fusion: Use evidence fusion algorithms (such as DS evidence theory) to integrate the fault correlation evidence output from each path and generate a zeroing report that includes investigation conclusions from each direction, cross-validation results and comprehensive improvement suggestions.

[0052] In one embodiment, step S5 above: When generating a report from a large model, the zeroing report generation module utilizes a Bayesian network to fuse multi-source evidence to ensure that the troubleshooting path and solution conform to the fault tree constraints, specifically including:

[0053] Step S51: Generate an initial troubleshooting path based on the fault tree structure and the fault causal relationships in the domain knowledge graph;

[0054] Step S52: Utilize Bayesian networks to fuse process parameters, sensor data, and solutions for similar faults in the knowledge graph to select the optimal troubleshooting path;

[0055] Step S53: Ensure that the troubleshooting path conforms to the fault tree logic constraints (e.g., the "blade wear" fault needs to be associated with the replacement process of the corresponding component in the knowledge graph), and generate a standard zeroing report.

[0056] Figure 2The diagram illustrates the framework of the method of this invention. First, historical troubleshooting data and relevant domain knowledge bases are collected. A large language model is used for information recognition to obtain knowledge of fault phenomena, faulty components, and fault causes. This is combined with a troubleshooting collaborative network for relationship extraction and information extraction, resulting in a fault investigation knowledge graph. Next, diverse raw fault data, including fault-related descriptions, images, and time-series data, are input. A pre-trained language model and predefined fault description templates are used for information extraction to obtain formatted fault descriptions. Based on the formatted fault descriptions, the fault investigation knowledge graph, and the BOM tree and fault tree, a troubleshooting collaborative agent generates fault investigation paths, enabling simultaneous multi-path troubleshooting. After investigation, the fault tree weights are updated, and a zeroing report is generated.

[0057] Example 2

[0058] like Figure 3 As shown, this embodiment of the invention provides an intelligent troubleshooting collaborative system that integrates large model and fault tree enhancement, including the following modules:

[0059] The data acquisition and preprocessing module 61 receives fault images, language descriptions, and time series data through the multimodal input module. It extracts features through image recognition, natural language processing, and signal processing, and transcribes the large model into a formatted fault description.

[0060] The knowledge base construction module 62 is used to build a knowledge base for the collaborative troubleshooting of aero-engines. It combines aero-engine-related knowledge and fault case knowledge to construct a knowledge graph of engine components, fault phenomena, causes, and solutions, and supports complex relationship queries.

[0061] The fault tree dynamic update module 63 is used to locate the fault tree position based on the formatted fault description and BOM tree, and dynamically update the fault tree structure and edge weights by combining historical and real-time data.

[0062] The multi-path parallel troubleshooting module 64 is used to achieve parallel reasoning and result integration in different troubleshooting directions through a distributed computing framework, thereby improving troubleshooting efficiency.

[0063] The zeroing report generation module 65 is used to integrate multi-source evidence (process parameters, batch processing information, etc.) to generate a zeroing report, ensuring that the troubleshooting path and solution comply with the fault tree constraints.

Claims

1. A collaborative intelligent troubleshooting method integrating large-scale models and fault tree enhancement, characterized in that, include: Step S1: Receive fault images, language descriptions, and time series data through the multimodal input module. Extract features through image recognition, natural language processing, and signal processing, and transcribe the large model into a formatted fault description. Step S2: Domain knowledge base construction. Combining knowledge of the aero-engine domain and fault case knowledge, a knowledge graph is constructed to represent engine components, fault phenomena, causes, and solutions, supporting complex relationship queries. Step S3: Based on the formatted fault description and BOM tree, locate the fault tree position using the knowledge graph, and update the edge weights of the fault tree using the gradient descent algorithm by combining historical and real-time data, thereby dynamically updating the knowledge graph. Step S4: Based on the identified faulty parts and fault conditions, trace the fault along multiple directions, including process design, product design, procurement and supply, production and manufacturing, service support, and simulation experiments, using the fault tree troubleshooting path. Parallel processing and result integration of different troubleshooting paths are achieved through distributed computing. The process design team is responsible for reviewing process documents, performing process simulation verification, and providing process optimization suggestions. The product design team is responsible for reviewing design drawings, performing finite element analysis, and outputting design improvement plans. The procurement and supply side conducts supplier evaluation, raw material inspection, and procurement process review. In the manufacturing sector, the tasks include reviewing production process monitoring data, checking operator procedures, and assessing the production environment; in the service support sector, the tasks include reviewing maintenance records, analyzing fault reporting and handling records, and providing suggestions for optimizing service support processes. The simulation experiment team is responsible for reviewing experimental plans, verifying experimental data, and improving experimental models. Step S5: When generating a report from a large model, use a Bayesian network to fuse multi-source evidence to ensure that the multiple investigation paths and generated solutions in step S4 conform to the fault tree constraints.

2. The intelligent troubleshooting collaborative method integrating large model and fault tree enhancement according to claim 1, characterized in that, Step S1 specifically includes: Step S11: Image recognition processing, using convolutional neural networks and ResNet models to identify faulty parts in fault images and extract the shape and color change features of the faulty parts; Step S12: Natural Language Processing. Using the BERT pre-trained model, fine-tuning is performed by combining domain knowledge and historical troubleshooting collaboration data to identify the fault intent, analyze the time, phenomenon, and location information of the fault occurrence, and quickly locate similar cases by semantic search matching historical fault descriptions. Step S13: Time series data processing, extracting vibration frequency and amplitude characteristic parameters from the time series data through Fourier transform or wavelet analysis; Step S14: Large model transcribing, generating a formatted fault description containing fault location, phenomenon, and characteristic parameters according to a preset format, and integrating multimodal features.

3. The intelligent troubleshooting collaborative method integrating large model and fault tree enhancement according to claim 1, characterized in that, Step S2 specifically includes: Step S21: Domain data collection, collecting data from various domains including aircraft engine component manuals, historical failure cases, historical zeroing reports, maintenance records, and BOM tree structures; Step S22: Integrate unstructured data with structured data; Step S23: Knowledge extraction and modeling. Based on named entity recognition, extract the entities of engine parts, fault phenomena, troubleshooting paths, fault causes, and solutions. Define the relationships between entities through relation extraction, including "occurred at", "caused by", and "solution". Step S24: Knowledge graph construction. An aero-engine knowledge graph is constructed using a graph database and stored in the form of triples. Step S25: Establish complex relationships between entities.

4. The intelligent troubleshooting collaborative method integrating large-scale model and fault tree enhancement according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Traverse the BOM tree, match component nodes according to the formatted fault description, and locate the initial position of the fault tree; Step S32: Combining the fault occurrence frequency, historical maintenance data, and initial values ​​of the fault tree edge weights in the domain knowledge graph, the fault nodes in the domain knowledge graph are analyzed. To the fault phenomenon node The number of occurrences is expressed as Fault cause node The total number of times it appears in all records is expressed as Right now: Fault Cause Node To the fault phenomenon node The number of times it was repaired is expressed as , Indicate the cause of the malfunction The total number of repairs in historical maintenance records is weighted by a coefficient. and and , Represents a node in the fault tree To the node The initial weights are calculated using the following formula: Step S33: Collect equipment operation data in real time, dynamically update the fault tree edge weights using gradient descent to reflect the current fault propagation probability, and use the mean squared error (MSE) as the loss function. The formula is: in, This represents the probability of fault propagation predicted by edge weights. This represents the actual observed probability of fault propagation.

5. The intelligent troubleshooting collaborative method integrating large model and fault tree enhancement according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Parallel traceability initialization: Based on the determined process design, product design, procurement and supply, production and manufacturing, service support, and simulation experiment investigation paths, computing resources are allocated to each path processing node through a distributed computing platform; Step S42: Multi-path parallel execution, each investigation path is traced in parallel; Step S43: Real-time information interaction: Each path processing node synchronizes and checks data in real time through the message queue or shared memory of the distributed computing platform, and dynamically adjusts the tracing depth based on cross-path association rules; Step S44: Multi-dimensional result fusion: The evidence fusion algorithm is used to integrate the fault correlation evidence output from each path to generate a zeroing report that includes investigation conclusions from each direction, cross-validation results, and comprehensive improvement suggestions.

6. The intelligent troubleshooting collaborative method integrating large model and fault tree enhancement according to claim 1, characterized in that, Step S5 specifically includes: Step S51: Generate an initial troubleshooting path based on the fault tree structure and the fault causal relationships in the domain knowledge graph; Step S52: Utilize Bayesian networks to fuse process parameters, sensor data, and solutions for similar faults in the knowledge graph to select the optimal troubleshooting path; Step S53: Ensure that the troubleshooting path conforms to the fault tree logic constraints and generate a standardized zeroing report.

7. The intelligent troubleshooting collaborative method integrating large model and fault tree enhancement according to claim 3, characterized in that, Unstructured data includes maintenance reports and technical documents, while structured data includes sensor logs and fault codes.

8. A collaborative intelligent troubleshooting device integrating large model and fault tree enhancement, comprising the method described in claim 1, characterized in that, include: The formatted fault description module receives fault images, language descriptions, and time series data through the multimodal input module. It extracts features through image recognition, natural language processing, and signal processing, and transcribes the large model into a formatted fault description. The knowledge graph construction module builds a domain knowledge base, combining knowledge of the aero-engine field and fault case knowledge to construct a knowledge graph of engine components, fault phenomena, causes, and solutions, supporting complex relationship queries; The fault tree dynamic update module locates the fault tree position based on the formatted fault description and BOM tree, combined with the knowledge graph, and updates the edge weights of the fault tree using the gradient descent algorithm by combining historical and real-time data, thereby dynamically updating the knowledge graph. The multi-path investigation synchronization module traces the faulty parts and fault conditions based on the fault tree investigation path, and traces the faults along multiple directions such as process design, product design, procurement and supply, production and manufacturing, service support and simulation experiments. It achieves parallel processing and result integration of different investigation paths through distributed computing. The zeroing report generation module uses Bayesian networks to fuse multi-source evidence when generating reports from large models, ensuring that the multiple investigation paths and generated solutions in step S4 conform to the fault tree constraints.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1 to 7.

Images